{ "name": "Embodied AI Radar - curated index", "source": "https://hu-zlatan.github.io/embodied-ai-radar/", "license": "CC BY 4.0 (curated text and data); arXiv metadata CC0", "built": "2026-10-29", "count": 463, "fields": [ "id", "title", "type", "arxiv", "year", "venue", "org", "url", "code", "project", "topics", "tags", "robot_types", "highlight", "added", "takeaway_en", "takeaway_zh", "github_stars", "page", "labs" ], "items": [ { "id": "arxiv-2212.06817", "title": "RT-1: Robotics Transformer for Real-World Control at Scale", "type": "paper", "arxiv": "2212.06817", "year": 2022, "venue": null, "org": "Google", "url": null, "code": "https://github.com/google-research/robotics_transformer", "project": null, "topics": [ "vla" ], "tags": [ "real", "arm", "transformer" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Tokenized actions plus a Transformer trained on 130k real demos showed that robot policies scale with data diversity.", "takeaway_zh": "将动作离散化为 token，用 13 万条真实演示训练 Transformer，证明机器人策略会随数据多样性扩展。", "github_stars": 1739, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2212.06817.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2307.15818", "title": "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control", "type": "paper", "arxiv": "2307.15818", "year": 2023, "venue": null, "org": "Google DeepMind", "url": null, "code": null, "project": "https://robotics-transformer2.github.io/", "topics": [ "vla" ], "tags": [ "real", "arm", "vlm" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "Coined \"VLA\" — co-fine-tuning a web-scale VLM to output action tokens transfers semantic knowledge to robot control.", "takeaway_zh": "提出 \"VLA\" 概念：把网络规模的视觉语言模型联合微调为输出动作 token，把语义知识迁移到机器人控制。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2307.15818.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2405.12213", "title": "Octo: An Open-Source Generalist Robot Policy", "type": "paper", "arxiv": "2405.12213", "year": 2024, "venue": null, "org": "UC Berkeley, Stanford, CMU", "url": null, "code": "https://github.com/octo-models/octo", "project": null, "topics": [ "vla" ], "tags": [ "real", "arm", "diffusion", "cross-embodiment", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A small open diffusion-head Transformer pretrained on Open X-Embodiment that fine-tunes quickly to new robots and sensors.", "takeaway_zh": "在 Open X-Embodiment 上预训练的开源扩散头 Transformer 策略，可快速微调到新机器人和新传感器。", "github_stars": 1811, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2405.12213.html", "labs": [ "berkeley", "stanford", "cmu" ] }, { "id": "arxiv-2406.09246", "title": "OpenVLA: An Open-Source Vision-Language-Action Model", "type": "paper", "arxiv": "2406.09246", "year": 2024, "venue": null, "org": "Stanford, UC Berkeley, TRI", "url": null, "code": "https://github.com/openvla/openvla", "project": null, "topics": [ "vla" ], "tags": [ "real", "arm", "vlm", "open-weights" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "A 7B open VLA built on Llama 2 + DINOv2/SigLIP; the de-facto open baseline for VLA research and LoRA fine-tuning.", "takeaway_zh": "基于 Llama 2 与 DINOv2/SigLIP 的 7B 开源 VLA，是 VLA 研究与 LoRA 微调的事实开源基线。", "github_stars": 7231, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2406.09246.html", "labs": [ "berkeley", "stanford", "tri" ] }, { "id": "arxiv-2502.19645", "title": "Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success", "type": "paper", "arxiv": "2502.19645", "year": 2025, "venue": null, "org": "Stanford", "url": null, "code": "https://github.com/moojink/openvla-oft", "project": null, "topics": [ "vla" ], "tags": [ "real", "sim", "arm", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "OpenVLA-OFT — parallel decoding, action chunking and continuous actions make VLA fine-tuning much faster and more accurate.", "takeaway_zh": "OpenVLA-OFT：并行解码、动作分块与连续动作让 VLA 微调更快、成功率更高。", "github_stars": 1437, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.19645.html", "labs": [ "stanford" ] }, { "id": "arxiv-2410.24164", "title": "$π_0$: A Vision-Language-Action Flow Model for General Robot Control", "type": "paper", "arxiv": "2410.24164", "year": 2024, "venue": null, "org": "Physical Intelligence", "url": null, "code": "https://github.com/Physical-Intelligence/openpi", "project": null, "topics": [ "vla" ], "tags": [ "real", "arm", "bimanual", "flow-matching", "cross-embodiment", "open-weights" ], "robot_types": [ "arm", "hand" ], "highlight": true, "added": null, "takeaway_en": "A VLM backbone with a flow-matching action expert, trained across many robots, doing long dexterous tasks like laundry folding.", "takeaway_zh": "VLM 主干加流匹配动作专家，跨多种机器人训练，能完成叠衣服等长时程灵巧任务。", "github_stars": 14452, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.24164.html", "labs": [ "physical-intelligence" ] }, { "id": "arxiv-2504.16054", "title": "$π_{0.5}$: a Vision-Language-Action Model with Open-World Generalization", "type": "paper", "arxiv": "2504.16054", "year": 2025, "venue": null, "org": "Physical Intelligence", "url": null, "code": "https://github.com/Physical-Intelligence/openpi", "project": null, "topics": [ "vla" ], "tags": [ "real", "mobile", "co-training" ], "robot_types": [ "arm", "mobile" ], "highlight": false, "added": null, "takeaway_en": "Co-training on heterogeneous data lets a mobile manipulator clean kitchens and bedrooms in homes it has never seen.", "takeaway_zh": "通过异构数据联合训练，让移动操作机器人在从未见过的家庭中整理厨房和卧室。", "github_stars": 14452, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2504.16054.html", "labs": [ "physical-intelligence" ] }, { "id": "arxiv-2501.09747", "title": "FAST: Efficient Action Tokenization for Vision-Language-Action Models", "type": "paper", "arxiv": "2501.09747", "year": 2025, "venue": null, "org": "Physical Intelligence, UC Berkeley, Stanford", "url": null, "code": null, "project": null, "topics": [ "vla" ], "tags": [ "real", "tokenization" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "Compresses action chunks with a DCT-based tokenizer so autoregressive VLAs can learn high-frequency, dexterous control.", "takeaway_zh": "用基于 DCT 的分词器压缩动作块，使自回归 VLA 能学习高频灵巧控制。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2501.09747.html", "labs": [ "berkeley", "stanford", "physical-intelligence" ] }, { "id": "arxiv-2410.07864", "title": "RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation", "type": "paper", "arxiv": "2410.07864", "year": 2024, "venue": null, "org": "Tsinghua University", "url": null, "code": "https://github.com/thu-ml/RoboticsDiffusionTransformer", "project": null, "topics": [ "vla" ], "tags": [ "real", "bimanual", "diffusion", "open-weights" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "A 1.2B diffusion Transformer with a unified action space across robots, targeted at bimanual manipulation.", "takeaway_zh": "12 亿参数扩散 Transformer，采用跨机器人统一动作空间，面向双臂操作。", "github_stars": 1815, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.07864.html", "labs": [ "tsinghua" ] }, { "id": "arxiv-2503.14734", "title": "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots", "type": "paper", "arxiv": "2503.14734", "year": 2025, "venue": null, "org": "NVIDIA", "url": null, "code": "https://github.com/NVIDIA/Isaac-GR00T", "project": null, "topics": [ "vla" ], "tags": [ "real", "sim", "humanoid", "open-weights" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "A dual-system (VLM \"slow\" + diffusion \"fast\") humanoid policy trained on real, human-video and synthetic data.", "takeaway_zh": "双系统（VLM 慢思考 + 扩散快执行）人形机器人策略，用真实、人类视频与合成数据训练。", "github_stars": 8288, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2503.14734.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2506.01844", "title": "SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics", "type": "paper", "arxiv": "2506.01844", "year": 2025, "venue": null, "org": "Hugging Face", "url": null, "code": "https://github.com/huggingface/lerobot", "project": null, "topics": [ "vla" ], "tags": [ "real", "arm", "open-weights", "low-cost" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A ~450M-parameter VLA trained on community LeRobot data that runs on consumer GPUs.", "takeaway_zh": "约 4.5 亿参数的 VLA，用社区 LeRobot 数据训练，可在消费级 GPU 上运行。", "github_stars": 28561, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2506.01844.html", "labs": [ "hugging-face" ] }, { "id": "arxiv-2503.20020", "title": "Gemini Robotics: Bringing AI into the Physical World", "type": "paper", "arxiv": "2503.20020", "year": 2025, "venue": null, "org": "Google DeepMind", "url": null, "code": null, "project": null, "topics": [ "vla" ], "tags": [ "real", "bimanual", "vlm" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "Gemini 2.0-based VLA and embodied-reasoning models for dexterous, general robot control.", "takeaway_zh": "基于 Gemini 2.0 的 VLA 与具身推理模型，面向灵巧、通用的机器人控制。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2503.20020.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2303.03378", "title": "PaLM-E: An Embodied Multimodal Language Model", "type": "paper", "arxiv": "2303.03378", "year": 2023, "venue": null, "org": "Google", "url": null, "code": null, "project": null, "topics": [ "vla" ], "tags": [ "real", "vlm", "planning" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Injects continuous sensor embeddings into a 562B language model for embodied reasoning and planning.", "takeaway_zh": "将连续传感器嵌入注入 5620 亿参数语言模型，用于具身推理与规划。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2303.03378.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2205.06175", "title": "A Generalist Agent", "type": "paper", "arxiv": "2205.06175", "year": 2022, "venue": null, "org": "DeepMind", "url": null, "code": null, "project": null, "topics": [ "vla" ], "tags": [ "sim", "real", "transformer" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Gato — one Transformer with one set of weights plays games, captions images and stacks blocks with a real arm.", "takeaway_zh": "Gato：同一套权重的 Transformer 既能玩游戏、写图像描述，也能用真实机械臂堆积木。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2205.06175.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2610.22178", "title": "REKI: Registering Knowledge in VLAs", "type": "paper", "arxiv": "2610.22178", "year": 2026, "venue": null, "org": "Kyutai", "url": null, "code": "https://github.com/kyutai-labs/reki", "project": null, "topics": [ "vla" ], "tags": [ "sim", "flow-matching", "memory", "open-weights", "public-data" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "A 2.4B VLM + flow-matching VLA pretrained only on public data; reports beating all but ~2× larger models, plus cheap register-token memory.", "takeaway_zh": "仅用公开数据预训练的 2.4B VLM + 流匹配 VLA；报告称只输给约两倍大的模型，并用少量寄存器 token 实现低成本记忆。", "github_stars": 34, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.22178.html", "labs": [] }, { "id": "arxiv-2610.21362", "title": "LTM: Latent Transition Based In-Context Memory for Vision-Language-Action Models", "type": "paper", "arxiv": "2610.21362", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/buaa-colalab/LTM-pi", "project": null, "topics": [ "vla" ], "tags": [ "sim", "memory", "pi0.5" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "Gives π0.5 memory by caching compact latent transitions between observations, aimed at tasks where key objects go out of view.", "takeaway_zh": "通过缓存相邻观测之间的紧凑潜在转移，为 π0.5 加入记忆，面向关键物体移出视野的任务。", "github_stars": 3, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.21362.html", "labs": [] }, { "id": "arxiv-2610.22146", "title": "FastVLA: Task-Conditioned and Temporally Adaptive Visual Token Compression for Efficient Vision-Language-Action Models", "type": "paper", "arxiv": "2610.22146", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/jiuyixu25/FastVLA", "project": null, "topics": [ "vla" ], "tags": [ "efficiency", "token-compression", "inference" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "Speeds up VLA inference by keeping only the visual tokens the current task needs, with a token budget that changes from step to step.", "takeaway_zh": "只保留当前任务所需的视觉 token、并让 token 预算随时间步变化，以加快 VLA 推理。", "github_stars": 4, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.22146.html", "labs": [] }, { "id": "arxiv-2610.23271", "title": "NestedVLA: Learning to Consolidate and Generate Skills for Vision-Language-Action Models", "type": "paper", "arxiv": "2610.23271", "year": 2026, "venue": "NeurIPS 2026", "org": null, "url": null, "code": "https://github.com/WanyuGroup/NeurIPS2026-NestedVLA", "project": null, "topics": [ "vla" ], "tags": [ "multitask", "skill-consolidation", "hypernetwork" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "Learns skills separately, then consolidates them after the fact with a nested hypernetwork that generates VLA parameters instead of merging weights.", "takeaway_zh": "先独立学习各项技能，再用嵌套超网络“生成”VLA 参数来事后整合技能，而不是直接合并权重。", "github_stars": 0, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.23271.html", "labs": [] }, { "id": "arxiv-2411.19650", "title": "CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation", "type": "paper", "arxiv": "2411.19650", "year": 2024, "venue": null, "org": "Microsoft Research Asia, Tsinghua", "url": null, "code": "https://github.com/microsoft/CogACT", "project": null, "topics": [ "vla" ], "tags": [ "real", "sim", "diffusion", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Instead of emitting action tokens from the VLM, a separate diffusion action transformer is conditioned on the VLM's \"cognition\" feature; reports clear gains over OpenVLA in sim and on real robots.", "takeaway_zh": "不再让 VLM 直接输出动作 token，而是用独立的扩散动作 Transformer 以 VLM 的“认知”特征为条件生成动作；报告在仿真和真机上明显优于 OpenVLA。", "github_stars": 434, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2411.19650.html", "labs": [ "microsoft", "tsinghua" ] }, { "id": "arxiv-2507.02029", "title": "RoboBrain 2.0 Technical Report", "type": "paper", "arxiv": "2507.02029", "year": 2025, "venue": null, "org": "BAAI", "url": null, "code": "https://github.com/FlagOpen/RoboBrain2.0", "project": null, "topics": [ "vla" ], "tags": [ "vlm", "planning", "spatial-reasoning", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open embodied VLMs (7B and 32B) trained for spatial grounding, affordances, trajectories and multi-step task planning — a \"brain\" that sits above low-level policies.", "takeaway_zh": "开源具身视觉语言模型（7B 与 32B），面向空间定位、可供性、轨迹预测与多步任务规划，作为底层策略之上的“大脑”。", "github_stars": 1155, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2507.02029.html", "labs": [] }, { "id": "arxiv-2610.21306", "title": "VLA-$Δ$: Learning Physical State Transition for Vision-Language-Action Models", "type": "paper", "arxiv": "2610.21306", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/ZJU-OmniAI/VLA-Delta", "project": "https://zju-omniai.github.io/VLA-Delta/", "topics": [ "vla" ], "tags": [ "state-transition", "auxiliary-supervision" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "Argues VLAs overfit observation→action pairs and adds explicit supervision of the physical state change an action should cause, rather than predicting future states in language or pixels.", "takeaway_zh": "认为 VLA 容易过拟合“观测→动作”的对应关系，因此显式监督动作应引起的物理状态变化，而不是在语言或像素空间里预测未来状态。", "github_stars": 28, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.21306.html", "labs": [] }, { "id": "arxiv-2610.22951", "title": "MambaAct: Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models", "type": "paper", "arxiv": "2610.22951", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/giakhuyendihoc/MambaAct", "project": "https://giakhuyendihoc.github.io/mambaact-page/", "topics": [ "vla" ], "tags": [ "efficiency", "mamba", "flow-matching", "pi0.5" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "Keeps π0.5's pretrained vision-language encoder but swaps its Transformer action expert — run many times per denoising loop — for a compact Mamba expert to cut inference cost.", "takeaway_zh": "保留 π0.5 预训练的视觉语言编码器，把在去噪循环中被反复调用的 Transformer 动作专家换成紧凑的 Mamba 专家，以降低推理开销。", "github_stars": 4, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.22951.html", "labs": [] }, { "id": "arxiv-2502.19417", "title": "Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models", "type": "paper", "arxiv": "2502.19417", "year": 2025, "venue": null, "org": "Physical Intelligence", "url": null, "code": null, "project": null, "topics": [ "vla" ], "tags": [ "real", "hierarchical", "language-feedback" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A high-level VLM reads open-ended prompts and mid-task corrections and turns them into short language commands for a low-level π0 policy.", "takeaway_zh": "高层 VLM 理解开放式指令与执行中的纠正，并将其转化为简短语言指令交给底层 π0 策略执行。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.19417.html", "labs": [ "physical-intelligence" ] }, { "id": "arxiv-2210.03094", "title": "VIMA: General Robot Manipulation with Multimodal Prompts", "type": "paper", "arxiv": "2210.03094", "year": 2022, "venue": "ICML 2023", "org": "Stanford, NVIDIA, Caltech", "url": null, "code": "https://github.com/vimalabs/VIMA", "project": null, "topics": [ "vla" ], "tags": [ "multimodal-prompt", "transformer", "benchmark" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Casts many tabletop tasks as interleaved text-and-image prompts to one Transformer policy, with the VIMA-Bench suite for systematic generalization tests.", "takeaway_zh": "把多种桌面任务统一成图文交错的提示，交给一个 Transformer 策略执行，并配套 VIMA-Bench 用于系统性泛化评测。", "github_stars": 857, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2210.03094.html", "labs": [ "stanford", "nvidia" ] }, { "id": "arxiv-2610.25927", "title": "CondenseVLA: Learnable History Condensation for Efficient Multi-Frame VLA", "type": "paper", "arxiv": "2610.25927", "year": 2026, "venue": "NeurIPS 2026", "org": null, "url": null, "code": "https://github.com/FeiyangHong/CondenseVLA", "project": null, "topics": [ "vla" ], "tags": [ "history", "memory", "efficiency" ], "robot_types": [], "highlight": false, "added": "2026-10-20", "takeaway_en": "Instead of feeding a VLA a longer frame history, learns to condense past frames into a few tokens that keep short-lived cues such as brief contacts and occlusions.", "takeaway_zh": "不靠简单加长历史帧窗口，而是学习把过去的帧压缩成少量 token，保留短暂接触、遮挡等转瞬即逝的线索。", "github_stars": 0, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25927.html", "labs": [] }, { "id": "arxiv-2507.15493", "title": "GR-3 Technical Report", "type": "paper", "arxiv": "2507.15493", "year": 2025, "venue": null, "org": "ByteDance Seed", "url": null, "code": null, "project": null, "topics": [ "vla" ], "tags": [ "real", "bimanual", "mobile-manipulation", "long-horizon" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "ByteDance Seed's generalist VLA: robot imitation data co-trained with web vision-language data, adapted quickly with VR-collected human trajectories; shown on long-horizon bimanual mobile tasks.", "takeaway_zh": "字节跳动 Seed 的通用 VLA——机器人数据模仿学习并与网络视觉-语言数据联合训练，再用 VR 采集的人类轨迹快速适配；在 ByteMini 双臂移动机器人上展示长时程与衣物操作任务。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2507.15493.html", "labs": [ "bytedance" ] }, { "id": "arxiv-2505.06111", "title": "UniVLA: Learning to Act Anywhere with Task-centric Latent Actions", "type": "paper", "arxiv": "2505.06111", "year": 2025, "venue": "RSS 2025", "org": "HKU, OpenDriveLab, AgiBot", "url": null, "code": "https://github.com/OpenDriveLab/UniVLA", "project": null, "topics": [ "vla", "learning" ], "tags": [ "latent-actions", "cross-embodiment", "video-pretraining" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Learns task-centric latent actions from videos (in DINO feature space, filtered by language) so one policy can be pretrained across robots and human video; reports beating OpenVLA with far less pretraining compute.", "takeaway_zh": "在 DINO 特征空间中、借助语言指令过滤，从视频里学习“任务相关”的隐动作，使策略可跨机器人与人类视频预训练；报告以少得多的预训练算力超过 OpenVLA。", "github_stars": 1141, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2505.06111.html", "labs": [ "hku", "agibot" ] }, { "id": "arxiv-2501.15830", "title": "SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model", "type": "paper", "arxiv": "2501.15830", "year": 2025, "venue": "RSS 2025", "org": "Shanghai AI Lab, ShanghaiTech, TeleAI", "url": null, "code": "https://github.com/SpatialVLA/SpatialVLA", "project": null, "topics": [ "vla" ], "tags": [ "3d", "spatial", "action-tokenization" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Adds 3D-aware position encoding to visual tokens and adaptive action grids, pretrained on 1.1M real robot episodes; new robot setups are handled by re-discretizing the grids.", "takeaway_zh": "为视觉 token 加入 3D 位置编码，并用自适应动作网格表示动作，在 110 万条真机轨迹上预训练；迁移到新机器人时只需重新划分动作网格。", "github_stars": 734, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2501.15830.html", "labs": [ "shanghai-ai-lab" ] }, { "id": "arxiv-2610.29730", "title": "Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies", "type": "paper", "arxiv": "2610.29730", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/post-hoc-vla-uncertainty/post-hoc-vla-uncertainty", "project": null, "topics": [ "vla" ], "tags": [ "uncertainty", "safety", "gaussian-process", "failure-detection" ], "robot_types": [], "highlight": false, "added": "2026-10-22", "takeaway_en": "Post-hoc kernel-variance heads (Gaussian-process variance over training-demo embeddings) estimate one VLA policy's epistemic uncertainty without an ensemble, so the robot knows when to ask for help.", "takeaway_zh": "在训练完成后为单个 VLA 策略加上轻量的核方差头（基于训练示教嵌入的高斯过程方差）来估计认知不确定性，无需训练和运行策略集成，让机器人知道何时该求助或补充示教。", "github_stars": 0, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.29730.html", "labs": [] }, { "id": "arxiv-2610.30735", "title": "When One Step Is Enough: An Empirical Study of Euler Integration in Flow-Matching Policies", "type": "paper", "arxiv": "2610.30735", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": null, "topics": [ "vla", "learning" ], "tags": [ "flow-matching", "inference-speed", "empirical-study" ], "robot_types": [], "highlight": false, "added": "2026-10-23", "takeaway_en": "Tests six flow-matching VLAs and reports that one-step Euler action generation largely matches multi-step sampling at scale — a useful check before adding special one-shot objectives to speed up control (CoRL 2026).", "takeaway_zh": "在六个基于流匹配的 VLA 上实验，报告单步 Euler 生成动作在规模足够时与多步采样效果基本相当——在为提速引入专门的单步生成目标之前值得先看（CoRL 2026）。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30735.html", "labs": [] }, { "id": "arxiv-2610.25043", "title": "VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse", "type": "paper", "arxiv": "2610.25043", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/ZJU-OmniAI/VLA-Rolling", "project": "https://chenbao123ww.github.io/VLA-Rolling/", "topics": [ "vla" ], "tags": [ "flow-matching", "inference-speed", "action-chunking", "training-free-backbone" ], "robot_types": [], "highlight": false, "added": "2026-10-23", "takeaway_en": "Reuses the unexecuted tail of a flow-matching VLA's action chunk to start the next one, with a small confidence model setting how much to re-denoise; reports lower inference cost on π0.5, SmolVLA and others.", "takeaway_zh": "流匹配 VLA 每次只执行动作块的前几步，剩余部分通常丢弃；本文把这段\"动作尾巴\"作为下一块的起点，并用小型置信度模型决定还需去噪多少步；论文报告在 π0.5、SmolVLA 等模型上降低推理开销。", "github_stars": 13, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25043.html", "labs": [] }, { "id": "arxiv-2610.22471", "title": "Has GPT-6-Astra Achieved Embodied Intelligence?", "type": "paper", "arxiv": "2610.22471", "year": 2026, "venue": null, "org": "SenseNova, ACE Robotics, Ropedia, NTU, UTokyo", "url": null, "code": null, "project": "https://github.com/OpenSenseNova/SenseNova-RoboRSI", "topics": [ "vla", "manipulation", "navigation" ], "tags": [ "evaluation", "llm-as-policy", "zero-shot", "sim", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-23", "takeaway_en": "Drives robots with a frozen frontier LLM plus in-context 'knowledge packages' on 12 sim benchmarks + 8 real robots; reports beating prior bests on 9 of 12, but weak real-robot results (competence, not yet intelligence).", "takeaway_zh": "用冻结的前沿大模型加上下文“知识包”直接控制机器人，覆盖 12 个仿真基准和 8 台真机；报告在 12 个基准中的 9 个超过此前最好结果，但真机表现较弱——结论是具备了具身“能力”，尚未达到具身“智能”。", "github_stars": 314, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.22471.html", "labs": [] }, { "id": "arxiv-2405.14093", "title": "A Survey on Vision-Language-Action Models for Embodied AI", "type": "paper", "arxiv": "2405.14093", "year": 2024, "venue": null, "org": "CUHK, Huawei Noah's Ark Lab", "url": null, "code": null, "project": "https://github.com/yueen-ma/Awesome-VLA", "topics": [ "vla" ], "tags": [ "survey", "vla", "taxonomy", "awesome-list" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "An early survey of VLA models, sorting them into component work (vision, language, world models), low-level control policies and high-level task planners, plus datasets, simulators and benchmarks; companion paper list.", "takeaway_zh": "较早的 VLA 综述：把相关工作分为组件研究（视觉、语言、世界模型等）、底层控制策略与高层任务规划器，并汇总数据集、仿真器与基准；附配套论文列表。", "github_stars": 688, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2405.14093.html", "labs": [ "cuhk" ] }, { "id": "arxiv-2312.07843", "title": "Foundation Models in Robotics: Applications, Challenges, and the Future", "type": "paper", "arxiv": "2312.07843", "year": 2023, "venue": "IJRR 2024", "org": "Stanford, Princeton, UT Austin, Google DeepMind and others", "url": null, "code": null, "project": "https://github.com/robotics-survey/Awesome-Robotics-Foundation-Models", "topics": [ "vla" ], "tags": [ "survey", "foundation-models", "llm", "vlm" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Survey of how pretrained foundation models are used in robot perception, decision-making and control, and what holds them back — scarce robot data, safety and uncertainty, real-time execution.", "takeaway_zh": "综述预训练基础模型如何用于机器人的感知、决策与控制，以及主要瓶颈：机器人数据稀缺、安全性与不确定性量化、实时执行。", "github_stars": 1422, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2312.07843.html", "labs": [ "stanford", "google-deepmind", "ut-austin", "princeton" ] }, { "id": "arxiv-2407.06886", "title": "Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI", "type": "paper", "arxiv": "2407.06886", "year": 2024, "venue": null, "org": "Sun Yat-sen University (HCPLab)", "url": null, "code": null, "project": "https://github.com/HCPLab-SYSU/Embodied_AI_Paper_List", "topics": [ "vla", "sim", "navigation", "perception" ], "tags": [ "survey", "embodied-ai", "sim2real", "world-models" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Broad embodied-AI survey — robots, simulators, embodied perception, interaction, agents and sim-to-real — with the role of multimodal and world models; backed by the widely used HCPLab paper list.", "takeaway_zh": "覆盖面很广的具身智能综述：机器人本体、仿真器、具身感知、交互、具身智能体与 sim-to-real，并讨论多模态大模型与世界模型的作用；配套中山大学 HCPLab 维护的常用论文列表。", "github_stars": 2189, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2407.06886.html", "labs": [] }, { "id": "arxiv-2610.18568", "title": "WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction", "type": "paper", "arxiv": "2610.18568", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/AutoLab-SAI-SJTU/WAVE-0", "project": "https://hmwang2002.github.io/release/wave0/", "topics": [ "vla", "world-models" ], "tags": [ "world-action-model", "latent-world-model", "sim", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Adds a 257M latent world model to a pretrained VLA so predicted future features condition the action expert, with no pixel generation or video pretraining; reports 98.7% on LIBERO and real-robot gains over π0.5.", "takeaway_zh": "在预训练 VLA 上加一个 2.57 亿参数的潜空间世界模型，用预测的未来表征去条件化动作专家，无需像素生成或大规模视频预训练；报告 LIBERO 成功率 98.7%，真机任务上明显优于 π0.5。", "github_stars": 34, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.18568.html", "labs": [] }, { "id": "arxiv-2610.20019", "title": "MorphAct: Robot Policies Morphing with the Worlds They Act In", "type": "paper", "arxiv": "2610.20019", "year": 2026, "venue": null, "org": "Shanghai Jiao Tong University", "url": null, "code": "https://github.com/MINT-SJTU/MorphAct", "project": "https://mint-sjtu.github.io/MorphAct.io/", "topics": [ "vla", "learning" ], "tags": [ "lora", "contextual-parameters", "multi-task", "sim", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Instead of one fixed set of VLA weights, generates context-dependent low-rank updates for both the vision-language and action modules; reports gains up to 16.3% on LIBERO/Meta-World and up to 14% over π0.5 on real tasks.", "takeaway_zh": "不再让 VLA 在所有交互情境下共用固定参数，而是按当前情境为视觉语言模块和动作模块生成耦合的低秩参数更新；报告在 LIBERO/Meta-World 上最高提升 16.3%，真机上相对 π0.5 最高提升 14%。", "github_stars": 41, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.20019.html", "labs": [ "sjtu" ] }, { "id": "arxiv-2610.21453", "title": "CarryOn: A Dual-Loop Harness for Personalized and Reliable Robot Assistance", "type": "paper", "arxiv": "2610.21453", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/JIA-Lab-research/CarryOn", "project": "https://vincent2311.github.io/Projects/CarryOn/", "topics": [ "vla", "manipulation" ], "tags": [ "agentic", "harness", "personalization", "safety", "failure-recovery", "sim", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Wraps a frozen VLA in two loops — one keeps user preferences and scene memory for planning, one blocks hazardous actions and calls recovery skills; reports violations down from 52.6% to 6.1%, recovery up from 13% to 80%.", "takeaway_zh": "给冻结的 VLA 套上双环框架：规划环保存用户偏好与场景记忆以做个性化任务分解，执行环拦截危险接触并调用可复用的恢复技能；报告安全违规率从 52.6% 降到 6.1%，失败恢复成功率从 13% 提升到 80%。", "github_stars": 5, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.21453.html", "labs": [] }, { "id": "arxiv-2610.20442", "title": "Towards Agentic Robots: Model, Data, Environment, and Harness", "type": "paper", "arxiv": "2610.20442", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://github.com/chonghanqin/Awesome-Agentic-Robots", "topics": [ "vla" ], "tags": [ "survey", "agentic", "awesome-list", "world-action-model" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Survey of 'agentic robots' — foundation-model agents that plan, collect data, build environments and recover from errors — organised into 17 categories under model, data, environment and harness; companion paper list.", "takeaway_zh": "关于\"智能体化机器人\"的综述：基础模型智能体负责规划、采集与筛选数据、构建环境与测试并自主纠错；把相关研究归为模型、数据、环境、执行框架四大领域下的 17 个类别，附配套论文列表。", "github_stars": 138, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.20442.html", "labs": [] }, { "id": "arxiv-2610.30108", "title": "Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use", "type": "paper", "arxiv": "2610.30108", "year": 2026, "venue": null, "org": "Show Lab, National University of Singapore", "url": null, "code": null, "project": "https://github.com/showlab/Awesome-Multimodal-Embodied-Agent", "topics": [ "vla" ], "tags": [ "survey", "agentic", "awesome-list", "benchmark" ], "robot_types": [], "highlight": false, "added": "2026-10-25", "takeaway_en": "Survey framing computer-use and robot agents with five functions — perceive, anticipate, plan, act, verify; across 64 benchmarks it finds only 5 test anticipation and 3 test verification. Companion paper list.", "takeaway_zh": "用\"感知、预判、规划、执行、验证\"五种能力统一看待电脑操作智能体和机器人智能体的综述；统计 64 个基准后发现只有 5 个评测预判、3 个评测验证。附配套论文列表。", "github_stars": 207, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30108.html", "labs": [ "nus" ] }, { "id": "arxiv-2508.07917", "title": "MolmoAct: Action Reasoning Models that can Reason in Space", "type": "paper", "arxiv": "2508.07917", "year": 2025, "venue": null, "org": "Ai2, University of Washington", "url": null, "code": "https://github.com/allenai/molmoact", "project": "https://allenai.org/blog/molmoact", "topics": [ "vla" ], "tags": [ "real", "arm", "vlm", "open-weights", "reasoning", "spatial" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "An open 7B VLA that first outputs depth-aware perception tokens, then an editable 2D trajectory, then actions; reports 70.5% zero-shot on SimplerEnv and 86.6% on LIBERO, with weights, code and data released.", "takeaway_zh": "开源 7B VLA，先输出深度感知 token，再生成可编辑的二维轨迹，最后输出动作；论文报告 SimplerEnv 零样本 70.5%、LIBERO 86.6%，并开放权重、代码和数据。", "github_stars": 393, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2508.07917.html", "labs": [ "uw", "ai2" ] }, { "id": "arxiv-2505.03233", "title": "GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data", "type": "paper", "arxiv": "2505.03233", "year": 2025, "venue": "CoRL 2025", "org": "Peking University (PKU-EPIC)", "url": null, "code": "https://github.com/PKU-EPIC/GraspVLA", "project": null, "topics": [ "vla", "manipulation", "sim" ], "tags": [ "sim2real", "arm", "grasping", "synthetic-data", "flow-matching" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Pretrains a grasping VLA almost only on SynGrasp-1B, a billion-frame photorealistic sim dataset, co-trained with web data; reports zero-shot sim-to-real open-vocabulary grasping and few-shot adaptation.", "takeaway_zh": "几乎只用十亿帧级的照片级仿真抓取数据集 SynGrasp-1B（并联合网络语义数据）预训练抓取 VLA；论文报告可零样本从仿真迁移到真实开放词汇抓取，并支持少样本适配。", "github_stars": 427, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2505.03233.html", "labs": [ "pku" ] }, { "id": "arxiv-2511.14759", "title": "π*0.6: a VLA That Learns From Experience", "type": "paper", "arxiv": "2511.14759", "year": 2025, "venue": "RSS 2026", "org": "Physical Intelligence", "url": null, "code": null, "project": "https://pi.website/blog/pistar06", "topics": [ "vla", "learning" ], "tags": [ "real", "arm", "mobile-manipulation", "rl", "offline-rl", "interventions" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "RECAP trains a VLA with RL via advantage conditioning on demos, on-robot rollouts and teleop corrections; reports it more than doubles throughput on hard tasks like laundry folding and espresso making.", "takeaway_zh": "RECAP 通过优势条件化，用演示、机器人自主运行数据和遥操作纠正对 VLA 做强化学习；论文报告在叠衣服、做意式咖啡等难任务上吞吐量提升一倍以上、失败率约减半。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2511.14759.html", "labs": [ "physical-intelligence" ] }, { "id": "arxiv-2510.10274", "title": "X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model", "type": "paper", "arxiv": "2510.10274", "year": 2025, "venue": "ICLR 2026", "org": "Tsinghua AIR, Shanghai AI Lab", "url": null, "code": "https://github.com/2toinf/X-VLA", "project": "https://thu-air-dream.github.io/X-VLA/", "topics": [ "vla" ], "tags": [ "real", "sim", "cross-embodiment", "flow-matching", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Gives each data source its own small set of learnable soft-prompt embeddings, so one plain-Transformer flow-matching policy (0.9B) trains across embodiments; evaluated on 6 sim benchmarks and 3 real robots.", "takeaway_zh": "为每个数据来源配一组可学习的软提示嵌入，使一个基于标准 Transformer 的流匹配策略（0.9B）能跨本体联合训练；在 6 个仿真基准和 3 台真实机器人上评测。", "github_stars": 754, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2510.10274.html", "labs": [ "shanghai-ai-lab", "tsinghua" ] }, { "id": "arxiv-2510.13778", "title": "InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot Policy", "type": "paper", "arxiv": "2510.13778", "year": 2025, "venue": null, "org": "Shanghai AI Lab (Intern Robotics)", "url": null, "code": "https://github.com/InternRobotics/InternVLA-M1", "project": null, "topics": [ "vla" ], "tags": [ "real", "sim", "arm", "spatial", "grounding", "synthetic-data" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Pretrains spatial grounding on 2.3M samples ('where to act'), then post-trains actions with spatial prompts ('how to act'); reports +14.6% on SimplerEnv Google Robot over the same model without spatial guidance.", "takeaway_zh": "先在 230 万条空间推理样本上做空间定位预训练（\"在哪里做\"），再用空间提示做动作后训练（\"怎么做\"）；论文报告在 SimplerEnv Google Robot 上比无空间引导的同一模型高 14.6%。", "github_stars": 436, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2510.13778.html", "labs": [ "shanghai-ai-lab" ] }, { "id": "arxiv-2502.13130", "title": "Magma: A Foundation Model for Multimodal AI Agents", "type": "paper", "arxiv": "2502.13130", "year": 2025, "venue": "CVPR 2025", "org": "Microsoft Research", "url": null, "code": "https://github.com/microsoft/Magma", "project": "https://microsoft.github.io/Magma", "topics": [ "vla" ], "tags": [ "real", "sim", "arm", "vlm", "agents", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "One model for UI navigation and robot manipulation, pretrained on images, videos and robot data labelled with Set-of-Mark (what to act on) and Trace-of-Mark (how things move); model and code are public.", "takeaway_zh": "同一个模型同时做界面操作与机器人操作：在图像、视频和机器人数据上预训练，用 Set-of-Mark 标注可操作对象、用 Trace-of-Mark 标注运动轨迹；模型与代码公开。", "github_stars": 1950, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.13130.html", "labs": [ "microsoft" ] }, { "id": "arxiv-2507.15597", "title": "Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos", "type": "paper", "arxiv": "2507.15597", "year": 2025, "venue": "ICML 2026", "org": "BeingBeyond", "url": null, "code": "https://github.com/BeingBeyond/Being-H0", "project": "https://beingbeyond.github.io/Being-H0", "topics": [ "vla", "manipulation" ], "tags": [ "real", "dexterous-hand", "human-video", "motion-tokenization" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "A dexterous VLA pretrained on human hand motion from mocap, VR and RGB video ('physical instruction tuning'), with part-level motion tokens, then aligned to 3D space and post-trained for real robot hands.", "takeaway_zh": "灵巧手 VLA：先在动捕、VR 与普通视频中的人手动作上预训练（\"物理指令微调\"），使用部件级动作 token，再做三维空间对齐，并针对真实机器人灵巧手后训练。", "github_stars": 59, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2507.15597.html", "labs": [] }, { "id": "arxiv-2509.11766", "title": "Igniting VLMs toward the Embodied Space", "type": "paper", "arxiv": "2509.11766", "year": 2025, "venue": null, "org": "X Square Robot", "url": null, "code": "https://github.com/X-Square-Robot/wall-x", "project": null, "topics": [ "vla" ], "tags": [ "real", "arm", "vlm", "reasoning", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "WALL-OSS: an open embodied foundation model that links instruction reasoning, subgoal decomposition and action generation in one differentiable model ('unified cross-level CoT'); training code and checkpoints are public.", "takeaway_zh": "WALL-OSS：开源具身基础模型，在同一可微模型中串联指令推理、子目标分解与细粒度动作生成（\"统一跨层级思维链\"）；训练代码与模型权重已开源。", "github_stars": 1305, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2509.11766.html", "labs": [] }, { "id": "arxiv-2409.20537", "title": "Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers", "type": "paper", "arxiv": "2409.20537", "year": 2024, "venue": "NeurIPS 2024", "org": "MIT", "url": null, "code": "https://github.com/liruiw/HPT", "project": "https://liruiw.github.io/hpt/", "topics": [ "vla", "learning" ], "tags": [ "real", "sim", "cross-embodiment", "pretraining", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "HPT: per-robot stems turn proprioception and vision into a few tokens for one shared transformer trunk, pretrained on 52 datasets (robot, sim, human video); reports 20%+ gains on unseen tasks.", "takeaway_zh": "HPT：为每种机器人设计“输入头”，把本体感知与视觉映射成短 token 序列，送入共享的 Transformer 主干；在 52 个数据集（真机、仿真、人类视频）上预训练，报告在未见任务上微调提升 20% 以上。", "github_stars": 546, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2409.20537.html", "labs": [ "mit" ] }, { "id": "arxiv-2410.11758", "title": "Latent Action Pretraining from Videos", "type": "paper", "arxiv": "2410.11758", "year": 2024, "venue": "ICLR 2025", "org": "KAIST / UW / Microsoft Research / NVIDIA", "url": null, "code": "https://github.com/LatentActionPretraining/LAPA", "project": "https://latentactionpretraining.github.io/", "topics": [ "vla", "learning" ], "tags": [ "real", "sim", "arm", "latent-actions", "human-video", "pretraining" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "LAPA pretrains a VLA without action labels: it learns discrete latent actions between video frames, predicts them from image + text, then maps them to robot actions; reports beating OpenVLA on real tasks.", "takeaway_zh": "LAPA 无需动作标注即可预训练 VLA：先从视频相邻帧间学习离散隐动作，再训练模型由图像和指令预测隐动作，最后用少量机器人数据映射到真实动作；报告在真机任务上优于 OpenVLA。", "github_stars": 585, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.11758.html", "labs": [ "nvidia", "uw", "microsoft", "kaist" ] }, { "id": "arxiv-2502.05485", "title": "HAMSTER: Hierarchical Action Models for Open-World Robot Manipulation", "type": "paper", "arxiv": "2502.05485", "year": 2025, "venue": "ICLR 2025", "org": "NVIDIA / UW / USC", "url": null, "code": null, "project": "https://hamster-robo.github.io/", "topics": [ "vla", "manipulation" ], "tags": [ "real", "arm", "hierarchical", "vlm-planner", "2d-paths" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A hierarchical VLA in which a fine-tuned VLM draws a coarse 2D path for the end effector on the image and a low-level 3D policy follows it, so the VLM can be trained on cheap off-domain data such as sim or videos.", "takeaway_zh": "分层 VLA：微调后的 VLM 在图像上画出末端执行器的粗略二维路径，再由底层三维策略跟随执行；因此 VLM 可以用仿真、视频等低成本的非本域数据训练。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.05485.html", "labs": [ "nvidia", "uw", "usc" ] }, { "id": "arxiv-2403.09631", "title": "3D-VLA: A 3D Vision-Language-Action Generative World Model", "type": "paper", "arxiv": "2403.09631", "year": 2024, "venue": "ICML 2024", "org": "UMass Amherst / MIT", "url": null, "code": "https://github.com/UMass-Embodied-AGI/3D-VLA", "project": "https://vis-www.cs.umass.edu/3dvla", "topics": [ "vla", "world-models" ], "tags": [ "3d", "point-cloud", "goal-generation", "dataset" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A 3D-aware LLM with scene, object and action tokens, aligned with diffusion models so it can imagine goal images and point clouds before planning actions; trained on ~2M 3D-language-action pairs the authors assembled.", "takeaway_zh": "具备三维感知的大语言模型，引入场景、物体与动作交互 token，并与扩散模型对齐，可在规划动作前“想象”目标图像和点云；训练数据为作者整理的约 200 万条三维-语言-动作样本。", "github_stars": 639, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2403.09631.html", "labs": [ "mit" ] }, { "id": "arxiv-2602.14979", "title": "RynnBrain: Open Embodied Foundation Models", "type": "paper", "arxiv": "2602.14979", "year": 2026, "venue": null, "org": "Alibaba DAMO Academy", "url": null, "code": "https://github.com/alibaba-damo-academy/RynnBrain", "project": "https://alibaba-damo-academy.github.io/RynnBrain.github.io", "topics": [ "vla", "perception", "navigation" ], "tags": [ "vlm", "embodied-brain", "spatiotemporal", "planning", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open spatiotemporal embodied foundation models (2B, 8B, 30B-A3B MoE) for egocentric understanding, localization, grounded reasoning and planning, plus post-trained Nav, Plan, VLA and CoP variants.", "takeaway_zh": "开源时空具身基础模型（2B、8B 与 30B-A3B MoE），覆盖第一视角理解、时空定位、物理落地推理与规划，并提供导航、规划、VLA、CoP 等后训练版本。", "github_stars": 899, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2602.14979.html", "labs": [ "alibaba" ] }, { "id": "arxiv-2610.31859", "title": "Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates", "type": "paper", "arxiv": "2610.31859", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": null, "topics": [ "vla", "learning" ], "tags": [ "fine-tuning", "training-recipe", "mixed-precision", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Practical warning: fine-tuning VLA weights in bf16 without an fp32 master copy rounds away AdamW updates below 1/512 of a weight; finds this recipe in 9 of 12 open VLA codebases and 79%+ of updates lost on LIBERO.", "takeaway_zh": "实用警示：VLA 权重以 bf16 微调且不保留 fp32 主副本时，小于权重 1/512 的 AdamW 更新会被舍入抹掉；报告 12 个开源 VLA 代码库中 9 个的默认或文档配方如此训练，LIBERO 实验中至少 79% 的更新被抹除。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31859.html", "labs": [] }, { "id": "arxiv-2410.06158", "title": "GR-2: A Generative Video-Language-Action Model with Web-Scale Knowledge for Robot Manipulation", "type": "paper", "arxiv": "2410.06158", "year": 2024, "venue": null, "org": "ByteDance Research", "url": null, "code": null, "project": "https://gr2-manipulation.github.io", "topics": [ "vla", "world-models" ], "tags": [ "real", "arm", "video-pretraining", "generalist" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Pretrained to generate future video on 38M clips, then fine-tuned to predict frames and actions together; reports 97.7% average success over 100+ real tabletop tasks and roughly double GR-1's success in unseen scenes.", "takeaway_zh": "先在 3800 万段视频上预训练未来帧生成，再在机器人数据上微调，同时预测视频帧与动作轨迹；报告在 100 多个真机桌面任务上平均成功率 97.7%，在未见场景中成功率约为 GR-1 的两倍。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.06158.html", "labs": [ "bytedance" ] }, { "id": "arxiv-2412.10345", "title": "TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies", "type": "paper", "arxiv": "2412.10345", "year": 2024, "venue": "ICLR 2025", "org": "UMD / Microsoft Research", "url": null, "code": "https://github.com/umd-huang-lab/tracevla", "project": "https://tracevla.github.io/", "topics": [ "vla" ], "tags": [ "real", "sim", "arm", "visual-prompting", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Draws the robot's recent motion traces on the input image ('visual trace prompting') and fine-tunes OpenVLA on 150K trajectories; reports ~10% gains on SimplerEnv and 3.5x on real WidowX tasks; 4B variant too.", "takeaway_zh": "把机器人近期的运动轨迹画在输入图像上（“视觉轨迹提示”），并在 15 万条轨迹上微调 OpenVLA；报告在 SimplerEnv 上提升约 10%、真实 WidowX 任务上提升 3.5 倍，另有基于 Phi-3-Vision 的 4B 精简版。", "github_stars": 76, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2412.10345.html", "labs": [ "microsoft" ] }, { "id": "arxiv-2505.11917", "title": "OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning", "type": "paper", "arxiv": "2505.11917", "year": 2025, "venue": "ICLR 2026", "org": "Tsinghua / Shanghai Qi Zhi / Shanghai AI Lab / Fudan / Spirit AI", "url": null, "code": "https://github.com/Fanqi-Lin/OneTwoVLA", "project": "https://one-two-vla.github.io", "topics": [ "vla" ], "tags": [ "real", "arm", "reasoning", "long-horizon", "hri" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "One model that decides when to think and when to act: it reasons at key moments (subtask done, error, human input) and otherwise acts on its latest reasoning; co-trained with synthetic reasoning-rich VL data.", "takeaway_zh": "单一模型自主决定何时“想”、何时“做”：在子任务完成、出错或人类插话等关键时刻进行推理，其余时间依据最新推理结果生成动作；并与合成的推理型视觉-语言数据共同训练。", "github_stars": 242, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2505.11917.html", "labs": [ "shanghai-ai-lab", "tsinghua", "qizhi", "fudan" ] }, { "id": "arxiv-2502.05855", "title": "DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control", "type": "paper", "arxiv": "2502.05855", "year": 2025, "venue": "CoRL 2025", "org": "Midea Group / East China Normal University", "url": null, "code": null, "project": "https://dex-vla.github.io/", "topics": [ "vla" ], "tags": [ "real", "sim", "cross-embodiment", "diffusion-policy", "bimanual", "dexterous-hand" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "Pairs a 2B Qwen2-VL with a ~1B plug-in diffusion action expert with one head per robot, trained in three stages (cross-embodiment, alignment, task); reports shirt folding without task-specific tuning and 97.3 on LIBERO.", "takeaway_zh": "将 2B 的 Qwen2-VL 与约 1B 参数、按机器人分头输出的可插拔扩散动作专家结合，分三阶段训练（跨本体预训练、本体对齐、任务适配）；报告无需任务专门微调即可叠衬衫，LIBERO 平均 97.3。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.05855.html", "labs": [ "midea", "ecnu" ] }, { "id": "arxiv-2409.12514", "title": "TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation", "type": "paper", "arxiv": "2409.12514", "year": 2024, "venue": null, "org": "East China Normal University / Midea Group", "url": null, "code": null, "project": "https://tiny-vla.github.io/", "topics": [ "vla" ], "tags": [ "real", "sim", "arm", "bimanual", "efficient", "diffusion-policy" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "A compact VLA built from a small multimodal model (up to 1.4B) with LoRA and a diffusion policy head, without large-scale robot pretraining; reports higher real-robot success than OpenVLA and ~20x faster inference.", "takeaway_zh": "由小型多模态模型（至多 1.4B）加 LoRA 微调与扩散策略头构成的紧凑 VLA，无需大规模机器人预训练；报告真机成功率高于 OpenVLA，推理速度约快 20 倍。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2409.12514.html", "labs": [ "midea", "ecnu" ] }, { "id": "arxiv-2502.14420", "title": "ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model", "type": "paper", "arxiv": "2502.14420", "year": 2025, "venue": "EMNLP 2025", "org": "Midea Group / East China Normal University / Shanghai University / Beijing Innovation Center of Humanoid Robotics / Tsinghua", "url": null, "code": "https://github.com/tutujingyugang1/ChatVLA_public", "project": "https://chatvla.github.io", "topics": [ "vla" ], "tags": [ "real", "arm", "vlm", "mixture-of-experts" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Keeps a VLA able to chat: phased alignment training plus a mixture-of-experts split to fight forgetting and task interference; reports strong VQA/MMMU scores and beats OpenVLA on 25 real tasks.", "takeaway_zh": "让 VLA 保留“对话”能力：通过分阶段对齐训练和混合专家结构缓解遗忘与任务干扰；论文报告在保持 VQA/MMMU 等多模态理解成绩的同时，在 25 个真实操作任务上优于 OpenVLA。", "github_stars": 15, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.14420.html", "labs": [ "tsinghua", "midea", "ecnu" ] }, { "id": "arxiv-2505.21432", "title": "Hume: Introducing System-2 Thinking in Visual-Language-Action Model", "type": "paper", "arxiv": "2505.21432", "year": 2025, "venue": null, "org": "Shanghai AI Lab / SJTU / Fudan / AgiBot", "url": null, "code": null, "project": "https://hume-vla.github.io", "topics": [ "vla" ], "tags": [ "real", "sim", "dual-system", "flow-matching" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Dual-system VLA: a value-query head scores sampled action-chunk candidates (System 2, ~4 Hz) and a light cascaded denoiser refines it (System 1, ~90 Hz); reports gains over π0 on LIBERO, SIMPLER and real robots.", "takeaway_zh": "双系统 VLA：价值查询头对多组候选动作块打分并择优（System 2，约 4 Hz），轻量级级联去噪策略再将其细化为流畅动作（System 1，约 90 Hz）；论文报告在 LIBERO、SIMPLER 和真机上均优于 π0。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2505.21432.html", "labs": [ "shanghai-ai-lab", "sjtu", "agibot", "fudan" ] }, { "id": "arxiv-2511.05275", "title": "TwinVLA: Data-Efficient Bimanual Manipulation with Twin Single-Arm Vision-Language-Action Models", "type": "paper", "arxiv": "2511.05275", "year": 2025, "venue": "ICLR 2026", "org": "Yonsei University / Microsoft Research", "url": null, "code": "https://github.com/jellyho/TwinVLA", "project": "https://jellyho.github.io/TwinVLA/", "topics": [ "vla" ], "tags": [ "real", "sim", "bimanual" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "Builds a bimanual VLA by composing two copies of a pretrained single-arm VLA instead of pretraining on bimanual data; reports beating a similar-size RDT-1B and narrowing the gap to π0 on real and sim tasks.", "takeaway_zh": "不做双臂预训练，而是把两份预训练单臂 VLA 组合成协同的双臂 VLA；论文报告在真实与仿真双臂任务上优于同规模 RDT-1B，并缩小了与依赖大量双臂数据的 π0 的差距。", "github_stars": 22, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2511.05275.html", "labs": [ "microsoft" ] }, { "id": "arxiv-2507.01016", "title": "VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers", "type": "paper", "arxiv": "2507.01016", "year": 2025, "venue": "ICCV 2025", "org": "Shanghai AI Lab / SJTU", "url": null, "code": "https://github.com/xiaoxiao0406/VQ-VLA", "project": null, "topics": [ "vla" ], "tags": [ "real", "sim", "arm", "action-tokenizer" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "Replaces OpenVLA's per-dimension action bins with a convolutional residual VQ-VAE tokenizer trained on 100x more (real + synthetic) data; reports higher LIBERO-90 and real Franka success and ~2.8x faster inference.", "takeaway_zh": "用卷积残差 VQ-VAE 动作分词器（以百倍规模的真实+合成数据训练）替换 OpenVLA 的逐维离散分箱；论文报告 LIBERO-90 和真实 Franka 任务成功率提升，推理速度约快 2.8 倍。", "github_stars": 135, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2507.01016.html", "labs": [ "shanghai-ai-lab", "sjtu" ] }, { "id": "arxiv-2510.19430", "title": "GigaBrain-0: A World Model-Powered Vision-Language-Action Model", "type": "paper", "arxiv": "2510.19430", "year": 2025, "venue": null, "org": "GigaAI", "url": null, "code": "https://github.com/open-gigaai/giga-brain-0", "project": "https://gigabrain0.github.io", "topics": [ "vla" ], "tags": [ "real", "bimanual", "mobile-manipulation", "world-model-data", "flow-matching", "edge-deployment" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "Trains a VLA largely on world-model-generated data (re-rendered real videos, new viewpoints, sim2real, human-to-robot video); reports gains over π0 on dexterous and mobile tasks, plus a 402M variant for Jetson Orin.", "takeaway_zh": "VLA 训练大量依赖世界模型生成的数据（真实视频重渲染、新视角、仿真转真实、人类视频转机器人）；论文报告在灵巧与移动操作任务上优于 π0，并提供可在 Jetson Orin 上运行的 4.02 亿参数小模型。", "github_stars": 2666, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2510.19430.html", "labs": [] }, { "id": "arxiv-2509.00576", "title": "Galaxea Open-World Dataset and G0 Dual-System VLA Model", "type": "paper", "arxiv": "2509.00576", "year": 2025, "venue": null, "org": "Galaxea", "url": null, "code": "https://github.com/OpenGalaxea/GalaxeaVLA", "project": null, "topics": [ "vla", "datasets" ], "tags": [ "real", "mobile-manipulation", "bimanual", "dual-system", "flow-matching", "dataset" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "~500 h / 100K trajectories on one mobile bimanual robot (R1 Lite) across 50 real scenes, plus G0: a VLM planner over a flow-matching VLA; reports that single-embodiment pretraining matters most for few-shot transfer.", "takeaway_zh": "在单一移动双臂机器人（R1 Lite）上采集约 500 小时、10 万条轨迹、覆盖 50 个真实场景，并提出 G0：VLM 规划器 + 流匹配 VLA 执行器的双系统模型；论文报告单本体预训练对少样本迁移最关键。", "github_stars": 820, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2509.00576.html", "labs": [] }, { "id": "arxiv-2505.03673", "title": "RoboOS: A Hierarchical Embodied Framework for Cross-Embodiment and Multi-Agent Collaboration", "type": "paper", "arxiv": "2505.03673", "year": 2025, "venue": null, "org": "BAAI / Peking University", "url": null, "code": "https://github.com/FlagOpen/RoboOS", "project": null, "topics": [ "vla" ], "tags": [ "real", "multi-robot", "cross-embodiment", "planning", "system" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open-source 'brain–cerebellum' stack: a RoboBrain MLLM plans and assigns tasks across heterogeneous robots, per-robot skill libraries execute, and a shared spatial/temporal/robot memory keeps them in sync.", "takeaway_zh": "开源的“大脑–小脑”分层系统：RoboBrain 多模态大模型负责跨异构机器人的任务规划与分配，各机器人上的技能库负责执行，共享的空间/时间/本体记忆保持多机协同。", "github_stars": 632, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2505.03673.html", "labs": [ "pku" ] }, { "id": "arxiv-2610.30051", "title": "From Reasoning to Action: Benchmarking Frontier AI Models on LIBERO-10 Robot Manipulation Tasks", "type": "paper", "arxiv": "2610.30051", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/IRVLUTD/LLM_LIBERO", "project": "https://irvlutd.github.io/LLM-LIBERO/", "topics": [ "vla", "manipulation" ], "tags": [ "benchmark", "sim", "llm", "tool-use", "zero-shot", "reasoning" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Six frontier general-purpose models drive a LIBERO-10 arm through tool calls (end-effector poses, gripper), no robot training; best reported success is 44%. More reasoning effort usually helps but costs more.", "takeaway_zh": "让六个前沿通用大模型通过工具调用（指定末端位姿与夹爪开合）直接控制 LIBERO-10 机械臂，无任何机器人专门训练；报告的最高成功率为 44%。提高推理强度通常有帮助，但推理成本更高。", "github_stars": 8, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30051.html", "labs": [] }, { "id": "arxiv-2610.31427", "title": "FastHarness: Experience-Driven Review Scheduling for Vision-Language-Action Policies", "type": "paper", "arxiv": "2610.31427", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/FastHarness/fast-harness", "project": null, "topics": [ "vla", "manipulation" ], "tags": [ "sim", "agentic", "harness", "llm", "failure-recovery", "frozen-policy" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Harness where a frontier LLM reviews a frozen π0.5 policy less often as experience builds up and focuses on recovery after errors; reports ~57% fewer reviewer tokens and higher success on four sim tasks.", "takeaway_zh": "让前沿大模型监督冻结的 π0.5 策略的执行框架：随着经验积累逐步减少审查频率，并把监督集中在出错后的恢复上；报告在四个仿真任务上审查 token 平均减少约 57%，成功率也更高。", "github_stars": 9, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31427.html", "labs": [] }, { "id": "arxiv-2610.25874", "title": "Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation", "type": "paper", "arxiv": "2610.25874", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/baidu-baige/LoongForge/tree/master/cookbook/loongforge-vlda", "project": null, "topics": [ "vla", "world-models", "manipulation" ], "tags": [ "latent-dynamics", "action-refinement", "open-framework" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Between direct action prediction and full video prediction: VLDA learns a compact latent 'Delta' of task-relevant state change and uses it to seed action denoising. Built on Baidu's open LoongForge.", "takeaway_zh": "介于直接预测动作与完整预测未来画面之间：VLDA 学习一个紧凑的隐变量 Delta 表示与任务相关的状态变化，并用它引导动作去噪；实现于百度开源训练框架 LoongForge。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25874.html", "labs": [] }, { "id": "arxiv-2610.32102", "title": "Coupling Execution Horizon and Sampling Depth in Flow-Matching VLA Policies", "type": "paper", "arxiv": "2610.32102", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://jiabingyang01.github.io/Horizon-Depth-VLA/", "topics": [ "vla", "manipulation" ], "tags": [ "deployment", "inference", "flow-matching", "training-free", "sim", "real", "arm" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Deployment knob for flow-matching VLAs: fewer solver steps hurt under fast replanning but help with long open-loop chunks. Tuning both together, training-free, reports higher success and lower latency for π0/π0.5/GR00T.", "takeaway_zh": "流匹配 VLA 的部署调参发现：求解步数少在频繁重规划时会降低精度，但在长时开环执行动作块时反而有利。报告在 π0、π0.5、GR00T 上联合调整两者，无需训练即可提高成功率并降低推理延迟。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.32102.html", "labs": [] }, { "id": "arxiv-2610.35832", "title": "Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy", "type": "paper", "arxiv": "2610.35832", "year": 2026, "venue": null, "org": "InternRobotics", "url": null, "code": null, "project": "https://internrobotics.github.io/EBench-Astra-Tech-Report/", "topics": [ "vla", "manipulation" ], "tags": [ "benchmark", "llm", "harness", "evaluation", "frontier-model", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Report: GPT-6-Astra drives a robot on EBench via waypoints or by calling a VLA/WAM policy. Direct control leads on short tasks, trails on long-horizon and contact-rich ones; calling a policy cuts token cost.", "takeaway_zh": "评测报告：在 EBench 上让 GPT-6-Astra 直接输出路径点控制机器人，或通过执行框架调用 VLA/WAM 策略。直接控制在短时任务上领先，但在长时程与接触密集任务上落后；调用策略可显著减少 token 开销。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35832.html", "labs": [ "shanghai-ai-lab" ] }, { "id": "arxiv-2610.35626", "title": "Robot-Aligned 3D Gaussian Splatting for Precise Reaching with Wrist-Camera VLA Policies", "type": "paper", "arxiv": "2610.35626", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://001-wang.github.io/PrecisionReach-GS/", "topics": [ "vla", "perception" ], "tags": [ "gaussian-splatting", "wrist-camera", "data-augmentation", "real", "arm" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Renders wrist-camera views from a 3D Gaussian Splatting scan of the real workspace to fine-tune π0.5 for precise reaching in e-waste disassembly; reports 80% vs. 58.7% with 150 teleop demos.", "takeaway_zh": "用真实工作台的 3D 高斯泼溅重建渲染腕部相机训练视角，微调 π0.5 以完成电子垃圾拆解中的精确抵达；报告成功率 80%，而 150 条遥操作示教为 58.7%，人工采集不到 10 分钟。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35626.html", "labs": [] }, { "id": "arxiv-2610.33074", "title": "AGQ: Action-Guided Ultra-Low-Bit Quantization for Vision-Language-Action Models", "type": "paper", "arxiv": "2610.33074", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/Qiuxu00/AGQ", "project": null, "topics": [ "vla", "manipulation" ], "tags": [ "quantization", "compression", "deployment", "edge", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Mixed-precision quantization for VLAs that sets each weight matrix's bit width by how much it affects actions; reports 92.7% LIBERO success at 3 bits (above AWQ and GPTQ) and 70.5% at 2.5 bits.", "takeaway_zh": "面向 VLA 的混合精度量化：根据每个权重矩阵对输出动作的影响分配比特数；报告在 3 比特下 LIBERO 成功率 92.7%（高于 AWQ 与 GPTQ），2.5 比特下仍有 70.5%。", "github_stars": 0, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.33074.html", "labs": [] }, { "id": "arxiv-2610.34356", "title": "RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions", "type": "paper", "arxiv": "2610.34356", "year": 2026, "venue": null, "org": "Alibaba DAMO Academy", "url": null, "code": "https://github.com/alibaba-damo-academy/RynnVLA-Latent", "project": "https://alibaba-damo-academy.github.io/RynnVLA-Latent.github.io/", "topics": [ "vla", "manipulation", "learning" ], "tags": [ "latent-actions", "pretraining", "human-video", "flow-matching", "cross-embodiment" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Pretrains a 4B VLA on 3D-grounded latent actions mined from 30K hours of human, robot and sim video, then adapts to real robot actions. Reports 98.8% on LIBERO and 88.6% on RoboTwin 2.0.", "takeaway_zh": "从 3 万小时人类、机器人与仿真视频中学习带 3D 约束的连续潜在动作（RynnLAM），用于预训练 4B 参数 VLA，再适配到具体机器人动作空间。报告 LIBERO 98.8%、RoboTwin 2.0 平均 88.6%。", "github_stars": 11, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.34356.html", "labs": [ "alibaba" ] }, { "id": "arxiv-2610.35051", "title": "VISTA: Active Exploration for Self-Improvement in Embodied Agents", "type": "paper", "arxiv": "2610.35051", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/AgenticEI/VISTA", "project": "https://agenticei.github.io/VISTA-page/", "topics": [ "vla", "learning" ], "tags": [ "harness", "agent", "self-improvement", "libero", "robotwin", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Meta-agent that improves a robot's harness by exploring: keeps a lineage graph of harness versions and runs new rollouts to probe failures before editing. Reports more success with fewer rollouts than static reflection.", "takeaway_zh": "通过主动探索改进机器人 harness 的元智能体：维护 harness 版本谱系图，先主动生成新回放探查失败与恢复行为再修改；报告在 LIBERO-Pro、RoboCasa、RoboTwin 上比静态反思用更少仿真回放取得更高成功率。", "github_stars": 2, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35051.html", "labs": [] }, { "id": "arxiv-2610.31569", "title": "Echo: Coupled Action and Task Status Signal Denoising for Vision Language Action Models", "type": "paper", "arxiv": "2610.31569", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/lyfeng001/Echo-Sim", "project": "https://lyfeng001.github.io/Echo/", "topics": [ "vla", "manipulation" ], "tags": [ "vla", "failure-detection", "recovery", "diffusion", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Teaches a VLA to know when it is done or failing: stop and failure signals are two extra action dimensions, denoised faster than the actions so they can guide them; no frame-level progress labels. Real-robot tests.", "takeaway_zh": "让 VLA 自己判断任务完成或失败：把停止与失败信号作为动作块的两个额外维度，与动作一起去噪且去噪更快，从而引导后续动作生成；无需逐帧进度标注，在真机操作实验中验证。", "github_stars": 2, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31569.html", "labs": [] }, { "id": "arxiv-2610.27794", "title": "Multi-View Latent Action Pretraining for Cross-Embodiment Robot Learning from Videos", "type": "paper", "arxiv": "2610.27794", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://learnsyslab.github.io/multiview_latent_action_pretraining/", "topics": [ "vla", "learning" ], "tags": [ "latent-actions", "multi-view", "cross-embodiment", "human-video", "flow-matching" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Latent action model that fuses synchronized camera views to resolve depth and occlusion. Reports 17% better 6D action decodability than single-view and +13.5% policy success (vs +6.2%); human-hand demos help on hardware.", "takeaway_zh": "融合多路同步相机视角的潜在动作模型（LAM），用交叉注意力解决深度与遮挡问题；报告 6D 动作可解码性比单视角高 17%，下游策略成功率提升 13.5%（单视角为 6.2%），真机上用人手演示也比纯动作策略高 20%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.27794.html", "labs": [] }, { "id": "arxiv-2610.37219", "title": "Vision-Language-Action or World-Action? Just Action Models", "type": "paper", "arxiv": "2610.37219", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/y0urOy/JAM", "project": "https://y0uroy.github.io/JAM", "topics": [ "vla", "world-models" ], "tags": [ "vla", "wam", "foresight-loss", "libero", "robotwin" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Splits the VLA vs WAM debate into backbone, future loss and inference machinery: a foresight loss predicts frozen future features, no video generation. Reports beating FastWAM on LIBERO-Plus (64.5% vs 55.3%).", "takeaway_zh": "把 VLA 与世界动作模型（WAM）之争拆成骨干初始化、未来预测损失和推理时是否保留预测模块三个选择：用预测冻结未来特征的前瞻损失，训练与推理都不生成视频；报告在 LIBERO-Plus 上 64.5%，FastWAM 为 55.3%。", "github_stars": 4, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.37219.html", "labs": [] }, { "id": "arxiv-2610.25961", "title": "PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models", "type": "paper", "arxiv": "2610.25961", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://pacevla.github.io/", "topics": [ "vla", "manipulation" ], "tags": [ "vla", "inference-speed", "action-chunking", "real", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Lets a VLA pick, at each step, between fine and temporally coarser action decoders: redundant motion runs faster while precision phases stay fine-grained. Tested on 3 sim benchmarks and 6 real tasks.", "takeaway_zh": "让 VLA 在每一步自行选择细粒度或时间上更粗的动作解码器：冗余动作阶段加速执行，精细接触阶段保持细粒度控制；在 3 个仿真基准和 6 个真实任务上验证效率提升且成功率不降。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25961.html", "labs": [] }, { "id": "arxiv-2610.36013", "title": "OmniBrain-0: A Unified Agentic Model for Embodied Intelligence", "type": "paper", "arxiv": "2610.36013", "year": 2026, "venue": null, "org": "Xiaomi", "url": null, "code": null, "project": "https://public-bots.github.io/OmniBrain-0/", "topics": [ "vla", "navigation" ], "tags": [ "embodied-agent", "planning", "distillation", "self-improvement", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Xiaomi's compact embodied agent that plans, tracks progress and recovers from failures across manipulation and navigation, distilled from stronger teachers with a rubric-driven self-improvement loop.", "takeaway_zh": "小米具身智能团队的轻量统一具身智能体：在操作与导航中负责任务理解、规划、进度跟踪和失败恢复，由更强的教师模型多教师在线蒸馏而来，并配有基于评分细则的自我改进数据飞轮。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36013.html", "labs": [] }, { "id": "arxiv-2610.37666", "title": "ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning", "type": "paper", "arxiv": "2610.37666", "year": 2026, "venue": null, "org": "NVIDIA", "url": null, "code": null, "project": "https://thinkworld-project.github.io/", "topics": [ "vla", "world-models", "learning" ], "tags": [ "vla", "human-video", "motion-tokens", "reinforcement-learning", "cross-embodiment", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "VLA that reasons in discrete motion tokens (a chain of dynamics) shared across embodiments; predicted futures are scored against what actually happened, so action-free human videos give RL signal. Sim and real tests.", "takeaway_zh": "VLA 用跨本体共享的离散运动 token 链（Chain-of-Dynamics）表达动作思考；预测每条链导致的未来并与真实结果比对来强化，使没有动作标注的人类视频也能提供监督。在仿真与真机上验证人到机器人的迁移。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.37666.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2610.37215", "title": "Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution", "type": "paper", "arxiv": "2610.37215", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://retrieve2act.github.io/r2a/", "topics": [ "vla", "manipulation" ], "tags": [ "test-time", "retrieval", "robustness", "ood", "gnn", "real", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Test-time add-on for frozen VLAs: a GNN retrieves future motion primitives from a graph of training demos and gates them into the actions. Reports up to +12% in sim and +40% real (GR00T N1.7, π0.5, StableVLA).", "takeaway_zh": "面向冻结 VLA 的测试时插件：把训练示范组织成运动基元图，用 GNN 定位当前执行并检索后续基元，再按维度门控融合到预测动作中；报告在 GR00T N1.7、π0.5 等上仿真最高 +12%、真机 +40%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.37215.html", "labs": [] }, { "id": "arxiv-2610.37381", "title": "PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations", "type": "paper", "arxiv": "2610.37381", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": null, "topics": [ "vla", "world-models" ], "tags": [ "vla", "physical-dynamics", "future-prediction", "ecot", "public-data", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Light module turns VLM features into 'physical tokens' trained to predict future RGB, depth and flow features (with world-model supervision); no online video generation. Reports 99.3% LIBERO, 93.6% RoboTwin 2.0.", "takeaway_zh": "轻量物理动力学模块从视觉-语言特征中提取“物理 token”，通过预测未来 RGB、深度和光流特征（并辅以冻结视频世界模型监督）来学习，推理时无需在线生成视频；仅用公开数据训练，报告 LIBERO 99.3%、RoboTwin 2.0 93.6%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.37381.html", "labs": [] }, { "id": "arxiv-2610.31423", "title": "FocusVLA: Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models", "type": "paper", "arxiv": "2610.31423", "year": 2026, "venue": "CoRL 2026", "org": "ISCAS", "url": null, "code": null, "project": null, "topics": [ "vla", "manipulation" ], "tags": [ "vla", "multi-view", "view-routing", "intent-prediction", "real", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Finds VLAs over-rely on the wrist camera; predicts future interaction intent from past actions to route between global and wrist views per stage. Reports +10.1 pp over π0.5 on RoboTwin 2.0 and 71.7% → 85.0% real.", "takeaway_zh": "发现现有 VLA 过度依赖腕部相机；根据历史动作预测未来交互意图，按任务阶段在全局与腕部视角间动态路由；报告在 RoboTwin 2.0 上比 π0.5 高 10.1 个百分点，真机成功率由 71.7% 提升到 85.0%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31423.html", "labs": [] }, { "id": "arxiv-2303.04137", "title": "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion", "type": "paper", "arxiv": "2303.04137", "year": 2023, "venue": null, "org": "Columbia, TRI, MIT", "url": null, "code": "https://github.com/real-stanford/diffusion_policy", "project": null, "topics": [ "manipulation" ], "tags": [ "real", "sim", "arm", "diffusion", "imitation" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "Generating action sequences by denoising handles multimodal demos far better than regression; now a standard IL baseline.", "takeaway_zh": "以去噪方式生成动作序列，比回归更能处理多模态演示，已成为模仿学习标准基线。", "github_stars": 4693, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2303.04137.html", "labs": [ "mit", "columbia", "tri" ] }, { "id": "arxiv-2304.13705", "title": "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware", "type": "paper", "arxiv": "2304.13705", "year": 2023, "venue": null, "org": "Stanford", "url": null, "code": "https://github.com/tonyzhaozh/act", "project": null, "topics": [ "manipulation", "hardware" ], "tags": [ "real", "bimanual", "imitation", "low-cost" ], "robot_types": [ "arm" ], "highlight": true, "added": null, "takeaway_en": "ALOHA + ACT — a ~$20k bimanual teleop rig and an action-chunking Transformer that learns precise tasks from ~50 demos.", "takeaway_zh": "ALOHA + ACT：约 2 万美元的双臂遥操作平台，加上动作分块 Transformer，约 50 条演示即可学会精细任务。", "github_stars": 2304, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2304.13705.html", "labs": [ "stanford" ] }, { "id": "arxiv-2401.02117", "title": "Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation", "type": "paper", "arxiv": "2401.02117", "year": 2024, "venue": null, "org": "Stanford", "url": null, "code": "https://github.com/MarkFzp/mobile-aloha", "project": null, "topics": [ "manipulation", "hardware" ], "tags": [ "real", "bimanual", "mobile", "imitation", "low-cost" ], "robot_types": [ "arm", "mobile" ], "highlight": false, "added": null, "takeaway_en": "Puts ALOHA on a mobile base; co-training with static ALOHA data boosts mobile tasks like cooking and opening cabinets.", "takeaway_zh": "将 ALOHA 装上移动底盘；与静态 ALOHA 数据联合训练显著提升做饭、开柜等移动任务表现。", "github_stars": 4477, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2401.02117.html", "labs": [ "stanford" ] }, { "id": "arxiv-2403.03954", "title": "3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations", "type": "paper", "arxiv": "2403.03954", "year": 2024, "venue": null, "org": "Shanghai Qi Zhi, Tsinghua", "url": null, "code": "https://github.com/YanjieZe/3D-Diffusion-Policy", "project": null, "topics": [ "manipulation" ], "tags": [ "real", "sim", "point-cloud", "diffusion" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Conditions a diffusion policy on a compact point-cloud encoding; more data-efficient and robust to viewpoint than images.", "takeaway_zh": "用紧凑点云编码作为扩散策略条件，比图像输入更省数据、对视角更鲁棒。", "github_stars": 1467, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2403.03954.html", "labs": [ "tsinghua", "qizhi" ] }, { "id": "arxiv-2402.10329", "title": "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots", "type": "paper", "arxiv": "2402.10329", "year": 2024, "venue": null, "org": "Stanford, Columbia, TRI", "url": null, "code": "https://github.com/real-stanford/universal_manipulation_interface", "project": null, "topics": [ "manipulation", "hardware", "datasets" ], "tags": [ "real", "arm", "handheld", "data-collection" ], "robot_types": [ "arm" ], "highlight": true, "added": null, "takeaway_en": "A handheld gripper with a GoPro collects demos anywhere; policies transfer directly to robot arms.", "takeaway_zh": "手持夹爪加 GoPro 可随处采集演示，训练出的策略可直接迁移到机械臂。", "github_stars": 1622, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2402.10329.html", "labs": [ "stanford", "columbia", "tri" ] }, { "id": "arxiv-2403.07788", "title": "DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation", "type": "paper", "arxiv": "2403.07788", "year": 2024, "venue": null, "org": "Stanford", "url": null, "code": null, "project": null, "topics": [ "manipulation", "hardware" ], "tags": [ "real", "hand", "data-collection" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "A wearable mocap glove + camera rig records human hand motion that is retargeted to train dexterous robot hands.", "takeaway_zh": "可穿戴动捕手套与相机采集人手动作，重定向后用于训练灵巧机器人手。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2403.07788.html", "labs": [ "stanford" ] }, { "id": "arxiv-1808.00177", "title": "Learning Dexterous In-Hand Manipulation", "type": "paper", "arxiv": "1808.00177", "year": 2018, "venue": null, "org": "OpenAI", "url": null, "code": null, "project": null, "topics": [ "manipulation" ], "tags": [ "sim2real", "hand", "rl" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "A Shadow Hand reorients a cube using RL trained purely in randomized simulation — a sim-to-real milestone.", "takeaway_zh": "完全在随机化仿真中用强化学习训练，Shadow 灵巧手在真实世界翻转方块，是 sim-to-real 里程碑。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1808.00177.html", "labs": [ "openai" ] }, { "id": "arxiv-1910.07113", "title": "Solving Rubik's Cube with a Robot Hand", "type": "paper", "arxiv": "1910.07113", "year": 2019, "venue": null, "org": "OpenAI", "url": null, "code": null, "project": null, "topics": [ "manipulation" ], "tags": [ "sim2real", "hand", "rl" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "Automatic domain randomization lets a sim-trained hand policy manipulate a Rubik's cube on real hardware.", "takeaway_zh": "自动域随机化让仿真训练的灵巧手策略在真实硬件上操作魔方。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1910.07113.html", "labs": [ "openai" ] }, { "id": "arxiv-2211.11744", "title": "Visual Dexterity: In-Hand Reorientation of Novel and Complex Object Shapes", "type": "paper", "arxiv": "2211.11744", "year": 2022, "venue": null, "org": "MIT", "url": null, "code": null, "project": null, "topics": [ "manipulation" ], "tags": [ "sim2real", "hand", "rl", "depth" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "Reorients unseen objects in-hand using only a depth camera, with a policy trained in simulation.", "takeaway_zh": "仅凭深度相机即可在手中翻转未见过的物体，策略在仿真中训练。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2211.11744.html", "labs": [ "mit" ] }, { "id": "arxiv-2307.05973", "title": "VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models", "type": "paper", "arxiv": "2307.05973", "year": 2023, "venue": null, "org": "Stanford", "url": null, "code": null, "project": null, "topics": [ "manipulation" ], "tags": [ "real", "arm", "llm", "zero-shot" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "LLMs write code that composes 3D affordance/constraint maps, which a planner turns into trajectories — no robot training.", "takeaway_zh": "大模型编写代码组合 3D 可供性/约束地图，再由规划器生成轨迹，无需机器人训练。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2307.05973.html", "labs": [ "stanford" ] }, { "id": "arxiv-2203.12601", "title": "R3M: A Universal Visual Representation for Robot Manipulation", "type": "paper", "arxiv": "2203.12601", "year": 2022, "venue": null, "org": "Stanford, Meta AI", "url": null, "code": null, "project": null, "topics": [ "manipulation" ], "tags": [ "representation", "egocentric-video" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Pretrains a visual encoder on Ego4D human videos that makes downstream imitation far more data-efficient.", "takeaway_zh": "在 Ego4D 人类视频上预训练视觉编码器，显著提高下游模仿学习的数据效率。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2203.12601.html", "labs": [ "stanford", "meta" ] }, { "id": "arxiv-2610.17350", "title": "Can a Frontier VLM Serve as a Robot Policy for Memory-Dependent Manipulation?", "type": "paper", "arxiv": "2610.17350", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/jackbjed/embodied-astra-robomme", "project": "https://embodied-astra.github.io", "topics": [ "manipulation" ], "tags": [ "sim", "vlm", "memory", "zero-shot", "benchmark" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "A frozen frontier VLM, prompted with the whole episode, outputs waypoints directly and reports 92.1% on RoboMME memory tasks with no task training.", "takeaway_zh": "冻结的前沿 VLM 以整段回合为上下文直接输出路径点，在 RoboMME 记忆类任务上报告 92.1% 成功率，且无需任务训练。", "github_stars": 3, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.17350.html", "labs": [] }, { "id": "arxiv-2610.23849", "title": "Blind Precise Assembly: Proprioceptive Sim-to-Real Peg-in-Hole for Diverse Parts", "type": "paper", "arxiv": "2610.23849", "year": 2026, "venue": null, "org": "MIT", "url": null, "code": null, "project": null, "topics": [ "manipulation", "sim" ], "tags": [ "sim-to-real", "assembly", "proprioception", "contact-rich" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "Asks whether joint readings alone suffice for millimetre-precise insertion — trains sim-to-real policies with no camera, force or tactile input.", "takeaway_zh": "探究仅靠关节读数能否完成毫米级精密插装：训练不用相机、力或触觉输入的 sim-to-real 策略。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.23849.html", "labs": [ "mit" ] }, { "id": "arxiv-2610.17802", "title": "Beyond the End-Effector: Learning Whole-Arm Contact Manipulation from Human Demonstrations", "type": "paper", "arxiv": "2610.17802", "year": 2026, "venue": "CoRL 2026", "org": "USC", "url": null, "code": null, "project": "https://wacl-robot.github.io", "topics": [ "manipulation" ], "tags": [ "whole-arm", "non-prehensile", "human-demos" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "WACL learns from human demos to push and hold with the whole arm, choosing where on the arm to make contact while keeping the gripper's constraints.", "takeaway_zh": "WACL 从人类示范中学习用整条手臂推、抵物体，自主选择手臂上的接触位置，同时满足末端执行器的约束。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.17802.html", "labs": [ "usc" ] }, { "id": "arxiv-2610.23479", "title": "HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents", "type": "paper", "arxiv": "2610.23479", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/HexaFuture/HexaShell", "project": null, "topics": [ "manipulation", "vla" ], "tags": [ "real", "bimanual", "vlm-agent", "code-as-policy" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-19", "takeaway_en": "A coding agent drives a bimanual robot through one command line; after each failure it reads the logs and writes or fixes small reusable tool programs.", "takeaway_zh": "编程智能体通过一条命令行驱动双臂机器人；每次失败后读取日志，编写或修复可复用的小工具程序。", "github_stars": 207, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.23479.html", "labs": [] }, { "id": "arxiv-2210.02697", "title": "DexGraspNet: A Large-Scale Robotic Dexterous Grasp Dataset for General Objects Based on Simulation", "type": "dataset", "arxiv": "2210.02697", "year": 2022, "venue": null, "org": "PKU, BIGAI", "url": null, "code": "https://github.com/PKU-EPIC/DexGraspNet", "project": null, "topics": [ "manipulation", "datasets" ], "tags": [ "sim", "dexterous-hand", "grasping", "large-scale" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "Over a million physically validated ShadowHand grasps on thousands of objects, synthesized by optimizing a differentiable force-closure energy; a standard training set for dexterous grasp models.", "takeaway_zh": "通过优化可微的力封闭能量合成、并经物理仿真验证的百万级 ShadowHand 抓取，覆盖数千个物体，是灵巧抓取模型的常用训练集。", "github_stars": 478, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2210.02697.html", "labs": [ "pku" ] }, { "id": "arxiv-2311.16098", "title": "On Bringing Robots Home", "type": "paper", "arxiv": "2311.16098", "year": 2023, "venue": null, "org": "NYU, Meta", "url": null, "code": "https://github.com/notmahi/dobb-e", "project": null, "topics": [ "manipulation" ], "tags": [ "home-robot", "low-cost-data", "imitation", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Dobb-E collects home demonstrations with a cheap reacher-grabber \"stick\", pretrains on them, then adapts to a new household task from a few minutes of data.", "takeaway_zh": "Dobb-E 用低成本的手持夹取“棒”采集家庭示范数据并预训练，在新家庭任务上只需几分钟数据即可适配。", "github_stars": 621, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2311.16098.html", "labs": [ "meta", "nyu" ] }, { "id": "arxiv-2610.23888", "title": "MoDex: Towards Universal Multi-Object Dexterous Grasping", "type": "paper", "arxiv": "2610.23888", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": "https://github.com/Aurora-Xiong/MoDex", "project": null, "topics": [ "manipulation" ], "tags": [ "dexterous-hand", "grasping", "multi-object", "residual-learning" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-20", "takeaway_en": "Grasps several objects at once with a dexterous hand by learning one-step residual corrections to a reference trajectory, with a strategy that adapts to how the objects are laid out.", "takeaway_zh": "让灵巧手一次抓取多个物体：学习对参考轨迹的单步残差修正，并根据物体的空间布局调整抓取策略。", "github_stars": 3, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.23888.html", "labs": [] }, { "id": "arxiv-2610.26507", "title": "Learn the Task, Not the Schedule: Reducing Spurious Cross-Arm Coupling in Bimanual Imitation Learning", "type": "paper", "arxiv": "2610.26507", "year": 2026, "venue": null, "org": "SJTU", "url": null, "code": null, "project": null, "topics": [ "manipulation", "learning" ], "tags": [ "bimanual", "imitation", "data-augmentation" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-20", "takeaway_en": "CTR retimes the two arms in bimanual demos at training time while keeping task-required constraints, so the policy stops using the other arm's progress as a shortcut cue.", "takeaway_zh": "CTR 在训练时对双臂示教中两只手臂的动作时序重新组合，同时保留任务必需的约束，避免策略把另一只手臂的进度当作捷径线索。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.26507.html", "labs": [ "sjtu" ] }, { "id": "arxiv-2010.14406", "title": "Transporter Networks: Rearranging the Visual World for Robotic Manipulation", "type": "paper", "arxiv": "2010.14406", "year": 2020, "venue": "CoRL 2020", "org": "Google", "url": null, "code": "https://github.com/google-research/ravens", "project": null, "topics": [ "manipulation" ], "tags": [ "sim", "real", "pick-and-place", "rearrangement", "equivariance" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Transporter Nets — pick-and-place as spatial matching of deep features cropped around the pick point; very sample-efficient on rearrangement tasks, released with the Ravens benchmark.", "takeaway_zh": "Transporter 网络：把抓取-放置建模为以抓取点为中心的深度特征的空间匹配，在重排类任务上样本效率很高，并随论文发布 Ravens 基准。", "github_stars": 629, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2010.14406.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2109.12098", "title": "CLIPort: What and Where Pathways for Robotic Manipulation", "type": "paper", "arxiv": "2109.12098", "year": 2021, "venue": "CoRL 2021", "org": "University of Washington, NVIDIA", "url": null, "code": "https://github.com/cliport/cliport", "project": null, "topics": [ "manipulation", "vla" ], "tags": [ "sim", "real", "language", "clip", "pick-and-place" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "CLIPort — combines CLIP's semantic \"what\" stream with Transporter's spatial \"where\" stream for language-conditioned tabletop manipulation; an early bridge from web vision-language models to robot actions.", "takeaway_zh": "CLIPort：把 CLIP 提供的语义\"是什么\"通路与 Transporter 的空间\"在哪里\"通路结合，实现语言条件的桌面操作；是较早把网络视觉语言模型接入机器人动作的工作。", "github_stars": 547, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2109.12098.html", "labs": [ "nvidia", "uw" ] }, { "id": "arxiv-2209.05451", "title": "Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation", "type": "paper", "arxiv": "2209.05451", "year": 2022, "venue": "CoRL 2022", "org": "University of Washington, NVIDIA", "url": null, "code": "https://github.com/peract/peract", "project": null, "topics": [ "manipulation", "vla" ], "tags": [ "sim", "real", "voxel", "rlbench", "language", "keyframe" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "PerAct — a Perceiver Transformer over voxelized RGB-D observations that predicts the next keyframe gripper pose as voxel classification; one language-conditioned policy for many RLBench and real tasks.", "takeaway_zh": "PerAct：在体素化的 RGB-D 观测上运行 Perceiver Transformer，把下一关键帧夹爪位姿当作体素分类来预测；一个语言条件策略即可覆盖众多 RLBench 与真机任务。", "github_stars": 497, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2209.05451.html", "labs": [ "nvidia", "uw" ] }, { "id": "arxiv-2306.14896", "title": "RVT: Robotic View Transformer for 3D Object Manipulation", "type": "paper", "arxiv": "2306.14896", "year": 2023, "venue": "CoRL 2023", "org": "NVIDIA", "url": null, "code": "https://github.com/NVlabs/RVT", "project": null, "topics": [ "manipulation" ], "tags": [ "sim", "real", "multi-view", "rlbench", "keyframe" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "RVT — re-renders the point cloud into a few virtual views and uses a multi-view Transformer to predict keyframe poses; reported to train and run much faster than voxel-based PerAct on RLBench. Repo also hosts RVT-2.", "takeaway_zh": "RVT：把点云重新渲染为若干虚拟视角，再用多视角 Transformer 预测关键帧位姿；论文报告在 RLBench 上训练与推理都比基于体素的 PerAct 快得多。代码库同时包含 RVT-2。", "github_stars": 409, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2306.14896.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2409.01652", "title": "ReKep: Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation", "type": "paper", "arxiv": "2409.01652", "year": 2024, "venue": "CoRL 2024", "org": "Stanford, Columbia", "url": null, "code": "https://github.com/huangwl18/ReKep", "project": null, "topics": [ "manipulation", "vla" ], "tags": [ "real", "zero-shot", "keypoints", "vlm", "optimization" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "ReKep — a VLM writes manipulation tasks as Python cost functions over 3D keypoints (relational keypoint constraints), solved by a hierarchical optimizer in real time; zero-shot, no task-specific training.", "takeaway_zh": "ReKep：由视觉语言模型把操作任务写成定义在三维关键点上的 Python 代价函数（关系关键点约束），再由分层优化器实时求解；零样本，无需针对任务训练。", "github_stars": 992, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2409.01652.html", "labs": [ "stanford", "columbia" ] }, { "id": "arxiv-2103.14127", "title": "Contact-GraspNet: Efficient 6-DoF Grasp Generation in Cluttered Scenes", "type": "paper", "arxiv": "2103.14127", "year": 2021, "venue": "ICRA 2021", "org": "NVIDIA", "url": null, "code": "https://github.com/NVlabs/contact_graspnet", "project": null, "topics": [ "manipulation", "perception" ], "tags": [ "grasping", "point-cloud", "6-dof", "clutter" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Contact-GraspNet — predicts 6-DoF parallel-jaw grasps for a whole cluttered scene from one depth view by treating observed points as grasp contacts, which shrinks the problem to 4 DoF per point.", "takeaway_zh": "Contact-GraspNet：从单帧深度图为整个杂乱场景预测 6 自由度平行夹爪抓取；把观测到的点当作抓取接触点，使每个点只需学习 4 个自由度。", "github_stars": 542, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2103.14127.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2212.08333", "title": "AnyGrasp: Robust and Efficient Grasp Perception in Spatial and Temporal Domains", "type": "paper", "arxiv": "2212.08333", "year": 2022, "venue": null, "org": "SJTU", "url": null, "code": "https://github.com/graspnet/anygrasp_sdk", "project": null, "topics": [ "manipulation", "perception" ], "tags": [ "grasping", "real", "dynamic", "tracking" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "AnyGrasp — dense 7-DoF parallel-gripper grasp detection trained on real sensor data with analytic labels and center-of-mass awareness, plus grasp tracking across frames for moving objects; released as an SDK.", "takeaway_zh": "AnyGrasp：基于真实传感器数据与解析标注训练、考虑物体质心的稠密 7 自由度平行夹爪抓取检测，并能跨帧跟踪抓取以应对运动物体；以 SDK 形式发布。", "github_stars": 1046, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2212.08333.html", "labs": [ "sjtu" ] }, { "id": "arxiv-2610.29997", "title": "DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations", "type": "paper", "arxiv": "2610.29997", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/Ropedia/DONIC", "project": "https://ropedia.github.io/DONIC/", "topics": [ "manipulation", "learning", "datasets" ], "tags": [ "sim", "dexterous", "bimanual", "tracking", "ppo", "human-demos" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": "2026-10-23", "takeaway_en": "One two-hand policy (PPO + BC on its own verified successes) tracks retargeted human hand-object motions; ships BiHOI (1,575 sequences, 229 objects). Reports 86.8% on unseen trajectories, 57.3% on new objects.", "takeaway_zh": "用一个双手跟踪策略（PPO 结合对自身成功执行的行为克隆）复现重定向后的人手-物体交互动作，并发布 BiHOI 数据集（1575 段双手序列、229 个物体）。论文报告在未见轨迹上成功率 86.8%、新物体上 57.3%。", "github_stars": 51, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.29997.html", "labs": [] }, { "id": "arxiv-2610.29560", "title": "From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation", "type": "paper", "arxiv": "2610.29560", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/StevenKKXS/SERA", "project": "https://stevenkkxs.github.io/sera-webpage/", "topics": [ "manipulation", "vla" ], "tags": [ "agentic", "harness", "memory", "long-horizon", "vlm-planner", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "SERA turns the high-level VLM into an agent with an editable harness that tracks task progress for a fixed VLA executor, and revises the harness from rollouts; reports 51.4% vs 44.4% overall success on RoboMME.", "takeaway_zh": "SERA 把高层 VLM 规划器做成带可编辑执行框架的智能体，为固定的底层 VLA 记录任务进度与执行反馈，并利用环境回放自动改进框架；报告在 RoboMME 上总体成功率 51.4%，未自我演化时为 44.4%。", "github_stars": 6, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.29560.html", "labs": [] }, { "id": "arxiv-2410.24221", "title": "EgoMimic: Scaling Imitation Learning via Egocentric Video", "type": "paper", "arxiv": "2410.24221", "year": 2024, "venue": null, "org": "Georgia Tech", "url": null, "code": "https://github.com/SimarKareer/EgoMimic", "project": "https://egomimic.github.io/", "topics": [ "manipulation", "learning" ], "tags": [ "real", "bimanual", "human-video", "egocentric", "imitation" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "Treats egocentric human video with 3D hand tracking (Project Aria glasses) as robot data and co-trains one policy on both; reports an extra hour of human hand data helps more than an extra hour of robot data.", "takeaway_zh": "把佩戴 Project Aria 眼镜采集、带三维手部跟踪的第一视角人类视频当作机器人数据，与机器人数据共同训练同一个策略；论文报告多 1 小时人手数据比多 1 小时机器人数据更有用。", "github_stars": 230, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.24221.html", "labs": [ "georgia-tech" ] }, { "id": "arxiv-2412.15109", "title": "Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation", "type": "paper", "arxiv": "2412.15109", "year": 2024, "venue": "ICLR 2025", "org": "Shanghai AI Lab / PKU / CUHK", "url": null, "code": "https://github.com/OpenRobotLab/Seer", "project": "https://nimolty.github.io/Seer/", "topics": [ "manipulation", "world-models" ], "tags": [ "real", "sim", "arm", "inverse-dynamics", "foresight", "pretraining" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Seer forecasts the next visual state and feeds it to an inverse dynamics model that outputs actions, trained end to end and pretrained on DROID; reports 4.28 avg. length on CALVIN ABC-D plus LIBERO and real-robot gains.", "takeaway_zh": "Seer 先预测未来视觉状态，再由逆动力学模型据此输出动作，端到端训练并在 DROID 上预训练；报告 CALVIN ABC-D 平均任务长度 4.28，在 LIBERO-LONG 和真机任务上也有提升。", "github_stars": 313, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2412.15109.html", "labs": [ "shanghai-ai-lab", "pku", "cuhk" ] }, { "id": "arxiv-2502.20900", "title": "DexGraspVLA: A Vision-Language-Action Framework Towards General Dexterous Grasping", "type": "paper", "arxiv": "2502.20900", "year": 2025, "venue": "AAAI 2026", "org": "Peking University / PsiBot", "url": null, "code": "https://github.com/Psi-Robot/DexGraspVLA", "project": "https://dexgraspvla.github.io", "topics": [ "manipulation", "vla" ], "tags": [ "real", "dexterous-hand", "grasping", "hierarchical", "diffusion-policy" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "A pretrained VLM plans and a diffusion controller grasps with a dexterous hand, working on domain-invariant features from foundation models; reports 90%+ grasp success across thousands of unseen cluttered scenes.", "takeaway_zh": "由预训练 VLM 做高层规划、扩散控制器驱动灵巧手抓取，底层输入为基础模型提取的域不变特征；报告在数千个未见过的杂乱场景中抓取成功率超过 90%。", "github_stars": 578, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.20900.html", "labs": [ "pku" ] }, { "id": "arxiv-2610.31841", "title": "FORCA: A Force-Aware Coding Agent for Contact-Rich Robot Manipulation", "type": "paper", "arxiv": "2610.31841", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://tanoom.github.io/forca/", "topics": [ "manipulation", "tactile" ], "tags": [ "real", "arm", "force-torque", "contact-rich", "llm-agent", "code-as-policy" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "A coding agent that reads wrist force/torque as well as images, writes probing and force-feedback programs, and revises them from force logs; no demos or training. Beats its vision-only variant on 5 real tasks.", "takeaway_zh": "同时读取腕部力/力矩与图像的编程智能体：编写接触探测与力反馈控制程序，并在每次执行后依据力记录诊断失败、修改代码；无需示教或训练。在 5 个真实任务上优于仅视觉版本。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31841.html", "labs": [] }, { "id": "arxiv-2610.24512", "title": "ContactBench 1.0: A Benchmark for Everyday Contact-Rich Robotic Manipulation", "type": "paper", "arxiv": "2610.24512", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://jingyi-xiang.github.io/contactbench", "topics": [ "manipulation", "tactile", "datasets" ], "tags": [ "benchmark", "contact-rich", "real", "sim", "3d-printable", "visuotactile" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "36 contact-rich tasks built around eight everyday skills, with 3D-printable objects that each have a matching sim model. Reports that tested visuo-tactile policies did not clearly beat vision-only ones.", "takeaway_zh": "围绕八种日常接触技能设计的 36 个接触密集型任务，40 多个物体可用消费级 3D 打印机复现，且每个都有对应仿真模型。报告称所测的视触觉策略并未明显优于纯视觉策略。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.24512.html", "labs": [] }, { "id": "arxiv-2610.28948", "title": "COWBOY: A Scalable Sim-to-Real Framework for Learning Contextual Whole-Body Manipulation", "type": "paper", "arxiv": "2610.28948", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://cowboy-manip.com", "topics": [ "manipulation", "learning" ], "tags": [ "mobile-manipulation", "whole-body", "dexterous-hand", "sim2real", "distillation", "point-cloud", "active-vision" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "Mobile dexterous grasping learned in sim: a whole-body controller does reaching and active vision, local RL experts the object, all distilled into one point-cloud policy; reports zero-shot real transfer.", "takeaway_zh": "在仿真中学习移动灵巧抓取：全身控制器负责底盘-手臂协同到达与主动视觉，局部强化学习专家负责物体交互，二者蒸馏成一个基于点云的视觉策略；报告可零样本迁移到真实世界的数百个物体与场景。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.28948.html", "labs": [] }, { "id": "arxiv-2610.34453", "title": "Scaling Dexterous Manipulation via Action Representation Pretraining", "type": "paper", "arxiv": "2610.34453", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://arpg-project.github.io/", "topics": [ "manipulation", "learning" ], "tags": [ "dexterous", "bimanual", "human-data", "pretraining", "action-representation" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "ARPG pretrains an action encoder on human manipulation data, then trains a 130M generative policy in its space. Reports bimanual dexterous success up from 10.1% (π0.5) to 44.7% on DexJoCo.", "takeaway_zh": "ARPG 把动作表征学习与动作生成解耦：先在大规模人类操作数据上预训练动作编码器，再在其表征空间训练 1.3 亿参数的生成策略。报告在 DexJoCo 双手灵巧任务上平均成功率从 π0.5 的 10.1% 提升到 44.7%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.34453.html", "labs": [] }, { "id": "arxiv-2610.33852", "title": "AdaDexTrack: Dynamic Modulation for Adaptive and Generalizable Dexterous Manipulation Tracking", "type": "paper", "arxiv": "2610.33852", "year": 2026, "venue": "CVPR 2026", "org": null, "url": null, "code": null, "project": "https://janebek.github.io/AdaDexTrack/", "topics": [ "manipulation", "learning" ], "tags": [ "dexterous", "motion-tracking", "language-conditioned", "rl", "distillation", "sim2real" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "Language-to-dexterous-manipulation as modulated tracking: a distilled generalist tracker plus an RL-trained modulator that keeps adjusting the reference, object latent and position target; zero-shot sim-to-real.", "takeaway_zh": "把语言条件灵巧操作变成“可调制的跟踪”：蒸馏得到的通用跟踪器加 RL 训练的调制器，在执行中持续修正参考轨迹、物体隐表征与位置目标，抑制误差累积；零样本迁移到真机。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.33852.html", "labs": [] }, { "id": "arxiv-2610.34280", "title": "AdaPolicy: Action-Driven Anticipatory Robot Policies via Next-Chunk Prediction", "type": "paper", "arxiv": "2610.34280", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://teee111.github.io/AdaPolicy/", "topics": [ "manipulation", "learning" ], "tags": [ "imitation-learning", "action-chunking", "auxiliary-loss", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Foresight without predicting future frames: a training-only head forecasts the next action chunk and backpropagates through the policy's latent. No inference cost; reports up to +14 pts on mm-precision real tasks.", "takeaway_zh": "不预测未来图像也能获得“预见性”：仅在训练时用级联头预测下一个动作块，并把损失回传到在线策略的隐状态，推理零额外开销；报告在耳机装配、穿鞋带等毫米级真机任务上最高提升 14 个百分点。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.34280.html", "labs": [] }, { "id": "arxiv-2610.25785", "title": "Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions", "type": "paper", "arxiv": "2610.25785", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": "https://distilled-search.github.io", "topics": [ "manipulation", "learning" ], "tags": [ "rearrangement", "mcts", "long-horizon", "distillation", "sim", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "DISCO runs progressive-widening MCTS in GPU-parallel simulation, with a learned policy proposing continuous actions, then distills plans into a policy-value model used without search. Reports wins over BC, RL and search.", "takeaway_zh": "DISCO 在 GPU 并行物理仿真中用渐进展宽 MCTS 搜索长程重排计划，由学习的策略提出连续动作、价值函数打分，并周期性把搜索结果蒸馏回策略-价值模型；部署时无需仿真搜索。报告在仿真与真机重排任务上优于行为克隆、强化学习与搜索基线（CoRL 2026 Oral）。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25785.html", "labs": [] }, { "id": "arxiv-2610.28880", "title": "Real2Sem2Real: 3D Semantic Augmentations for Robot Learning from a Single Human RGB Demonstration", "type": "paper", "arxiv": "2610.28880", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://real2sem2real.github.io/", "topics": [ "manipulation", "learning" ], "tags": [ "human-video", "data-generation", "gaussian-splatting", "digital-cousins", "vla", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Turns one monocular RGB human demo into thousands of robot demos across object instances: Gaussian-splat pose tracking, aligned 'digital cousins', replay in IsaacSim or MuJoCo. Reports VLAs matching teleop-trained ones.", "takeaway_zh": "把一段单目 RGB 人类演示扩展为同类别不同物体上的数千条机器人演示：用 DINOv3 蒸馏的高斯溅射跟踪物体位姿，对齐由照片生成的“数字表亲”，在 IsaacSim 或 MuJoCo 中生成数据（约 2 小时 2000+ 条）；报告所训练的 VLA 在真机上达到或超过遥操作数据训练的策略。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.28880.html", "labs": [] }, { "id": "arxiv-2610.37652", "title": "Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well", "type": "paper", "arxiv": "2610.37652", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://ian-chuang.github.io/dexsafe/", "topics": [ "manipulation", "learning", "sim" ], "tags": [ "dexterous", "in-hand", "leap-hand", "sim2real", "safety", "real" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "Sim-only RL recipe for in-hand object pose reaching on a LEAP hand, zero-shot to hardware; reports 95.2% of 4,800 unseen goals reached and a tunable residual safety filter that cuts drops by 75%.", "takeaway_zh": "在仿真中无需人类数据训练 LEAP 灵巧手的手内物体位姿到达策略，零样本迁移到真机；报告在 4800 个未见目标上 95.2% 到达，可调节的残差安全滤波器使掉落减少 75%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.37652.html", "labs": [] }, { "id": "arxiv-2610.36539", "title": "CARD: Contact-Anchored Retargeting for Dexterous Manipulation", "type": "paper", "arxiv": "2610.36539", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://card-retargeting.github.io/", "topics": [ "manipulation", "tactile" ], "tags": [ "dexterous", "retargeting", "contact", "tactile-glove", "human-video", "real" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "Retargets human hand demos (video, mocap, tactile gloves) by keeping contact-force targets, not just kinematics; a tactile tracking controller replays them. Reports contact kept on 43–78% of frames vs <10%.", "takeaway_zh": "把人手示教（视频、动捕、触觉手套）重定向到灵巧手时显式保留接触力目标而不只是运动学，并用触觉跟踪控制器回放；报告 43–78% 的接触帧保持接触，纯运动学方法不到 10%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36539.html", "labs": [] }, { "id": "arxiv-2610.33863", "title": "RoboLongR: Causally Consistent Rewards Teach Robots to Progressively Master Long-Horizon Manipulation", "type": "paper", "arxiv": "2610.33863", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://chenzixuan99.github.io/RoboLongR-public/", "topics": [ "manipulation", "learning" ], "tags": [ "long-horizon", "reward-learning", "offline-to-online", "rlbench", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Learns history-aware progress rewards for long-horizon manipulation so missing prerequisites are not rewarded, then refines policies from BC to online RL. Reports 31.8% on 18 held-out RLBench tasks (+9.2 pts).", "takeaway_zh": "为长时程操作学习依赖执行历史的进度奖励，避免跳过前置步骤也得高分，再从行为克隆过渡到在线强化学习微调策略；报告在 18 个未参与预训练的 RLBench 任务上平均 31.8%（高出最强基线 9.2 个百分点），真机 57.5%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.33863.html", "labs": [] }, { "id": "arxiv-2610.36685", "title": "RoboSTD: Zero-Shot Single-to-Dual Transfer via Sagittal Mirroring for Bimanual Learning", "type": "paper", "arxiv": "2610.36685", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://huahungy.github.io/RoboSTD/", "topics": [ "manipulation", "learning" ], "tags": [ "bimanual", "data-augmentation", "mirroring", "aloha", "llm", "real", "sim" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "Builds pseudo-bimanual demos from single-arm data by sagittal-plane mirroring plus LLM-written coordination constraints. On AgileX ALOHA it reports 2.11× cups collected vs. a single-arm baseline.", "takeaway_zh": "将单臂示范沿矢状面镜像（观测、本体感觉与动作），再由 LLM 生成时空协调约束，构造伪双臂示范；在 AgileX ALOHA（RoboTwin 仿真与真机）上报告收集杯子数达到单臂基线的 2.11 倍。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36685.html", "labs": [] }, { "id": "arxiv-2610.36250", "title": "cuArms: GPU-Accelerated Conflict-Based Planning for Multi-Arm Manipulation", "type": "paper", "arxiv": "2610.36250", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://cuarms.github.io/", "topics": [ "manipulation", "control" ], "tags": [ "multi-arm", "motion-planning", "gpu", "task-scheduling", "real", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "GPU-parallel conflict-based multi-arm planner: reusable per-robot roadmaps, lazy conflict/deadlock resolution and a scheduler proposing task skeletons. Reports real-time first solutions that keep improving.", "takeaway_zh": "为 GPU 并行而设计的多臂冲突搜索规划器：每台机器人使用可复用的多模态路线图，惰性解决冲突与死锁，并由高层调度器提出任务骨架；报告可实时给出初始解并持续改进运动质量与调度。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36250.html", "labs": [] }, { "id": "arxiv-2005.14679", "title": "DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand Manipulation", "type": "paper", "arxiv": "2005.14679", "year": 2020, "venue": null, "org": "Meta AI", "url": null, "code": "https://github.com/facebookresearch/digit-interface", "project": null, "topics": [ "tactile", "hardware" ], "tags": [ "real", "sensor", "vision-based-tactile", "low-cost" ], "robot_types": [ "hand" ], "highlight": true, "added": null, "takeaway_en": "An inexpensive, compact GelSight-style fingertip sensor that made vision-based touch widely accessible.", "takeaway_zh": "低成本、紧凑的 GelSight 式指尖传感器，让视触觉传感器得以普及。", "github_stars": 94, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2005.14679.html", "labs": [ "meta" ] }, { "id": "arxiv-2012.08456", "title": "TACTO: A Fast, Flexible, and Open-source Simulator for High-Resolution Vision-based Tactile Sensors", "type": "paper", "arxiv": "2012.08456", "year": 2020, "venue": null, "org": "Meta AI", "url": null, "code": "https://github.com/facebookresearch/tacto", "project": null, "topics": [ "tactile", "sim" ], "tags": [ "sim", "vision-based-tactile" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Renders DIGIT/GelSight-style tactile images in simulation so touch-based policies can be trained at scale.", "takeaway_zh": "在仿真中渲染 DIGIT/GelSight 式触觉图像，便于大规模训练基于触觉的策略。", "github_stars": 472, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2012.08456.html", "labs": [ "meta" ] }, { "id": "arxiv-2410.24090", "title": "Sparsh: Self-supervised touch representations for vision-based tactile sensing", "type": "paper", "arxiv": "2410.24090", "year": 2024, "venue": null, "org": "Meta AI", "url": null, "code": "https://github.com/facebookresearch/sparsh", "project": null, "topics": [ "tactile" ], "tags": [ "representation", "vision-based-tactile", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Self-supervised encoders pretrained on many tactile images that work across different vision-based touch sensors.", "takeaway_zh": "在大量触觉图像上自监督预训练的编码器，可跨不同视触觉传感器使用。", "github_stars": 225, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.24090.html", "labs": [ "meta" ] }, { "id": "arxiv-2401.18084", "title": "Binding Touch to Everything: Learning Unified Multimodal Tactile Representations", "type": "paper", "arxiv": "2401.18084", "year": 2024, "venue": null, "org": "Yale, UMich", "url": null, "code": null, "project": null, "topics": [ "tactile" ], "tags": [ "representation", "multimodal" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "UniTouch aligns tactile signals with vision, language and audio embeddings, enabling zero-shot touch understanding.", "takeaway_zh": "UniTouch 将触觉信号与视觉、语言、音频嵌入对齐，实现零样本触觉理解。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2401.18084.html", "labs": [] }, { "id": "arxiv-2410.24091", "title": "3D-ViTac: Learning Fine-Grained Manipulation with Visuo-Tactile Sensing", "type": "paper", "arxiv": "2410.24091", "year": 2024, "venue": null, "org": "Columbia, UIUC", "url": null, "code": null, "project": null, "topics": [ "tactile", "manipulation" ], "tags": [ "real", "bimanual", "visuotactile", "diffusion" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "Fuses dense tactile arrays with 3D vision in a diffusion policy for fragile, contact-rich bimanual tasks.", "takeaway_zh": "在扩散策略中融合密集触觉阵列与 3D 视觉，完成易碎、接触丰富的双臂任务。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.24091.html", "labs": [ "columbia" ] }, { "id": "arxiv-2409.08276", "title": "AnySkin: Plug-and-play Skin Sensing for Robotic Touch", "type": "paper", "arxiv": "2409.08276", "year": 2024, "venue": null, "org": "NYU, Meta AI", "url": null, "code": null, "project": null, "topics": [ "tactile", "hardware" ], "tags": [ "real", "sensor", "magnetic-skin" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A replaceable magnetic tactile skin whose learned policies keep working after the skin is swapped.", "takeaway_zh": "可更换的磁性触觉皮肤，更换皮肤后训练好的策略仍能继续使用。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2409.08276.html", "labs": [ "meta", "nyu" ] }, { "id": "arxiv-2111.00071", "title": "ReSkin: versatile, replaceable, lasting tactile skins", "type": "paper", "arxiv": "2111.00071", "year": 2021, "venue": null, "org": "CMU, Meta AI", "url": null, "code": "https://github.com/raunaqbhirangi/reskin_sensor", "project": null, "topics": [ "tactile", "hardware" ], "tags": [ "sensor", "magnetic-skin", "low-cost" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A thin magnetic-elastomer skin read by magnetometers; cheap to make and swap, with learned calibration that carries over to new skins.", "takeaway_zh": "由磁力计读取的薄磁性弹性体皮肤，制作和更换成本低，学习得到的标定可迁移到新皮肤上。", "github_stars": 138, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2111.00071.html", "labs": [ "meta", "cmu" ] }, { "id": "arxiv-2211.12498", "title": "Touch and Go: Learning from Human-Collected Vision and Touch", "type": "dataset", "arxiv": "2211.12498", "year": 2022, "venue": null, "org": "University of Michigan", "url": null, "code": null, "project": null, "topics": [ "tactile" ], "tags": [ "dataset", "visuotactile", "gelsight" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A dataset of paired video and GelSight touch collected by people probing objects in the wild, for cross-modal tactile learning.", "takeaway_zh": "由人手持 GelSight 在真实场景中触碰物体采集的视觉-触觉配对数据集，用于跨模态触觉学习。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2211.12498.html", "labs": [] }, { "id": "arxiv-2411.02479", "title": "Digitizing Touch with an Artificial Multimodal Fingertip", "type": "paper", "arxiv": "2411.02479", "year": 2024, "venue": null, "org": "Meta FAIR, TU Dresden, UC Berkeley", "url": null, "code": "https://github.com/facebookresearch/digit360", "project": null, "topics": [ "tactile", "hardware" ], "tags": [ "real", "sensor", "vision-based-tactile", "multimodal", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Digit 360: an open-design, finger-shaped sensor that senses force, fine texture, vibration and heat, with an on-board neural accelerator for reflex-like local processing.", "takeaway_zh": "Digit 360：开放设计的指尖形触觉传感器，可感知力、细微纹理、振动和温度，并用板载神经网络加速器做类似反射弧的本地处理。", "github_stars": 269, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2411.02479.html", "labs": [ "berkeley", "meta" ] }, { "id": "arxiv-2109.04027", "title": "Taxim: An Example-based Simulation Model for GelSight Tactile Sensors", "type": "simulator", "arxiv": "2109.04027", "year": 2021, "venue": null, "org": "CMU", "url": null, "code": "https://github.com/Robo-Touch/Taxim", "project": null, "topics": [ "tactile", "sim" ], "tags": [ "simulator", "gelsight", "vision-based-tactile" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Simulates GelSight images from calibration examples (optical lookup + linear elastic deformation) instead of physics-based rendering, so tactile data can be generated quickly.", "takeaway_zh": "用标定样本（光学查找表 + 线弹性形变）而不是基于物理的渲染来模拟 GelSight 图像，可快速生成触觉数据。", "github_stars": 126, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2109.04027.html", "labs": [ "cmu" ] }, { "id": "arxiv-2207.10763", "title": "Tactile Gym 2.0: Sim-to-real Deep Reinforcement Learning for Comparing Low-cost High-Resolution Robot Touch", "type": "simulator", "arxiv": "2207.10763", "year": 2022, "venue": null, "org": "University of Bristol", "url": null, "code": "https://github.com/dexterousrobot/tactile_gym", "project": null, "topics": [ "tactile", "sim", "learning" ], "tags": [ "simulator", "sim-to-real", "rl", "tactip", "digit" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open-source tactile RL environments with sim-to-real transfer across three low-cost optical sensors (TacTip, DIGIT, DigiTac), for comparing them on the same tasks.", "takeaway_zh": "开源触觉强化学习环境，支持三种低成本光学触觉传感器（TacTip、DIGIT、DigiTac）的仿真到真机迁移，便于在相同任务上比较。", "github_stars": 44, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2207.10763.html", "labs": [] }, { "id": "arxiv-2103.12269", "title": "GelSlim 3.0: High-Resolution Measurement of Shape, Force and Slip in a Compact Tactile-Sensing Finger", "type": "paper", "arxiv": "2103.12269", "year": 2021, "venue": null, "org": "MIT", "url": null, "code": null, "project": null, "topics": [ "tactile", "hardware" ], "tags": [ "gelsight", "sensor-design", "finger", "force", "slip" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A slim GelSight-style finger that measures contact shape, force and slip at high resolution, designed to fit robot grippers.", "takeaway_zh": "纤薄的 GelSight 类触觉手指，可高分辨率测量接触形状、力与滑移，便于装在机器人夹爪上。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2103.12269.html", "labs": [ "mit" ] }, { "id": "arxiv-2209.13916", "title": "DTact: A Vision-Based Tactile Sensor that Measures High-Resolution 3D Geometry Directly from Darkness", "type": "paper", "arxiv": "2209.13916", "year": 2022, "venue": null, "org": null, "url": null, "code": null, "project": null, "topics": [ "tactile", "hardware" ], "tags": [ "sensor-design", "low-cost", "3d-geometry", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A low-cost visuotactile sensor that reads contact depth from how dark each pixel gets as the gel is pressed, giving high-resolution 3D geometry without photometric stereo.", "takeaway_zh": "低成本视觉触觉传感器：根据凝胶受压时每个像素变暗的程度直接读出接触深度，无需光度立体即可获得高分辨率三维形状。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2209.13916.html", "labs": [] }, { "id": "arxiv-2308.14277", "title": "9DTact: A Compact Vision-Based Tactile Sensor for Accurate 3D Shape Reconstruction and Generalizable 6D Force Estimation", "type": "paper", "arxiv": "2308.14277", "year": 2023, "venue": null, "org": null, "url": null, "code": "https://github.com/linchangyi1/9DTact", "project": null, "topics": [ "tactile", "hardware" ], "tags": [ "sensor-design", "open-source", "force", "3d-geometry" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Successor to DTact - a compact open-source visuotactile sensor that reconstructs 3D contact shape and estimates 6D force/torque with a model that generalizes to unseen objects.", "takeaway_zh": "DTact 的后续：紧凑的开源视觉触觉传感器，可重建三维接触形状，并用可泛化到新物体的模型估计六维力/力矩。", "github_stars": 195, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2308.14277.html", "labs": [] }, { "id": "arxiv-1810.10191", "title": "Making Sense of Vision and Touch: Self-Supervised Learning of Multimodal Representations for Contact-Rich Tasks", "type": "paper", "arxiv": "1810.10191", "year": 2018, "venue": "ICRA 2019", "org": "Stanford", "url": null, "code": "https://github.com/stanford-iprl-lab/multimodal_representation", "project": null, "topics": [ "tactile", "manipulation", "learning" ], "tags": [ "real", "visuotactile", "force-torque", "representation", "rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Learns a compact vision + force/torque representation with self-supervised objectives, then trains a peg-insertion policy with RL on top of it on a real robot.", "takeaway_zh": "用自监督目标学习视觉 + 力/力矩的紧凑联合表征，再在其上用强化学习在真实机器人上训练插孔策略。", "github_stars": 103, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1810.10191.html", "labs": [ "stanford" ] }, { "id": "arxiv-2212.03858", "title": "See, Hear, and Feel: Smart Sensory Fusion for Robotic Manipulation", "type": "paper", "arxiv": "2212.03858", "year": 2022, "venue": "CoRL 2022", "org": "Stanford", "url": null, "code": "https://github.com/JunzheJosephZhu/see_hear_feel", "project": null, "topics": [ "tactile", "manipulation" ], "tags": [ "real", "visuotactile", "audio", "multimodal", "imitation-learning" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Fuses camera, contact microphone and GelSight touch with self-attention in an imitation-learned policy, studied on dense packing and pouring tasks.", "takeaway_zh": "在模仿学习策略中用自注意力融合相机、接触麦克风和 GelSight 触觉，在密集装箱与倾倒任务上做了研究。", "github_stars": 46, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2212.03858.html", "labs": [ "stanford" ] }, { "id": "arxiv-2402.13232", "title": "A Touch, Vision, and Language Dataset for Multimodal Alignment", "type": "paper", "arxiv": "2402.13232", "year": 2024, "venue": "ICML 2024", "org": "UC Berkeley", "url": null, "code": "https://github.com/Max-Fu/tvl", "project": null, "topics": [ "tactile", "datasets" ], "tags": [ "dataset", "vision-based-tactile", "language", "representation" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "TVL pairs tactile and visual observations with language descriptions and trains a touch-vision-language encoder plus a model that describes how things feel.", "takeaway_zh": "TVL 数据集将触觉与视觉观测配上语言描述，并训练触觉-视觉-语言编码器以及能描述物体触感的模型。", "github_stars": 102, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2402.13232.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2309.09979", "title": "General In-Hand Object Rotation with Vision and Touch", "type": "paper", "arxiv": "2309.09979", "year": 2023, "venue": "CoRL 2023", "org": "UC Berkeley, Meta AI", "url": null, "code": null, "project": null, "topics": [ "tactile", "manipulation" ], "tags": [ "dexterous-hand", "visuotactile", "sim2real", "rl" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "RotateIt trains in simulation with privileged object information, then distills to a policy that rotates diverse objects in hand using vision, touch and proprioception on a real hand.", "takeaway_zh": "RotateIt 先在仿真中利用特权物体信息训练，再蒸馏为仅用视觉、触觉和本体感觉的策略，在真实灵巧手上旋转多种物体。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.09979.html", "labs": [ "berkeley", "meta" ] }, { "id": "arxiv-2408.06506", "title": "TacSL: A Library for Visuotactile Sensor Simulation and Learning", "type": "paper", "arxiv": "2408.06506", "year": 2024, "venue": null, "org": "NVIDIA", "url": null, "code": null, "project": null, "topics": [ "tactile", "sim" ], "tags": [ "sim", "gpu", "vision-based-tactile", "sim2real" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "GPU-based simulation of GelSight-style tactile images and contact forces in Isaac Gym, plus learning tools for sim-to-real transfer of visuotactile policies.", "takeaway_zh": "在 Isaac Gym 中用 GPU 模拟 GelSight 式触觉图像与接触力，并提供用于视触觉策略仿真到真实迁移的学习工具。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2408.06506.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2303.10880", "title": "Rotating without Seeing: Towards In-hand Dexterity through Touch", "type": "paper", "arxiv": "2303.10880", "year": 2023, "venue": "RSS 2023", "org": "UC San Diego", "url": null, "code": null, "project": "https://touchdexterity.github.io", "topics": [ "tactile", "manipulation" ], "tags": [ "dexterous-hand", "binary-touch", "low-cost", "sim2real", "rl" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "Covers a robot hand with cheap binary (contact / no contact) touch sensors and trains an RL policy in simulation that rotates unseen objects in hand on a real robot without any vision.", "takeaway_zh": "在灵巧手上铺设廉价的二值（有/无接触）触觉传感器，在仿真中用强化学习训练策略，迁移到真机后无需视觉即可在手中旋转未见过的物体。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2303.10880.html", "labs": [ "ucsd" ] }, { "id": "arxiv-2312.01853", "title": "Robot Synesthesia: In-Hand Manipulation with Visuotactile Sensing", "type": "paper", "arxiv": "2312.01853", "year": 2023, "venue": null, "org": "UC San Diego", "url": null, "code": null, "project": "https://yingyuan0414.github.io/visuotactile/", "topics": [ "tactile", "manipulation" ], "tags": [ "dexterous-hand", "visuotactile", "point-cloud", "sim2real", "rl" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "Turns tactile readings into points in the same 3D frame as the camera point cloud and the hand model, so one visuotactile point-cloud input drives in-hand object rotation.", "takeaway_zh": "将触觉读数转换为与相机点云、手部模型处于同一三维坐标系的点，用统一的视触觉点云输入驱动手内物体旋转。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2312.01853.html", "labs": [ "ucsd" ] }, { "id": "arxiv-2312.13469", "title": "Neural feels with neural fields: Visuo-tactile perception for in-hand manipulation", "type": "paper", "arxiv": "2312.13469", "year": 2023, "venue": "Science Robotics 2024", "org": "Meta AI, CMU", "url": null, "code": "https://github.com/facebookresearch/neuralfeels", "project": null, "topics": [ "tactile", "perception", "manipulation" ], "tags": [ "visuotactile", "neural-fields", "pose-tracking", "dexterous-hand" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "NeuralFeels fuses vision and touch from a multi-fingered hand into an online neural field, reconstructing an object's shape and tracking its pose while the hand manipulates it.", "takeaway_zh": "NeuralFeels 将多指手的视觉与触觉在线融合为神经场，在操作过程中同时重建物体形状并跟踪其位姿。", "github_stars": 162, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2312.13469.html", "labs": [ "meta", "cmu" ] }, { "id": "arxiv-2503.02881", "title": "Reactive Diffusion Policy: Slow-Fast Visual-Tactile Policy Learning for Contact-Rich Manipulation", "type": "paper", "arxiv": "2503.02881", "year": 2025, "venue": "RSS 2025", "org": "SJTU, Tsinghua", "url": null, "code": "https://github.com/xiaoxiaoxh/reactive_diffusion_policy", "project": "https://reactive-diffusion-policy.github.io/", "topics": [ "tactile", "manipulation", "learning" ], "tags": [ "visuotactile", "force", "diffusion-policy", "teleoperation", "real" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A slow latent diffusion policy plans action chunks while a fast policy reacts to tactile/force feedback; reports >35% higher scores than visual imitation baselines on three contact-rich tasks.", "takeaway_zh": "慢速扩散策略在隐空间规划动作块，快速策略根据触觉/力反馈高频调整；论文报告在三个接触丰富任务上比纯视觉模仿学习基线总分高 35% 以上。", "github_stars": 394, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2503.02881.html", "labs": [ "tsinghua", "sjtu" ] }, { "id": "arxiv-2610.31661", "title": "OpenTac: An Open Ecosystem for Customizable, Scalable, and Transferable 3D Tactile Sensing", "type": "paper", "arxiv": "2610.31661", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://opentac.github.io", "topics": [ "tactile", "hardware" ], "tags": [ "tactile-skin", "shear-sensing", "open-source-hardware", "design-tool" ], "robot_types": [], "highlight": false, "added": "2026-10-23", "takeaway_en": "Open-source resistive tactile skin whose structured overlay turns shear into pressure patterns, so neighbouring taxels decode shear and normal force; ships with a design tool for custom skins plus hardware and software.", "takeaway_zh": "开源电阻式触觉皮肤：结构化覆层把切向力转成空间压力差异，相邻触点即可解码剪切与法向力；附带可定制皮肤的设计工具，硬件和软件均开源。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31661.html", "labs": [] }, { "id": "arxiv-2610.31334", "title": "GelSlim 5.0: Unifying Vision and Touch with a Single Fingertip Camera", "type": "paper", "arxiv": "2610.31334", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": null, "topics": [ "tactile", "perception" ], "tags": [ "visuotactile", "gelsight", "fingertip-sensor", "multimodal-transformer" ], "robot_types": [], "highlight": false, "added": "2026-10-23", "takeaway_en": "Fingertip sensor whose one camera switches illumination to alternate between see-through-membrane vision and tactile images — one shared viewpoint, no extrinsic calibration; fused by a transformer (VisTacFusion).", "takeaway_zh": "指尖传感器只用一个摄像头，通过切换照明交替获取“透过膜的视觉图像”和触觉图像，视觉与触觉共享同一视角、无需相机外参标定；再用 Transformer（VisTacFusion）融合两者。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31334.html", "labs": [] }, { "id": "arxiv-2610.30438", "title": "Tactile-Based Robotic Manipulation: From Mechanics-Driven to Data-Driven Perspectives", "type": "paper", "arxiv": "2610.30438", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": null, "topics": [ "tactile", "manipulation" ], "tags": [ "survey", "tactile-manipulation", "contact-mechanics", "imitation", "rl", "vla" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Review (Annual Review of Control, Robotics, and Autonomous Systems) tracing tactile manipulation from contact-mechanics models to learned policies: sensing basics, state estimation, IL, RL, VLAs, predictive models.", "takeaway_zh": "综述（Annual Review of Control, Robotics, and Autonomous Systems 录用稿）：从接触力学建模到数据驱动学习，梳理触觉操作的发展，涵盖传感基础、状态表征、模仿学习、强化学习、VLA 与预测模型。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30438.html", "labs": [] }, { "id": "arxiv-2610.31472", "title": "VecTile: A Low-Cost Flexible Tactile Array for 3-DoF Force Sensing", "type": "paper", "arxiv": "2610.31472", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://jiannnh.github.io/VecTile/", "topics": [ "tactile", "hardware" ], "tags": [ "tactile-skin", "shear-sensing", "low-cost", "piezoresistive", "diffusion-policy" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "6 mm flexible array: silicone nubs over a piezoresistive layer turn shear into readable pressure patterns, giving 3-axis force per taxel; under 35 USD for 4x4. Reports it helps a Diffusion Policy insert a plug.", "takeaway_zh": "6 毫米厚的柔性触觉阵列：压阻层上的硅胶凸点把切向力转成可区分的压力分布，实现每个触点的三轴力测量；4x4 阵列成本不到 35 美元。报告称其提升了扩散策略插插头任务的表现。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31472.html", "labs": [] }, { "id": "arxiv-2610.24285", "title": "ACTS: Learning Visuo-Tactile World Models from Human Physical Interaction", "type": "paper", "arxiv": "2610.24285", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://acts-anon.github.io/", "topics": [ "tactile", "world-models" ], "tags": [ "gelsight", "human-data", "world-model", "bimanual", "dataset" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "Visuo-tactile world model trained only on humans handling GelSight sensors (4.9 h React dataset). Reports that contact dynamics transfer to robot rollouts more reliably than visual appearance does.", "takeaway_zh": "只用人手持 GelSight 传感器的交互数据（4.9 小时 React 数据集）训练视触觉世界模型。报告在机器人回放中接触动态比视觉外观更容易迁移，部分任务的相机视角预测甚至不如直接复制帧。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.24285.html", "labs": [] }, { "id": "arxiv-2109.11978", "title": "Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning", "type": "paper", "arxiv": "2109.11978", "year": 2021, "venue": null, "org": "ETH Zurich, NVIDIA", "url": null, "code": "https://github.com/leggedrobotics/legged_gym", "project": null, "topics": [ "locomotion", "sim" ], "tags": [ "sim2real", "quadruped", "rl", "gpu-parallel" ], "robot_types": [ "legged" ], "highlight": true, "added": null, "takeaway_en": "Thousands of parallel GPU environments train a quadruped walking policy in minutes; legged_gym became a standard starter.", "takeaway_zh": "数千个 GPU 并行环境几分钟就能训练出四足行走策略；legged_gym 成为标准入门框架。", "github_stars": 3164, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2109.11978.html", "labs": [ "nvidia", "eth" ] }, { "id": "arxiv-2010.11251", "title": "Learning Quadrupedal Locomotion over Challenging Terrain", "type": "paper", "arxiv": "2010.11251", "year": 2020, "venue": null, "org": "ETH Zurich", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "quadruped", "rl", "proprioception" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "A teacher-student, proprioception-only policy that walks ANYmal over mud, snow and rubble in the wild.", "takeaway_zh": "仅依赖本体感知的师生策略，让 ANYmal 在泥地、雪地和碎石等野外地形行走。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2010.11251.html", "labs": [ "eth" ] }, { "id": "arxiv-1901.08652", "title": "Learning Agile and Dynamic Motor Skills for Legged Robots", "type": "paper", "arxiv": "1901.08652", "year": 2019, "venue": null, "org": "ETH Zurich", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "quadruped", "rl", "actuator-net" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "A learned actuator model closes the sim-to-real gap so RL policies run and recover from falls on ANYmal.", "takeaway_zh": "学习的执行器模型缩小仿真与现实差距，使强化学习策略能让 ANYmal 奔跑并摔倒后起身。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1901.08652.html", "labs": [ "eth" ] }, { "id": "arxiv-2107.04034", "title": "RMA: Rapid Motor Adaptation for Legged Robots", "type": "paper", "arxiv": "2107.04034", "year": 2021, "venue": null, "org": "UC Berkeley, CMU, Meta AI", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "quadruped", "rl", "adaptation" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "An adaptation module infers environment parameters online, so one policy handles new terrains and payloads in real time.", "takeaway_zh": "自适应模块在线估计环境参数，使同一策略实时适应新地形和负载。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2107.04034.html", "labs": [ "berkeley", "meta", "cmu" ] }, { "id": "arxiv-2309.05665", "title": "Robot Parkour Learning", "type": "paper", "arxiv": "2309.05665", "year": 2023, "venue": null, "org": "Shanghai Qi Zhi, Stanford", "url": null, "code": "https://github.com/ZiwenZhuang/parkour", "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "quadruped", "rl", "vision" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "Distills specialist parkour skills into a single vision-based policy that climbs, leaps and crawls on low-cost quadrupeds.", "takeaway_zh": "将多个跑酷专家技能蒸馏为单一视觉策略，让低成本四足机器人攀爬、跳跃和匍匐。", "github_stars": 1142, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.05665.html", "labs": [ "stanford", "qizhi" ] }, { "id": "arxiv-2309.14341", "title": "Extreme Parkour with Legged Robots", "type": "paper", "arxiv": "2309.14341", "year": 2023, "venue": null, "org": "CMU", "url": null, "code": "https://github.com/chengxuxin/extreme-parkour", "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "quadruped", "rl", "vision" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "An end-to-end depth-camera policy makes a small quadruped do long jumps, high jumps and handstands.", "takeaway_zh": "端到端深度相机策略让小型四足机器人完成跳远、跳高和倒立。", "github_stars": 1188, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.14341.html", "labs": [ "cmu" ] }, { "id": "arxiv-2303.03381", "title": "Real-World Humanoid Locomotion with Reinforcement Learning", "type": "paper", "arxiv": "2303.03381", "year": 2023, "venue": null, "org": "UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "humanoid", "rl", "transformer" ], "robot_types": [ "humanoid" ], "highlight": true, "added": null, "takeaway_en": "A causal Transformer policy trained in simulation walks a full-size humanoid outdoors, zero-shot.", "takeaway_zh": "在仿真中训练的因果 Transformer 策略零样本驱动全尺寸人形机器人在户外行走。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2303.03381.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2402.16796", "title": "Expressive Whole-Body Control for Humanoid Robots", "type": "paper", "arxiv": "2402.16796", "year": 2024, "venue": null, "org": "UC San Diego", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "humanoid", "rl", "motion-imitation" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Imitates upper-body human motion capture while keeping robust legs, giving humanoids expressive gestures and dance.", "takeaway_zh": "上半身模仿人类动捕、下肢保持鲁棒行走，让人形机器人做出富有表现力的动作和舞蹈。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2402.16796.html", "labs": [ "ucsd" ] }, { "id": "arxiv-2403.04436", "title": "Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation", "type": "paper", "arxiv": "2403.04436", "year": 2024, "venue": null, "org": "CMU", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "humanoid", "teleoperation" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "H2O — an RL tracking policy lets an RGB camera stream a person's whole-body motion onto a humanoid in real time.", "takeaway_zh": "H2O：强化学习跟踪策略借助 RGB 相机，把人的全身动作实时映射到人形机器人。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2403.04436.html", "labs": [ "cmu" ] }, { "id": "arxiv-2406.10454", "title": "HumanPlus: Humanoid Shadowing and Imitation from Humans", "type": "paper", "arxiv": "2406.10454", "year": 2024, "venue": null, "org": "Stanford", "url": null, "code": "https://github.com/MarkFzp/humanplus", "project": null, "topics": [ "locomotion" ], "tags": [ "real", "humanoid", "teleoperation", "imitation" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "A humanoid shadows a human from one RGB camera, then learns autonomous skills from the collected data.", "takeaway_zh": "人形机器人通过单个 RGB 相机实时模仿人类，再用采集的数据学习自主技能。", "github_stars": 851, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2406.10454.html", "labs": [ "stanford" ] }, { "id": "arxiv-2502.01143", "title": "ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills", "type": "paper", "arxiv": "2502.01143", "year": 2025, "venue": null, "org": "CMU, NVIDIA", "url": null, "code": "https://github.com/LeCAR-Lab/ASAP", "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "humanoid", "rl", "delta-action" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Learns a residual \"delta action\" model from real rollouts to fix sim physics, enabling agile athletic humanoid motions.", "takeaway_zh": "从真实轨迹学习残差\"增量动作\"模型来修正仿真物理，实现敏捷的人形运动技能。", "github_stars": 2135, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.01143.html", "labs": [ "nvidia", "cmu" ] }, { "id": "unitree-rl-gym", "title": "Unitree RL Gym", "type": "code", "arxiv": null, "year": 2024, "venue": null, "org": "Unitree", "url": null, "code": "https://github.com/unitreerobotics/unitree_rl_gym", "project": null, "topics": [ "locomotion", "hardware", "learning" ], "tags": [ "sim2real", "humanoid", "quadruped", "rl", "isaac-gym" ], "robot_types": [ "humanoid", "legged" ], "highlight": false, "added": null, "takeaway_en": "Unitree's official starting point for RL locomotion on Go2, H1 and G1 — train in Isaac Gym, check sim-to-sim in MuJoCo, then deploy to the real robot.", "takeaway_zh": "宇树官方的强化学习运动控制入门仓库，支持 Go2、H1、G1：在 Isaac Gym 中训练，在 MuJoCo 中做 sim2sim 验证，再部署到真机。", "github_stars": 3649, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/unitree-rl-gym.html", "labs": [] }, { "id": "arxiv-2610.28480", "title": "SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model", "type": "paper", "arxiv": "2610.28480", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": "https://supermimic.github.io", "topics": [ "locomotion", "world-models", "learning" ], "tags": [ "humanoid", "motion-tracking", "whole-body-control", "differentiable-model" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-21", "takeaway_en": "Rather than distilling per-motion experts or brute-forcing one RL policy, trains a humanoid tracking policy jointly with one differentiable world model of shared dynamics, fit from simulator logs (CoRL 2026 oral).", "takeaway_zh": "SuperMimic 不再蒸馏逐动作专家或用大算力硬训单一 RL 策略，而是把人形机器人动作跟踪策略与一个从仿真日志拟合的、共享动力学的可微世界模型联合训练（CoRL 2026 口头报告）。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.28480.html", "labs": [] }, { "id": "arxiv-2201.08117", "title": "Learning robust perceptive locomotion for quadrupedal robots in the wild", "type": "paper", "arxiv": "2201.08117", "year": 2022, "venue": "Science Robotics 2022", "org": "ETH Zurich, Intel", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "quadruped", "rl", "perception", "real", "sim-to-real" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "ANYmal fuses proprioception with noisy height-map perception through a learned belief encoder, and is shown hiking an alpine trail and other rough terrain.", "takeaway_zh": "ANYmal 通过学习的信念编码器融合本体感知与带噪声的高程图感知，在高山徒步路线等复杂地形上完成行走。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2201.08117.html", "labs": [ "eth" ] }, { "id": "arxiv-2406.08858", "title": "OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning", "type": "paper", "arxiv": "2406.08858", "year": 2024, "venue": "CoRL 2024", "org": "CMU", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "humanoid", "teleoperation", "whole-body", "rl", "sim-to-real" ], "robot_types": [ "hand", "humanoid" ], "highlight": false, "added": null, "takeaway_en": "One RL whole-body controller that a humanoid can follow from a VR headset, verbal commands or an RGB camera, and that also runs learned autonomous policies.", "takeaway_zh": "一个强化学习全身控制器：人形机器人可跟随 VR 头显、语音指令或 RGB 相机的输入，也可执行学习得到的自主策略。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2406.08858.html", "labs": [ "cmu" ] }, { "id": "arxiv-2410.21229", "title": "HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots", "type": "paper", "arxiv": "2410.21229", "year": 2024, "venue": "ICRA 2025", "org": "NVIDIA, CMU", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "humanoid", "whole-body", "rl", "distillation" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Distills one policy that supports several humanoid control modes (root velocity, joint angles, keypoint tracking) instead of training a separate controller per mode.", "takeaway_zh": "蒸馏出单一策略，支持多种人形控制模式（根部速度、关节角、关键点跟踪），不必为每种模式各训练一个控制器。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.21229.html", "labs": [ "nvidia", "cmu" ] }, { "id": "arxiv-2212.03238", "title": "Walk These Ways: Tuning Robot Control for Generalization with Multiplicity of Behavior", "type": "paper", "arxiv": "2212.03238", "year": 2022, "venue": "CoRL 2022", "org": "MIT", "url": null, "code": "https://github.com/Improbable-AI/walk-these-ways", "project": null, "topics": [ "locomotion" ], "tags": [ "real", "quadruped", "sim2real", "rl", "unitree-go1" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "Trains one quadruped policy that exposes gait parameters (frequency, foot swing, posture) as commands, so a human can tune behavior at deploy time for new tasks; open sim-to-real stack for the Unitree Go1.", "takeaway_zh": "训练一个把步态参数（频率、抬脚高度、姿态等）作为指令开放的四足策略，部署时可由人调节行为以适应新任务；附带面向 Unitree Go1 的开源仿真到真实工具链。", "github_stars": 1470, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2212.03238.html", "labs": [ "mit" ] }, { "id": "arxiv-2301.10602", "title": "DreamWaQ: Learning Robust Quadrupedal Locomotion With Implicit Terrain Imagination via Deep Reinforcement Learning", "type": "paper", "arxiv": "2301.10602", "year": 2023, "venue": "ICRA 2023", "org": "KAIST", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "real", "quadruped", "blind", "sim2real", "rl" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "A blind quadruped policy that infers terrain properties implicitly from proprioception history with a context-aided estimator, enabling robust walking on stairs and rough ground without cameras.", "takeaway_zh": "盲视四足策略：借助上下文辅助估计器从本体感觉历史中隐式推断地形特性，无需相机即可稳健地走楼梯和崎岖地面。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2301.10602.html", "labs": [ "kaist" ] }, { "id": "arxiv-2306.14874", "title": "ANYmal Parkour: Learning Agile Navigation for Quadrupedal Robots", "type": "paper", "arxiv": "2306.14874", "year": 2023, "venue": null, "org": "ETH Zurich, NVIDIA", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "real", "quadruped", "parkour", "hierarchical", "perception" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "A learned navigation policy sequences pre-trained skills (walk, jump, climb, crouch) using a learned obstacle reconstruction; reports up to 2 m/s on hardware (Science Robotics).", "takeaway_zh": "分层的全学习系统：导航策略结合学习到的障碍物重建来调度预训练技能（行走、跳跃、攀爬、下蹲）；报告实机速度可达 2 米/秒。发表于 Science Robotics。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2306.14874.html", "labs": [ "nvidia", "eth" ] }, { "id": "arxiv-2406.10759", "title": "Humanoid Parkour Learning", "type": "paper", "arxiv": "2406.10759", "year": 2024, "venue": "CoRL 2024", "org": "Shanghai Qi Zhi, ShanghaiTech, Tsinghua", "url": null, "code": null, "project": null, "topics": [ "locomotion" ], "tags": [ "real", "humanoid", "parkour", "vision", "unitree-h1", "sim2real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Learns a vision-based whole-body parkour policy for the Unitree H1 without motion references, covering skills such as jumping onto platforms and leaping over gaps, and transfers it to the real robot.", "takeaway_zh": "无需动作参考，为 Unitree H1 学习基于视觉的全身跑酷策略，涵盖跳上平台、跨越沟壑等技能，并迁移到真实机器人。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2406.10759.html", "labs": [ "tsinghua", "qizhi" ] }, { "id": "arxiv-2610.29612", "title": "Watch Your Step: Anticipated Perception for Humanoid Locomotion", "type": "paper", "arxiv": "2610.29612", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://leggedrobotics.github.io/watch-your-step", "topics": [ "locomotion", "perception" ], "tags": [ "humanoid", "perceptive-locomotion", "elevation-map", "rl" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-22", "takeaway_en": "A lightweight foothold prediction model strips terrain irrelevant to the next steps from the elevation map before the humanoid policy sees it, aiming at faster training and better transfer to new terrain.", "takeaway_zh": "不再把整张高程图输入人形机器人运动策略，而是先用轻量的落脚点预测模型去掉与接下来几步无关的地形信息，以加快训练并提升对新地形的泛化。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.29612.html", "labs": [] }, { "id": "arxiv-2104.02180", "title": "AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control", "type": "paper", "arxiv": "2104.02180", "year": 2021, "venue": "SIGGRAPH 2021", "org": "UC Berkeley, Shanghai Jiao Tong University", "url": null, "code": null, "project": null, "topics": [ "locomotion", "learning" ], "tags": [ "sim", "humanoid", "rl", "motion-prior", "imitation" ], "robot_types": [ "humanoid", "legged" ], "highlight": false, "added": null, "takeaway_en": "A discriminator trained on unstructured mocap clips acts as a style reward, so RL learns tasks with natural motion without hand-picked reference clips; widely reused as a gait prior for legged robots.", "takeaway_zh": "用在无结构动捕片段上训练的判别器作为\"风格奖励\"，强化学习无需逐段指定参考动作即可学到自然的动作；后来被广泛用作足式机器人的步态先验。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2104.02180.html", "labs": [ "berkeley", "sjtu" ] }, { "id": "arxiv-2401.17583", "title": "Agile But Safe: Learning Collision-Free High-Speed Legged Locomotion", "type": "paper", "arxiv": "2401.17583", "year": 2024, "venue": "RSS 2024", "org": "CMU, ETH Zurich", "url": null, "code": "https://github.com/LeCAR-Lab/ABS", "project": null, "topics": [ "locomotion", "navigation" ], "tags": [ "sim2real", "quadruped", "rl", "safety", "reach-avoid" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "ABS pairs an agile RL locomotion policy with a learned reach-avoid value network and a recovery policy that takes over near obstacles, so a quadruped can run fast without crashing.", "takeaway_zh": "ABS 将敏捷的强化学习运动策略与学习的可达-规避价值网络及恢复策略结合，接近障碍时由恢复策略接管，让四足机器人高速奔跑而不碰撞。", "github_stars": 630, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2401.17583.html", "labs": [ "cmu", "eth" ] }, { "id": "arxiv-2610.18950", "title": "$λ_0$: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation", "type": "paper", "arxiv": "2610.18950", "year": 2026, "venue": null, "org": "Alibaba (Lambda-humanoid Team)", "url": null, "code": "https://github.com/LambdaHumanoid/openlambda", "project": "https://lambdahumanoid.github.io/", "topics": [ "locomotion", "manipulation", "datasets" ], "tags": [ "real", "humanoid", "unitree-g1", "human-video", "egocentric", "whole-body", "foundation-model" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-24", "takeaway_en": "HumanVerse, 500 h of egocentric human video with body and hand motion, aligned to the Unitree G1 for mid-training a humanoid model; reports 90% on SIMPLE and 62.5% on real loco-manipulation tasks.", "takeaway_zh": "构建 500 小时的第一视角人类数据集 HumanVerse（含同步的身体与手部动作），对齐到 Unitree G1 后用于人形基础模型 λ0 的中期训练；论文报告 SIMPLE 上成功率 90%、真实移动操作任务 62.5%，人类数据越多下游表现越好。", "github_stars": 121, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.18950.html", "labs": [ "alibaba" ] }, { "id": "arxiv-2508.08241", "title": "BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion", "type": "paper", "arxiv": "2508.08241", "year": 2025, "venue": null, "org": "UC Berkeley, Stanford", "url": null, "code": "https://github.com/HybridRobotics/whole_body_tracking", "project": null, "topics": [ "locomotion" ], "tags": [ "sim2real", "humanoid", "motion-tracking", "diffusion", "whole-body" ], "robot_types": [ "humanoid" ], "highlight": true, "added": null, "takeaway_en": "One motion-tracking recipe with shared hyperparameters learns agile skills (cartwheels, flip-kicks); a guided latent diffusion model composes them for new goals like joystick control, zero-shot on hardware.", "takeaway_zh": "用同一套超参数的运动跟踪方案学会侧手翻、飞踢、冲刺等敏捷技能，再用带引导的潜空间扩散模型组合这些技能以完成摇杆遥控、避障等新目标，零样本迁移到真机。", "github_stars": 2444, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2508.08241.html", "labs": [ "berkeley", "stanford" ] }, { "id": "arxiv-2610.32966", "title": "ViT-H: Generative Video-To-Humanoid Control with Motion Tracking Feedback", "type": "paper", "arxiv": "2610.32966", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://awfuact.github.io/vit-h/", "topics": [ "locomotion" ], "tags": [ "real", "humanoid", "motion-tracking", "human-video", "diffusion", "teleoperation" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "Skips reconstruction and retargeting: a diffusion model turns monocular human video straight into humanoid reference motion, post-trained with DPO on tracking outcomes; real-time variant teleoperates a G1 under 400 ms.", "takeaway_zh": "跳过人体重建与重定向：扩散模型直接把单目人类视频生成为人形机器人参考动作，并以运动跟踪结果做 DPO 后训练；自回归版本可在 Unitree G1 上实现低于 400 ms 延迟的实时视频遥操作。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.32966.html", "labs": [] }, { "id": "arxiv-2610.24610", "title": "Solving a Giant Rubik's Cube with a Quadruped Robot", "type": "paper", "arxiv": "2610.24610", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://leggedrobotics.github.io/rubik-skills/", "topics": [ "locomotion", "manipulation", "perception" ], "tags": [ "real", "quadruped", "loco-manipulation", "rl", "sim2real", "distillation", "rgbd" ], "robot_types": [ "legged" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "A quadruped solves a giant Rubik's Cube with its feet: onboard RGB-D estimates cube pose and colours, a planner sequences skills, and RL whole-body skills turn layers. Reports three full autonomous solves.", "takeaway_zh": "四足机器人只用脚完成巨型魔方：机载 RGB-D 估计魔方位姿与颜色状态，规划器把解法转成技能序列，强化学习训练的全身技能负责翻转魔方、一脚固定一脚推动转层。报告完成三次完全自主求解。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.24610.html", "labs": [] }, { "id": "arxiv-2610.35862", "title": "Learning Force-Reactive Humanoid Loco-Manipulation", "type": "paper", "arxiv": "2610.35862", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://louis-leee.github.io/FLoRe/", "topics": [ "locomotion", "manipulation", "control" ], "tags": [ "humanoid", "force-control", "whole-body", "diffusion-policy", "umi", "real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "FLoRe: robot-free demos from force-sensing UMI-style grippers + body trackers train a diffusion policy that outputs per-hand force commands for a compliant whole-body controller; reports 80% vs. 26.7% for position-only.", "takeaway_zh": "FLoRe：用带力传感的 UMI 式手持夹爪与人体追踪器采集无机器人示教，训练扩散策略输出每只手的力指令，由柔顺全身控制器执行；报告在三个接触密集任务上成功率 80%，仅位置控制为 26.7%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35862.html", "labs": [] }, { "id": "arxiv-2610.35075", "title": "Zero-HLM: Generalizable Humanoid Loco-Manipulation without Real Robot Data", "type": "paper", "arxiv": "2610.35075", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://young-bit.github.io/zero-hlm/", "topics": [ "locomotion", "manipulation", "sim" ], "tags": [ "humanoid", "real-to-sim", "gaussian-splatting", "data-generation", "unitree-g1", "sim2real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "Real-to-sim-to-real: smartphone scans become physics-ready sim scenes, whole-body motions are corrected by rollouts, giving 4,800 demos; a sim-only GR00T N1.7 on Unitree G1 reports 66.7% vs. 46.7% trained on real data.", "takeaway_zh": "真实到仿真再到真实：用手机拍摄重建可物理仿真的场景，并根据仿真回放修正全身动作，生成 4800 条示教；仅用仿真数据训练的 GR00T N1.7 在宇树 G1 上报告平均成功率 66.7%，真实数据训练为 46.7%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35075.html", "labs": [] }, { "id": "arxiv-2610.34126", "title": "HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model", "type": "paper", "arxiv": "2610.34126", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": null, "topics": [ "locomotion", "learning" ], "tags": [ "humanoid", "motion-tracking", "real-world-finetuning", "neural-dynamics", "unitree-g1" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "Fine-tunes humanoid motion-tracking policies online on the real robot: a proprioceptive neural dynamics model learned from real data supplies gradients, no motion capture needed. Shown on Unitree G1.", "takeaway_zh": "在真机上在线微调人形机器人动作跟踪策略：用真实数据训练的本体感知神经动力学模型提供可反传梯度，无需动捕系统。在 Unitree G1 上验证对多种扰动的快速适应。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.34126.html", "labs": [] }, { "id": "arxiv-2610.35852", "title": "Differentiable Contact Dynamics for Legged Locomotion: Policy Learning and Contact-Implicit Trajectory Optimization", "type": "paper", "arxiv": "2610.35852", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://diffcontact.github.io", "topics": [ "locomotion", "control", "learning" ], "tags": [ "quadruped", "differentiable-simulation", "contact", "trajectory-optimization", "sim2real", "real" ], "robot_types": [ "legged" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "Smooth analytic contact gradients only in the backward pass, hard-contact forward rollout unchanged: a quadruped trotting policy trains in under 2 min on a desktop and transfers to hardware without fine-tuning.", "takeaway_zh": "只在反向传播中换用平滑的解析摩擦接触梯度，前向仍是硬接触仿真：四足小跑策略在台式机上不到 2 分钟训练完成，无需微调即可迁移到真机；在接触隐式轨迹优化中也能找到更丰富的接触解。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35852.html", "labs": [] }, { "id": "arxiv-2610.29903", "title": "HuMiT: Low-Latency Whole-Body Humanoid Teleoperation via Compact Reference Tracking", "type": "paper", "arxiv": "2610.29903", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://xiaohu-art.github.io/Track/", "topics": [ "locomotion", "manipulation" ], "tags": [ "humanoid", "teleoperation", "motion-tracking", "unitree-g1", "real", "vr" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "Whole-body humanoid teleoperation from a compact target (root height and velocity + 4 keypoint poses, current frame only): ~20 ms headset-to-policy latency; a Unitree G1 catches a tossed ball.", "takeaway_zh": "只用紧凑参考目标（根高度、根速度与 4 个稀疏关键点的当前帧位姿）实现人形机器人全身遥操作：头显到策略输出延迟约 20 ms，Unitree G1 能接住抛来的球并完成移动操作。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.29903.html", "labs": [] }, { "id": "arxiv-2610.25711", "title": "SWAP: Test-Time Steerable Humanoid Whole-Body Collision Avoidance", "type": "paper", "arxiv": "2610.25711", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://iscoyizj.github.io/SWAP-steerable-humanoid-oa-website/", "topics": [ "locomotion", "control" ], "tags": [ "humanoid", "collision-avoidance", "diffusion-policy", "whole-body-control", "unitree-g1", "real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "Diffusion planner proposes multimodal collision-aware motions; a whole-body RL controller tracks them. Trained by placing obstacles outside each motion's swept space. Real G1; steering reports +34% on target rewards.", "takeaway_zh": "分层框架：感知条件扩散策略生成多模态、考虑碰撞的动作提议，底层全身 RL 控制器负责跟踪与平衡；训练数据靠事后标注获得（只在已执行动作未扫过的空间放置障碍物）。在真实 Unitree G1 上验证，测试时引导使目标奖励平均提升 34%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25711.html", "labs": [] }, { "id": "arxiv-2610.24675", "title": "Action-Aligned Latent Planning for Versatile Humanoid Loco-Manipulation", "type": "paper", "arxiv": "2610.24675", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://fiberaa.github.io/", "topics": [ "locomotion", "manipulation" ], "tags": [ "humanoid", "loco-manipulation", "behavior-foundation-model", "latent-planning", "unitree-g1", "sim-to-real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "FIBER plans in the latent space of a frozen behavioral foundation model: skills are projected into it, one perceptive flow planner is trained offline. Reports 42.0% mean sim success (latent replay: 11.6%); real G1.", "takeaway_zh": "FIBER 以冻结的行为基础模型（BFM）潜在空间作为统一规划表示，把各任务专用策略投影进该空间，再离线训练单一感知流匹配规划器；仿真中搬箱、开门、抓取平均成功率 42.0%（潜在回放基线 11.6%），并在真实 Unitree G1 上验证。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.24675.html", "labs": [] }, { "id": "arxiv-2610.36546", "title": "REACH: An RL-based Hierarchical Humanoid Crawling Framework for Adaptable Confined Space Traversal", "type": "paper", "arxiv": "2610.36546", "year": 2026, "venue": "Humanoids 2026", "org": "SUSTech", "url": null, "code": null, "project": "https://aislab-sustech.github.io/REACH/", "topics": [ "locomotion" ], "tags": [ "humanoid", "crawling", "rl", "unitree-g1", "lidar", "real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "Unitree G1 walks at variable height (1.27–0.50 m) and crawls (0.50–0.20 m), with a learned transition policy between modes and a LiDAR planner that sets height targets for confined-space traversal on hardware.", "takeaway_zh": "让 Unitree G1 在 1.27–0.50 米高度范围内变高步行、在 0.50–0.20 米范围内爬行，并用学习到的过渡策略在两种模式间切换；结合 LiDAR 规划器给出高度目标，在真机上穿越低矮受限空间。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36546.html", "labs": [] }, { "id": "arxiv-2610.36507", "title": "GRAFT: Geometry-Aware Residual Adaptation for Foundation Trackers", "type": "paper", "arxiv": "2610.36507", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://graft-humanoid.github.io/", "topics": [ "locomotion", "learning" ], "tags": [ "humanoid", "behavior-foundation-model", "residual", "terrain", "depth", "unitree-g1", "real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "Keeps a humanoid behavior foundation model frozen and trains a small lower-body residual adapter for stairs, gaps and platforms, distilled to onboard depth; tested on four trackers and on a Unitree G1 with SONIC.", "takeaway_zh": "冻结人形行为基础模型，仅训练一个作用于下肢目标的轻量残差适配器来应对台阶、沟壑和平台，并从特权地形信息蒸馏到机载深度；在四种跟踪器上验证，并在 Unitree G1（SONIC）上真机部署。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36507.html", "labs": [] }, { "id": "arxiv-2610.36452", "title": "Whole-Body Humanoid Control via Stochastic Motion Tokenization", "type": "paper", "arxiv": "2610.36452", "year": 2026, "venue": "Humanoids 2026", "org": null, "url": null, "code": null, "project": null, "topics": [ "locomotion", "learning" ], "tags": [ "humanoid", "motion-tokenizer", "whole-body-tracking", "text-to-motion", "unitree-g1", "real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "A stochastic motion tokenizer gives a high-level generator a compact interface to a whole-body tracking policy; reports better G1 motion reconstruction than FSQ/VQ-VAE and real dance/fight/walk sequences.", "takeaway_zh": "用随机运动分词器把行为标签或文本生成的动作压缩成离散表示，直接作为全身跟踪策略的参考；报告在 Unitree G1 动作重建上优于 FSQ 与 VQ-VAE，并在真机上执行舞蹈、格斗、行走及 35 秒的格斗转舞蹈序列。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36452.html", "labs": [] }, { "id": "arxiv-2610.30061", "title": "Test-Time Motion Steering for Perceptive Humanoid Fall Recovery", "type": "paper", "arxiv": "2610.30061", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://osheraz.github.io/RISE/", "topics": [ "locomotion" ], "tags": [ "humanoid", "fall-recovery", "diffusion-planner", "perception", "unitree-g1", "real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "RISE learns a vocabulary of recovery movements plus a visual diffusion planner whose predicted intent is steered at test time by task costs (heading, speed, path), no retraining; sim and real Unitree G1.", "takeaway_zh": "RISE 学习一组可组合的起身动作词表，并用视觉扩散规划器预测未来身体状态作为动作意图；推理时用任务代价（朝向、起身速度、路径）引导该意图，无需重新训练即可选择不同起身方式；仿真实验并在 Unitree G1 真机演示。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30061.html", "labs": [] }, { "id": "arxiv-1803.10122", "title": "World Models", "type": "paper", "arxiv": "1803.10122", "year": 2018, "venue": null, "org": "Google Brain, IDSIA", "url": null, "code": null, "project": "https://worldmodels.github.io/", "topics": [ "world-models" ], "tags": [ "sim", "model-based-rl", "vae", "rnn" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The classic VAE + RNN agent that learns a compressed model of its environment and trains a controller inside the \"dream\".", "takeaway_zh": "经典的 VAE + RNN 智能体，学习环境的压缩模型并在\"梦境\"中训练控制器。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1803.10122.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2301.04104", "title": "Mastering Diverse Domains through World Models", "type": "paper", "arxiv": "2301.04104", "year": 2023, "venue": null, "org": "Google DeepMind", "url": null, "code": "https://github.com/danijar/dreamerv3", "project": null, "topics": [ "world-models" ], "tags": [ "sim", "model-based-rl" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "DreamerV3 — one set of hyperparameters works across 150+ tasks, including collecting diamonds in Minecraft from scratch.", "takeaway_zh": "DreamerV3：一套超参数适用于 150 多个任务，包括从零开始在《我的世界》中挖到钻石。", "github_stars": 3915, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2301.04104.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2206.14176", "title": "DayDreamer: World Models for Physical Robot Learning", "type": "paper", "arxiv": "2206.14176", "year": 2022, "venue": null, "org": "UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "world-models" ], "tags": [ "real", "quadruped", "arm", "model-based-rl" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "Dreamer trained directly on real robots — a quadruped learns to walk in about an hour with no simulator.", "takeaway_zh": "直接在真实机器人上训练 Dreamer——四足机器人无需仿真器，约一小时学会行走。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2206.14176.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2310.06114", "title": "Learning Interactive Real-World Simulators", "type": "paper", "arxiv": "2310.06114", "year": 2023, "venue": null, "org": "UC Berkeley, Google DeepMind", "url": null, "code": null, "project": null, "topics": [ "world-models" ], "tags": [ "video-generation", "simulator" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "UniSim — an action-conditioned video model used as a simulator to train policies that transfer to the real world.", "takeaway_zh": "UniSim：以动作为条件的视频生成模型充当仿真器，训练出的策略可迁移到真实世界。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2310.06114.html", "labs": [ "berkeley", "google-deepmind" ] }, { "id": "arxiv-2402.15391", "title": "Genie: Generative Interactive Environments", "type": "paper", "arxiv": "2402.15391", "year": 2024, "venue": null, "org": "Google DeepMind", "url": null, "code": null, "project": null, "topics": [ "world-models" ], "tags": [ "video-generation", "latent-actions" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Learns playable environments and latent actions from unlabeled internet videos alone.", "takeaway_zh": "仅从无标注的网络视频中学习可交互环境与潜在动作。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2402.15391.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2501.03575", "title": "Cosmos World Foundation Model Platform for Physical AI", "type": "paper", "arxiv": "2501.03575", "year": 2025, "venue": null, "org": "NVIDIA", "url": null, "code": null, "project": null, "topics": [ "world-models" ], "tags": [ "video-generation", "open-weights", "synthetic-data" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open video world foundation models plus tokenizers and data pipelines for generating physical-AI training data.", "takeaway_zh": "开放的视频世界基础模型及分词器与数据管线，用于生成物理 AI 训练数据。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2501.03575.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2506.09985", "title": "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning", "type": "paper", "arxiv": "2506.09985", "year": 2025, "venue": null, "org": "Meta AI", "url": null, "code": null, "project": null, "topics": [ "world-models" ], "tags": [ "representation", "planning", "real", "arm" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A self-supervised video encoder, post-trained with a little robot data, plans zero-shot pick-and-place by latent prediction.", "takeaway_zh": "自监督视频编码器经少量机器人数据后训练，通过潜空间预测实现零样本抓放规划。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2506.09985.html", "labs": [ "meta" ] }, { "id": "arxiv-1912.01603", "title": "Dream to Control: Learning Behaviors by Latent Imagination", "type": "paper", "arxiv": "1912.01603", "year": 2019, "venue": null, "org": "University of Toronto, Google", "url": null, "code": null, "project": null, "topics": [ "world-models" ], "tags": [ "model-based-rl", "latent-dynamics" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The first Dreamer — learns a latent world model from pixels and trains the policy entirely on imagined rollouts.", "takeaway_zh": "第一代 Dreamer：从像素学习潜空间世界模型，并完全在想象的轨迹上训练策略。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1912.01603.html", "labs": [ "google-deepmind", "toronto" ] }, { "id": "arxiv-2310.16828", "title": "TD-MPC2: Scalable, Robust World Models for Continuous Control", "type": "paper", "arxiv": "2310.16828", "year": 2023, "venue": null, "org": "UC San Diego", "url": null, "code": "https://github.com/nicklashansen/tdmpc2", "project": null, "topics": [ "world-models", "control" ], "tags": [ "model-based-rl", "mpc", "multitask" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "An implicit (decoder-free) world model planned with MPC; one set of hyperparameters works across many continuous-control tasks and scales to a multitask agent.", "takeaway_zh": "无解码器的隐式世界模型配合 MPC 规划；同一组超参数适用于大量连续控制任务，并可扩展为多任务智能体。", "github_stars": 992, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2310.16828.html", "labs": [ "ucsd" ] }, { "id": "arxiv-2411.04983", "title": "DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning", "type": "paper", "arxiv": "2411.04983", "year": 2024, "venue": null, "org": "NYU", "url": null, "code": "https://github.com/gaoyuezhou/dino_wm", "project": null, "topics": [ "world-models" ], "tags": [ "planning", "representation", "offline" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Predicts future DINOv2 patch features instead of pixels, so a world model trained on offline data can plan to new visual goals at test time.", "takeaway_zh": "预测未来的 DINOv2 图像块特征而非像素，使基于离线数据训练的世界模型能在测试时规划到新的视觉目标。", "github_stars": 614, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2411.04983.html", "labs": [ "nyu" ] }, { "id": "arxiv-2610.26855", "title": "Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience", "type": "paper", "arxiv": "2610.26855", "year": 2026, "venue": null, "org": null, "url": "https://ropedia.github.io/ropedia-w0/", "code": "https://github.com/Ropedia/ropedia-w0", "project": "https://ropedia.github.io/ropedia-w0/", "topics": [ "world-models" ], "tags": [ "egocentric", "human-data", "cross-embodiment", "loco-manipulation" ], "robot_types": [], "highlight": false, "added": "2026-10-20", "takeaway_en": "An egocentric world model covering both walking and manipulating, trained on ~10k hours of human video plus ~3k hours of robot data under one shared kinematic action representation.", "takeaway_zh": "同时覆盖移动与操作的第一人称世界模型，在约 1 万小时人类视频和约 3 千小时机器人数据上训练，用统一的运动学表示描述不同本体的动作。", "github_stars": 147, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.26855.html", "labs": [] }, { "id": "arxiv-2601.07823", "title": "Robotic Video World Models: A Survey of Applications, Research Challenges, Future Directions", "type": "paper", "arxiv": "2601.07823", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/irom-princeton/awesome-robotics-video-world-model-papers", "project": null, "topics": [ "world-models" ], "tags": [ "survey", "video-generation", "awesome-list" ], "robot_types": [], "highlight": false, "added": "2026-10-21", "takeaway_en": "A survey of video world models in robotics - where they replace or complement physics simulators (e.g. deformable objects), their applications, and open challenges - with a maintained paper list.", "takeaway_zh": "机器人视频世界模型综述：它们在哪些场景可替代或补充物理仿真器（如可变形物体）、主要应用与开放挑战，并附持续更新的论文列表。", "github_stars": 2, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2601.07823.html", "labs": [] }, { "id": "arxiv-1811.04551", "title": "Learning Latent Dynamics for Planning from Pixels", "type": "paper", "arxiv": "1811.04551", "year": 2019, "venue": "ICML 2019", "org": "Google Brain, DeepMind", "url": null, "code": "https://github.com/google-research/planet", "project": null, "topics": [ "world-models" ], "tags": [ "sim", "model-based-rl", "planning", "latent-dynamics" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "PlaNet — learns a latent dynamics model (the RSSM later reused by Dreamer) from images and plans actions directly in latent space with CEM.", "takeaway_zh": "PlaNet：从图像中学习潜空间动力学模型（即后来 Dreamer 沿用的 RSSM），并用 CEM 直接在潜空间中规划动作。", "github_stars": 1263, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1811.04551.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2010.02193", "title": "Mastering Atari with Discrete World Models", "type": "paper", "arxiv": "2010.02193", "year": 2021, "venue": "ICLR 2021", "org": "Google Research, DeepMind", "url": null, "code": "https://github.com/danijar/dreamerv2", "project": null, "topics": [ "world-models" ], "tags": [ "sim", "model-based-rl", "discrete-latents" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "DreamerV2 — switches Dreamer to categorical latent variables and learns Atari behaviour purely inside the world model, reaching human-level scores.", "takeaway_zh": "DreamerV2：把 Dreamer 的潜变量改为离散类别分布，仅在世界模型的想象中学习 Atari 策略，达到人类水平。", "github_stars": 1061, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2010.02193.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2209.00588", "title": "Transformers are Sample-Efficient World Models", "type": "paper", "arxiv": "2209.00588", "year": 2023, "venue": "ICLR 2023", "org": "University of Geneva", "url": null, "code": "https://github.com/eloialonso/iris", "project": null, "topics": [ "world-models" ], "tags": [ "sim", "model-based-rl", "transformer", "tokenizer" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "IRIS — a discrete autoencoder turns frames into tokens and a Transformer predicts the next tokens; the agent learns in imagination with only ~2 hours of Atari gameplay.", "takeaway_zh": "IRIS：离散自编码器把画面变成 token，Transformer 预测后续 token；智能体只用约 2 小时的 Atari 游戏经验便在想象中学会策略。", "github_stars": 903, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2209.00588.html", "labs": [] }, { "id": "arxiv-2302.00111", "title": "Learning Universal Policies via Text-Guided Video Generation", "type": "paper", "arxiv": "2302.00111", "year": 2023, "venue": "NeurIPS 2023", "org": "MIT, Google", "url": null, "code": null, "project": null, "topics": [ "world-models" ], "tags": [ "video-generation", "planning", "language" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "UniPi — casts decision making as text-conditioned video generation, then recovers actions from the generated frames with an inverse dynamics model.", "takeaway_zh": "UniPi：把决策问题转化为文本条件的视频生成，再用逆动力学模型从生成的画面中恢复动作。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2302.00111.html", "labs": [ "google-deepmind", "mit" ] }, { "id": "arxiv-2203.04955", "title": "Temporal Difference Learning for Model Predictive Control", "type": "paper", "arxiv": "2203.04955", "year": 2022, "venue": "ICML 2022", "org": "UC San Diego", "url": null, "code": "https://github.com/nicklashansen/tdmpc", "project": null, "topics": [ "world-models" ], "tags": [ "sim", "model-based-rl", "mpc", "planning" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "TD-MPC — a task-oriented latent model used for short-horizon planning, with a learned value function covering the long-term return; the predecessor of TD-MPC2.", "takeaway_zh": "TD-MPC：用面向任务的潜空间模型做短视距规划，由学习到的价值函数估计长期回报；是 TD-MPC2 的前身。", "github_stars": 529, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2203.04955.html", "labs": [ "ucsd" ] }, { "id": "arxiv-2610.30117", "title": "CW-WAM: A Structured Contact-Wrench World Action Model for Force-Sensitive Manipulation", "type": "paper", "arxiv": "2610.30117", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://cw-wam-project.github.io/", "topics": [ "world-models", "manipulation" ], "tags": [ "world-action-model", "force-control", "contact-rich", "video-prediction" ], "robot_types": [], "highlight": false, "added": "2026-10-22", "takeaway_en": "World action model that jointly predicts video, motion and contact wrench, outputting time-aligned motion–wrench chunks as control references (hybrid motion-force style) for force-sensitive manipulation.", "takeaway_zh": "一个同时预测未来视频、运动与接触力旋量的世界动作模型，借鉴力位混合控制思路，输出时间对齐的运动-力旋量片段作为控制参考，面向仅靠视觉无法感知接触与负载的力敏感操作任务。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30117.html", "labs": [] }, { "id": "arxiv-2312.13139", "title": "Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation", "type": "paper", "arxiv": "2312.13139", "year": 2024, "venue": "ICLR 2024", "org": "ByteDance Research", "url": null, "code": "https://github.com/bytedance/GR-1", "project": null, "topics": [ "world-models", "manipulation" ], "tags": [ "world-action-model", "video-prediction", "pretraining", "gpt" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "GR-1 — a GPT-style model that predicts both actions and future frames; pre-training on Ego4D video prediction before robot fine-tuning lifts CALVIN unseen-scene success from 53.3% to 85.4% (as reported).", "takeaway_zh": "GR-1：GPT 式模型同时预测动作与未来画面；先在 Ego4D 上做视频预测预训练、再用机器人数据微调，论文报告 CALVIN 未见场景成功率从 53.3% 提升到 85.4%。", "github_stars": 308, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2312.13139.html", "labs": [ "bytedance" ] }, { "id": "arxiv-2412.14803", "title": "Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations", "type": "paper", "arxiv": "2412.14803", "year": 2025, "venue": "ICML 2025", "org": "Tsinghua University, Shanghai Qi Zhi Institute, RobotEra", "url": null, "code": "https://github.com/roboterax/video-prediction-policy", "project": null, "topics": [ "world-models", "manipulation" ], "tags": [ "world-action-model", "video-diffusion", "diffusion-policy", "dexterous" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "VPP — fine-tunes Stable Video Diffusion on robot and human videos and feeds its one-pass predictive features to a diffusion policy; reports 4.33 average length on CALVIN ABC→D and real Franka / dexterous-hand results.", "takeaway_zh": "VPP：在机器人与人类操作视频上微调 Stable Video Diffusion，只做一次前向提取\"预测性\"视觉特征再交给扩散策略；论文报告 CALVIN ABC→D 平均完成长度 4.33，并有 Franka 与灵巧手真机实验。", "github_stars": 418, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2412.14803.html", "labs": [ "tsinghua", "qizhi" ] }, { "id": "arxiv-2503.00200", "title": "Unified Video Action Model", "type": "paper", "arxiv": "2503.00200", "year": 2025, "venue": "RSS 2025", "org": "Stanford University", "url": null, "code": "https://github.com/ShuangLI59/unified_video_action", "project": null, "topics": [ "world-models", "manipulation" ], "tags": [ "world-action-model", "video-generation", "diffusion", "umi" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "UVA — one shared latent for video and actions with two light diffusion heads; masked training lets the same model act as policy, video generator and forward/inverse dynamics model, while keeping policy inference fast.", "takeaway_zh": "UVA：视频与动作共享一个潜表示，配两个轻量扩散头解码；通过掩码训练，同一模型既能当策略，也能做视频生成和正/逆动力学，且策略推理仍然较快。", "github_stars": 415, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2503.00200.html", "labs": [ "stanford" ] }, { "id": "arxiv-2506.21539", "title": "WorldVLA: Towards Autoregressive Action World Model", "type": "paper", "arxiv": "2506.21539", "year": 2025, "venue": null, "org": "Alibaba DAMO Academy, Zhejiang University", "url": null, "code": "https://github.com/alibaba-damo-academy/WorldVLA", "project": null, "topics": [ "world-models", "vla" ], "tags": [ "world-action-model", "autoregressive", "libero" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "One autoregressive model that is both a VLA policy and an image-predicting world model; the two help each other on LIBERO, and an attention mask hiding earlier actions curbs error build-up in action chunks.", "takeaway_zh": "单个自回归模型同时充当 VLA 策略与预测未来图像的世界模型，二者在 LIBERO 上相互促进；论文还发现自回归生成动作块会累积误差，用屏蔽先前动作的注意力掩码可以缓解。", "github_stars": 1142, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2506.21539.html", "labs": [ "alibaba" ] }, { "id": "arxiv-2508.05635", "title": "Genie Envisioner: A Unified World Foundation Platform for Robotic Manipulation", "type": "paper", "arxiv": "2508.05635", "year": 2025, "venue": "ICLR 2026", "org": "AgiBot, NUS, Beihang University", "url": null, "code": "https://github.com/AgibotTech/Genie-Envisioner-V1", "project": null, "topics": [ "world-models", "manipulation", "sim" ], "tags": [ "world-action-model", "video-diffusion", "neural-simulator", "benchmark" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Genie Envisioner (AgiBot) — a video-diffusion world model (GE-Base) plus a flow-matching action decoder (GE-Act), an action-conditioned neural simulator (GE-Sim) and the EWMBench benchmark in one platform.", "takeaway_zh": "智元 Genie Envisioner：一个平台集成视频扩散世界模型 GE-Base、流匹配动作解码器 GE-Act、动作条件神经仿真器 GE-Sim，以及世界模型评测基准 EWMBench。", "github_stars": 589, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2508.05635.html", "labs": [ "agibot", "nus" ] }, { "id": "arxiv-2610.30991", "title": "UniFlowWAM: A Unified World-Action Model with Visual Semantics and 4D Geometry", "type": "paper", "arxiv": "2610.30991", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://kksham111.github.io/UniFlowWAM_ProjectPage/", "topics": [ "world-models", "vla" ], "tags": [ "world-action-model", "3d-flow", "4d-geometry" ], "robot_types": [], "highlight": false, "added": "2026-10-23", "takeaway_en": "World action model that brings scene-level 3D flow (tracked 3D points) inside the policy rather than as a separate module or extra head, so 4D geometry and visual semantics are modelled together with actions.", "takeaway_zh": "世界动作模型：把场景级三维流（跟踪的三维点）纳入策略内部，而不是作为独立模块或额外输出头，让 4D 几何、视觉语义与动作一起建模。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30991.html", "labs": [] }, { "id": "arxiv-2610.18774", "title": "GreenWAM: A World Foundation Model for Robotics", "type": "paper", "arxiv": "2610.18774", "year": 2026, "venue": null, "org": "Sber Robotics Center", "url": null, "code": "https://github.com/greenvla/GreenWAM", "project": "https://greenvla.github.io/GreenWAM", "topics": [ "world-models", "vla" ], "tags": [ "world-action-model", "cross-embodiment", "humanoid", "technical-report" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-23", "takeaway_en": "One world model that is also the policy, across 20 robot families mapped to a shared joint-space action format; reports strong WorldArena, RoboTwin 2.0 and LIBERO-plus results and a one-step distilled humanoid policy.", "takeaway_zh": "同一个世界模型兼作策略，用共享的关节坐标动作空间覆盖 20 类机器人；论文报告其在 WorldArena、RoboTwin 2.0、LIBERO-plus 上表现突出，并蒸馏出可实时运行的单步人形机器人策略。", "github_stars": 11, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.18774.html", "labs": [] }, { "id": "arxiv-2610.30320", "title": "Self-Distilling Future Video Representations for World Action Models", "type": "paper", "arxiv": "2610.30320", "year": 2026, "venue": null, "org": "KAIST AI", "url": null, "code": "https://github.com/suhyeok-jang/FRED", "project": "https://huiwon-jang.github.io/FRED/", "topics": [ "world-models" ], "tags": [ "world-action-model", "video-diffusion", "self-distillation", "human-to-robot", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "FRED trains a world action model to predict better one-step future video features via self-distillation plus history context; reports large gains on long-horizon, out-of-domain and human-video-only tasks.", "takeaway_zh": "FRED 通过自蒸馏并引入历史观测，让世界动作模型在单步去噪中预测出更好的未来视频表征；报告在长时程、域外环境以及只有人类视频示范的任务上提升明显。", "github_stars": 10, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30320.html", "labs": [ "kaist" ] }, { "id": "arxiv-2610.30136", "title": "Agentic World Action Models", "type": "paper", "arxiv": "2610.30136", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://agentic-wam.github.io/", "topics": [ "world-models", "manipulation" ], "tags": [ "world-action-model", "agentic", "code-as-policy", "skill-library", "bimanual", "sim", "real" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-25", "takeaway_en": "AW-0 jointly generates code, future video and actions in one mixture-of-transformers; code structures planning and becomes reusable skills. Reports gains over VLA and WAM baselines on RoboTwin 2.0 and real two-arm tasks.", "takeaway_zh": "AW-0 在同一个混合 Transformer 中联合生成代码、未来视频和机器人动作三路输出：代码既组织多步预测与控制，又沉淀为可复用技能库；报告在 RoboTwin 2.0 和真机双臂任务上优于 VLA 与 WAM 基线。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30136.html", "labs": [] }, { "id": "arxiv-2610.31474", "title": "Skeleton-Grounded Latent Action Model for Robot World Models", "type": "paper", "arxiv": "2610.31474", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://aopolin-lv.github.io/SkeLAM/", "topics": [ "world-models" ], "tags": [ "latent-action", "video-world-model", "droid", "policy-evaluation" ], "robot_types": [], "highlight": false, "added": "2026-10-25", "takeaway_en": "SkeLAM splits latent actions into a robot-skeleton-grounded kinematic part and a visual residual, so world-model rollouts follow robot controls; on DROID reports 19.5% lower end-effector error than DreamDojo.", "takeaway_zh": "SkeLAM 把潜在动作拆成由机器人骨架监督的运动学部分和保留其余视觉动态的残差部分，使世界模型生成更贴合机器人控制指令；报告在 DROID 上末端轨迹误差比 DreamDojo 低 19.5%，预测成功率与真机评测一致。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31474.html", "labs": [] }, { "id": "arxiv-2610.32965", "title": "RynnUnify: Towards Unified Embodied Brain", "type": "paper", "arxiv": "2610.32965", "year": 2026, "venue": null, "org": "Alibaba DAMO Academy", "url": null, "code": "https://github.com/alibaba-damo-academy/RynnUnify", "project": "https://alibaba-damo-academy.github.io/RynnUnify.github.io/", "topics": [ "world-models", "vla", "perception" ], "tags": [ "unified-model", "video-prediction", "visual-planning", "world-action-model", "benchmark" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "A 16B model interleaving video prediction and language reasoning for embodied understanding and long-horizon visual planning, plus a world-action extension for control; reports 51.5% on 20 RoboCasa365 tasks.", "takeaway_zh": "16B 统一具身模型，交替进行视频预测与语言推理，兼顾具身理解与长程视觉规划，并扩展出用于机器人控制的世界-动作版本；另提出 Rynn-VP、Rynn-GEU 评测，报告在 20 个 RoboCasa365 组合任务上成功率 51.5%。", "github_stars": 44, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.32965.html", "labs": [ "alibaba" ] }, { "id": "arxiv-2505.12705", "title": "DreamGen: Unlocking Generalization in Robot Learning through Video World Models", "type": "paper", "arxiv": "2505.12705", "year": 2025, "venue": "CoRL 2025", "org": "NVIDIA", "url": null, "code": "https://github.com/NVIDIA/GR00T-Dreams", "project": "https://research.nvidia.com/labs/gear/dreamgen", "topics": [ "world-models", "learning", "locomotion" ], "tags": [ "synthetic-data", "video-world-model", "inverse-dynamics", "humanoid", "real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Fine-tunes video world models on a robot to dream new demos, labels them with latent actions or an inverse dynamics model, and trains policies on them; a humanoid learns 20+ new behaviors from one-scene teleop data.", "takeaway_zh": "把视频世界模型微调到目标机器人上“做梦”生成新示范，再用隐动作或逆动力学模型补上动作标签用于训练策略；人形机器人仅凭单一场景的遥操作数据学会 20 多种新行为。", "github_stars": 626, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2505.12705.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2610.28613", "title": "Physical Canvas: A Vision Tactile World Model For Robot Policy Evaluation And Learning", "type": "paper", "arxiv": "2610.28613", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://zebinx.github.io/PhysCanvas_HomePage/", "topics": [ "world-models", "tactile" ], "tags": [ "visuotactile", "video-generation", "policy-evaluation", "grpo", "real", "contact-rich" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Action-conditioned world model that predicts multi-view RGB and tactile images together in one video space, post-trained for state consistency; used for closed-loop policy evaluation and GRPO updates.", "takeaway_zh": "动作条件的视触觉世界模型：在统一的视频空间中同时预测多视角 RGB 与触觉图像，并经一致性后训练保证与真实任务状态吻合；可用于闭环策略评估和 GRPO 策略更新。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.28613.html", "labs": [] }, { "id": "arxiv-2610.26813", "title": "PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models", "type": "paper", "arxiv": "2610.26813", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/YudeZou/PICA-WAM-release", "project": null, "topics": [ "world-models", "vla" ], "tags": [ "wam", "robustness", "self-supervised", "ema-teacher", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Auxiliary training task for world action models: an EMA teacher sees the clean frame, the online model sees a perturbed one and aligns its action features. Reports better robustness for four WAMs.", "takeaway_zh": "为世界-动作模型加一个自监督辅助任务：EMA 教师看原始画面，在线模型看外观扰动后的画面并对齐其动作表征；报告在 RoboTwin 2.0、LIBERO-Plus 等基准上提升四种 WAM 对视觉扰动的鲁棒性，且不损失原有性能。", "github_stars": 7, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.26813.html", "labs": [] }, { "id": "arxiv-2610.35850", "title": "OSWAM: An Effective World Action Model with One-Step Action Generation", "type": "paper", "arxiv": "2610.35850", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://gwenzhang.github.io/OSWAM/", "topics": [ "world-models", "vla" ], "tags": [ "wam", "one-step", "real-time", "distillation", "driving", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "World action model that acts in one diffusion step: 'imagination tokens' replace random-noise initialisation as an action prior. Reports ~3× speedup over direct-policy WAMs, 40 ms latency on an RTX 5090.", "takeaway_zh": "单步扩散即可生成动作的世界动作模型：用“想象 token”代替随机噪声初始化，为动作提供先验。报告比直接策略式 WAM 快约 3 倍，编译后在 RTX 5090 上延迟低至 40 毫秒，覆盖自动驾驶与具身控制。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35850.html", "labs": [] }, { "id": "arxiv-2610.34663", "title": "Video World Models Need Predictable Latents", "type": "paper", "arxiv": "2610.34663", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/mattia-scardecchia/predictable-latents", "project": null, "topics": [ "world-models", "learning" ], "tags": [ "video-world-model", "latent-space", "autoencoder", "policy-evaluation" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Train the video world model's autoencoder for temporal predictability, not just reconstruction, with a past-conditioned decoder. Reports up to 5× cheaper inference, 7× faster training and lower long-horizon / OOD error.", "takeaway_zh": "视频世界模型的潜空间不应只为重建而训练：加入时间可预测性目标并配合以历史为条件的解码器。报告推理成本最多降低 5 倍、训练快 7 倍以上，长时域与分布外误差降低 15–24%，与真实策略评测更一致。", "github_stars": 4, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.34663.html", "labs": [] }, { "id": "arxiv-2610.30781", "title": "DeXplicit: Making Physical State Explicit in Dexterous Manipulation World Models", "type": "paper", "arxiv": "2610.30781", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://dexplicit.github.io", "topics": [ "world-models", "manipulation" ], "tags": [ "dexterous", "3d-dynamics", "video-generation", "hand-object" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-27", "takeaway_en": "Dexterous-manipulation world model that first predicts object motion and deformation in metric 3D from a hand trajectory, then renders video. Reports 27–40% lower moved-point error than PointWorld.", "takeaway_zh": "灵巧操作世界模型：先根据手部轨迹在度量 3D 空间中显式预测物体运动与形变，再据此渲染视频。报告在 HOT3D、OakInk2、PhysTwin 上运动点误差比 PointWorld 低 27–40%，两步因果生成大幅提速。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30781.html", "labs": [] }, { "id": "arxiv-2610.27387", "title": "Training Controllable World Models via Action Influence Maximization", "type": "paper", "arxiv": "2610.27387", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": null, "topics": [ "world-models", "learning" ], "tags": [ "world-model", "video-generation", "data-collection", "controllability", "active-learning" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Why action-conditioned video models ignore actions: expert demos carry too little action influence. AIM actively collects data to raise it; 500 such episodes + 3k demos match a 300k-demo model on controllability.", "takeaway_zh": "解释动作条件视频世界模型为何忽略动作：专家示教的“动作影响力”太低，模型容量再大也无法弥补。AIM 主动采集数据提高这一指标；报告 500 条主动采集数据加 3000 条示教即可在末端可控性上追平 30 万条专家数据。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.27387.html", "labs": [] }, { "id": "arxiv-2610.25763", "title": "TrajWAM: An Efficient Trajectory-Guided World Action Model for Dynamic Robotic Manipulation", "type": "paper", "arxiv": "2610.25763", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://kevingeng3.github.io/TrajWAM/", "topics": [ "world-models", "manipulation" ], "tags": [ "wam", "dynamic-manipulation", "video-generation", "distillation", "one-step", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "World action model for moving targets: forecasts the object's future masks, steers a video model with a Trajectory ControlNet, and distills the pipeline to one-step video-action inference to cut latency.", "takeaway_zh": "面向动态目标的世界动作模型：先预测目标物体的未来掩码轨迹，用 Trajectory ControlNet 引导视频模型生成与轨迹一致的未来，再通过解耦蒸馏实现一步视频-动作推理以降低延迟。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25763.html", "labs": [] }, { "id": "arxiv-2610.37091", "title": "GAWM: A Compact Latent World-Action Model for Multi-Task Robot Manipulation", "type": "paper", "arxiv": "2610.37091", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/gaoxiang12/ga-wm", "project": null, "topics": [ "world-models", "manipulation" ], "tags": [ "wam", "dinov3", "latent-prediction", "robotwin", "libero", "bimanual" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "Compact (<100M) world-action model: predicts frozen-DINOv3 tokens at two future horizons and decodes action chunks, with no robot or video pretraining. Reports beating billion-parameter WAMs on RoboTwin 2.0.", "takeaway_zh": "参数不到 1 亿的紧凑世界动作模型：在冻结 DINOv3 特征空间中预测两个未来时刻的 token 并直接回归动作块，无需机器人或视频预训练；报告在 RoboTwin 2.0 上超过十亿参数级 WAM。", "github_stars": 10, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.37091.html", "labs": [] }, { "id": "arxiv-2610.36823", "title": "TAPESTRY: A Single Backbone for Video Generation and Robot Control", "type": "paper", "arxiv": "2610.36823", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": "https://tapestry-wam.github.io", "topics": [ "world-models", "manipulation" ], "tags": [ "wam", "video-diffusion", "action-tokens", "real", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Turns a pretrained video diffusion transformer into a policy by encoding each action step as one patch in the video latent space; at test time it outputs an action chunk in one DDIM step (84 ms), no video generation.", "takeaway_zh": "把预训练视频扩散 Transformer 改造成策略：每个时间步的动作编码为视频潜空间中的一个 patch，联合训练视频、视频+动作与纯动作三种扩散目标；测试时一步 DDIM 生成动作块（84 毫秒），无需生成视频。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36823.html", "labs": [] }, { "id": "arxiv-2610.36791", "title": "RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions", "type": "paper", "arxiv": "2610.36791", "year": 2026, "venue": null, "org": "Alibaba DAMO Academy", "url": null, "code": "https://github.com/alibaba-damo-academy/RynnWorld-Latent", "project": "https://alibaba-damo-academy.github.io/RynnWorld-Latent.github.io", "topics": [ "world-models", "learning" ], "tags": [ "world-model", "latent-actions", "cross-embodiment", "human-video", "data-generation" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Cross-embodiment video world model conditioned on RynnLAM's 3D latent actions, pretrained on 23 datasets (~553M frames) of human, robot and sim video; its generated data improves LIBERO and real-robot policies.", "takeaway_zh": "以 RynnLAM 的三维潜在动作为统一条件接口的跨本体视频世界模型，在 23 个数据集（约 5.5 亿帧）的人类、机器人和仿真视频上预训练；生成的数据提升了 LIBERO 与真机策略成功率。", "github_stars": 11, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36791.html", "labs": [ "alibaba" ] }, { "id": "arxiv-2610.36068", "title": "Rollback: Experience Backtracking in World Models for Policy Improvement", "type": "paper", "arxiv": "2610.36068", "year": 2026, "venue": null, "org": "AgiBot", "url": null, "code": null, "project": "https://rollback-wm.github.io/", "topics": [ "world-models", "learning" ], "tags": [ "world-model", "counterfactual", "policy-improvement", "real-world-rl", "lora", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Rewinds an action-conditioned world model to key decision points to try alternative actions, turning outcomes into action-level supervision; real data keeps correcting the model. Reports +10.7–11.9 pts on 3 real tasks.", "takeaway_zh": "把动作条件世界模型当作反事实引擎：回到关键决策点尝试其他动作并对比结果，得到动作级监督而无需单独的价值模型；真机经验再通过 LoRA 持续修正世界模型。报告在 3 个真实任务上比最强基线高 10.7–11.9 个百分点。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36068.html", "labs": [ "agibot" ] }, { "id": "arxiv-2610.36840", "title": "Trace-WAM: Learning to Act via 3D Motion Prediction for World Action Models", "type": "paper", "arxiv": "2610.36840", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://trace-wam.github.io/", "topics": [ "world-models", "manipulation" ], "tags": [ "wam", "3d-motion", "point-tracking", "mixture-of-transformers", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Adds future 3D point-motion prediction as an auxiliary task to a world action model (video/motion/action MoT); inference stays one visual pass. Reports gains over Fast-WAM with few demos and under shift.", "takeaway_zh": "在世界动作模型中加入未来三维点运动预测作为辅助目标（视频/运动/动作三专家 MoT），推理仍只需一次视觉前向；报告在少量示范和分布偏移下明显优于 Fast-WAM（RoboCasa-GR1、LIBERO、LIBERO-Plus）。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36840.html", "labs": [] }, { "id": "arxiv-2610.36624", "title": "Adapting Video Models for Robot Control via Motion-guided Asymmetric Prediction", "type": "paper", "arxiv": "2610.36624", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://ismty0805.github.io/mogap/", "topics": [ "world-models", "manipulation" ], "tags": [ "wam", "video-diffusion", "auxiliary-loss", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "MoGAP adds a training-only auxiliary path to world action models that corrupts moving regions more than static ones, focusing visual supervision on motion. Reports GR-1 tabletop gains (Fast-WAM 38.8% → 47.5%).", "takeaway_zh": "MoGAP 为世界动作模型增加一条仅在训练时使用的辅助预测分支：对运动区域加更重的噪声、以静态区域为上下文重建，使视觉监督聚焦于动态部分；报告在 GR-1 桌面基准上 Fast-WAM 从 38.8% 提升到 47.5%，Cosmos3-Nano 从 52.3% 到 62.3%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36624.html", "labs": [] }, { "id": "arxiv-2610.35877", "title": "Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention", "type": "paper", "arxiv": "2610.35877", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/noumenal-ai/conformal-realrobot-lab", "project": null, "topics": [ "world-models", "learning" ], "tags": [ "world-model", "uncertainty", "conformal-prediction", "distribution-shift", "theory", "real" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "Shows a fixed world model's uncertainty can break when the deployment policy changes; reweights conformal calibration by a learned state-action ratio. Reports 90% sets covering 55.0% vs 90.9% on real Franka data.", "takeaway_zh": "指出部署策略一变，固定世界模型的不确定性校准就可能失效；用域分类器学到的状态-动作密度比做加权共形校准，并给出覆盖率保证（部分用 Lean 4 形式化验证）。报告在真实 Franka 轨迹上，普通 90% 共形集合仅覆盖 55.0%，加权后为 90.9%。", "github_stars": 0, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35877.html", "labs": [] }, { "id": "arxiv-2204.01691", "title": "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances", "type": "paper", "arxiv": "2204.01691", "year": 2022, "venue": null, "org": "Google", "url": null, "code": null, "project": null, "topics": [ "navigation" ], "tags": [ "real", "mobile", "llm", "planning" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "SayCan — an LLM proposes steps and learned value functions score what the robot can actually do right now.", "takeaway_zh": "SayCan：大模型提出步骤，学习到的价值函数评估机器人当下实际能做什么。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2204.01691.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2209.07753", "title": "Code as Policies: Language Model Programs for Embodied Control", "type": "paper", "arxiv": "2209.07753", "year": 2022, "venue": null, "org": "Google", "url": null, "code": null, "project": null, "topics": [ "navigation", "manipulation" ], "tags": [ "real", "llm", "code-generation" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "LLMs write Python that calls perception and control APIs, turning instructions into reactive robot programs.", "takeaway_zh": "大模型编写调用感知与控制 API 的 Python 代码，把指令变成可响应的机器人程序。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2209.07753.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2207.05608", "title": "Inner Monologue: Embodied Reasoning through Planning with Language Models", "type": "paper", "arxiv": "2207.05608", "year": 2022, "venue": null, "org": "Google", "url": null, "code": null, "project": null, "topics": [ "navigation" ], "tags": [ "real", "llm", "planning", "feedback" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Feeding success detection and scene descriptions back to the LLM planner lets it replan after failures.", "takeaway_zh": "将成功检测与场景描述反馈给大模型规划器，使其在失败后重新规划。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2207.05608.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-1711.07280", "title": "Vision-and-Language Navigation: Interpreting visually-grounded navigation instructions in real environments", "type": "paper", "arxiv": "1711.07280", "year": 2017, "venue": null, "org": "Australian National University", "url": null, "code": null, "project": null, "topics": [ "navigation", "datasets" ], "tags": [ "sim", "benchmark", "vln" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Introduced the VLN task and the Room-to-Room (R2R) benchmark in Matterport3D homes.", "takeaway_zh": "提出视觉语言导航任务与基于 Matterport3D 的 Room-to-Room (R2R) 基准。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1711.07280.html", "labs": [] }, { "id": "arxiv-2210.05714", "title": "Visual Language Maps for Robot Navigation", "type": "paper", "arxiv": "2210.05714", "year": 2022, "venue": null, "org": "Freiburg, Google", "url": null, "code": null, "project": null, "topics": [ "navigation" ], "tags": [ "real", "mapping", "open-vocabulary" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "VLMaps fuses visual-language features into a 3D map so robots can navigate to open-vocabulary spatial goals.", "takeaway_zh": "VLMaps 将视觉语言特征融合进 3D 地图，使机器人可按开放词汇空间目标导航。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2210.05714.html", "labs": [ "google-deepmind", "freiburg" ] }, { "id": "arxiv-2310.07896", "title": "NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration", "type": "paper", "arxiv": "2310.07896", "year": 2023, "venue": null, "org": "UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "navigation" ], "tags": [ "real", "diffusion", "exploration" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "One diffusion policy handles both goal-directed navigation and exploration by masking the goal.", "takeaway_zh": "通过对目标做掩码，单一扩散策略同时完成目标导航与探索。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2310.07896.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2401.12202", "title": "OK-Robot: What Really Matters in Integrating Open-Knowledge Models for Robotics", "type": "paper", "arxiv": "2401.12202", "year": 2024, "venue": null, "org": "NYU, Meta AI", "url": null, "code": null, "project": null, "topics": [ "navigation", "manipulation" ], "tags": [ "real", "mobile", "open-vocabulary", "zero-shot" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Combines off-the-shelf open-vocabulary models into a pick-and-drop system that works in real homes without training.", "takeaway_zh": "组合现成的开放词汇模型，构建无需训练即可在真实家庭中抓取放置的系统。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2401.12202.html", "labs": [ "meta", "nyu" ] }, { "id": "arxiv-2610.19600", "title": "Cracking Mobile Manipulation: From Failure Diagnosis to Agentic Robot System", "type": "paper", "arxiv": "2610.19600", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/PKU-VaLuE-Lab/Cracking-Mobile-Manipulation", "project": null, "topics": [ "navigation", "manipulation", "vla" ], "tags": [ "real", "mobile", "failure-analysis", "agentic" ], "robot_types": [ "mobile" ], "highlight": false, "added": "2026-10-19", "takeaway_en": "Finds a VLA at 80% on atomic skills drops to 19% when composed into mobile manipulation, diagnoses why, and builds an agentic system around it.", "takeaway_zh": "发现 VLA 在原子技能上 80% 成功率、组合成移动操作后仅 19%，分析失败原因并据此构建智能体式系统。", "github_stars": 7, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.19600.html", "labs": [] }, { "id": "arxiv-2610.20607", "title": "Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance", "type": "paper", "arxiv": "2610.20607", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": "https://hire-moma.github.io/", "topics": [ "navigation", "manipulation" ], "tags": [ "real", "mobile-manipulation", "residual-rl", "fine-tuning" ], "robot_types": [ "mobile" ], "highlight": false, "added": "2026-10-19", "takeaway_en": "HiRe-MoMa fine-tunes a frozen mobile-manipulation policy in the real world by learning residual corrections, with only occasional human guidance.", "takeaway_zh": "HiRe-MoMa 在真实环境中通过学习残差修正来微调冻结的移动操作策略，只需偶尔的人类引导。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.20607.html", "labs": [] }, { "id": "arxiv-2306.14846", "title": "ViNT: A Foundation Model for Visual Navigation", "type": "paper", "arxiv": "2306.14846", "year": 2023, "venue": "CoRL 2023", "org": "UC Berkeley", "url": null, "code": "https://github.com/robodhruv/visualnav-transformer", "project": null, "topics": [ "navigation" ], "tags": [ "foundation-model", "image-goal", "cross-embodiment", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A Transformer navigation policy trained on many robots' driving data that adapts to new goal types and robots with light fine-tuning.", "takeaway_zh": "在多种机器人的导航数据上训练的 Transformer 导航策略，经少量微调即可适配新的目标形式和新机器人。", "github_stars": 1335, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2306.14846.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2210.03370", "title": "GNM: A General Navigation Model to Drive Any Robot", "type": "paper", "arxiv": "2210.03370", "year": 2022, "venue": "ICRA 2023", "org": "UC Berkeley", "url": null, "code": "https://github.com/robodhruv/visualnav-transformer", "project": null, "topics": [ "navigation" ], "tags": [ "foundation-model", "image-goal", "cross-embodiment", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The predecessor of ViNT — one goal-conditioned navigation policy trained on pooled data from several different robots, which transfers to new robots including ones not seen in training.", "takeaway_zh": "ViNT 的前作：在多种不同机器人汇集的数据上训练一个目标条件导航策略，可迁移到包括训练中未见过的新机器人。", "github_stars": 1335, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2210.03370.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2312.03275", "title": "VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation", "type": "paper", "arxiv": "2312.03275", "year": 2023, "venue": "ICRA 2024", "org": "Boston Dynamics AI Institute, Georgia Tech", "url": null, "code": "https://github.com/rai-opensource/vlfm", "project": null, "topics": [ "navigation" ], "tags": [ "zero-shot", "object-goal", "vlm", "frontier-exploration", "real", "sim" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Zero-shot object-goal navigation that scores exploration frontiers by vision-language similarity to the target, with no task-specific training; tested in Habitat and on a Spot robot.", "takeaway_zh": "零样本目标物体导航：用视觉-语言模型对探索前沿与目标的相关度打分来决定往哪走，无需针对任务训练；在 Habitat 和 Spot 机器人上验证。", "github_stars": 829, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2312.03275.html", "labs": [ "georgia-tech" ] }, { "id": "arxiv-2402.15852", "title": "NaVid: Video-based VLM Plans the Next Step for Vision-and-Language Navigation", "type": "paper", "arxiv": "2402.15852", "year": 2024, "venue": "RSS 2024", "org": "Peking University", "url": null, "code": "https://github.com/jzhzhang/NaVid-VLN-CE", "project": null, "topics": [ "navigation" ], "tags": [ "vln", "video-vlm", "monocular", "sim-to-real" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A video VLM that reads the monocular RGB history plus the instruction and outputs the next navigation action directly, without maps, odometry or depth.", "takeaway_zh": "基于视频的视觉语言模型：只看单目 RGB 历史帧和语言指令，直接输出下一步导航动作，不依赖地图、里程计或深度。", "github_stars": 457, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2402.15852.html", "labs": [ "pku" ] }, { "id": "arxiv-2412.06224", "title": "Uni-NaVid: A Video-based Vision-Language-Action Model for Unifying Embodied Navigation Tasks", "type": "paper", "arxiv": "2412.06224", "year": 2024, "venue": "RSS 2025", "org": "Peking University", "url": null, "code": "https://github.com/jzhzhang/NaVid-VLN-CE", "project": null, "topics": [ "navigation", "vla" ], "tags": [ "vln", "object-goal", "following", "video-vlm" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Extends NaVid into one video VLA model for several navigation tasks (instruction following, object search, question answering, human following).", "takeaway_zh": "把 NaVid 扩展为单一视频 VLA 模型，统一处理多种导航任务（指令导航、目标搜索、具身问答、跟随人）。", "github_stars": 457, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2412.06224.html", "labs": [ "pku" ] }, { "id": "arxiv-2003.00368", "title": "The Marathon 2: A Navigation System", "type": "paper", "arxiv": "2003.00368", "year": 2020, "venue": "IROS 2020", "org": null, "url": null, "code": "https://github.com/ros-navigation/navigation2", "project": null, "topics": [ "navigation" ], "tags": [ "ros2", "nav2", "behavior-trees", "open-source", "mobile" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The paper behind Nav2, the standard ROS 2 navigation stack - behavior-tree orchestration of planners, controllers and recovery behaviors, validated by long marathon runs on real robots.", "takeaway_zh": "ROS 2 标准导航栈 Nav2 的论文：用行为树编排规划器、控制器和恢复行为，并在真实机器人上长距离“马拉松”测试验证。", "github_stars": 4943, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2003.00368.html", "labs": [] }, { "id": "arxiv-2412.04453", "title": "NaVILA: Legged Robot Vision-Language-Action Model for Navigation", "type": "paper", "arxiv": "2412.04453", "year": 2024, "venue": "RSS 2025", "org": "UC San Diego, USC, NVIDIA", "url": null, "code": "https://github.com/AnjieCheng/NaVILA", "project": "https://navila-bot.github.io/", "topics": [ "navigation", "vla", "locomotion" ], "tags": [ "vln", "legged", "hierarchical" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "A VLA outputs mid-level navigation commands in language that an RL locomotion policy executes, so legged robots follow natural-language routes; adds Isaac Lab VLN benchmarks.", "takeaway_zh": "VLA 先用语言输出中层导航指令，再由强化学习运动策略执行，让腿式机器人按自然语言指令导航；并提供 Isaac Lab 上的 VLN 基准。", "github_stars": 726, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2412.04453.html", "labs": [ "nvidia", "ucsd", "usc" ] }, { "id": "arxiv-2004.02857", "title": "Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments", "type": "benchmark", "arxiv": "2004.02857", "year": 2020, "venue": "ECCV 2020", "org": "Oregon State, Georgia Tech, Meta AI", "url": null, "code": "https://github.com/jacobkrantz/VLN-CE", "project": null, "topics": [ "navigation", "datasets" ], "tags": [ "vln", "habitat", "benchmark" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "VLN-CE moves Room-to-Room instructions from a fixed navigation graph into continuous Habitat scenes with low-level actions; the standard setting for realistic VLN.", "takeaway_zh": "VLN-CE 把 Room-to-Room 指令从固定导航图搬到 Habitat 连续环境中、使用底层动作执行，是更贴近真实的 VLN 标准设定。", "github_stars": 899, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2004.02857.html", "labs": [ "meta", "georgia-tech" ] }, { "id": "arxiv-2207.04429", "title": "LM-Nav: Robotic Navigation with Large Pre-Trained Models of Language, Vision, and Action", "type": "paper", "arxiv": "2207.04429", "year": 2022, "venue": "CoRL 2022", "org": "UC Berkeley, Google", "url": null, "code": "https://github.com/blazejosinski/lm_nav", "project": null, "topics": [ "navigation", "vla" ], "tags": [ "real", "wheeled", "language", "foundation-models", "zero-shot" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Chains an LLM (parses landmarks from the instruction), CLIP (finds them in images) and a visual navigation model to follow route instructions on a real robot without language-annotated robot data.", "takeaway_zh": "将大语言模型（从指令中解析地标）、CLIP（在图像中找到地标）与视觉导航模型串联，无需带语言标注的机器人数据即可在真实机器人上按路线指令导航。", "github_stars": 272, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2207.04429.html", "labs": [ "berkeley", "google-deepmind" ] }, { "id": "arxiv-2007.00643", "title": "Object Goal Navigation using Goal-Oriented Semantic Exploration", "type": "paper", "arxiv": "2007.00643", "year": 2020, "venue": "NeurIPS 2020", "org": "CMU, Meta AI", "url": null, "code": "https://github.com/devendrachaplot/Object-Goal-Navigation", "project": null, "topics": [ "navigation", "perception" ], "tags": [ "sim", "objectnav", "semantic-map", "modular", "rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "SemExp builds an episodic semantic map and learns a policy that picks long-term goals on it, a modular recipe that became a standard ObjectNav baseline in Habitat.", "takeaway_zh": "SemExp 构建情景语义地图，并学习在地图上选择长期目标的策略；这一模块化方案成为 Habitat 中 ObjectNav 的常用基线。", "github_stars": 463, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2007.00643.html", "labs": [ "meta", "cmu" ] }, { "id": "arxiv-2004.05155", "title": "Learning to Explore using Active Neural SLAM", "type": "paper", "arxiv": "2004.05155", "year": 2020, "venue": "ICLR 2020", "org": "CMU, Meta AI", "url": null, "code": "https://github.com/devendrachaplot/Neural-SLAM", "project": null, "topics": [ "navigation", "perception" ], "tags": [ "sim", "exploration", "slam", "modular", "rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Splits exploration into a learned neural SLAM module, a global policy that chooses long-term goals and a local policy that reaches them, which is far more sample-efficient than end-to-end RL.", "takeaway_zh": "将探索拆为可学习的神经 SLAM 模块、选择长期目标的全局策略和到达目标的局部策略，比端到端强化学习样本效率高得多。", "github_stars": 853, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2004.05155.html", "labs": [ "meta", "cmu" ] }, { "id": "arxiv-2203.10421", "title": "CoWs on Pasture: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation", "type": "paper", "arxiv": "2203.10421", "year": 2022, "venue": "CVPR 2023", "org": "Columbia University", "url": null, "code": null, "project": "https://cow.cs.columbia.edu/", "topics": [ "navigation" ], "tags": [ "zero-shot", "object-goal", "clip", "benchmark", "sim" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "CLIP on Wheels pairs open-vocabulary localization (CLIP-style models, no fine-tuning) with classical exploration for zero-shot object navigation; the Pasture benchmark adds rare, attribute-described and hidden objects.", "takeaway_zh": "CoWs on Pasture：CLIP on Wheels 把开放词汇定位（CLIP 等模型，无需微调）与经典探索结合，实现零样本目标物体导航；配套的 Pasture 基准加入了少见物体、按属性描述的物体和被遮挡物体。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2203.10421.html", "labs": [ "columbia" ] }, { "id": "arxiv-2301.07302", "title": "PIRLNav: Pretraining with Imitation and RL Finetuning for ObjectNav", "type": "paper", "arxiv": "2301.07302", "year": 2023, "venue": "CVPR 2023", "org": "Georgia Tech, Meta AI", "url": null, "code": "https://github.com/Ram81/pirlnav", "project": null, "topics": [ "navigation", "learning" ], "tags": [ "object-goal", "imitation", "rl-finetuning", "habitat", "sim" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "PIRLNav — behavior cloning on human ObjectNav demonstrations, then RL fine-tuning of the same policy; the paper reports 65.0% success on Habitat ObjectNav, 5 points above the previous best.", "takeaway_zh": "PIRLNav：先在人类 ObjectNav 演示上做行为克隆，再用强化学习微调同一策略；论文报告在 Habitat ObjectNav 上成功率 65.0%，比此前最佳高 5 个百分点。", "github_stars": 92, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2301.07302.html", "labs": [ "meta", "georgia-tech" ] }, { "id": "arxiv-2610.25690", "title": "ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation", "type": "paper", "arxiv": "2610.25690", "year": 2026, "venue": null, "org": "Zhejiang University", "url": null, "code": "https://github.com/MachEmbodied/ScaleNav", "project": "https://zju3dv.github.io/ScaleNav/", "topics": [ "navigation" ], "tags": [ "internet-video", "language-navigation", "quadruped", "pretraining" ], "robot_types": [ "legged" ], "highlight": false, "added": "2026-10-23", "takeaway_en": "Pretrains on internet egocentric video to learn how the view changes when approaching a described target, then aligns to actions in sim; reports 94.9% success on its HabitatGS-LN benchmark plus a real quadruped.", "takeaway_zh": "先在互联网第一视角视频上预训练，学习\"朝描述的目标走近时画面如何变化\"，再用仿真数据对齐到动作；论文报告在自建 HabitatGS-LN 基准上成功率 94.9%，并在真实四足机器人上部署。", "github_stars": 48, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25690.html", "labs": [] }, { "id": "arxiv-2305.16986", "title": "NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models", "type": "paper", "arxiv": "2305.16986", "year": 2023, "venue": "AAAI 2024", "org": "University of Adelaide", "url": null, "code": "https://github.com/GengzeZhou/NavGPT", "project": null, "topics": [ "navigation" ], "tags": [ "vln", "llm", "zero-shot", "reasoning", "r2r" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A purely LLM-based VLN agent — GPT models read text descriptions of what the robot sees, its history and possible directions, then reason step by step about where to go; zero-shot on R2R, still behind trained models.", "takeaway_zh": "纯大语言模型的视觉语言导航智能体：GPT 读取视觉观测的文字描述、历史轨迹和可走方向，逐步显式推理下一步去哪；在 R2R 上零样本运行，但仍落后于专门训练的模型。", "github_stars": 355, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2305.16986.html", "labs": [] }, { "id": "arxiv-2304.03047", "title": "ETPNav: Evolving Topological Planning for Vision-Language Navigation in Continuous Environments", "type": "paper", "arxiv": "2304.03047", "year": 2023, "venue": "TPAMI 2024", "org": null, "url": null, "code": "https://github.com/MarSaKi/ETPNav", "project": null, "topics": [ "navigation" ], "tags": [ "vln-ce", "topological-map", "planning", "r2r-ce", "rxr-ce" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "For VLN in continuous spaces, builds a topological map online from predicted waypoints, plans on it with a cross-modal transformer and moves with an obstacle-avoiding controller; reports >10%/>20% gains on R2R-CE/RxR-CE.", "takeaway_zh": "面向连续环境视觉语言导航：根据预测的路点在线构建拓扑地图，用跨模态 Transformer 在图上规划，再由避障控制器执行；报告在 R2R-CE、RxR-CE 上分别比此前最好方法提升 10% 和 20% 以上。", "github_stars": 498, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2304.03047.html", "labs": [] }, { "id": "arxiv-2610.30043", "title": "TARSphere: A Robot-Oriented Full-Stack Framework for Long-Horizon Embodied Agents in Evolving Physical and Social Environments", "type": "paper", "arxiv": "2610.30043", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/xuyicheng-zju/TARSphere", "project": "https://xuyicheng-zju.github.io/TARSphere/", "topics": [ "navigation", "datasets" ], "tags": [ "benchmark", "mllm-agent", "long-horizon", "hri", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-25", "takeaway_en": "Simulator with embodied NPCs, dialogue and event-driven task states, plus a skill harness for frozen MLLMs; its 360-episode bench reports general MLLMs can beat embodied ones; task updates and HRI stay hard.", "takeaway_zh": "带具身 NPC、受限对话和事件驱动任务状态机的物理仿真框架，并提供让冻结多模态大模型调用多级技能的执行框架；基于 30 个室内场景构建 360 回合基准，评测 8 个模型，报告通用模型可胜过具身专用模型，任务变更和人机协作仍是难点。", "github_stars": 23, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.30043.html", "labs": [] }, { "id": "arxiv-2610.25959", "title": "EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models", "type": "paper", "arxiv": "2610.25959", "year": 2026, "venue": "NeurIPS 2026", "org": null, "url": null, "code": null, "project": "https://xiaokunfeng.github.io/EmbodiedWorld-200K/", "topics": [ "navigation", "datasets", "world-models" ], "tags": [ "dataset", "navigation-planning", "video-world-model", "rl-finetuning", "open-world" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "200K+ navigation-planning samples annotated from camera trajectories up to instructions; RL post-training inside a video world model, used as a cheap stand-in environment, reportedly helps on unseen scenes.", "takeaway_zh": "包含 20 万余条导航规划样本的数据集，场景多样，标注从相机运动轨迹到高层语义指令；并把视频世界模型当作低成本交互环境做强化学习后训练，报告称在未见场景上有明显提升。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25959.html", "labs": [] }, { "id": "arxiv-2610.23896", "title": "InJePA: Intention-First Latent Planning with Coupled JEPA World Models for Visual Navigation", "type": "paper", "arxiv": "2610.23896", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/zhuxun132/InJepa", "project": null, "topics": [ "navigation", "world-models" ], "tags": [ "jepa", "latent-planning", "image-goal", "habitat", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Two coupled JEPAs share a latent space: one proposes goal-conditioned latent intentions, the other predicts consequences, closing an intention-action-consequence loop. Evaluated on image-goal navigation in Habitat MP3D.", "takeaway_zh": "两个耦合的 JEPA 共享隐空间：一个按目标提出隐式“意图”，另一个预测动作后果，形成意图-动作-后果闭环，让世界模型直接参与构造规划；在 Habitat MP3D 图像目标导航上评测。", "github_stars": 52, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.23896.html", "labs": [] }, { "id": "arxiv-2610.24122", "title": "HAVANA: A Hierarchical Video-Action Model for Vision-Language Navigation", "type": "paper", "arxiv": "2610.24122", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://github.com/Anastasiawd/HAVANA", "topics": [ "navigation", "world-models" ], "tags": [ "vln", "video-action-model", "dual-system", "sim", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Dual-system video-action model for VLN: a slow system turns history and instruction into a compact latent plan, a fast one turns plan + latest view into local trajectories; no video is generated at inference.", "takeaway_zh": "面向视觉语言导航的双系统视频-动作模型：慢系统根据历史观测与指令生成紧凑的隐式规划，快系统结合该规划与最新观测输出局部轨迹；推理时不生成未来视频。代码与权重称将在该仓库发布。", "github_stars": 2, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.24122.html", "labs": [] }, { "id": "arxiv-2610.33546", "title": "MOON: Learning When to Forget and How to Move On in Vision-and-Language Navigation", "type": "paper", "arxiv": "2610.33546", "year": 2026, "venue": null, "org": "SJTU", "url": null, "code": "https://github.com/AutoLab-SAI-SJTU/MOON", "project": null, "topics": [ "navigation", "learning" ], "tags": [ "vln", "recovery", "memory", "grpo", "r2r-ce" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Off-route VLN agents keep following their stale history ('behavioral inertia'). MOON adds a deviation judge that clears it and a GRPO-trained recovery policy (<1% params); reports +5.2 / +6.0 points on R2R-CE / RxR-CE.", "takeaway_zh": "发现 VLN 智能体偏离路线后会被历史记忆“惯性”拖住。MOON 给冻结的导航策略加上偏航判断器（判断何时清空历史）和 GRPO 训练的恢复策略，参数增加不到 1%；报告在 R2R-CE / RxR-CE 上分别提升 5.2 / 6.0 个百分点。", "github_stars": 10, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.33546.html", "labs": [ "sjtu" ] }, { "id": "arxiv-2610.33794", "title": "One Scan is Enough: Demonstration-Free Adaptation for Language-Guided Navigation", "type": "paper", "arxiv": "2610.33794", "year": 2026, "venue": "NeurIPS 2026", "org": null, "url": null, "code": null, "project": null, "topics": [ "navigation", "vla" ], "tags": [ "vln", "scene-adaptation", "real-to-sim", "iphone-scan", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Adapts a language-guided navigation policy to a new building from one iPhone scan, compiled into a training environment with synthetic instructed episodes. Reports 85% real success vs 49% for a demo-fine-tuned baseline.", "takeaway_zh": "只需用 iPhone 扫描一次新建筑：扫描被编译成训练环境并自动合成带语言指令的导航回合，再用无评论家的闭式 Pareto 效用适配策略；报告真机成功率 85%，用人类示教后训练的基线为 49%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.33794.html", "labs": [] }, { "id": "arxiv-2610.31762", "title": "See What I Need, Dream What I Want: A Navigation Foundation Model on a Self-Constructed Representation Space", "type": "paper", "arxiv": "2610.31762", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://wang-yinxi.github.io/representnav/", "topics": [ "navigation", "world-models" ], "tags": [ "vln", "foundation-model", "jepa", "latent-prediction", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "RepresentNav fuses several vision foundation models into a policy-learned space and predicts the next latent before acting (JEPA-style); reports 83.8% on VLNVerse fine-grained (+20.05 pts), real-robot tests.", "takeaway_zh": "RepresentNav 用一个小型 Transformer 融合多个视觉基础模型各层特征，构建由导航目标自行塑造的表示空间，并在每步动作前预测下一时刻的潜在表示（JEPA 式）；报告 VLNVerse 细粒度成功率 83.8%（领先 20.05 个百分点），并在真机上验证。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31762.html", "labs": [] }, { "id": "arxiv-2007.11898", "title": "ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM", "type": "paper", "arxiv": "2007.11898", "year": 2020, "venue": null, "org": "Universidad de Zaragoza", "url": null, "code": "https://github.com/UZ-SLAMLab/ORB_SLAM3", "project": null, "topics": [ "perception" ], "tags": [ "slam", "visual-inertial", "open-source" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "The go-to feature-based SLAM library — monocular, stereo and RGB-D, with IMU fusion and multi-map relocalization.", "takeaway_zh": "最常用的特征点 SLAM 库：支持单目、双目与 RGB-D，可融合 IMU，并支持多地图重定位。", "github_stars": 9195, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2007.11898.html", "labs": [] }, { "id": "arxiv-2007.00258", "title": "LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping", "type": "paper", "arxiv": "2007.00258", "year": 2020, "venue": null, "org": "MIT", "url": null, "code": "https://github.com/TixiaoShan/LIO-SAM", "project": null, "topics": [ "perception" ], "tags": [ "lidar", "odometry", "factor-graph" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "LiDAR-inertial odometry built on a factor graph, so GPS and loop closures slot in as extra factors.", "takeaway_zh": "基于因子图的激光-惯性里程计，GPS 与回环可以作为额外因子直接加入。", "github_stars": 4957, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2007.00258.html", "labs": [ "mit" ] }, { "id": "arxiv-2107.06829", "title": "FAST-LIO2: Fast Direct LiDAR-inertial Odometry", "type": "paper", "arxiv": "2107.06829", "year": 2021, "venue": null, "org": "HKU MaRS Lab", "url": null, "code": "https://github.com/hku-mars/FAST_LIO", "project": null, "topics": [ "perception" ], "tags": [ "lidar", "odometry", "real-time" ], "robot_types": [ "aerial" ], "highlight": false, "added": null, "takeaway_en": "Registers raw LiDAR points directly to the map with an incremental k-d tree; fast enough for onboard drone and robot computers.", "takeaway_zh": "用增量 k-d 树把原始激光点直接配准到地图，速度足以在无人机和机器人机载计算机上运行。", "github_stars": 5315, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2107.06829.html", "labs": [ "hku" ] }, { "id": "arxiv-2108.10869", "title": "DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras", "type": "paper", "arxiv": "2108.10869", "year": 2021, "venue": null, "org": "Princeton", "url": null, "code": "https://github.com/princeton-vl/DROID-SLAM", "project": null, "topics": [ "perception" ], "tags": [ "slam", "learned", "visual" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "End-to-end learned visual SLAM — recurrent optical-flow updates plus a differentiable bundle-adjustment layer.", "takeaway_zh": "端到端学习的视觉 SLAM：循环光流更新加上可微分的光束法平差层。", "github_stars": 2717, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2108.10869.html", "labs": [ "princeton" ] }, { "id": "arxiv-2312.06741", "title": "Gaussian Splatting SLAM", "type": "paper", "arxiv": "2312.06741", "year": 2023, "venue": null, "org": "Imperial College London", "url": null, "code": "https://github.com/muskie82/MonoGS", "project": null, "topics": [ "perception" ], "tags": [ "slam", "gaussian-splatting", "dense" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "MonoGS — runs SLAM directly on a 3D Gaussian map, giving camera tracking and a photorealistic dense map from one camera.", "takeaway_zh": "MonoGS：直接在三维高斯地图上做 SLAM，仅用单个相机即可同时完成相机跟踪与照片级稠密建图。", "github_stars": 2189, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2312.06741.html", "labs": [] }, { "id": "arxiv-2312.14132", "title": "DUSt3R: Geometric 3D Vision Made Easy", "type": "paper", "arxiv": "2312.14132", "year": 2023, "venue": null, "org": "NAVER LABS Europe", "url": null, "code": "https://github.com/naver/dust3r", "project": null, "topics": [ "perception" ], "tags": [ "3d-reconstruction", "foundation-model" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Regresses 3D point maps straight from image pairs without known camera intrinsics or poses; started the \"3D foundation model\" wave.", "takeaway_zh": "不需要已知相机内参或位姿，直接从图像对回归三维点图，开启了“三维基础模型”这一方向。", "github_stars": 7362, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2312.14132.html", "labs": [] }, { "id": "arxiv-2503.11651", "title": "VGGT: Visual Geometry Grounded Transformer", "type": "paper", "arxiv": "2503.11651", "year": 2025, "venue": null, "org": "Oxford VGG · Meta AI", "url": null, "code": "https://github.com/facebookresearch/vggt", "project": null, "topics": [ "perception" ], "tags": [ "3d-reconstruction", "foundation-model", "feed-forward" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "One feed-forward transformer predicts cameras, depth, point maps and tracks for many views at once, in about a second.", "takeaway_zh": "单个前馈 Transformer 一次性为多视角图像预测相机参数、深度、点图与跟踪，约一秒完成。", "github_stars": 14584, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2503.11651.html", "labs": [ "meta", "oxford" ] }, { "id": "arxiv-2312.08344", "title": "FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects", "type": "paper", "arxiv": "2312.08344", "year": 2023, "venue": null, "org": "NVIDIA", "url": null, "code": "https://github.com/NVlabs/FoundationPose", "project": null, "topics": [ "perception", "manipulation" ], "tags": [ "6d-pose", "tracking", "zero-shot" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "6D pose estimation and tracking of unseen objects from a CAD model or a few reference images, with no per-object training.", "takeaway_zh": "只需 CAD 模型或少量参考图像即可对未见过的物体做 6D 位姿估计与跟踪，无需逐物体训练。", "github_stars": 3666, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2312.08344.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2401.10891", "title": "Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data", "type": "paper", "arxiv": "2401.10891", "year": 2024, "venue": null, "org": "HKU · TikTok", "url": null, "code": "https://github.com/LiheYoung/Depth-Anything", "project": null, "topics": [ "perception" ], "tags": [ "depth", "monocular", "foundation-model" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Monocular depth model trained on tens of millions of pseudo-labelled images; a common off-the-shelf depth prior for robots.", "takeaway_zh": "在数千万张伪标注图像上训练的单目深度模型，常被机器人系统直接拿来作为深度先验。", "github_stars": 8246, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2401.10891.html", "labs": [ "hku" ] }, { "id": "arxiv-2309.16650", "title": "ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning", "type": "paper", "arxiv": "2309.16650", "year": 2023, "venue": null, "org": "University of Toronto · MIT · Mila", "url": null, "code": null, "project": "https://concept-graphs.github.io/", "topics": [ "perception", "navigation" ], "tags": [ "scene-graph", "open-vocabulary", "llm" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Builds an object-level 3D scene graph labelled by vision-language models, which an LLM planner can query in natural language.", "takeaway_zh": "用视觉语言模型构建带语义标注的物体级三维场景图，大模型规划器可以用自然语言查询。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.16650.html", "labs": [ "mit", "toronto" ] }, { "id": "arxiv-2304.02643", "title": "Segment Anything", "type": "paper", "arxiv": "2304.02643", "year": 2023, "venue": "ICCV 2023", "org": "Meta AI", "url": null, "code": "https://github.com/facebookresearch/segment-anything", "project": null, "topics": [ "perception" ], "tags": [ "segmentation", "foundation-model", "promptable" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "SAM — a promptable segmentation model (points, boxes, masks) trained on 1B+ masks; a standard building block for object masks in robot perception pipelines.", "takeaway_zh": "SAM：可用点、框、掩码提示的分割模型，在 10 亿级掩码上训练；已成为机器人感知流程中获取物体掩码的常用基础组件。", "github_stars": 55073, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2304.02643.html", "labs": [ "meta" ] }, { "id": "arxiv-2408.00714", "title": "SAM 2: Segment Anything in Images and Videos", "type": "paper", "arxiv": "2408.00714", "year": 2024, "venue": "ICLR 2025", "org": "Meta AI", "url": null, "code": "https://github.com/facebookresearch/sam2", "project": null, "topics": [ "perception" ], "tags": [ "segmentation", "video", "tracking" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Extends SAM to video with a streaming memory, so a mask prompted once is tracked across frames — handy for following objects during manipulation.", "takeaway_zh": "用流式记忆把 SAM 扩展到视频：一次提示得到的掩码可在后续帧中持续跟踪，适合在操作过程中跟随目标物体。", "github_stars": 20076, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2408.00714.html", "labs": [ "meta" ] }, { "id": "arxiv-2303.05499", "title": "Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection", "type": "paper", "arxiv": "2303.05499", "year": 2023, "venue": "ECCV 2024", "org": "IDEA Research · Tsinghua", "url": null, "code": "https://github.com/IDEA-Research/GroundingDINO", "project": null, "topics": [ "perception" ], "tags": [ "open-vocabulary", "detection", "language-grounding" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open-set detector that finds objects from a free-text prompt; often paired with SAM (\"Grounded-SAM\") to turn a language instruction into object masks.", "takeaway_zh": "开放集目标检测器，可根据任意文本提示找到物体；常与 SAM 组合（Grounded-SAM），把语言指令变成物体掩码。", "github_stars": 10716, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2303.05499.html", "labs": [ "tsinghua" ] }, { "id": "arxiv-2501.09898", "title": "FoundationStereo: Zero-Shot Stereo Matching", "type": "paper", "arxiv": "2501.09898", "year": 2025, "venue": "CVPR 2025", "org": "NVIDIA", "url": null, "code": "https://github.com/NVlabs/FoundationStereo", "project": null, "topics": [ "perception" ], "tags": [ "stereo", "depth", "zero-shot" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Stereo depth model trained on large synthetic data that reports strong zero-shot accuracy on real scenes — a drop-in depth source for stereo-camera robots.", "takeaway_zh": "在大规模合成数据上训练的双目深度模型，报告称在真实场景上零样本精度很强；可直接作为双目相机机器人的深度来源。", "github_stars": 2971, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2501.09898.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2610.25665", "title": "SpaRV: Sparse Radiance Voxels for Scalable Geometrically Consistent LiDAR-Visual Mapping and Rendering", "type": "paper", "arxiv": "2610.25665", "year": 2026, "venue": null, "org": "HKU MaRS Lab", "url": null, "code": "https://github.com/hku-mars/SpaRV", "project": null, "topics": [ "perception" ], "tags": [ "mapping", "lidar", "radiance-field" ], "robot_types": [], "highlight": false, "added": "2026-10-20", "takeaway_en": "A sparse-voxel radiance map built from LiDAR and camera data, aimed at robot logs — few viewpoints, imperfect poses and large scenes — while keeping geometry accurate and GPU memory in check.", "takeaway_zh": "用激光雷达与相机数据构建稀疏体素辐射场地图，针对机器人数据视角少、位姿有误差、场景大的问题，在保证几何精度的同时控制显存占用。", "github_stars": 34, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25665.html", "labs": [ "hku" ] }, { "id": "arxiv-2610.25288", "title": "Vision as Proprioception: Estimating Humanoid Robot Pose from Video", "type": "paper", "arxiv": "2610.25288", "year": 2026, "venue": null, "org": "ETH Zurich", "url": null, "code": "https://github.com/eth-siplab/HumanoidPose", "project": null, "topics": [ "perception", "locomotion" ], "tags": [ "humanoid", "pose-estimation", "video" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-20", "takeaway_en": "HumanoidPose recovers a humanoid's full-body joint configuration from video alone — useful when onboard state is unavailable, and for turning online robot videos into training data.", "takeaway_zh": "HumanoidPose 仅凭视频恢复人形机器人的全身关节构型：适用于拿不到机载状态的场景，也能把网上的机器人视频变成训练数据。", "github_stars": 6, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25288.html", "labs": [ "eth" ] }, { "id": "arxiv-2610.27199", "title": "Which Depth Estimator for Jetson AGX Orin? A 2026 Benchmark of Monocular and Stereo Methods", "type": "benchmark", "arxiv": "2610.27199", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/nodarhub/orin-depth-benchmark", "project": null, "topics": [ "perception" ], "tags": [ "depth", "stereo", "edge", "jetson", "benchmark" ], "robot_types": [], "highlight": false, "added": "2026-10-21", "takeaway_en": "Measured numbers for robot builders: nine monocular and stereo depth families (Depth Anything V2/3, Depth Pro, MoGe-2, Fast-FoundationStereo, ...) run on a Jetson AGX Orin, scored for metric accuracy and speed.", "takeaway_zh": "给机器人开发者的实测数据：在 Jetson AGX Orin 上用 TensorRT 运行九类单目与双目深度估计模型（Depth Anything V2/3、Depth Pro、MoGe-2、Fast-FoundationStereo 等），并评测度量精度与速度。", "github_stars": 1, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.27199.html", "labs": [] }, { "id": "arxiv-2304.07193", "title": "DINOv2: Learning Robust Visual Features without Supervision", "type": "paper", "arxiv": "2304.07193", "year": 2023, "venue": "TMLR 2024", "org": "Meta AI", "url": null, "code": "https://github.com/facebookresearch/dinov2", "project": null, "topics": [ "perception", "vla" ], "tags": [ "self-supervised", "vision-backbone", "foundation-model" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Self-supervised image features that work frozen for depth, segmentation and retrieval; a common visual backbone in VLAs (e.g. OpenVLA) and latent world models.", "takeaway_zh": "自监督学习的图像特征，冻结即可用于深度估计、分割和检索；是 VLA（如 OpenVLA）和潜空间世界模型常用的视觉骨干。", "github_stars": 13489, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2304.07193.html", "labs": [ "meta" ] }, { "id": "arxiv-2307.07635", "title": "CoTracker: It is Better to Track Together", "type": "paper", "arxiv": "2307.07635", "year": 2023, "venue": "ECCV 2024", "org": "Meta AI · University of Oxford (VGG)", "url": null, "code": "https://github.com/facebookresearch/co-tracker", "project": null, "topics": [ "perception" ], "tags": [ "point-tracking", "video", "transformer" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Transformer that tracks many video points jointly rather than one at a time, staying robust through occlusions; widely used to extract motion tracks from human or robot videos for policy learning.", "takeaway_zh": "联合跟踪视频中大量点（而非逐点独立跟踪）的 Transformer，遮挡下依然稳健；常用于从人类或机器人视频中提取运动轨迹来学习策略。", "github_stars": 5178, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2307.07635.html", "labs": [ "meta", "oxford" ] }, { "id": "arxiv-2406.09756", "title": "Grounding Image Matching in 3D with MASt3R", "type": "paper", "arxiv": "2406.09756", "year": 2024, "venue": "ECCV 2024", "org": "NAVER LABS Europe", "url": null, "code": "https://github.com/naver/mast3r", "project": null, "topics": [ "perception" ], "tags": [ "3d-reconstruction", "image-matching", "pointmap" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Extends DUSt3R with a dense local-feature head, so one network gives both 3D pointmaps and accurate pixel matches between views — useful for calibration-free reconstruction and localization.", "takeaway_zh": "在 DUSt3R 基础上加入稠密局部特征头，同一网络既输出三维点图又给出视角间精确的像素匹配，适用于免标定的三维重建与定位。", "github_stars": 3145, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2406.09756.html", "labs": [] }, { "id": "arxiv-2408.14035", "title": "FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry", "type": "paper", "arxiv": "2408.14035", "year": 2024, "venue": null, "org": "HKU MaRS Lab", "url": null, "code": "https://github.com/hku-mars/FAST-LIVO2", "project": null, "topics": [ "perception" ], "tags": [ "lidar", "visual", "odometry", "sensor-fusion", "real-time" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "FAST-LIVO2 — fuses LiDAR, camera and IMU in one filter with direct (feature-free) updates for both sensors, sharing a single voxel map where image patches are attached to LiDAR points.", "takeaway_zh": "FAST-LIVO2：在同一滤波器中融合激光雷达、相机与 IMU，两种传感器都采用直接法（无需提取特征）更新，并共用一张体素地图，图像块挂载在激光点上。", "github_stars": 4812, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2408.14035.html", "labs": [ "hku" ] }, { "id": "arxiv-2406.09414", "title": "Depth Anything V2", "type": "paper", "arxiv": "2406.09414", "year": 2024, "venue": "NeurIPS 2024", "org": "HKU, TikTok", "url": null, "code": "https://github.com/DepthAnything/Depth-Anything-V2", "project": null, "topics": [ "perception" ], "tags": [ "monocular-depth", "foundation-model", "synthetic-data" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Trains a large teacher only on synthetic images with precise depth, then pseudo-labels many real images to train smaller students; gives sharper, more robust monocular depth than V1.", "takeaway_zh": "先只用深度精确的合成图像训练大教师模型，再给大量真实图像打伪标签来训练小模型；单目深度比 V1 更锐利、更鲁棒。", "github_stars": 9050, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2406.09414.html", "labs": [ "hku" ] }, { "id": "arxiv-2303.14158", "title": "BundleSDF: Neural 6-DoF Tracking and 3D Reconstruction of Unknown Objects", "type": "paper", "arxiv": "2303.14158", "year": 2023, "venue": "CVPR 2023", "org": "NVIDIA", "url": null, "code": "https://github.com/NVlabs/BundleSDF", "project": null, "topics": [ "perception", "manipulation" ], "tags": [ "real", "6d-pose", "tracking", "reconstruction", "rgbd", "model-free" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Tracks the 6-DoF pose of a never-seen object in RGB-D video while building its neural 3D model online, with no CAD model needed; useful for manipulating novel objects.", "takeaway_zh": "在 RGB-D 视频中跟踪从未见过物体的六自由度位姿，同时在线构建其神经三维模型，无需 CAD 模型；适用于操作新物体。", "github_stars": 1426, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2303.14158.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2610.27040", "title": "ACE-Brain-0.6: Scaling Spatial Grounding for Physical Foundation Models", "type": "paper", "arxiv": "2610.27040", "year": 2026, "venue": null, "org": "ACE Robotics", "url": null, "code": "https://github.com/DAXIAORobotics/ACE-Brain-0.6", "project": "https://daxiaorobotics.github.io/ACE-Brain-0.6/", "topics": [ "perception", "vla", "manipulation" ], "tags": [ "visual-grounding", "vlm", "embodied-brain", "data-pipeline" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Fine-tunes Qwen3.5-9B on ~1.5M grounding samples auto-built from robot recordings plus other visual data; reports a 75.41 average over seven grounding benchmarks vs 73.68 for GPT-6 Astra.", "takeaway_zh": "用自动标注的机器人操作录像加多来源视觉数据构建约 150 万条可验证的定位监督，微调 Qwen3.5-9B；报告在七个公开视觉定位基准上平均 75.41 分，GPT-6 Astra 为 73.68。", "github_stars": 8, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.27040.html", "labs": [] }, { "id": "arxiv-2511.16719", "title": "SAM 3: Segment Anything with Concepts", "type": "paper", "arxiv": "2511.16719", "year": 2025, "venue": null, "org": "Meta", "url": null, "code": "https://github.com/facebookresearch/sam3", "project": null, "topics": [ "perception" ], "tags": [ "segmentation", "tracking", "open-vocabulary", "foundation-model", "open-weights" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Segments and tracks every instance of a concept given a short noun phrase or image exemplar, in images and video; trained with a data engine covering 4M concept labels, with open weights and the SA-Co benchmark.", "takeaway_zh": "给定短名词短语或示例图像，即可在图像和视频中分割并跟踪某一概念的所有实例；数据引擎覆盖 400 万个概念标签，开放权重与 SA-Co 基准，可直接用作机器人开放词汇感知模块。", "github_stars": 12146, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2511.16719.html", "labs": [ "meta" ] }, { "id": "arxiv-2310.17274", "title": "cuRobo: Parallelized Collision-Free Minimum-Jerk Robot Motion Generation", "type": "paper", "arxiv": "2310.17274", "year": 2023, "venue": null, "org": "NVIDIA", "url": null, "code": "https://github.com/NVlabs/curobo", "project": null, "topics": [ "control", "manipulation" ], "tags": [ "motion-planning", "gpu", "trajectory-optimization" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "GPU-parallel motion generation for arms — many trajectory optimizations run at once, giving collision-free, smooth plans in milliseconds.", "takeaway_zh": "面向机械臂的 GPU 并行运动生成：同时运行大量轨迹优化，在毫秒级给出无碰撞且平滑的轨迹。", "github_stars": 1944, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2310.17274.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2210.12209", "title": "Motion Policy Networks", "type": "paper", "arxiv": "2210.12209", "year": 2022, "venue": null, "org": "NVIDIA · University of Washington", "url": null, "code": "https://github.com/NVlabs/motion-policy-networks", "project": null, "topics": [ "control", "manipulation" ], "tags": [ "motion-planning", "point-cloud", "imitation" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A neural policy trained on millions of planner-generated trajectories that moves an arm to goals straight from point clouds.", "takeaway_zh": "在数百万条规划器生成的轨迹上训练的神经策略，可直接根据点云把机械臂移动到目标位置。", "github_stars": 189, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2210.12209.html", "labs": [ "nvidia", "uw" ] }, { "id": "arxiv-2205.09991", "title": "Planning with Diffusion for Flexible Behavior Synthesis", "type": "paper", "arxiv": "2205.09991", "year": 2022, "venue": null, "org": "UC Berkeley", "url": null, "code": "https://github.com/jannerm/diffuser", "project": null, "topics": [ "control", "world-models" ], "tags": [ "diffusion", "planning", "offline-rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Diffuser — treats planning as denoising whole trajectories, so goals and rewards can steer sampling; an early link between diffusion and control.", "takeaway_zh": "Diffuser：把规划视为对整条轨迹去噪，目标与奖励可以引导采样；扩散模型与控制结合的早期代表作。", "github_stars": 1315, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2205.09991.html", "labs": [ "berkeley" ] }, { "id": "arxiv-1903.11199", "title": "Control Barrier Functions: Theory and Applications", "type": "paper", "arxiv": "1903.11199", "year": 2019, "venue": null, "org": "Caltech", "url": null, "code": null, "project": null, "topics": [ "control" ], "tags": [ "safety", "control-barrier-function", "tutorial" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Tutorial on control barrier functions — the standard way to wrap any controller in a minimal-change filter that keeps the system safe.", "takeaway_zh": "控制障碍函数教程：为任意控制器加一层“尽量少改动”的安全滤波器、保证系统安全的标准方法。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1903.11199.html", "labs": [] }, { "id": "arxiv-2108.06266", "title": "Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning", "type": "paper", "arxiv": "2108.06266", "year": 2021, "venue": null, "org": "University of Toronto", "url": null, "code": null, "project": null, "topics": [ "control" ], "tags": [ "safety", "survey", "safe-rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Survey connecting control-theory and RL views of safety, from learned dynamics with guarantees to constrained RL and safety filters.", "takeaway_zh": "综述，连接控制理论与强化学习两种安全视角：从带保证的学习动力学到约束强化学习与安全滤波器。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2108.06266.html", "labs": [ "toronto" ] }, { "id": "arxiv-2109.06325", "title": "safe-control-gym: a Unified Benchmark Suite for Safe Learning-based Control and Reinforcement Learning in Robotics", "type": "benchmark", "arxiv": "2109.06325", "year": 2021, "venue": null, "org": "University of Toronto", "url": null, "code": "https://github.com/utiasDSL/safe-control-gym", "project": null, "topics": [ "control", "datasets" ], "tags": [ "safety", "benchmark", "safe-rl" ], "robot_types": [ "aerial" ], "highlight": false, "added": null, "takeaway_en": "Gym-style benchmark with cart-pole and quadrotor tasks, constraints and disturbances for comparing safe control and safe RL methods.", "takeaway_zh": "Gym 风格基准，含倒立摆与四旋翼任务、约束与扰动，用于比较安全控制与安全强化学习方法。", "github_stars": 931, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2109.06325.html", "labs": [ "toronto" ] }, { "id": "arxiv-1909.04947", "title": "Crocoddyl: An Efficient and Versatile Framework for Multi-Contact Optimal Control", "type": "paper", "arxiv": "1909.04947", "year": 2019, "venue": null, "org": "LAAS-CNRS, University of Edinburgh", "url": null, "code": "https://github.com/loco-3d/crocoddyl", "project": null, "topics": [ "control", "locomotion" ], "tags": [ "optimal-control", "ddp", "legged", "open-source" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "An open-source DDP-based solver for contact-rich trajectory optimization, widely used for legged and humanoid motion generation and MPC.", "takeaway_zh": "基于 DDP 的开源接触丰富轨迹优化求解器，广泛用于足式与人形机器人的运动生成和 MPC。", "github_stars": 1320, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1909.04947.html", "labs": [] }, { "id": "arxiv-1910.13753", "title": "acados: a modular open-source framework for fast embedded optimal control", "type": "paper", "arxiv": "1910.13753", "year": 2019, "venue": null, "org": "University of Freiburg", "url": null, "code": "https://github.com/acados/acados", "project": null, "topics": [ "control" ], "tags": [ "mpc", "optimal-control", "embedded", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Fast nonlinear MPC and estimation solvers with Python/MATLAB interfaces that generate C code for embedded targets.", "takeaway_zh": "高速非线性 MPC 与状态估计求解器，提供 Python/MATLAB 接口并可为嵌入式平台生成 C 代码。", "github_stars": 1512, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1910.13753.html", "labs": [ "freiburg" ] }, { "id": "arxiv-2610.21913", "title": "Sub-Millisecond Reflexes for Continuous Physical Human-Robot Interaction", "type": "paper", "arxiv": "2610.21913", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/RoboRef/franka_reflex_engine", "project": null, "topics": [ "control", "manipulation" ], "tags": [ "real", "arm", "phri", "ros2", "safety" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-19", "takeaway_en": "An open-source ROS 2 multimode controller that switches control modes and triggers safety reflexes in under a millisecond, without stopping the task; shown on a Franka arm.", "takeaway_zh": "开源的 ROS 2 多模式控制器，可在亚毫秒内切换控制模式并触发安全反射而无需中断任务；在 Franka 机械臂上验证。", "github_stars": 2, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.21913.html", "labs": [] }, { "id": "arxiv-2610.23884", "title": "PrismPlan: A Shared GPU Substrate for Motion Planning Across Paradigms", "type": "paper", "arxiv": "2610.23884", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/XiaoyubotMotionTeam/PrismPlan", "project": null, "topics": [ "control" ], "tags": [ "motion-planning", "gpu", "benchmark" ], "robot_types": [], "highlight": false, "added": "2026-10-20", "takeaway_en": "Builds batched GPU collision checking and forward kinematics once and runs sampling, search and optimization planners on top, so the three paradigms can be compared fairly.", "takeaway_zh": "一次性实现批量 GPU 碰撞检测与正运动学，在其上运行采样、搜索和优化三类规划器，使三种范式可以公平比较。", "github_stars": 3, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.23884.html", "labs": [] }, { "id": "arxiv-2212.00541", "title": "Predictive Sampling: Real-time Behaviour Synthesis with MuJoCo", "type": "paper", "arxiv": "2212.00541", "year": 2022, "venue": null, "org": "Google DeepMind", "url": null, "code": "https://github.com/google-deepmind/mujoco_mpc", "project": null, "topics": [ "control" ], "tags": [ "sim", "mpc", "sampling-based", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Introduces MuJoCo MPC (MJPC), an interactive tool for real-time predictive control in MuJoCo, and shows that a simple sampling-based planner is a strong baseline for contact-rich behaviours.", "takeaway_zh": "推出 MuJoCo MPC（MJPC），可在 MuJoCo 中交互式地做实时预测控制，并表明简单的基于采样的规划器在接触丰富的行为上是很强的基线。", "github_stars": 1755, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2212.00541.html", "labs": [ "google-deepmind" ] }, { "id": "pinocchio", "title": "Pinocchio", "type": "code", "arxiv": null, "year": 2019, "venue": null, "org": "LAAS-CNRS, INRIA", "url": null, "code": "https://github.com/stack-of-tasks/pinocchio", "project": null, "topics": [ "control" ], "tags": [ "library", "dynamics", "kinematics", "open-source" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Fast C++/Python rigid-body dynamics library with analytical derivatives; a common backbone for kinematics, optimal control (e.g. Crocoddyl) and legged/humanoid control stacks.", "takeaway_zh": "高性能 C++/Python 刚体动力学库，提供解析导数；是运动学、最优控制（如 Crocoddyl）以及足式/人形机器人控制栈的常用底座。", "github_stars": 3828, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/pinocchio.html", "labs": [] }, { "id": "arxiv-2610.25476", "title": "CusADi2: Latency-Optimized Batch Computation of Robot Kinematics, Jacobians, and Hessians on the GPU", "type": "paper", "arxiv": "2610.25476", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/johnviljoen/cusadi2", "project": null, "topics": [ "control" ], "tags": [ "gpu", "kinematics", "trajectory-optimization" ], "robot_types": [], "highlight": false, "added": "2026-10-20", "takeaway_en": "Evaluates batches of kinematics, Jacobians and Hessians on the GPU with low wall-clock latency, targeting online sampling- and gradient-based trajectory optimization.", "takeaway_zh": "在 GPU 上以低延迟批量计算运动学、雅可比与海森矩阵，面向在线运行的采样式与梯度式轨迹优化。", "github_stars": 0, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25476.html", "labs": [] }, { "id": "arxiv-1711.08013", "title": "OSQP: An Operator Splitting Solver for Quadratic Programs", "type": "paper", "arxiv": "1711.08013", "year": 2020, "venue": null, "org": "University of Oxford, ETH Zurich, IMT Lucca, Stanford", "url": null, "code": "https://github.com/osqp/osqp", "project": null, "topics": [ "control" ], "tags": [ "optimization", "qp-solver", "mpc", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A fast, robust ADMM-based QP solver with warm starting and embedded C code generation; widely used inside MPC and whole-body controllers.", "takeaway_zh": "基于 ADMM 的快速、鲁棒二次规划求解器，支持热启动与嵌入式 C 代码生成，广泛用于 MPC 与全身控制器。", "github_stars": 2216, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1711.08013.html", "labs": [ "stanford", "eth", "oxford" ] }, { "id": "ompl-the-open-motion-planning-library", "title": "OMPL: The Open Motion Planning Library", "type": "code", "arxiv": null, "year": 2012, "venue": null, "org": "Rice University (Kavraki Lab)", "url": "https://ompl.kavrakilab.org/", "code": "https://github.com/ompl/ompl", "project": null, "topics": [ "control" ], "tags": [ "library", "motion-planning", "sampling-based", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The standard library of sampling-based motion planners (RRT, RRT-Connect, PRM, RRT* and many more); the default planning backend of MoveIt.", "takeaway_zh": "采样式运动规划器的标准库（RRT、RRT-Connect、PRM、RRT* 等众多算法），也是 MoveIt 默认的规划后端。", "github_stars": 2189, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/ompl-the-open-motion-planning-library.html", "labs": [] }, { "id": "moveit-2", "title": "MoveIt 2", "type": "code", "arxiv": null, "year": 2019, "venue": null, "org": "PickNik Robotics and community", "url": "https://moveit.ai/", "code": "https://github.com/moveit/moveit2", "project": null, "topics": [ "control", "manipulation" ], "tags": [ "ros2", "motion-planning", "manipulation", "open-source" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "The ROS 2 manipulation framework - motion planning, inverse kinematics, collision checking and real-time servoing for robot arms.", "takeaway_zh": "ROS 2 的机械臂操作框架：提供运动规划、逆运动学、碰撞检测与实时伺服控制。", "github_stars": 2094, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/moveit-2.html", "labs": [] }, { "id": "drake", "title": "Drake", "type": "code", "arxiv": null, "year": 2014, "venue": null, "org": "MIT, Toyota Research Institute", "url": "https://drake.mit.edu/", "code": "https://github.com/RobotLocomotion/drake", "project": null, "topics": [ "control", "sim" ], "tags": [ "library", "optimization", "simulation", "trajectory-optimization", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Model-based toolbox for robotics - multibody dynamics with contact simulation, mathematical programming and trajectory optimization; used in the MIT manipulation and underactuated courses.", "takeaway_zh": "面向机器人的基于模型的工具箱：带接触的多体动力学仿真、数学规划与轨迹优化；MIT 操作与欠驱动课程即使用它。", "github_stars": 4242, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/drake.html", "labs": [ "mit", "tri" ] }, { "id": "arxiv-2309.14545", "title": "Motions in Microseconds via Vectorized Sampling-Based Planning", "type": "paper", "arxiv": "2309.14545", "year": 2023, "venue": "ICRA 2024", "org": "Rice University (Kavraki Lab)", "url": null, "code": "https://github.com/KavrakiLab/vamp", "project": null, "topics": [ "control", "manipulation" ], "tags": [ "motion-planning", "simd", "sampling-based", "cpu" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "VAMP — vectorizes collision checking with CPU SIMD so sampling-based planners like RRT-Connect solve arm motion-planning queries in microseconds to milliseconds, no GPU needed.", "takeaway_zh": "VAMP：用 CPU SIMD 指令把碰撞检测向量化，使 RRT-Connect 等基于采样的规划器在微秒到毫秒级求解机械臂运动规划，无需 GPU。", "github_stars": 498, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.14545.html", "labs": [] }, { "id": "arxiv-1801.02854", "title": "Riemannian Motion Policies", "type": "paper", "arxiv": "1801.02854", "year": 2018, "venue": null, "org": "NVIDIA", "url": null, "code": null, "project": null, "topics": [ "control" ], "tags": [ "reactive-control", "motion-generation", "geometric-control" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "RMPs — describe each behaviour (reach, avoid obstacles, joint limits) as an acceleration policy plus a Riemannian metric, then combine them geometrically; the basis of later RMPflow and geometric fabrics controllers.", "takeaway_zh": "RMP：把每种行为（到达、避障、关节限位）表示为加速度策略加黎曼度量，再按几何方式组合；是后来 RMPflow 与几何织物（geometric fabrics）控制器的基础。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1801.02854.html", "labs": [ "nvidia" ] }, { "id": "arxiv-1709.07523", "title": "Hamilton-Jacobi Reachability: A Brief Overview and Recent Advances", "type": "paper", "arxiv": "1709.07523", "year": 2017, "venue": "CDC 2017", "org": "UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "control" ], "tags": [ "safety", "reachability", "tutorial" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Short tutorial on Hamilton-Jacobi reachability — computing the set of states from which a system can be kept safe under worst-case disturbances, and the tricks used to scale it up.", "takeaway_zh": "Hamilton-Jacobi 可达性分析简明教程：计算在最坏扰动下仍能保持系统安全的状态集合，以及将其扩展到更高维系统的方法。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1709.07523.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2105.04830", "title": "Jerk-limited Real-time Trajectory Generation with Arbitrary Target States", "type": "paper", "arxiv": "2105.04830", "year": 2021, "venue": "RSS 2021", "org": "Karlsruhe Institute of Technology (KIT)", "url": null, "code": "https://github.com/pantor/ruckig", "project": null, "topics": [ "control", "manipulation" ], "tags": [ "trajectory-generation", "real-time", "jerk-limited", "library" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Ruckig — open-source C++/Python library for time-optimal, jerk-limited online trajectories to any target position, velocity and acceleration; a common drop-in for smooth arm motion on real controllers.", "takeaway_zh": "Ruckig：开源 C++/Python 库，可实时生成到任意目标位置、速度和加速度的时间最优、限加加速度（jerk）轨迹；常被直接用于真实控制器上的机械臂平滑运动。", "github_stars": 1423, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2105.04830.html", "labs": [] }, { "id": "ocs2-optimal-control-for-switched-systems", "title": "OCS2: Optimal Control for Switched Systems", "type": "code", "arxiv": null, "year": 2021, "venue": null, "org": "ETH Zurich (Robotic Systems Lab)", "url": null, "code": "https://github.com/leggedrobotics/ocs2", "project": null, "topics": [ "control", "locomotion" ], "tags": [ "library", "mpc", "optimal-control", "legged", "mobile-manipulation" ], "robot_types": [ "arm", "legged", "mobile" ], "highlight": false, "added": null, "takeaway_en": "C++ toolbox for nonlinear MPC on switched systems (SLQ, DDP, SQP) with robot examples for legged robots and mobile manipulators; the base of many open-source legged MPC + whole-body control stacks.", "takeaway_zh": "面向切换系统的非线性 MPC C++ 工具箱（SLQ、DDP、SQP），附足式机器人与移动操作臂示例；许多开源足式 MPC + 全身控制方案都基于它。", "github_stars": 1530, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/ocs2-optimal-control-for-switched-systems.html", "labs": [ "eth" ] }, { "id": "dynamic-locomotion-in-the-mit-cheetah-3-through-convex-model", "title": "Dynamic Locomotion in the MIT Cheetah 3 Through Convex Model-Predictive Control", "type": "paper", "arxiv": null, "year": 2018, "venue": "IROS 2018", "org": "MIT (Biomimetic Robotics Lab)", "url": null, "code": "https://github.com/mit-biomimetics/Cheetah-Software", "project": null, "topics": [ "control", "locomotion" ], "tags": [ "mpc", "quadruped", "convex-optimization", "ground-reaction-forces" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "Convex MPC for quadrupeds — simplified rigid-body dynamics turn ground-reaction-force planning into a small QP solved in real time; the classic model-based baseline behind many open-source quadruped controllers.", "takeaway_zh": "四足机器人的凸 MPC：用简化刚体动力学把足端地面反作用力规划变成可实时求解的小规模 QP；众多开源四足控制器背后的经典基于模型的基线。", "github_stars": 3330, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/dynamic-locomotion-in-the-mit-cheetah-3-through-convex-model.html", "labs": [ "mit" ] }, { "id": "legged-control", "title": "legged_control", "type": "code", "arxiv": null, "year": 2022, "venue": null, "org": "Qiayuan Liao (open-source project)", "url": null, "code": "https://github.com/qiayuanl/legged_control", "project": null, "topics": [ "control", "locomotion" ], "tags": [ "mpc", "whole-body-control", "state-estimation", "ros", "legged" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "Open-source NMPC + whole-body control + state-estimation stack for quadrupeds built on OCS2 and ros-control, with a sim-to-real path; a popular starting point for model-based legged control.", "takeaway_zh": "基于 OCS2 与 ros-control 的开源四足控制栈：非线性 MPC + 全身控制 + 状态估计，并支持 sim2real；基于模型的足式控制常用的起点。", "github_stars": 1839, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/legged-control.html", "labs": [] }, { "id": "arxiv-2610.36961", "title": "Keeping Contact: Task-Adaptive Friction Margins for Bimanual Non-Prehensile Manipulation", "type": "paper", "arxiv": "2610.36961", "year": 2026, "venue": "RA-L", "org": null, "url": null, "code": null, "project": "https://keeping-contact.github.io/", "topics": [ "control", "manipulation" ], "tags": [ "contact", "friction", "socp", "bimanual", "non-prehensile", "real" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-28", "takeaway_en": "Signed friction distance as a task-adaptive slip margin for bimanual tilting about a pivot; a sequence of SOCPs gives optimal contact forces. Real tests show adaptive margins beat fixed or no margins.", "takeaway_zh": "以带符号摩擦距离度量防滑裕度，并在双臂绕支点倾转物体时按任务需求自适应调整；一系列二阶锥规划给出全局最优接触力。真机实验显示自适应裕度优于固定裕度和无裕度。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36961.html", "labs": [] }, { "id": "arxiv-2610.36535", "title": "A generalist policy for muscle-driven control across individual musculoskeletal models", "type": "paper", "arxiv": "2610.36535", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://woodater-thu.github.io/MyoAcross/", "topics": [ "control", "locomotion" ], "tags": [ "musculoskeletal", "morphology-conditioned", "motion-tracking", "biomechanics", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "MyoAcross trains one morphology-conditioned policy that tracks reference motions on unseen individual musculoskeletal models without fine-tuning. Reports 99.24% completion on 31 held-out models and 22.6% lower error.", "takeaway_zh": "MyoAcross 结合个体化肌骨模型构建、形态条件策略学习与虚拟模型增广，用一个策略驱动不同个体的肌骨模型跟踪参考动作而无需逐个微调；报告在 31 个未见个体模型上完成率 99.24%，跟踪误差比无条件 MLP 降低 22.6%，并与 120 人的地面反力和肌电数据做了对比。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36535.html", "labs": [] }, { "id": "arxiv-1707.06347", "title": "Proximal Policy Optimization Algorithms", "type": "paper", "arxiv": "1707.06347", "year": 2017, "venue": null, "org": "OpenAI", "url": null, "code": null, "project": null, "topics": [ "learning" ], "tags": [ "rl", "on-policy", "sim" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "The clipped policy-gradient method behind most GPU-parallel sim-to-real locomotion and dexterity training.", "takeaway_zh": "采用截断目标的策略梯度方法，是大多数 GPU 并行仿真训练（运动控制、灵巧操作）背后的算法。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1707.06347.html", "labs": [ "openai" ] }, { "id": "arxiv-1801.01290", "title": "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor", "type": "paper", "arxiv": "1801.01290", "year": 2018, "venue": null, "org": "UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "learning" ], "tags": [ "rl", "off-policy", "max-entropy" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Off-policy RL with an entropy bonus; sample-efficient enough to become the default for learning directly on real robots.", "takeaway_zh": "带熵奖励的离策略强化学习；样本效率较高，成为真机上直接学习的常用默认算法。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1801.01290.html", "labs": [ "berkeley" ] }, { "id": "arxiv-1011.0686", "title": "A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning", "type": "paper", "arxiv": "1011.0686", "year": 2010, "venue": null, "org": "CMU", "url": null, "code": null, "project": null, "topics": [ "learning" ], "tags": [ "imitation", "dagger", "interactive" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "DAgger — query the expert on the states the learner itself visits, fixing the compounding errors of plain behavior cloning.", "takeaway_zh": "DAgger：在学习者自己访问到的状态上向专家查询标注，解决单纯行为克隆的误差累积问题。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1011.0686.html", "labs": [ "cmu" ] }, { "id": "arxiv-2302.02948", "title": "Efficient Online Reinforcement Learning with Offline Data", "type": "paper", "arxiv": "2302.02948", "year": 2023, "venue": null, "org": "UC Berkeley", "url": null, "code": "https://github.com/ikostrikov/rlpd", "project": null, "topics": [ "learning" ], "tags": [ "rl", "offline-to-online", "sample-efficiency" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "RLPD — a few simple changes to off-policy RL (symmetric sampling of offline and online data, layer norm, ensembles) make prior data speed up online learning.", "takeaway_zh": "RLPD：对离策略强化学习做几处简单改动（离线/在线数据对称采样、层归一化、集成），让已有数据加速在线学习。", "github_stars": 420, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2302.02948.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2401.16013", "title": "SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning", "type": "paper", "arxiv": "2401.16013", "year": 2024, "venue": null, "org": "UC Berkeley", "url": null, "code": "https://github.com/rail-berkeley/serl", "project": null, "topics": [ "learning", "manipulation" ], "tags": [ "rl", "real", "arm", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "An open-source stack for training manipulation policies with RL on a real arm — learned reward classifiers, reset-free training and an impedance controller.", "takeaway_zh": "在真实机械臂上用强化学习训练操作策略的开源软件栈：学习式奖励分类器、免重置训练与阻抗控制器。", "github_stars": 865, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2401.16013.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2410.21845", "title": "Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning", "type": "paper", "arxiv": "2410.21845", "year": 2024, "venue": null, "org": "UC Berkeley", "url": null, "code": "https://github.com/rail-berkeley/hil-serl", "project": null, "topics": [ "learning", "manipulation" ], "tags": [ "rl", "real", "arm", "human-in-the-loop" ], "robot_types": [ "hand" ], "highlight": true, "added": null, "takeaway_en": "HIL-SERL — real-world RL with human corrections, reported to reach near-perfect success on precise assembly and dynamic tasks after a few hours of training.", "takeaway_zh": "HIL-SERL：结合人类纠正的真机强化学习，报告称在精密装配与动态任务上经数小时训练即可接近满分成功率。", "github_stars": 1564, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.21845.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2108.03298", "title": "What Matters in Learning from Offline Human Demonstrations for Robot Manipulation", "type": "paper", "arxiv": "2108.03298", "year": 2021, "venue": "CoRL 2021", "org": "Stanford", "url": null, "code": "https://github.com/ARISE-Initiative/robomimic", "project": null, "topics": [ "learning", "manipulation", "datasets" ], "tags": [ "imitation", "offline-rl", "benchmark", "human-demos" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "robomimic — a controlled study of imitation and offline RL on human-collected manipulation demos (demo quality, observation choices, history-aware BC), released with the framework and datasets.", "takeaway_zh": "robomimic：在人类采集的操作示教数据上系统比较模仿学习与离线强化学习（示教质量、观测选择、带历史的行为克隆），并开源框架与数据集。", "github_stars": 1591, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2108.03298.html", "labs": [ "stanford" ] }, { "id": "arxiv-2110.06169", "title": "Offline Reinforcement Learning with Implicit Q-Learning", "type": "paper", "arxiv": "2110.06169", "year": 2021, "venue": "ICLR 2022", "org": "UC Berkeley", "url": null, "code": "https://github.com/ikostrikov/implicit_q_learning", "project": null, "topics": [ "learning" ], "tags": [ "offline-rl", "expectile-regression", "offline-to-online" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "IQL — learns values with expectile regression without ever querying actions outside the dataset, then extracts a policy by advantage-weighted regression; a common base for offline-to-online robot RL.", "takeaway_zh": "IQL：用期望分位数回归学习价值函数、从不查询数据集外的动作，再用优势加权回归提取策略；是机器人离线到在线强化学习的常用基础。", "github_stars": 340, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2110.06169.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2206.11251", "title": "Behavior Transformers: Cloning $k$ modes with one stone", "type": "paper", "arxiv": "2206.11251", "year": 2022, "venue": "NeurIPS 2022", "org": "NYU", "url": null, "code": "https://github.com/notmahi/bet", "project": null, "topics": [ "learning" ], "tags": [ "imitation", "multimodal", "transformer" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "BeT — clusters actions into k bins plus a continuous offset so a transformer can imitate multimodal demonstrations instead of averaging them.", "takeaway_zh": "BeT：把动作聚成 k 个类别再加连续偏移，让 Transformer 能模仿多模态示教，而不是把不同模式平均掉。", "github_stars": 152, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2206.11251.html", "labs": [ "nyu" ] }, { "id": "arxiv-2409.00588", "title": "Diffusion Policy Policy Optimization", "type": "paper", "arxiv": "2409.00588", "year": 2024, "venue": "ICLR 2025", "org": "Princeton", "url": null, "code": "https://github.com/irom-princeton/dppo", "project": null, "topics": [ "learning", "manipulation" ], "tags": [ "rl", "diffusion-policy", "fine-tuning", "sim-to-real" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "DPPO — fine-tunes a pre-trained diffusion policy with PPO by treating the denoising steps as part of the MDP; reports stable gains over other RL fine-tuning methods and sim-to-real transfer.", "takeaway_zh": "DPPO：把去噪步骤视为 MDP 的一部分，用 PPO 微调预训练扩散策略；报告称比其他强化学习微调方法更稳定，并实现了仿真到真机迁移。", "github_stars": 866, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2409.00588.html", "labs": [ "princeton" ] }, { "id": "arxiv-2006.04779", "title": "Conservative Q-Learning for Offline Reinforcement Learning", "type": "paper", "arxiv": "2006.04779", "year": 2020, "venue": "NeurIPS 2020", "org": "UC Berkeley", "url": null, "code": "https://github.com/aviralkumar2907/CQL", "project": null, "topics": [ "learning" ], "tags": [ "offline-rl", "value-learning" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "CQL — penalizes Q-values on actions outside the dataset so the learned value is a lower bound; a standard offline-RL baseline for learning from fixed robot logs.", "takeaway_zh": "CQL：压低数据集外动作的 Q 值，使学到的价值成为下界；是从固定机器人日志中学习的标准离线强化学习基线。", "github_stars": 494, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2006.04779.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2106.01345", "title": "Decision Transformer: Reinforcement Learning via Sequence Modeling", "type": "paper", "arxiv": "2106.01345", "year": 2021, "venue": "NeurIPS 2021", "org": "UC Berkeley · Google · FAIR", "url": null, "code": "https://github.com/kzl/decision-transformer", "project": null, "topics": [ "learning" ], "tags": [ "offline-rl", "transformer", "sequence-modeling" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Recasts RL as sequence modeling — a GPT-style model predicts actions conditioned on the desired return; an early bridge between transformers and control.", "takeaway_zh": "把强化学习改写成序列建模：GPT 式模型以期望回报为条件预测动作；是 Transformer 与控制结合的早期代表。", "github_stars": 2845, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2106.01345.html", "labs": [ "berkeley", "google-deepmind", "meta" ] }, { "id": "arxiv-1802.09477", "title": "Addressing Function Approximation Error in Actor-Critic Methods", "type": "paper", "arxiv": "1802.09477", "year": 2018, "venue": "ICML 2018", "org": "McGill University", "url": null, "code": "https://github.com/sfujim/TD3", "project": null, "topics": [ "learning" ], "tags": [ "rl", "off-policy", "continuous-control" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "TD3 — fixes DDPG's value overestimation with twin critics, delayed policy updates and target smoothing; still a standard continuous-control baseline.", "takeaway_zh": "TD3：通过双 Critic、延迟策略更新和目标策略平滑解决 DDPG 的价值高估问题，至今仍是连续控制的常用基线。", "github_stars": 2118, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1802.09477.html", "labs": [] }, { "id": "arxiv-1606.03476", "title": "Generative Adversarial Imitation Learning", "type": "paper", "arxiv": "1606.03476", "year": 2016, "venue": "NeurIPS 2016", "org": "Stanford", "url": null, "code": null, "project": null, "topics": [ "learning" ], "tags": [ "imitation", "adversarial", "inverse-rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "GAIL — learns a policy whose state-action distribution a discriminator cannot tell apart from the expert's, skipping explicit reward recovery.", "takeaway_zh": "GAIL：训练策略使判别器无法区分其状态-动作分布与专家数据，无需显式恢复奖励函数。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1606.03476.html", "labs": [ "stanford" ] }, { "id": "arxiv-2109.00137", "title": "Implicit Behavioral Cloning", "type": "paper", "arxiv": "2109.00137", "year": 2021, "venue": "CoRL 2021", "org": "Google", "url": null, "code": null, "project": null, "topics": [ "learning" ], "tags": [ "imitation", "energy-based", "multimodal" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Represents the policy as an energy-based model over actions instead of a regressor, which handles multimodal and discontinuous demonstrations much better.", "takeaway_zh": "用基于能量的模型（而非直接回归）表示策略，更好地处理多模态、不连续的示范数据。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2109.00137.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2006.09359", "title": "AWAC: Accelerating Online Reinforcement Learning with Offline Datasets", "type": "paper", "arxiv": "2006.09359", "year": 2020, "venue": null, "org": "UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "learning" ], "tags": [ "rl", "offline-to-online", "real" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "Advantage-weighted actor-critic that pre-trains on prior demonstrations and keeps improving online, shown on real dexterous-hand tasks.", "takeaway_zh": "优势加权的 Actor-Critic：先用已有示范数据预训练，再在线持续提升，并在真实灵巧手任务上验证。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2006.09359.html", "labs": [ "berkeley" ] }, { "id": "arxiv-1806.10293", "title": "QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation", "type": "paper", "arxiv": "1806.10293", "year": 2018, "venue": "CoRL 2018", "org": "Google", "url": null, "code": null, "project": null, "topics": [ "learning", "manipulation" ], "tags": [ "real", "grasping", "off-policy", "q-learning", "large-scale" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Off-policy Q-learning with a cross-entropy-method action optimizer, trained on several hundred thousand real grasp attempts collected by a fleet of robots; reports 96% grasp success on unseen objects.", "takeaway_zh": "使用交叉熵方法优化动作的离策略 Q 学习，在机器人集群采集的数十万次真实抓取上训练；报告对未见物体的抓取成功率为 96%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1806.10293.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2107.09645", "title": "Mastering Visual Continuous Control: Improved Data-Augmented Reinforcement Learning", "type": "paper", "arxiv": "2107.09645", "year": 2021, "venue": "ICLR 2022", "org": "NYU, Meta AI", "url": null, "code": "https://github.com/facebookresearch/drqv2", "project": null, "topics": [ "learning" ], "tags": [ "sim", "pixels", "data-augmentation", "off-policy" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "DrQ-v2 - a simple model-free actor-critic with image augmentation that learns continuous control from pixels (incl. humanoid tasks in the DeepMind Control Suite) on a single GPU; a common visual-RL baseline.", "takeaway_zh": "DrQ-v2：带图像增强的简单无模型 actor-critic，单 GPU 即可从像素学习连续控制（包括 DeepMind Control Suite 的人形任务），是常用的视觉强化学习基线。", "github_stars": 438, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2107.09645.html", "labs": [ "meta", "nyu" ] }, { "id": "arxiv-2506.15799", "title": "Steering Your Diffusion Policy with Latent Space Reinforcement Learning", "type": "paper", "arxiv": "2506.15799", "year": 2025, "venue": "CoRL 2025", "org": "UC Berkeley, University of Washington", "url": null, "code": "https://github.com/ajwagen/dsrl", "project": "https://diffusion-steering.github.io", "topics": [ "learning", "manipulation" ], "tags": [ "real", "rl-finetuning", "diffusion-policy", "sample-efficient" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "DSRL — adapts a frozen diffusion policy by running RL over its input noise, not its weights; needs only forward passes, works with π0, and reports real-robot success rising from 20% to 90% in under 50 online episodes.", "takeaway_zh": "DSRL：不改扩散策略的权重，而是对其输入的初始噪声做强化学习来\"引导\"输出动作；只需前向推理、可用于 π0，论文报告真机任务在不到 50 次在线交互内成功率从 20% 升至 90%。", "github_stars": 236, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2506.15799.html", "labs": [ "berkeley", "uw" ] }, { "id": "arxiv-2502.02538", "title": "Flow Q-Learning", "type": "paper", "arxiv": "2502.02538", "year": 2025, "venue": "ICML 2025", "org": "UC Berkeley", "url": null, "code": "https://github.com/seohongpark/fql", "project": "https://seohong.me/projects/fql/", "topics": [ "learning" ], "tags": [ "sim", "offline-rl", "flow-matching", "ogbench" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "FQL — a flow-matching BC policy plus a one-step policy that maximizes Q while distilling from it; no backprop through flow steps, strong on 73 OGBench / D4RL offline and offline-to-online tasks.", "takeaway_zh": "FQL：先用行为克隆训练流匹配策略，再训练一个最大化 Q 值、同时向流策略蒸馏的单步策略；避免对多步流采样反向传播，在 73 个 OGBench / D4RL 离线及离线转在线任务上表现突出。", "github_stars": 341, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.02538.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2509.09674", "title": "SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning", "type": "paper", "arxiv": "2509.09674", "year": 2025, "venue": "ICLR 2026", "org": "Shanghai Jiao Tong University, Peking University, HKU, Shanghai AI Lab", "url": null, "code": "https://github.com/PRIME-RL/SimpleVLA-RL", "project": null, "topics": [ "learning", "vla" ], "tags": [ "sim", "rl-finetuning", "grpo", "libero", "robotwin" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Online RL for VLAs with GRPO and only binary success rewards on OpenVLA-OFT, plus exploration tweaks; reports 99.1% on LIBERO and 70.4% on RoboTwin 1.0 (vs 39.8% for the SFT baseline).", "takeaway_zh": "在 OpenVLA-OFT 上用 GRPO 与仅\"成功/失败\"二值奖励做在线强化学习，并加入若干促进探索的改动；论文报告 LIBERO 成功率 99.1%、RoboTwin 1.0 为 70.4%（SFT 基线 39.8%）。", "github_stars": 1900, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2509.09674.html", "labs": [ "shanghai-ai-lab", "pku", "sjtu", "hku" ] }, { "id": "arxiv-2303.05479", "title": "Cal-QL: Calibrated Offline RL Pre-Training for Efficient Online Fine-Tuning", "type": "paper", "arxiv": "2303.05479", "year": 2023, "venue": "NeurIPS 2023", "org": "UC Berkeley, Stanford", "url": null, "code": "https://github.com/nakamotoo/Cal-QL", "project": null, "topics": [ "learning" ], "tags": [ "offline-to-online", "rl", "conservative-q" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Keeps conservative offline Q-values from dropping below a reference policy's value, which avoids the early performance dip when an offline-pretrained policy starts online fine-tuning.", "takeaway_zh": "约束保守的离线 Q 值不低于参考策略的价值，从而避免离线预训练策略开始在线微调时常见的性能先降后升。", "github_stars": 124, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2303.05479.html", "labs": [ "berkeley", "stanford" ] }, { "id": "arxiv-2309.10150", "title": "Q-Transformer: Scalable Offline Reinforcement Learning via Autoregressive Q-Functions", "type": "paper", "arxiv": "2309.10150", "year": 2023, "venue": "CoRL 2023", "org": "Google DeepMind", "url": null, "code": null, "project": null, "topics": [ "learning", "manipulation" ], "tags": [ "real", "offline-rl", "transformer", "large-scale" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Discretizes each action dimension and predicts Q-values per dimension autoregressively with a Transformer, plus a conservative regularizer; trained offline on large real robot datasets mixing demos and autonomous data.", "takeaway_zh": "将每个动作维度离散化，用 Transformer 自回归地逐维预测 Q 值，并加入保守正则；在混合演示与自主采集数据的大规模真实机器人数据集上离线训练。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.10150.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2510.06710", "title": "RLinf-VLA: A Unified and Efficient Framework for Reinforcement Learning of Vision-Language-Action Models", "type": "paper", "arxiv": "2510.06710", "year": 2025, "venue": null, "org": null, "url": null, "code": "https://github.com/RLinf/RLinf", "project": null, "topics": [ "learning", "vla" ], "tags": [ "sim", "rl-finetuning", "ppo", "grpo", "infrastructure", "libero", "maniskill", "robotwin" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open RL training framework for VLAs (PPO/GRPO across GPU-parallel simulators); reports 98.11% on 130 LIBERO tasks and 97.66% on 25 ManiSkill tasks after RL — a common base for VLA+RL work.", "takeaway_zh": "面向 VLA 的开源强化学习训练框架（PPO/GRPO，支持 GPU 并行仿真器）；论文报告经 RL 后在 130 个 LIBERO 任务上成功率 98.11%、25 个 ManiSkill 任务上 97.66%，是 VLA+RL 研究常用的底座。", "github_stars": 5546, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2510.06710.html", "labs": [] }, { "id": "arxiv-2610.25736", "title": "Learning from Fleet Experience: Collective Experience Learning Across 96 Robots", "type": "paper", "arxiv": "2610.25736", "year": 2026, "venue": null, "org": "Tsinghua, AgiBot, Infinigence AI", "url": null, "code": "https://github.com/RLinf/RLinf", "project": "https://cel-robot-learning.github.io", "topics": [ "learning", "manipulation" ], "tags": [ "real", "fleet-learning", "deployment", "failure-retrieval" ], "robot_types": [], "highlight": false, "added": "2026-10-23", "takeaway_en": "Deployed robots pool their runs; failures retrieve matching successful segments from the fleet buffer to fine-tune on, with no new demos. Reports 93–99% on five real tasks; 96 robots lift phone packing 79% → 98%.", "takeaway_zh": "部署中的机器人共享执行经验：用失败片段在机群经验库中检索相似的成功片段来微调策略，无需新增示教或人工纠正。论文报告五个真机任务成功率 93–99%，96 台机器人把手机装盒任务从 79% 提升到 98%。", "github_stars": 5546, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25736.html", "labs": [ "tsinghua", "agibot" ] }, { "id": "arxiv-2610.18596", "title": "ACCRUE: Continual Off-Policy Reinforcement Learning for Robot Foundation Models", "type": "paper", "arxiv": "2610.18596", "year": 2026, "venue": null, "org": "NVIDIA, KAIST", "url": null, "code": null, "project": "https://accrue-rl.github.io", "topics": [ "learning", "vla", "manipulation" ], "tags": [ "real", "sim", "continual-learning", "off-policy-rl", "rl-finetuning", "assembly" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Continual off-policy RL for VLAs — small actor/critic networks carry over between tasks, distilled on a few kept episodes to avoid forgetting; reports 31–68.5% less real-robot time on new assembly tasks.", "takeaway_zh": "面向 VLA 的持续离策略强化学习：轻量 actor/critic 在任务间延续，用少量保留的旧任务轨迹做蒸馏以防遗忘。论文报告新的接触式装配任务所需真机训练时间减少 31–68.5%，旧任务几乎不遗忘。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.18596.html", "labs": [ "nvidia", "kaist" ] }, { "id": "arxiv-2610.28589", "title": "AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics", "type": "paper", "arxiv": "2610.28589", "year": 2026, "venue": null, "org": "University of Central Florida", "url": null, "code": null, "project": "https://xinxinliu-ai.github.io/AgenticRoboRL/", "topics": [ "learning", "vla" ], "tags": [ "llm-planner", "reinforcement-learning", "agentic", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Post-trains an open 9B robot planner (SFT + multi-turn RL) across harnesses that call VLA skills or write control code; reports LIBERO-PRO success rising from 18.1% to 71.0% without external memory.", "takeaway_zh": "像训练编程智能体一样对 9B 开源机器人规划器做后训练（SFT + 多轮强化学习），适配调用 VLA 技能和生成控制代码两类执行框架；报告在 LIBERO-PRO 上无外部记忆成功率从 18.1% 提升到 71.0%。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.28589.html", "labs": [] }, { "id": "arxiv-2610.28698", "title": "Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds", "type": "paper", "arxiv": "2610.28698", "year": 2026, "venue": null, "org": "Tsinghua Spatial Intelligence Lab", "url": null, "code": "https://github.com/THUSI-Lab/GameScaling", "project": "https://gamescaling.github.io", "topics": [ "learning" ], "tags": [ "scaling-law", "games", "virtual-worlds", "continued-pretraining", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Continued pre-training study for game-playing agents: 120 runs over 0.8B–27B models, 50–1,600 h of data and multi-world mixes; reports diversity aids transfer to related worlds, little to distant ones.", "takeaway_zh": "针对游戏中的具身智能体做持续预训练的对照研究：120 组配置覆盖 0.8B–27B 模型、50–1600 小时数据和多世界混合；报告数据多样性有助于迁移到相近世界，对差异大的世界帮助有限。", "github_stars": 440, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.28698.html", "labs": [ "tsinghua" ] }, { "id": "arxiv-2610.29677", "title": "Toward Recursive Self-Improvement in Physical AI", "type": "paper", "arxiv": "2610.29677", "year": 2026, "venue": null, "org": "National University of Singapore", "url": null, "code": null, "project": "https://github.com/Physical2RSI/Awesome-Physical2RSI", "topics": [ "learning" ], "tags": [ "survey", "self-improvement", "awesome-list" ], "robot_types": [], "highlight": false, "added": "2026-10-25", "takeaway_en": "Survey of recursive self-improvement in robot learning — improving skills, then the process that makes them — and when gains carry forward; notes no physical study has yet isolated gains from recursion itself.", "takeaway_zh": "把\"递归自我改进\"放到机器人学习语境下的综述：从提升技能到改进产生技能的过程，并总结增益能持续累积的条件；指出现有真机研究尚未单独验证\"递归\"本身带来的收益。附论文列表。", "github_stars": 79, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.29677.html", "labs": [ "nus" ] }, { "id": "arxiv-2610.26848", "title": "Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy", "type": "paper", "arxiv": "2610.26848", "year": 2026, "venue": null, "org": "Galbot", "url": null, "code": null, "project": "https://physrsi-eval.github.io/", "topics": [ "learning", "datasets" ], "tags": [ "benchmark", "self-improvement", "vlm-agent", "scaling-law", "technical-report" ], "robot_types": [], "highlight": false, "added": "2026-10-25", "takeaway_en": "PhysRSI-eval scores VLM agents on 18 manipulation tasks under a fixed trial budget on five axes, incl. learning from failure; reports similar benchmark averages hide different improvement speeds.", "takeaway_zh": "PhysRSI-eval 在固定交互预算下用 18 个操作任务、五个维度（含从失败中改进）评测 VLM 智能体；报告公开基准平均分相近的模型改进速度差别很大，且早期物理反馈可用于预测后续进展。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.26848.html", "labs": [] }, { "id": "arxiv-2610.25376", "title": "TILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models", "type": "paper", "arxiv": "2610.25376", "year": 2026, "venue": null, "org": "The University of Hong Kong", "url": null, "code": "https://github.com/hku-sail/TILDE", "project": "https://hku-sail.github.io/TILDE/", "topics": [ "learning", "vla" ], "tags": [ "distillation", "dagger", "recovery", "reward-model", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-25", "takeaway_en": "Distills a pretrained VLA teacher into a policy without action pretraining via on-policy supervision plus teacher-run recoveries, triggered by a reward model (AutoDAgger); reports up to 4.1x over BC on 3 real tasks.", "takeaway_zh": "把预训练具身基础模型（教师）蒸馏到未做动作预训练的策略：在学生访问的状态上做在线监督，并由奖励模型监控进度、自动触发教师接管采集纠错轨迹（AutoDAgger，无需人工看护）；报告在 3 个真机任务上比行为克隆最高提升 4.1 倍。", "github_stars": 9, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25376.html", "labs": [ "hku" ] }, { "id": "arxiv-2610.31845", "title": "EgoForesight: Effectively Scaling Policy Learning with Noisy Egocentric Actions", "type": "paper", "arxiv": "2610.31845", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://lin-shan.com/EgoForesight", "topics": [ "learning", "vla" ], "tags": [ "real", "sim", "human-video", "egocentric", "scaling", "dexterous-hand" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "Co-trains a VLA on robot demos and noisy egocentric-video actions by weighting each source by reliability and distilling from clean future frames; reports steeper scaling than the baseline up to 30k hours of data.", "takeaway_zh": "将机器人示教与动作标签噪声较大的第一视角视频联合训练 VLA：按数据源可靠性分配监督权重，并从干净的未来画面蒸馏表征；报告在最多 3 万小时预训练数据范围内扩展斜率高于基线。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31845.html", "labs": [] }, { "id": "arxiv-2610.29264", "title": "Training in Deployment: Human-in-the-loop Online Post-Training for Robot Manipulation", "type": "paper", "arxiv": "2610.29264", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://logosroboticsgroup.github.io/TIDE/", "topics": [ "learning", "manipulation" ], "tags": [ "human-in-the-loop", "online-learning", "post-training", "bimanual", "real" ], "robot_types": [ "arm" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "TIDE updates a pretrained policy online from human interventions and its own rollouts, with no critic or reward model; reports gains on four real bimanual tasks, each session finishing in under an hour.", "takeaway_zh": "TIDE 在部署过程中利用人类干预和策略自身的执行数据在线更新预训练策略，无需价值函数或奖励模型；报告在四个真实双臂任务上优于初始策略，每次在线后训练不到一小时完成。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.29264.html", "labs": [] }, { "id": "arxiv-2610.32093", "title": "DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning", "type": "paper", "arxiv": "2610.32093", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/DAVIAN-Robotics/DiscoDemo", "project": "https://davian-robotics.github.io/DiscoDemo", "topics": [ "learning", "manipulation" ], "tags": [ "rl", "data-generation", "imitation-learning", "diversity", "sim", "real" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Uses RL as a demo generator, not a controller: a latent-conditioned policy finds several distinct successful solutions, exported as an imitation dataset; reports 78% real-robot success vs. 39% for the best baseline.", "takeaway_zh": "把强化学习当作示教数据生成器而非部署控制器：以隐变量为条件的策略为同一任务找到多种不同的成功解法，导出为模仿学习数据集；报告真机下游成功率从最强基线的 39% 提升到 78%。", "github_stars": 15, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.32093.html", "labs": [] }, { "id": "arxiv-2610.36374", "title": "HIDRA: Hierarchical Dual-Routing Attention for Replay-Free Lifelong Imitation Learning", "type": "paper", "arxiv": "2610.36374", "year": 2026, "venue": "NeurIPS 2026", "org": null, "url": null, "code": "https://github.com/yfqi/HIDRA", "project": null, "topics": [ "learning", "manipulation" ], "tags": [ "lifelong-learning", "imitation-learning", "mixture-of-experts", "libero" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Replay-free lifelong imitation learning: hierarchical dual-routing attention retrieves experts by instruction, then refines by context, plus key regularization against drift. Reports beating replay baselines on LIBERO.", "takeaway_zh": "无回放的终身模仿学习：分层双路由注意力先按指令检索专家、再按上下文细化，并用键级几何正则抑制表征漂移带来的路由干扰；报告在多个 LIBERO 任务序列上优于无回放和有回放基线。", "github_stars": 0, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36374.html", "labs": [] }, { "id": "arxiv-2610.25692", "title": "Progress-Guided Preference Generation for Aligning Generative Control Policies", "type": "paper", "arxiv": "2610.25692", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": "https://pgpg2026.github.io/", "topics": [ "learning", "manipulation" ], "tags": [ "preference-learning", "diffusion-policy", "flow-matching", "fine-tuning", "long-horizon" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Fine-tunes diffusion and flow policies with automatic preference labels: rollouts are ranked by progress measured against a few expert demos, no human labels or reward design. Reports beating SFT on long-horizon tasks.", "takeaway_zh": "用自动生成的偏好标签微调扩散与流匹配策略：以少量专家示教为参照衡量自主回放的顺序进度并据此排序，无需人工标注或手工奖励；报告在长时程操作任务上优于监督微调基线。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25692.html", "labs": [] }, { "id": "mujoco", "title": "MuJoCo", "type": "simulator", "arxiv": null, "year": 2012, "venue": null, "org": "Google DeepMind", "url": "https://mujoco.org/", "code": "https://github.com/google-deepmind/mujoco", "project": null, "topics": [ "sim" ], "tags": [ "physics", "open-source" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "Fast, accurate contact physics; the default simulator for robot learning research, open-sourced by DeepMind.", "takeaway_zh": "快速准确的接触物理仿真，机器人学习研究的默认仿真器，由 DeepMind 开源。", "github_stars": 15956, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/mujoco.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2502.08844", "title": "MuJoCo Playground", "type": "paper", "arxiv": "2502.08844", "year": 2025, "venue": null, "org": "Google DeepMind, UC Berkeley", "url": null, "code": "https://github.com/google-deepmind/mujoco_playground", "project": null, "topics": [ "sim" ], "tags": [ "gpu-parallel", "sim2real", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "GPU-accelerated MJX environments for locomotion and manipulation that train in minutes and transfer to real robots.", "takeaway_zh": "基于 MJX 的 GPU 加速运动与操作环境，几分钟即可训练并迁移到真实机器人。", "github_stars": 2266, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.08844.html", "labs": [ "berkeley", "google-deepmind" ] }, { "id": "arxiv-2108.10470", "title": "Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning", "type": "paper", "arxiv": "2108.10470", "year": 2021, "venue": null, "org": "NVIDIA", "url": null, "code": null, "project": null, "topics": [ "sim" ], "tags": [ "gpu-parallel", "physics" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Keeps physics and policy on the GPU end-to-end, enabling thousands of parallel environments on one card.", "takeaway_zh": "物理仿真与策略全程在 GPU 上运行，单卡即可并行数千个环境。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2108.10470.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2301.04195", "title": "Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments", "type": "paper", "arxiv": "2301.04195", "year": 2023, "venue": null, "org": "ETH Zurich, NVIDIA, Toronto", "url": null, "code": "https://github.com/isaac-sim/IsaacLab", "project": null, "topics": [ "sim" ], "tags": [ "gpu-parallel", "framework", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The framework that became Isaac Lab — modular robots, sensors and tasks on top of Isaac Sim.", "takeaway_zh": "Isaac Lab 的前身，在 Isaac Sim 之上提供模块化的机器人、传感器与任务。", "github_stars": 8515, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2301.04195.html", "labs": [ "nvidia", "eth", "toronto" ] }, { "id": "arxiv-2410.00425", "title": "ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI", "type": "paper", "arxiv": "2410.00425", "year": 2024, "venue": null, "org": "UC San Diego", "url": null, "code": "https://github.com/haosulab/ManiSkill", "project": null, "topics": [ "sim" ], "tags": [ "gpu-parallel", "rendering", "benchmark", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "GPU-parallel simulation and rendering with many manipulation tasks; fast visual RL and data generation.", "takeaway_zh": "GPU 并行仿真与渲染，内置大量操作任务，适合快速视觉强化学习与数据生成。", "github_stars": 3478, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.00425.html", "labs": [ "ucsd" ] }, { "id": "genesis", "title": "Genesis", "type": "simulator", "arxiv": null, "year": 2024, "venue": null, "org": "Genesis Embodied AI", "url": null, "code": "https://github.com/Genesis-Embodied-AI/Genesis", "project": null, "topics": [ "sim" ], "tags": [ "physics", "gpu-parallel", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "An open-source, Python-native physics platform covering rigid bodies, soft bodies, fluids and fast GPU simulation.", "takeaway_zh": "开源、Python 原生的物理平台，覆盖刚体、软体、流体，支持高速 GPU 仿真。", "github_stars": 30139, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/genesis.html", "labs": [] }, { "id": "arxiv-2009.12293", "title": "robosuite: A Modular Simulation Framework and Benchmark for Robot Learning", "type": "paper", "arxiv": "2009.12293", "year": 2020, "venue": null, "org": "Stanford, UT Austin", "url": null, "code": "https://github.com/ARISE-Initiative/robosuite", "project": null, "topics": [ "sim", "datasets" ], "tags": [ "mujoco", "benchmark", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Modular MuJoCo-based manipulation environments behind many imitation-learning benchmarks.", "takeaway_zh": "基于 MuJoCo 的模块化操作环境，是众多模仿学习基准的底座。", "github_stars": 2687, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2009.12293.html", "labs": [ "stanford", "ut-austin" ] }, { "id": "arxiv-2406.02523", "title": "RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots", "type": "paper", "arxiv": "2406.02523", "year": 2024, "venue": null, "org": "UT Austin, NVIDIA", "url": null, "code": "https://github.com/robocasa/robocasa", "project": null, "topics": [ "sim", "datasets" ], "tags": [ "kitchen", "benchmark", "synthetic-data" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "120 kitchen scenes, 2,500+ objects and 100 tasks with generated demonstrations for training generalist policies.", "takeaway_zh": "120 个厨房场景、2500 多个物体与 100 个任务，并附生成演示，用于训练通用策略。", "github_stars": 1849, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2406.02523.html", "labs": [ "nvidia", "ut-austin" ] }, { "id": "arxiv-2003.08515", "title": "SAPIEN: A SimulAted Part-based Interactive ENvironment", "type": "paper", "arxiv": "2003.08515", "year": 2020, "venue": null, "org": "UC San Diego", "url": null, "code": null, "project": null, "topics": [ "sim" ], "tags": [ "articulated-objects", "physics" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A simulator with large-scale articulated objects (PartNet-Mobility) for interaction tasks like opening drawers.", "takeaway_zh": "包含大规模铰接物体（PartNet-Mobility）的仿真器，适合开抽屉等交互任务。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2003.08515.html", "labs": [ "ucsd" ] }, { "id": "arxiv-1703.06907", "title": "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World", "type": "paper", "arxiv": "1703.06907", "year": 2017, "venue": null, "org": "OpenAI", "url": null, "code": null, "project": null, "topics": [ "sim" ], "tags": [ "sim2real", "domain-randomization" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Randomizing textures, lighting and cameras in sim makes vision models work on real images with no real data.", "takeaway_zh": "在仿真中随机化纹理、光照和相机，使视觉模型无需真实数据即可在真实图像上工作。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1703.06907.html", "labs": [ "openai" ] }, { "id": "arxiv-1710.06537", "title": "Sim-to-Real Transfer of Robotic Control with Dynamics Randomization", "type": "paper", "arxiv": "1710.06537", "year": 2017, "venue": null, "org": "OpenAI, UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "sim" ], "tags": [ "sim2real", "domain-randomization", "rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Randomizing physical parameters during training yields control policies robust enough to run on a real arm.", "takeaway_zh": "训练中随机化物理参数，得到足以在真实机械臂上运行的鲁棒控制策略。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1710.06537.html", "labs": [ "berkeley", "openai" ] }, { "id": "arxiv-2610.19479", "title": "Active Generative Simulation: Meta-Learning What to Simulate for Sim-and-Real Policy Co-Training", "type": "paper", "arxiv": "2610.19479", "year": 2026, "venue": "CoRL 2026", "org": "Georgia Tech", "url": null, "code": null, "project": "https://meta-gensim.github.io", "topics": [ "sim" ], "tags": [ "sim-and-real", "co-training", "meta-learning" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "MetaGenSim learns which simulated data to generate, using meta-gradients from real-world performance, instead of sampling the simulator uniformly.", "takeaway_zh": "MetaGenSim 利用来自真实世界表现的元梯度，学习应该生成哪些仿真数据，而不是对仿真器做均匀采样。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.19479.html", "labs": [ "georgia-tech" ] }, { "id": "arxiv-2506.18088", "title": "RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation", "type": "paper", "arxiv": "2506.18088", "year": 2025, "venue": null, "org": null, "url": null, "code": "https://github.com/RoboTwin-Platform/RoboTwin", "project": null, "topics": [ "sim", "manipulation", "datasets" ], "tags": [ "sim", "dual-arm", "data-generation", "domain-randomization", "benchmark" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "Simulation pipeline that writes and checks task code with multimodal LLMs to mass-produce bimanual demos under heavy randomization (clutter, lighting, backgrounds, instructions); also a 50-task dual-arm benchmark.", "takeaway_zh": "用多模态大模型自动生成并校验任务代码，在强域随机化（杂物、光照、背景、指令）下批量产出双臂演示数据的仿真管线，同时提供 50 个双臂任务的基准。", "github_stars": 3016, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2506.18088.html", "labs": [] }, { "id": "arxiv-2310.17596", "title": "MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations", "type": "paper", "arxiv": "2310.17596", "year": 2023, "venue": "CoRL 2023", "org": "NVIDIA, UT Austin", "url": null, "code": "https://github.com/NVlabs/mimicgen", "project": null, "topics": [ "sim", "manipulation", "datasets" ], "tags": [ "sim", "data-generation", "imitation" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Turns a handful of human demos into thousands by splitting them into object-centric segments and replaying those segments in new scenes; widely reused to scale imitation data.", "takeaway_zh": "把少量人类演示切分为以物体为中心的片段，并在新场景中变换重放，从而扩增为数千条演示；被广泛用于扩充模仿学习数据。", "github_stars": 663, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2310.17596.html", "labs": [ "nvidia", "ut-austin" ] }, { "id": "arxiv-2403.09227", "title": "BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation", "type": "paper", "arxiv": "2403.09227", "year": 2022, "venue": "CoRL 2022", "org": "Stanford", "url": null, "code": "https://github.com/StanfordVL/BEHAVIOR-1K", "project": null, "topics": [ "sim", "datasets" ], "tags": [ "sim", "household", "long-horizon", "benchmark" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "1,000 household activities chosen from a survey of what people want robots to do, defined in a logic language and simulated in OmniGibson with realistic physics and rendering.", "takeaway_zh": "依据人们希望机器人做什么的调查挑选出 1000 项家务活动，用逻辑语言定义，并在具备真实物理与渲染的 OmniGibson 中仿真。", "github_stars": 1784, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2403.09227.html", "labs": [ "stanford" ] }, { "id": "arxiv-2511.04831", "title": "Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning", "type": "paper", "arxiv": "2511.04831", "year": 2025, "venue": null, "org": "NVIDIA", "url": null, "code": "https://github.com/isaac-sim/IsaacLab", "project": null, "topics": [ "sim" ], "tags": [ "gpu-parallel", "framework", "sensors", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The reference paper for Isaac Lab, Orbit's successor — GPU-parallel physics, rendering and sensor simulation with ready-made robots and tasks for RL and imitation.", "takeaway_zh": "Isaac Lab（Orbit 的后继）的正式论文：GPU 并行的物理、渲染与传感器仿真，内置机器人与任务，面向强化学习和模仿学习。", "github_stars": 8515, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2511.04831.html", "labs": [ "nvidia" ] }, { "id": "arxiv-1712.05474", "title": "AI2-THOR: An Interactive 3D Environment for Visual AI", "type": "paper", "arxiv": "1712.05474", "year": 2017, "venue": null, "org": "Allen Institute for AI", "url": null, "code": "https://github.com/allenai/ai2thor", "project": null, "topics": [ "sim" ], "tags": [ "household", "interactive", "embodied-agents", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Photorealistic interactive indoor scenes where agents open, pick up and change objects; the base of many embodied-AI benchmarks such as ALFRED.", "takeaway_zh": "具有可交互物体的逼真室内场景，智能体可开关、拾取和改变物体，是 ALFRED 等众多具身智能基准的基础。", "github_stars": 1821, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1712.05474.html", "labs": [ "ai2" ] }, { "id": "arxiv-2310.13724", "title": "Habitat 3.0: A Co-Habitat for Humans, Avatars and Robots", "type": "paper", "arxiv": "2310.13724", "year": 2023, "venue": "ICLR 2024", "org": "Meta AI", "url": null, "code": "https://github.com/facebookresearch/habitat-lab", "project": null, "topics": [ "sim", "navigation" ], "tags": [ "sim", "human-robot-collaboration", "humanoid-avatars", "social-navigation" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Adds simulated humanoid avatars and a human-in-the-loop interface to Habitat, so robots can be trained and evaluated on tasks shared with people, such as social navigation and social rearrangement.", "takeaway_zh": "为 Habitat 加入仿真人形化身和真人在环交互接口，可在与人协作的任务（如社交导航、社交重排）上训练和评测机器人。", "github_stars": 3174, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2310.13724.html", "labs": [ "meta" ] }, { "id": "arxiv-2610.28950", "title": "HyperDrone: Simulation-to-Reality Transfer at Millions of Frames per Second", "type": "paper", "arxiv": "2610.28950", "year": 2026, "venue": "CoRL 2026", "org": null, "url": null, "code": null, "project": null, "topics": [ "sim" ], "tags": [ "aerial", "quadrotor", "simulator", "gpu", "rendering", "sim-to-real" ], "robot_types": [ "aerial" ], "highlight": false, "added": "2026-10-22", "takeaway_en": "GPU batch-rendering quadrotor simulator; reports 20×+ the speed of existing batch quadrotor sims, with a new ray tracer rendering RGB, depth, segmentation, normals and optical flow (CoRL 2026).", "takeaway_zh": "以 GPU 批量渲染为核心的四旋翼仿真器；报告比现有批量四旋翼仿真器快 20 倍以上，新的光线追踪引擎（支持 OptiX、Vulkan、Metal、WebGPU）可渲染 RGB、深度、分割、法向和光流（CoRL 2026）。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.28950.html", "labs": [] }, { "id": "arxiv-2106.13281", "title": "Brax: A Differentiable Physics Engine for Large Scale Rigid Body Simulation", "type": "paper", "arxiv": "2106.13281", "year": 2021, "venue": "NeurIPS 2021 Datasets and Benchmarks", "org": "Google", "url": null, "code": "https://github.com/google/brax", "project": null, "topics": [ "sim", "learning" ], "tags": [ "physics", "jax", "gpu", "tpu", "differentiable" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "JAX physics engine that runs thousands of environments in parallel on GPU/TPU together with the RL training loop, cutting locomotion training from hours to minutes.", "takeaway_zh": "基于 JAX 的物理引擎，在 GPU/TPU 上与强化学习训练循环一起并行运行数千个环境，把运动控制训练从数小时缩短到数分钟。", "github_stars": 3252, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2106.13281.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2106.14405", "title": "Habitat 2.0: Training Home Assistants to Rearrange their Habitat", "type": "paper", "arxiv": "2106.14405", "year": 2021, "venue": "NeurIPS 2021", "org": "Meta AI", "url": null, "code": null, "project": null, "topics": [ "sim", "navigation", "manipulation" ], "tags": [ "rearrangement", "mobile-manipulation", "home", "benchmark" ], "robot_types": [ "arm", "mobile" ], "highlight": false, "added": null, "takeaway_en": "Adds fast rigid-body physics and an interactive apartment dataset (ReplicaCAD) to Habitat, plus a home-rearrangement benchmark for mobile manipulators.", "takeaway_zh": "为 Habitat 加入快速刚体物理与可交互公寓数据集 ReplicaCAD，并提出面向移动操作机器人的家居物品重排基准。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2106.14405.html", "labs": [ "meta" ] }, { "id": "arxiv-2404.05695", "title": "Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer", "type": "paper", "arxiv": "2404.05695", "year": 2024, "venue": null, "org": "RobotEra", "url": null, "code": "https://github.com/roboterax/humanoid-gym", "project": null, "topics": [ "sim", "locomotion" ], "tags": [ "humanoid", "rl", "isaac-gym", "sim-to-sim", "sim-to-real" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Isaac Gym–based RL framework for humanoid locomotion with a sim-to-sim check in MuJoCo before deploying zero-shot to real humanoids; a common starting point for humanoid RL.", "takeaway_zh": "基于 Isaac Gym 的人形机器人运动强化学习框架，先在 MuJoCo 中做 sim2sim 验证，再零样本部署到真实人形机器人，是人形 RL 的常用起点。", "github_stars": 2110, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2404.05695.html", "labs": [] }, { "id": "gazebo", "title": "Gazebo", "type": "simulator", "arxiv": null, "year": 2004, "venue": null, "org": "Open Robotics", "url": "https://gazebosim.org/", "code": "https://github.com/gazebosim/gz-sim", "project": null, "topics": [ "sim" ], "tags": [ "ros", "ros2", "physics", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The long-standing open-source robot simulator of the ROS ecosystem — sensors, plugins and ROS 2 integration; the usual choice for testing full robot software stacks rather than training policies.", "takeaway_zh": "ROS 生态中历史悠久的开源机器人仿真器，提供传感器、插件与 ROS 2 集成；通常用于测试完整的机器人软件栈，而非训练策略。", "github_stars": 1556, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/gazebo.html", "labs": [] }, { "id": "arxiv-2006.12983", "title": "dm_control: Software and Tasks for Continuous Control", "type": "paper", "arxiv": "2006.12983", "year": 2020, "venue": null, "org": "DeepMind", "url": null, "code": "https://github.com/google-deepmind/dm_control", "project": null, "topics": [ "sim", "learning" ], "tags": [ "mujoco", "benchmark", "rl", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "DeepMind's Python stack on top of MuJoCo — bindings, PyMJCF and Composer for building models and tasks, the standard Control Suite benchmarks, plus locomotion and arm-and-bricks manipulation tasks.", "takeaway_zh": "DeepMind 基于 MuJoCo 的 Python 工具栈：底层绑定、用于程序化建模与编写任务的 PyMJCF 和 Composer、作为标准基准的 Control Suite，以及运动控制和机械臂拼积木操作任务。", "github_stars": 4719, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2006.12983.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2011.07215", "title": "SoftGym: Benchmarking Deep Reinforcement Learning for Deformable Object Manipulation", "type": "paper", "arxiv": "2011.07215", "year": 2020, "venue": "CoRL 2020", "org": "Carnegie Mellon University", "url": null, "code": "https://github.com/Xingyu-Lin/softgym", "project": "https://sites.google.com/view/softgym", "topics": [ "sim", "manipulation" ], "tags": [ "deformable", "cloth", "rope", "liquid", "benchmark", "rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open-source Gym-style benchmark for manipulating deformable objects such as cloth, rope and liquids; shows that high-dimensional, partially observable states make these tasks hard for standard RL.", "takeaway_zh": "开源、Gym 接口的可变形物体操作基准（布料、绳子、液体等）；实验表明高维且部分可观测的状态让这类任务对常规强化学习方法很困难。", "github_stars": 356, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2011.07215.html", "labs": [ "cmu" ] }, { "id": "newton", "title": "Newton", "type": "simulator", "arxiv": null, "year": 2025, "venue": null, "org": "Disney Research, Google DeepMind, NVIDIA (Linux Foundation project)", "url": null, "code": "https://github.com/newton-physics/newton", "project": null, "topics": [ "sim" ], "tags": [ "physics", "gpu-parallel", "open-source", "warp", "openusd" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Open, GPU-accelerated physics engine built on NVIDIA Warp and OpenUSD with multiple solvers for contact-rich robot learning; co-developed by Disney Research, Google DeepMind and NVIDIA, hosted by the Linux Foundation.", "takeaway_zh": "基于 NVIDIA Warp 和 OpenUSD 的开源 GPU 物理引擎，支持多种求解器，面向接触丰富的机器人学习；由 Disney Research、Google DeepMind 和 NVIDIA 联合开发，由 Linux 基金会托管。", "github_stars": 5865, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/newton.html", "labs": [ "nvidia", "google-deepmind" ] }, { "id": "arxiv-2410.24185", "title": "DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning", "type": "paper", "arxiv": "2410.24185", "year": 2024, "venue": "ICRA 2025", "org": "NVIDIA, UT Austin", "url": null, "code": "https://github.com/NVlabs/dexmimicgen", "project": "https://dexmimicgen.github.io/", "topics": [ "sim", "manipulation", "datasets" ], "tags": [ "sim", "humanoid", "dexterous-hand", "bimanual", "data-generation", "imitation" ], "robot_types": [ "arm", "hand", "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Extends MimicGen to humanoids with two dexterous hands, turning 60 human demos into 21K simulated ones across bimanual tasks, plus a real-to-sim-to-real pipeline tested on real can sorting.", "takeaway_zh": "把 MimicGen 扩展到双灵巧手人形机器人，从 60 条人类演示生成 2.1 万条双臂仿真演示，并给出在真实分拣罐子任务上验证的 real-to-sim-to-real 流程。", "github_stars": 297, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2410.24185.html", "labs": [ "nvidia", "ut-austin" ] }, { "id": "arxiv-2610.32978", "title": "SceneBloom: Agentic Indoor Environment Generation with Realistic and Diverse Layouts for Embodied Learning", "type": "paper", "arxiv": "2610.32978", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/Tinaliu0123/scenebloom", "project": null, "topics": [ "sim", "navigation" ], "tags": [ "scene-generation", "diffusion", "simulation-ready", "object-navigation" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Generates sim-ready rooms by sampling furniture layouts with diffusion priors, then compiling them into collision-free 3D scenes; nav policies trained there beat ones trained on real scans on held-out real scans.", "takeaway_zh": "先用扩散先验在功能分区内采样家具布局，再“编译”为无碰撞、可直接仿真的 3D 场景；论文报告在其场景中训练的物体导航策略，在留出的真实扫描场景上优于用等量真实扫描训练的策略。", "github_stars": 15, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.32978.html", "labs": [] }, { "id": "arxiv-2506.10966", "title": "GenManip: LLM-driven Simulation for Generalizable Instruction-Following Manipulation", "type": "paper", "arxiv": "2506.10966", "year": 2025, "venue": "CVPR 2025", "org": "Shanghai AI Lab", "url": null, "code": "https://github.com/InternRobotics/GenManip", "project": null, "topics": [ "sim", "manipulation", "datasets" ], "tags": [ "sim", "isaac-sim", "benchmark", "llm", "generalization" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Isaac Sim platform where an LLM builds tasks and layouts from a task-oriented scene graph over ~10K objects; its 200-scenario benchmark finds modular foundation-model systems generalize better than end-to-end policies.", "takeaway_zh": "基于 Isaac Sim 的平台：LLM 依据任务导向场景图，从约 1 万个物体中生成任务与布局；其 200 个场景的基准发现，结合基础模型的模块化系统比端到端策略泛化更好。", "github_stars": 177, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2506.10966.html", "labs": [ "shanghai-ai-lab" ] }, { "id": "arxiv-2610.28465", "title": "MARE: An Open-Source Lunar Rover Simulator with Shadow Ground Truth for Perception and Autonomy", "type": "paper", "arxiv": "2610.28465", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/raulcastar/MARE-Lunar-Simulator", "project": null, "topics": [ "sim", "perception", "navigation" ], "tags": [ "sim", "space", "rover", "unreal-engine", "ros2", "synthetic-data" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Unreal Engine 5 + ROS 2 lunar rover simulator: generated cratered terrain, two rovers with stereo, lidar, event and SPAD cameras, pixel-accurate shadow ground truth for low-sun polar lighting, and a GNC stack.", "takeaway_zh": "基于 Unreal Engine 5 与 ROS 2 的开源月球车仿真：可生成带陨石坑的地形，两台月球车配备双目、激光雷达、事件相机与 SPAD 相机，提供极区低太阳角下像素级的阴影真值，并附 GNC 导航控制栈。", "github_stars": 8, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.28465.html", "labs": [] }, { "id": "arxiv-2310.08864", "title": "Open X-Embodiment: Robotic Learning Datasets and RT-X Models", "type": "dataset", "arxiv": "2310.08864", "year": 2023, "venue": null, "org": "Open X-Embodiment Collaboration", "url": null, "code": "https://github.com/google-deepmind/open_x_embodiment", "project": null, "topics": [ "datasets", "vla" ], "tags": [ "real", "cross-embodiment", "multi-robot" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "1M+ trajectories from 22 robot types pooled by 30+ labs; RT-X models trained on it beat single-robot baselines.", "takeaway_zh": "30 多个实验室汇集 22 种机器人的 100 多万条轨迹；在其上训练的 RT-X 优于单机器人基线。", "github_stars": 2086, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2310.08864.html", "labs": [] }, { "id": "arxiv-2403.12945", "title": "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset", "type": "dataset", "arxiv": "2403.12945", "year": 2024, "venue": null, "org": "DROID Collaboration", "url": null, "code": "https://github.com/droid-dataset/droid", "project": null, "topics": [ "datasets" ], "tags": [ "real", "arm", "in-the-wild" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "76k teleoperated Franka trajectories collected in hundreds of real scenes across many institutions.", "takeaway_zh": "由多家机构在数百个真实场景中采集的 7.6 万条 Franka 遥操作轨迹。", "github_stars": 461, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2403.12945.html", "labs": [] }, { "id": "arxiv-2308.12952", "title": "BridgeData V2: A Dataset for Robot Learning at Scale", "type": "dataset", "arxiv": "2308.12952", "year": 2023, "venue": null, "org": "UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "datasets" ], "tags": [ "real", "arm", "low-cost" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "60k trajectories on a low-cost WidowX arm across many kitchens and tasks; a common pretraining source.", "takeaway_zh": "在低成本 WidowX 机械臂上采集、涵盖众多厨房与任务的 6 万条轨迹，常用作预训练数据。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2308.12952.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2307.00595", "title": "RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot", "type": "dataset", "arxiv": "2307.00595", "year": 2023, "venue": null, "org": "Shanghai Jiao Tong University", "url": null, "code": null, "project": null, "topics": [ "datasets" ], "tags": [ "real", "arm", "multimodal", "force" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "110k+ contact-rich sequences with vision, force, audio and paired human demonstration videos.", "takeaway_zh": "11 万多条接触丰富的序列，包含视觉、力、音频及配对的人类演示视频。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2307.00595.html", "labs": [ "sjtu" ] }, { "id": "arxiv-2503.06669", "title": "AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems", "type": "dataset", "arxiv": "2503.06669", "year": 2025, "venue": null, "org": "AgiBot, Shanghai AI Lab", "url": null, "code": null, "project": null, "topics": [ "datasets" ], "tags": [ "real", "humanoid", "dual-arm", "large-scale" ], "robot_types": [ "arm", "humanoid" ], "highlight": false, "added": null, "takeaway_en": "A million-scale real-robot dataset collected in purpose-built scenes, with the GO-1 generalist policy trained on it.", "takeaway_zh": "在专门搭建的场景中采集的百万级真实机器人数据集，并基于其训练通用策略 GO-1。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2503.06669.html", "labs": [ "shanghai-ai-lab", "agibot" ] }, { "id": "arxiv-2110.07058", "title": "Ego4D: Around the World in 3,000 Hours of Egocentric Video", "type": "dataset", "arxiv": "2110.07058", "year": 2021, "venue": null, "org": "Meta AI and university partners", "url": null, "code": null, "project": null, "topics": [ "datasets" ], "tags": [ "egocentric-video", "human-data" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "3,670 hours of first-person daily-life video; widely used to pretrain visual representations for robots.", "takeaway_zh": "3670 小时第一视角日常生活视频，广泛用于为机器人预训练视觉表征。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2110.07058.html", "labs": [ "meta" ] }, { "id": "arxiv-2306.03310", "title": "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning", "type": "benchmark", "arxiv": "2306.03310", "year": 2023, "venue": null, "org": "UT Austin", "url": null, "code": "https://github.com/Lifelong-Robot-Learning/LIBERO", "project": null, "topics": [ "datasets" ], "tags": [ "sim", "manipulation", "lifelong-learning" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "130 language-conditioned tasks in four suites; now the most common simulation benchmark for VLA papers.", "takeaway_zh": "四个套件共 130 个语言条件任务，现已成为 VLA 论文最常用的仿真基准。", "github_stars": 2491, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2306.03310.html", "labs": [ "ut-austin" ] }, { "id": "arxiv-2112.03227", "title": "CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks", "type": "benchmark", "arxiv": "2112.03227", "year": 2021, "venue": null, "org": "University of Freiburg", "url": null, "code": "https://github.com/mees/calvin", "project": null, "topics": [ "datasets" ], "tags": [ "sim", "manipulation", "long-horizon" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Chains of five language instructions in a row test long-horizon, language-conditioned control.", "takeaway_zh": "连续执行五条语言指令，测试长时程、语言条件下的控制能力。", "github_stars": 999, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2112.03227.html", "labs": [ "freiburg" ] }, { "id": "arxiv-1909.12271", "title": "RLBench: The Robot Learning Benchmark & Learning Environment", "type": "benchmark", "arxiv": "1909.12271", "year": 2019, "venue": null, "org": "Imperial College London", "url": null, "code": "https://github.com/stepjam/RLBench", "project": null, "topics": [ "datasets" ], "tags": [ "sim", "manipulation" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "100 hand-designed manipulation tasks with demos; a staple for 3D policy methods like PerAct.", "takeaway_zh": "100 个手工设计的操作任务并附演示，是 PerAct 等 3D 策略方法的常用基准。", "github_stars": 1834, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1909.12271.html", "labs": [] }, { "id": "arxiv-1910.10897", "title": "Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning", "type": "benchmark", "arxiv": "1910.10897", "year": 2019, "venue": null, "org": "Stanford, UC Berkeley", "url": null, "code": "https://github.com/Farama-Foundation/Metaworld", "project": null, "topics": [ "datasets" ], "tags": [ "sim", "manipulation", "rl", "multi-task" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "50 Sawyer manipulation tasks for multi-task and meta-RL evaluation.", "takeaway_zh": "50 个 Sawyer 操作任务，用于多任务与元强化学习评测。", "github_stars": 1904, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1910.10897.html", "labs": [ "berkeley", "stanford" ] }, { "id": "arxiv-2405.05941", "title": "Evaluating Real-World Robot Manipulation Policies in Simulation", "type": "benchmark", "arxiv": "2405.05941", "year": 2024, "venue": null, "org": "UC San Diego, Stanford, UC Berkeley, Google DeepMind", "url": null, "code": "https://github.com/simpler-env/SimplerEnv", "project": null, "topics": [ "datasets", "vla" ], "tags": [ "sim", "real-to-sim", "evaluation" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "SimplerEnv — real-to-sim environments whose scores correlate with real-robot results for RT-1/Octo-style policies.", "takeaway_zh": "SimplerEnv：真实到仿真的评测环境，其得分与 RT-1/Octo 类策略的真机表现相关。", "github_stars": 1181, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2405.05941.html", "labs": [ "berkeley", "stanford", "google-deepmind", "ucsd" ] }, { "id": "arxiv-1904.01201", "title": "Habitat: A Platform for Embodied AI Research", "type": "paper", "arxiv": "1904.01201", "year": 2019, "venue": null, "org": "Meta AI", "url": null, "code": "https://github.com/facebookresearch/habitat-sim", "project": null, "topics": [ "datasets", "sim", "navigation" ], "tags": [ "sim", "navigation", "photorealistic" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "A very fast photorealistic 3D simulator and task suite that standardized embodied navigation research.", "takeaway_zh": "极速的逼真 3D 仿真器与任务集，统一了具身导航研究的标准。", "github_stars": 3852, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1904.01201.html", "labs": [ "meta" ] }, { "id": "lerobot", "title": "LeRobot", "type": "code", "arxiv": null, "year": 2024, "venue": null, "org": "Hugging Face", "url": null, "code": "https://github.com/huggingface/lerobot", "project": null, "topics": [ "datasets", "hardware", "manipulation" ], "tags": [ "library", "open-source", "low-cost", "datasets" ], "robot_types": [], "highlight": true, "added": null, "takeaway_en": "PyTorch library with datasets, pretrained policies (ACT, Diffusion, SmolVLA) and drivers for low-cost arms.", "takeaway_zh": "PyTorch 机器人学习库，提供数据集、预训练策略（ACT、Diffusion、SmolVLA）与低成本机械臂驱动。", "github_stars": 28561, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/lerobot.html", "labs": [ "hugging-face" ] }, { "id": "arxiv-2610.16238", "title": "Embodied Skill Bench: Benchmarking How Well Agents with Embodied Skills Perform on Physical Tasks", "type": "benchmark", "arxiv": "2610.16238", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/HanZhi306/EmbodiedSkill-Bench", "project": null, "topics": [ "datasets", "vla", "navigation" ], "tags": [ "llm-agents", "skills", "closed-loop" ], "robot_types": [], "highlight": false, "added": "2026-10-19", "takeaway_en": "Packages VLA policies, perception models and know-how as 250 reusable \"embodied skills\" and measures how much they help LLM agents on physical tasks.", "takeaway_zh": "将 VLA 策略、感知模型与经验知识封装成 250 个可复用的“具身技能”，评测它们对大模型智能体完成物理任务的帮助。", "github_stars": 3, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.16238.html", "labs": [] }, { "id": "arxiv-2412.13877", "title": "RoboMIND: Benchmark on Multi-embodiment Intelligence Normative Data for Robot Manipulation", "type": "dataset", "arxiv": "2412.13877", "year": 2024, "venue": null, "org": "Beijing Humanoid Robot Innovation Center, PKU", "url": null, "code": null, "project": "https://x-humanoid-robomind.github.io/", "topics": [ "datasets" ], "tags": [ "real", "multi-embodiment", "humanoid", "teleoperation" ], "robot_types": [ "arm", "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Large real-robot teleoperation dataset across several embodiments (single-arm, dual-arm, humanoid) collected under one standardized protocol, and it keeps failed demos with annotated causes.", "takeaway_zh": "在统一规范下采集的大规模真机遥操作数据集，覆盖单臂、双臂与人形等多种本体，并保留了标注失败原因的失败演示。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2412.13877.html", "labs": [ "pku" ] }, { "id": "arxiv-2610.27861", "title": "RoboCasa-Shift: Stage-Aware Robustness Evaluation of Vision-Language-Action Policies in Long-Horizon Household Manipulation", "type": "benchmark", "arxiv": "2610.27861", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/HaochengWang23/RoboCasa-Shift", "project": null, "topics": [ "datasets", "vla", "sim" ], "tags": [ "benchmark", "robustness", "long-horizon", "household", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-21", "takeaway_en": "VLA robustness benchmark on RoboCasa365: paired clean/perturbed rollouts differing in one shift, plus stage-level diagnostics of where long-horizon failures start; evaluated on GR00T N1.5.", "takeaway_zh": "基于 RoboCasa365 的 VLA 鲁棒性诊断基准：成对比较只差一个分布偏移因素的干净/扰动执行，并定位长时序任务中失败从哪个阶段开始，而不只看总成功率；在 GR00T N1.5 上评测。", "github_stars": 2, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.27861.html", "labs": [] }, { "id": "arxiv-2109.08238", "title": "Habitat-Matterport 3D Dataset (HM3D): 1000 Large-scale 3D Environments for Embodied AI", "type": "dataset", "arxiv": "2109.08238", "year": 2021, "venue": "NeurIPS 2021", "org": "Meta AI, Matterport", "url": "https://aihabitat.org/datasets/hm3d/", "code": "https://github.com/facebookresearch/habitat-matterport3d-dataset", "project": null, "topics": [ "datasets", "navigation", "sim" ], "tags": [ "3d-scans", "indoor", "habitat" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "1,000 3D scans of real indoor spaces for Habitat; the main scene set behind ObjectNav and many embodied navigation benchmarks.", "takeaway_zh": "1000 个真实室内空间的 3D 扫描，供 Habitat 使用；ObjectNav 等众多具身导航基准的主要场景集。", "github_stars": 216, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2109.08238.html", "labs": [ "meta" ] }, { "id": "arxiv-1910.11215", "title": "RoboNet: Large-Scale Multi-Robot Learning", "type": "dataset", "arxiv": "1910.11215", "year": 2019, "venue": "CoRL 2019", "org": "UC Berkeley, Stanford, UPenn", "url": null, "code": "https://github.com/SudeepDasari/RoboNet", "project": null, "topics": [ "datasets", "learning" ], "tags": [ "dataset", "real", "arm", "multi-embodiment", "video-prediction" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "An early shared dataset of robot interaction videos from seven robot platforms, used to pretrain visual dynamics models that transfer to new robots - a forerunner of Open X-Embodiment.", "takeaway_zh": "早期的共享机器人交互视频数据集，涵盖七种机器人平台，用于预训练可迁移到新机器人的视觉动力学模型，是 Open X-Embodiment 的先驱。", "github_stars": 178, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1910.11215.html", "labs": [ "berkeley", "stanford" ] }, { "id": "arxiv-2311.18259", "title": "Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives", "type": "dataset", "arxiv": "2311.18259", "year": 2023, "venue": "CVPR 2024", "org": "Meta AI and university partners", "url": null, "code": null, "project": null, "topics": [ "datasets" ], "tags": [ "dataset", "egocentric", "human-video", "multimodal" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Time-synced first-person and third-person video of people performing skilled activities (cooking, sports, repair...), with rich annotations; a large source for learning manipulation skills from humans.", "takeaway_zh": "人们进行技能活动（烹饪、运动、维修等）时同步采集的第一人称与第三人称视频，附丰富标注，是从人类视频学习操作技能的重要数据来源。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2311.18259.html", "labs": [ "meta" ] }, { "id": "arxiv-2212.08051", "title": "Objaverse: A Universe of Annotated 3D Objects", "type": "dataset", "arxiv": "2212.08051", "year": 2022, "venue": "CVPR 2023", "org": "Allen Institute for AI", "url": "https://huggingface.co/datasets/allenai/objaverse", "code": null, "project": null, "topics": [ "datasets", "sim" ], "tags": [ "dataset", "3d-assets", "objects" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "800K+ annotated 3D object models; widely used to populate simulators with diverse objects for grasping, rearrangement and procedural scene generation (Objaverse-XL extends it to 10M+).", "takeaway_zh": "80 万+ 带标注的 3D 物体模型，常用于为仿真器填充多样物体，用于抓取、重排和程序化场景生成（Objaverse-XL 扩展到 1000 万+）。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2212.08051.html", "labs": [ "ai2" ] }, { "id": "arxiv-1702.04405", "title": "ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes", "type": "dataset", "arxiv": "1702.04405", "year": 2017, "venue": "CVPR 2017", "org": "Stanford, Princeton, TUM", "url": null, "code": "https://github.com/ScanNet/ScanNet", "project": null, "topics": [ "datasets", "perception" ], "tags": [ "dataset", "rgbd", "3d-reconstruction", "semantic-segmentation", "indoor" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "RGB-D scans of 1,500+ indoor scenes with 3D reconstructions, camera poses and semantic labels; the standard benchmark for 3D scene understanding that embodied perception builds on.", "takeaway_zh": "1500+ 个室内场景的 RGB-D 扫描，含三维重建、相机位姿与语义标注；具身感知所依赖的三维场景理解标准基准。", "github_stars": 2360, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1702.04405.html", "labs": [ "stanford", "princeton" ] }, { "id": "arxiv-2403.10506", "title": "HumanoidBench: Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation", "type": "benchmark", "arxiv": "2403.10506", "year": 2024, "venue": "RSS 2024", "org": "UC Berkeley · Yonsei University", "url": null, "code": "https://github.com/carlosferrazza/humanoid-bench", "project": "https://humanoid-bench.github.io", "topics": [ "datasets", "locomotion", "sim" ], "tags": [ "sim", "humanoid", "whole-body", "rl-benchmark", "mujoco" ], "robot_types": [ "hand", "humanoid" ], "highlight": false, "added": null, "takeaway_en": "27 simulated tasks (12 locomotion, 15 whole-body manipulation) on a Unitree H1 with two dexterous hands; reports that DreamerV3, TD-MPC2, SAC and PPO fail most of them, especially long-horizon whole-body ones.", "takeaway_zh": "27 个仿真任务（12 个运动、15 个全身操作），机器人为带两只灵巧手的 Unitree H1；报告 DreamerV3、TD-MPC2、SAC、PPO 在大多数任务上未达标，长时程全身协调任务尤甚。", "github_stars": 802, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2403.10506.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2402.08191", "title": "THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulation", "type": "benchmark", "arxiv": "2402.08191", "year": 2024, "venue": "RSS 2024", "org": "UW · AI2 · USC · NVIDIA · UCSP", "url": null, "code": "https://github.com/robot-colosseum/robot-colosseum", "project": "https://robot-colosseum.github.io", "topics": [ "datasets", "manipulation" ], "tags": [ "sim", "generalization", "perturbations", "rlbench" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "20 RLBench tasks × 14 perturbation factors (color, texture, lighting, distractors, camera pose, friction, mass…); reports 30–50% success drops for leading policies, 75%+ when perturbations combine.", "takeaway_zh": "20 个 RLBench 任务 × 14 种扰动因素（颜色、纹理、光照、干扰物、相机位姿、摩擦、质量……）；报告主流策略成功率下降 30–50%，多种扰动叠加时下降 75% 以上。", "github_stars": 159, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2402.08191.html", "labs": [ "nvidia", "uw", "ai2", "usc" ] }, { "id": "arxiv-2104.04631", "title": "DexYCB: A Benchmark for Capturing Hand Grasping of Objects", "type": "dataset", "arxiv": "2104.04631", "year": 2021, "venue": "CVPR 2021", "org": "NVIDIA · University of Washington", "url": null, "code": "https://github.com/NVlabs/dex-ycb-toolkit", "project": null, "topics": [ "datasets", "manipulation", "perception" ], "tags": [ "real", "human-hands", "grasping", "hand-pose", "6d-pose", "handover" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "~582K RGB-D frames (1,000 sequences, 10 people, 20 YCB objects, 8 cameras) of human grasps with hand and object pose labels; benchmarks detection, 6D object pose, 3D hand pose and human-to-robot handover.", "takeaway_zh": "约 58.2 万帧 RGB-D（1000 段序列、10 人、20 个 YCB 物体、8 台相机）的人手抓取数据，带手部与物体位姿标注；评测检测、物体 6D 位姿、三维手部姿态及人到机器人的物体交接。", "github_stars": 214, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2104.04631.html", "labs": [ "nvidia", "uw" ] }, { "id": "arxiv-1502.03143", "title": "Benchmarking in Manipulation Research: The YCB Object and Model Set and Benchmarking Protocols", "type": "dataset", "arxiv": "1502.03143", "year": 2015, "venue": null, "org": "Yale · CMU · UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "datasets", "manipulation" ], "tags": [ "objects", "3d-models", "benchmark-protocols", "grasping" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "The YCB object set — 73 everyday objects with scans, textured meshes and simulation models, plus benchmarking protocols; widely used in grasping and pose-estimation work.", "takeaway_zh": "YCB 物体集：73 个日常物体，附扫描数据、带纹理网格与仿真模型，以及基准测试规程；被抓取与位姿估计研究广泛使用。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1502.03143.html", "labs": [ "berkeley", "cmu" ] }, { "id": "arxiv-2006.13256", "title": "Rescaling Egocentric Vision", "type": "dataset", "arxiv": "2006.13256", "year": 2020, "venue": "IJCV 2022", "org": "University of Bristol · University of Catania · University of Toronto", "url": null, "code": "https://github.com/epic-kitchens/epic-kitchens-100-annotations", "project": null, "topics": [ "datasets", "perception" ], "tags": [ "egocentric-video", "human-data", "kitchens", "action-recognition" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "EPIC-KITCHENS-100 — 100 hours (20M frames, ~90K action segments) of unscripted first-person kitchen video from 45 environments, with dense action labels and challenges such as action anticipation.", "takeaway_zh": "EPIC-KITCHENS-100：来自 45 个厨房环境、100 小时（2000 万帧、约 9 万个动作片段）的非脚本第一视角视频，动作标注密集，并设有动作预测等挑战任务。", "github_stars": 173, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2006.13256.html", "labs": [ "toronto" ] }, { "id": "arxiv-2203.01577", "title": "HOI4D: A 4D Egocentric Dataset for Category-Level Human-Object Interaction", "type": "dataset", "arxiv": "2203.01577", "year": 2022, "venue": "CVPR 2022", "org": "Tsinghua University · Shanghai Qi Zhi Institute", "url": null, "code": "https://github.com/leolyliu/HOI4D-Instructions", "project": null, "topics": [ "datasets", "manipulation", "perception" ], "tags": [ "egocentric-video", "human-data", "rgbd", "articulated-objects", "4d" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "2.4M egocentric RGB-D frames (4,000+ sequences, 800 object instances in 16 categories, 610 rooms) of people handling rigid and articulated objects, with segmentation, hand and object pose labels.", "takeaway_zh": "240 万帧第一视角 RGB-D（4000 多段序列、16 类 800 个物体实例、610 个房间），记录人操作刚体与铰接物体，带分割、手部与物体位姿标注。", "github_stars": 98, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2203.01577.html", "labs": [ "tsinghua", "qizhi" ] }, { "id": "arxiv-2203.15709", "title": "OakInk: A Large-scale Knowledge Repository for Understanding Hand-Object Interaction", "type": "dataset", "arxiv": "2203.15709", "year": 2022, "venue": "CVPR 2022", "org": "Shanghai Jiao Tong University", "url": null, "code": "https://github.com/oakink/OakInk", "project": null, "topics": [ "datasets", "manipulation" ], "tags": [ "human-hands", "grasping", "affordance", "mano" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "Affordance labels for ~1,800 household objects plus 50K intent-oriented hand-object interactions recorded on 100 of them; a common source for dexterous grasp synthesis. Follow-up OakInk2 covers bimanual tasks.", "takeaway_zh": "为约 1800 个日常物体标注可供性，并在其中 100 个物体上记录 5 万次带意图的手-物交互；常用于灵巧抓取生成。后续 OakInk2 扩展到双手任务。", "github_stars": 152, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2203.15709.html", "labs": [ "sjtu" ] }, { "id": "arxiv-2204.13662", "title": "ARCTIC: A Dataset for Dexterous Bimanual Hand-Object Manipulation", "type": "dataset", "arxiv": "2204.13662", "year": 2022, "venue": "CVPR 2023", "org": "ETH Zürich · MPI for Intelligent Systems", "url": null, "code": "https://github.com/zc-alexfan/arctic", "project": null, "topics": [ "datasets", "manipulation", "perception" ], "tags": [ "human-hands", "bimanual", "articulated-objects", "mocap", "mano" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "2.1M video frames of two hands manipulating articulated objects (scissors, laptops…) with accurate 3D hand and object meshes from motion capture; a reference for bimanual hand-object reconstruction.", "takeaway_zh": "210 万帧双手操作铰接物体（剪刀、笔记本电脑等）的视频，借助动作捕捉提供精确的手与物体三维网格；双手-物体重建的常用基准。", "github_stars": 525, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2204.13662.html", "labs": [ "eth" ] }, { "id": "arxiv-2412.18194", "title": "VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks", "type": "benchmark", "arxiv": "2412.18194", "year": 2024, "venue": null, "org": "Fudan University", "url": null, "code": "https://github.com/OpenMOSS/VLABench", "project": "https://vlabench.github.io/", "topics": [ "datasets", "vla", "manipulation" ], "tags": [ "sim", "language", "long-horizon", "reasoning", "vla-eval" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "100 task categories over 2,000+ objects that need common sense, world knowledge and multi-step reasoning, not just instruction following; evaluates VLAs, embodied agents and VLMs in one suite.", "takeaway_zh": "100 类任务、2000 多个物体，任务需要常识、世界知识与多步推理，而不只是照指令执行；在同一套件中评测 VLA、具身智能体与 VLM。", "github_stars": 482, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2412.18194.html", "labs": [ "fudan" ] }, { "id": "arxiv-2504.18904", "title": "RoboVerse: Towards a Unified Platform, Dataset and Benchmark for Scalable and Generalizable Robot Learning", "type": "benchmark", "arxiv": "2504.18904", "year": 2025, "venue": "RSS 2025", "org": "UC Berkeley · PKU · BIGAI · Stanford · CMU and others", "url": null, "code": "https://github.com/RoboVerseOrg/RoboVerse", "project": "https://roboverseorg.github.io/", "topics": [ "datasets", "sim" ], "tags": [ "sim", "multi-simulator", "synthetic-data", "imitation-learning", "rl" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "MetaSim puts many simulators (Isaac Sim/Gym, MuJoCo, Genesis, SAPIEN…) behind one interface; adds a large synthetic dataset migrated from public benchmarks and unified IL/RL benchmarks.", "takeaway_zh": "MetaSim 用统一接口封装多种仿真器（Isaac Sim/Gym、MuJoCo、Genesis、SAPIEN、PyBullet 等）；提供由公开基准迁移而来的大规模合成数据集，以及按泛化层级划分的统一模仿学习/强化学习基准。", "github_stars": 1877, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2504.18904.html", "labs": [ "berkeley", "stanford", "cmu", "pku" ] }, { "id": "arxiv-2610.18872", "title": "EgoExplorer: A Vision-Motion-Touch Observatory for Human Loco-Manipulation In Situ", "type": "dataset", "arxiv": "2610.18872", "year": 2026, "venue": null, "org": "InternRobotics", "url": null, "code": "https://github.com/InternRobotics/EgoExplorer", "project": "https://internrobotics.github.io/egoexplorer.github.io/", "topics": [ "datasets", "locomotion", "tactile", "manipulation" ], "tags": [ "real", "human-data", "egocentric", "wearable", "whole-body", "tactile", "humanoid" ], "robot_types": [ "humanoid" ], "highlight": false, "added": "2026-10-24", "takeaway_en": "100 h of wearable-captured human loco-manipulation (mostly professional work) aligning egocentric video, whole-body and finger motion, touch and step-level text; reports 71.9% lower body error than baselines.", "takeaway_zh": "用可穿戴设备采集的 100 小时人类移动操作数据，侧重职业场景，对齐第一视角视频、全身与手指动作、触觉和步骤级文字描述；其补全全身动作的方法报告身体关键点误差比最强基线低 71.9%。", "github_stars": 26, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.18872.html", "labs": [ "shanghai-ai-lab" ] }, { "id": "arxiv-2610.26765", "title": "HandsOn: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation", "type": "benchmark", "arxiv": "2610.26765", "year": 2026, "venue": null, "org": "Tsinghua University, BUPT, NatureWill, Peking University", "url": null, "code": "https://github.com/TsinghuaC3I/HandsOn", "project": "https://llltttwww.github.io/HandsOn-website/", "topics": [ "datasets", "manipulation", "sim" ], "tags": [ "sim", "dexterous", "cross-embodiment", "benchmark", "human-demos" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-24", "takeaway_en": "Isaac Sim benchmark with 5 arms, 6 dexterous hands, ~200 assets and 18k human demos over 66 tasks; reports the best trained policy (a WAM) at only 29.1% on Core-50, near 0% on precision-fit and deformable tasks.", "takeaway_zh": "基于 Isaac Sim 的灵巧操作基准：5 种机械臂、6 种灵巧手、约 200 个资产、1.8 万条人类演示、66 个任务；报告最好的训练策略（一种世界动作模型）在 Core-50 上仅 29.1%，精密装配和柔性物体任务接近 0%。", "github_stars": 53, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.26765.html", "labs": [ "tsinghua", "pku" ] }, { "id": "arxiv-2610.26663", "title": "RoboValue: A Fine-Grained Sim-and-Real Benchmark for Unified Evaluation of Robotic Value Models", "type": "benchmark", "arxiv": "2610.26663", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/RoboValue-Benchmark/RoboValue", "project": "https://robovalue-benchmark.github.io/", "topics": [ "datasets", "learning", "manipulation" ], "tags": [ "sim", "real", "benchmark", "value-model", "evaluation" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "Tests robot value (progress) models on 15 sim and 20 real tasks across task state, progress, failure/recovery and consistency; reports that good forward-progress scores hide weak history and recovery reasoning.", "takeaway_zh": "在 15 个仿真和 20 个真实任务上评测机器人价值（进度）模型，覆盖任务状态、进度监测、失败与恢复、一致性四个维度；报告称前向进度得分高的模型在依赖历史和恢复判断上仍然薄弱。", "github_stars": 17, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.26663.html", "labs": [] }, { "id": "arxiv-2610.16782", "title": "LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints", "type": "benchmark", "arxiv": "2610.16782", "year": 2026, "venue": null, "org": "Tongji University (IAIE Physical AI Group)", "url": null, "code": "https://github.com/IAIE-PhysicalAI/LIBERO-Industry", "project": "https://iaie-physicalai.github.io/LIBERO-Industry/", "topics": [ "datasets", "manipulation" ], "tags": [ "sim", "benchmark", "industrial", "conveyor", "libero" ], "robot_types": [], "highlight": false, "added": "2026-10-24", "takeaway_en": "216 LIBERO-based tasks with industrial constraints — conveyor loading/unloading/transfer, speed, time windows, tight tolerances; reports that VLAs and world action models degrade as constraints are combined.", "takeaway_zh": "基于 LIBERO 的 216 个工业约束任务：传送带上料、下料与转运，可调节物体形状、带速、交互时间窗与公差；报告 VLA 和世界动作模型在多重约束叠加时性能逐步下降。", "github_stars": 3, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.16782.html", "labs": [] }, { "id": "arxiv-2505.11709", "title": "EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video", "type": "dataset", "arxiv": "2505.11709", "year": 2025, "venue": "ICLR 2026", "org": "Apple", "url": null, "code": "https://github.com/apple-aiml-research/ml-egodex", "project": null, "topics": [ "datasets", "manipulation" ], "tags": [ "real", "human-video", "egocentric", "dexterous-hand", "hand-pose" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "829 hours of egocentric Apple Vision Pro video with tracked 3D poses for every hand joint, across 194 tabletop tasks from tying shoelaces to folding laundry; includes hand-trajectory prediction benchmarks.", "takeaway_zh": "用 Apple Vision Pro 采集的 829 小时第一视角视频，配有每个手部关节的三维姿态，覆盖从系鞋带到叠衣服的 194 种桌面任务，并提供手部轨迹预测基准。", "github_stars": 404, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2505.11709.html", "labs": [] }, { "id": "arxiv-2610.31779", "title": "FLORA: An Open Model Suite for Scaling Robotics Data Engines", "type": "paper", "arxiv": "2610.31779", "year": 2026, "venue": null, "org": "NVIDIA", "url": null, "code": "https://github.com/NVIDIA/Flora", "project": "https://research.nvidia.com/labs/gear/flora", "topics": [ "datasets", "learning" ], "tags": [ "data-engine", "reward-model", "captioning", "human-video", "open-weights" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "The open data engine behind GR00T training: models that write timestamped action captions, dense process rewards and episode outcome rewards for human and robot video; reports SOTA on WGO-Bench, RoboReward and Robometer.", "takeaway_zh": "GR00T 训练所用的开放数据引擎：一组模型为人类与机器人视频生成带时间戳的动作描述、稠密过程奖励和整段结果奖励；报告在 WGO-Bench、RoboReward 与 Robometer 上达到最优。", "github_stars": 50, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31779.html", "labs": [ "nvidia" ] }, { "id": "arxiv-2610.25459", "title": "BiDex-100: A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks", "type": "paper", "arxiv": "2610.25459", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://BiDex-100.github.io/BiDex-100/", "topics": [ "datasets", "manipulation" ], "tags": [ "benchmark", "sim", "bimanual", "dexterous-hand", "teleoperation", "vla" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "106 sim tasks for two dexterous hands along five axes (generalization, instructions, few demos, long horizon, memory), each with an out-of-domain twin; 8,080 teleoperated demos; seven VLAs benchmarked.", "takeaway_zh": "双灵巧手仿真基准：106 个任务按泛化、指令遵循、少示教、长时程、记忆五个维度组织，每个任务配有域外变体；提供 8,080 条遥操作示教，并在统一协议下评测了七个 VLA 模型。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.25459.html", "labs": [] }, { "id": "arxiv-2610.31064", "title": "AstroBench: Benchmarking Free-Floating Capture and Space Robotic Manipulation", "type": "paper", "arxiv": "2610.31064", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/AstroUpLab/AstroBench", "project": "https://astroup.com/astrobench/", "topics": [ "datasets", "manipulation", "sim" ], "tags": [ "benchmark", "sim", "space", "free-floating", "arm" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Benchmark for catching drifting, tumbling objects inside a spacecraft with a fixed-base arm, plus storage and servicing tasks. Reports a frozen generalist policy captured only 2/200 static targets.", "takeaway_zh": "太空机器人操作基准：用固定基座机械臂抓取舱内漂浮、旋转的物体，另含样本存储与在轨维护任务。报告称所测的冻结通用策略在 200 个静止目标中只抓住 2 个。", "github_stars": 1, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.31064.html", "labs": [] }, { "id": "arxiv-2610.27039", "title": "Embodied Playground: A Long-Horizon Agentic Robotics Benchmark for Physical, Strategic, and Spatial Reasoning", "type": "benchmark", "arxiv": "2610.27039", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://winniechen2002.github.io/Agent-bench-website/", "topics": [ "datasets", "vla" ], "tags": [ "benchmark", "long-horizon", "llm-agents", "planning", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-26", "takeaway_en": "Benchmark of 410 long-horizon task environments (manipulation, puzzles, household) for foundation-model robot agents, with a curated EP-100 split and a shared protocol checking goals, rules and physical validity.", "takeaway_zh": "面向基础模型机器人智能体的长程任务基准：410 个任务环境（基础操作、序列谜题、多阶段家务），精选 EP-100 子集，统一协议同时检查目标完成、任务规则与物理有效性。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.27039.html", "labs": [] }, { "id": "arxiv-2610.34051", "title": "Behavior Spatial Benchmark: Evaluating Spatial Understanding for Robotics", "type": "paper", "arxiv": "2610.34051", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://arielchen07.github.io/behavior-spatial/", "topics": [ "datasets", "vla", "sim" ], "tags": [ "benchmark", "spatial-reasoning", "vlm", "instruction-following", "behavior-1k" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "BEHAVIOR-Spatial, built on BEHAVIOR-1K: spatial QA, point grounding and robot action tests. A baseline policy drops from 41.6% on held-out layouts to 16.1% when the requested target changes; oracle target cues help most.", "takeaway_zh": "基于 BEHAVIOR-1K 的 BEHAVIOR-Spatial 基准：包含空间问答、指点定位与机器人动作评测。基线策略在未见布局上成功率 41.6%，但在同一布局中更换目标物体指令时降至 16.1%；提供真值目标提示带来的提升最大。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.34051.html", "labs": [] }, { "id": "arxiv-2610.35385", "title": "EmbodiedLab: Benchmarking Embodied Agents That Act to Know", "type": "paper", "arxiv": "2610.35385", "year": 2026, "venue": null, "org": "HexaFuture", "url": null, "code": "https://github.com/HexaFuture/EmbodiedLab", "project": "https://hexafuture.ai/benchmarks/embodiedlab/", "topics": [ "datasets", "sim", "vla" ], "tags": [ "benchmark", "agents", "scientific-discovery", "simulated-lab", "frontier-models" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Benchmark where a robot must run physics experiments to find an unknown answer: 30 experiments, 5 disciplines, parameters resampled per instance. Reports even the best frontier agent succeeds under half the time.", "takeaway_zh": "“为求知而行动”的基准：机器人在物理仿真实验室中自行操作仪器、读数，测出每次随机重采样的未知参数。共 30 个实验、5 个学科；报告最强的前沿模型智能体在原生控制下成功率也不到一半，失败多源于与仪器交互。", "github_stars": 16, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.35385.html", "labs": [] }, { "id": "arxiv-2610.34649", "title": "RobotAgentBench: Towards Open-world Robot Autonomy", "type": "paper", "arxiv": "2610.34649", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://aetherlabsai.github.io/RobotAgentBench/", "topics": [ "datasets", "vla" ], "tags": [ "benchmark", "robot-agent", "household", "real-to-sim", "long-horizon", "sim" ], "robot_types": [], "highlight": false, "added": "2026-10-27", "takeaway_en": "Benchmark for open-world robot agents: 50 household layouts built real-to-sim, tasks from 90 atomic ones to 4 open-ended scenarios, scoring discovery, reasoning, manipulation, recovery and robustness.", "takeaway_zh": "面向开放世界机器人智能体的基准：通过真实到仿真流水线构建 50 个家居布局，任务从 90 个原子任务到 4 个开放场景，评测任务发现、上下文推理、操作、失败恢复与鲁棒性，可统一比较 VLA、WAM 与智能体架构。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.34649.html", "labs": [] }, { "id": "arxiv-2610.36420", "title": "SO-101 Bench: Accessible Spatial Manipulation", "type": "paper", "arxiv": "2610.36420", "year": 2026, "venue": null, "org": null, "url": null, "code": "https://github.com/5hadytru/so101_bench", "project": "https://so101-bench.github.io/", "topics": [ "datasets", "sim", "manipulation" ], "tags": [ "benchmark", "so-101", "low-cost", "isaac-lab", "digital-twin", "vla" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Isaac Lab digital-twin benchmark for language-conditioned spatial tasks on the low-cost SO-101 arm (138 objects, 916 episodes); sim ranks policies exactly as the real arm. Best VLAs ~61%, an LLM-agent harness 83.2%.", "takeaway_zh": "基于 Isaac Lab 数字孪生的低成本 SO-101 机械臂语言条件空间操作基准（138 个物体、916 个评测回合）；仿真中的策略排名与真机完全一致。最好的 VLA 约 61%，由 LLM 写代码调用抓取策略的方案达 83.2%。", "github_stars": 34, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36420.html", "labs": [] }, { "id": "arxiv-2610.27782", "title": "RoboJudge: Multimodal Language Models as Judges for Embodied Video Generation", "type": "paper", "arxiv": "2610.27782", "year": 2026, "venue": null, "org": "The Chinese University of Hong Kong", "url": null, "code": "https://github.com/Sphere-AI-Lab/RoboJudge", "project": "https://spherelab.ai/RoboJudge/", "topics": [ "datasets", "world-models" ], "tags": [ "benchmark", "video-generation", "mllm-judge", "physical-adherence", "human-annotation" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Benchmark of 800 human-rated robot/hand videos from 10 generators to test MLLM judges on physical adherence and instruction alignment, plus a 12K-video training set and RoboJudge-9B (Pearson 0.719 with humans).", "takeaway_zh": "评测多模态大模型能否评判具身视频生成：800 条人工标注视频（10 个生成模型，夹爪/人手/灵巧手），考察物理合理性与指令一致性；另提供 1.2 万条训练视频及 RoboJudge-9B（与人类评分 Pearson 0.719）。", "github_stars": 2, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.27782.html", "labs": [] }, { "id": "arxiv-2309.06440", "title": "LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning", "type": "hardware", "arxiv": "2309.06440", "year": 2023, "venue": null, "org": "CMU", "url": null, "code": null, "project": "https://leaphand.com/", "topics": [ "hardware", "manipulation" ], "tags": [ "hand", "low-cost", "open-source" ], "robot_types": [ "hand" ], "highlight": true, "added": null, "takeaway_en": "A 16-DoF dexterous hand buildable for about $2,000 from off-the-shelf parts.", "takeaway_zh": "16 自由度灵巧手，用现成零件约 2000 美元即可搭建。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.06440.html", "labs": [ "cmu" ] }, { "id": "arxiv-2309.13037", "title": "GELLO: A General, Low-Cost, and Intuitive Teleoperation Framework for Robot Manipulators", "type": "hardware", "arxiv": "2309.13037", "year": 2023, "venue": null, "org": "UC Berkeley", "url": null, "code": "https://github.com/wuphilipp/gello_software", "project": null, "topics": [ "hardware" ], "tags": [ "teleoperation", "low-cost", "3d-printed" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "3D-printed kinematic replicas of robot arms for cheap, intuitive joint-space teleoperation.", "takeaway_zh": "3D 打印的机械臂运动学复刻件，实现低成本、直观的关节空间遥操作。", "github_stars": 561, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.13037.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2407.01512", "title": "Open-TeleVision: Teleoperation with Immersive Active Visual Feedback", "type": "hardware", "arxiv": "2407.01512", "year": 2024, "venue": null, "org": "UC San Diego, MIT", "url": null, "code": "https://github.com/OpenTeleVision/TeleVision", "project": null, "topics": [ "hardware", "locomotion" ], "tags": [ "teleoperation", "vr", "humanoid" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "VR teleoperation with an actively steered stereo camera, used to collect humanoid manipulation data.", "takeaway_zh": "带主动转向立体相机的 VR 遥操作系统，用于采集人形机器人操作数据。", "github_stars": 1321, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2407.01512.html", "labs": [ "mit", "ucsd" ] }, { "id": "arxiv-2407.21781", "title": "Berkeley Humanoid: A Research Platform for Learning-based Control", "type": "hardware", "arxiv": "2407.21781", "year": 2024, "venue": null, "org": "UC Berkeley", "url": null, "code": null, "project": null, "topics": [ "hardware", "locomotion" ], "tags": [ "humanoid", "low-cost" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "A mid-size, low-cost humanoid designed for easy sim-to-real transfer of RL locomotion.", "takeaway_zh": "中等尺寸、低成本的人形机器人，专为强化学习运动策略的 sim-to-real 迁移设计。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2407.21781.html", "labs": [ "berkeley" ] }, { "id": "arxiv-2504.17249", "title": "Demonstrating Berkeley Humanoid Lite: An Open-source, Accessible, and Customizable 3D-printed Humanoid Robot", "type": "hardware", "arxiv": "2504.17249", "year": 2025, "venue": "RSS 2025", "org": "UC Berkeley", "url": null, "code": "https://github.com/HybridRobotics/Berkeley-Humanoid-Lite", "project": "https://lite.berkeley-humanoid.org", "topics": [ "hardware", "locomotion" ], "tags": [ "humanoid", "open-source", "3d-printed", "low-cost" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Open-source humanoid with 3D-printed modular gearboxes, buildable for under US$5,000; hardware designs, embedded code and training code are released.", "takeaway_zh": "开源人形机器人，采用 3D 打印的模块化减速器，在美国可用不到 5000 美元搭建；硬件设计、嵌入式代码与训练框架均已开源。", "github_stars": 1998, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2504.17249.html", "labs": [ "berkeley" ] }, { "id": "so-arm100-so-101", "title": "SO-ARM100 / SO-101", "type": "hardware", "arxiv": null, "year": 2024, "venue": null, "org": "The Robot Studio, Hugging Face", "url": null, "code": "https://github.com/TheRobotStudio/SO-ARM100", "project": null, "topics": [ "hardware" ], "tags": [ "arm", "low-cost", "3d-printed", "open-source" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "An open 3D-printed 6-DoF arm (leader + follower) costing roughly $100–200 per arm; the default LeRobot starter kit.", "takeaway_zh": "开源 3D 打印 6 自由度机械臂（主从臂），每只约 100–200 美元，是 LeRobot 的默认入门套件。", "github_stars": 7914, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/so-arm100-so-101.html", "labs": [ "hugging-face" ] }, { "id": "arxiv-2502.00893", "title": "ToddlerBot: Open-Source ML-Compatible Humanoid Platform for Loco-Manipulation", "type": "hardware", "arxiv": "2502.00893", "year": 2025, "venue": null, "org": "Stanford", "url": null, "code": null, "project": "https://toddlerbot.github.io/", "topics": [ "hardware", "locomotion" ], "tags": [ "humanoid", "low-cost", "open-source", "3d-printed" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "A small, 3D-printed open-source humanoid with a matching simulation model, built so that ML policies and teleoperated data transfer to the real robot.", "takeaway_zh": "3D 打印的小型开源人形机器人，配套仿真模型，便于机器学习策略与遥操作数据迁移到真机。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.00893.html", "labs": [ "stanford" ] }, { "id": "arxiv-2503.05652", "title": "BEHAVIOR Robot Suite: Streamlining Real-World Whole-Body Manipulation for Everyday Household Activities", "type": "hardware", "arxiv": "2503.05652", "year": 2025, "venue": null, "org": "Stanford", "url": null, "code": null, "project": "https://behavior-robot-suite.github.io/", "topics": [ "hardware", "manipulation" ], "tags": [ "mobile-manipulation", "whole-body", "teleoperation", "bimanual" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "A whole-body teleoperation interface plus a learning method for a wheeled bimanual robot doing long household tasks that need reach, balance and navigation.", "takeaway_zh": "面向轮式双臂机器人的全身遥操作接口与学习方法，用于需要伸展、平衡与导航的长程家务任务。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2503.05652.html", "labs": [ "stanford" ] }, { "id": "arxiv-2405.02292", "title": "ALOHA 2: An Enhanced Low-Cost Hardware for Bimanual Teleoperation", "type": "hardware", "arxiv": "2405.02292", "year": 2024, "venue": null, "org": "Google DeepMind", "url": null, "code": null, "project": "https://aloha-2.github.io/", "topics": [ "hardware", "manipulation" ], "tags": [ "bimanual", "teleoperation", "low-cost", "open-source" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "Upgraded ALOHA bimanual rig (stronger grippers, gravity compensation, sturdier frame) with open hardware designs and a MuJoCo model for simulation.", "takeaway_zh": "ALOHA 双臂平台的升级版（更强夹爪、重力补偿、更稳固的框架），开源硬件设计并提供 MuJoCo 仿真模型。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2405.02292.html", "labs": [ "google-deepmind" ] }, { "id": "arxiv-2109.10892", "title": "The Design of Stretch: A Compact, Lightweight Mobile Manipulator for Indoor Human Environments", "type": "hardware", "arxiv": "2109.10892", "year": 2021, "venue": null, "org": "Hello Robot", "url": null, "code": null, "project": null, "topics": [ "hardware", "navigation" ], "tags": [ "mobile-manipulator", "home", "assistive" ], "robot_types": [ "arm", "mobile" ], "highlight": false, "added": null, "takeaway_en": "The design of Hello Robot's Stretch — a slim, lightweight mobile manipulator with a telescoping arm on a lift, built for homes and used widely in home-robot research.", "takeaway_zh": "Hello Robot 的 Stretch 设计：在升降柱上装伸缩臂的轻巧移动操作机器人，面向家庭环境，被大量家庭机器人研究采用。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2109.10892.html", "labs": [] }, { "id": "arxiv-2504.04259", "title": "ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning", "type": "hardware", "arxiv": "2504.04259", "year": 2025, "venue": null, "org": "ETH Zurich", "url": null, "code": null, "project": "https://www.orcahand.com", "topics": [ "hardware", "manipulation", "tactile" ], "tags": [ "hand", "open-source", "tendon-driven", "low-cost" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "A 3D-printable, 17-DoF tendon-driven hand with fingertip tactile sensing, under CHF 2,000 in materials; design files and code are open, and the paper reports long continuous runs for teleoperation and learning.", "takeaway_zh": "可 3D 打印的 17 自由度腱驱动灵巧手，指尖带触觉传感，材料成本低于 2000 瑞士法郎；设计文件与代码开源，论文报告了可长时间连续运行用于遥操作与学习。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2504.04259.html", "labs": [ "eth" ] }, { "id": "arxiv-2502.13013", "title": "HOMIE: Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit", "type": "hardware", "arxiv": "2502.13013", "year": 2025, "venue": "RSS 2025", "org": "Shanghai AI Lab · CUHK", "url": null, "code": null, "project": "https://homietele.github.io/", "topics": [ "hardware", "locomotion" ], "tags": [ "teleoperation", "humanoid", "exoskeleton", "low-cost" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Open-source whole-body humanoid teleoperation — an arm exoskeleton plus motion-sensing gloves, with a foot pedal steering an RL walking controller; the cockpit hardware costs about US$500.", "takeaway_zh": "开源的人形机器人全身遥操作系统：手臂外骨骼加动作捕捉手套，脚踏板驱动强化学习行走控制器；整套座舱硬件约 500 美元。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.13013.html", "labs": [ "shanghai-ai-lab", "cuhk" ] }, { "id": "arxiv-2307.04577", "title": "AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System", "type": "hardware", "arxiv": "2307.04577", "year": 2023, "venue": "RSS 2023", "org": "UC San Diego, NVIDIA", "url": null, "code": null, "project": null, "topics": [ "hardware", "manipulation" ], "tags": [ "teleoperation", "hand", "vision-based", "retargeting" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "Camera-only teleoperation for arm + dexterous hand pairs - tracks the operator's hand and retargets it to many different robot hands, in sim or on real hardware.", "takeaway_zh": "仅用相机即可遥操作\"机械臂 + 灵巧手\"：跟踪操作者手部并重定向到多种不同的机器人手，可用于仿真或真机。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2307.04577.html", "labs": [ "nvidia", "ucsd" ] }, { "id": "arxiv-2309.14975", "title": "AirExo: Low-Cost Exoskeletons for Learning Whole-Arm Manipulation in the Wild", "type": "hardware", "arxiv": "2309.14975", "year": 2023, "venue": "ICRA 2024", "org": "Shanghai Jiao Tong University", "url": null, "code": "https://github.com/AirExo/collector", "project": null, "topics": [ "hardware", "manipulation" ], "tags": [ "exoskeleton", "teleoperation", "low-cost", "data-collection" ], "robot_types": [ "arm" ], "highlight": false, "added": null, "takeaway_en": "A portable dual-arm exoskeleton (about $600 in parts, per the paper) for joint-level teleop and cheap in-the-wild demos; reports few-minute teleop plus wild data rivals much longer teleop-only training.", "takeaway_zh": "便携式双臂外骨骼（论文称零件成本约 600 美元），可做关节级遥操作并低成本采集野外示教；报告称少量遥操作数据加野外数据即可达到甚至超过长时间纯遥操作训练的效果。", "github_stars": 50, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2309.14975.html", "labs": [ "sjtu" ] }, { "id": "arxiv-2503.03081", "title": "AirExo-2: Scaling up Generalizable Robotic Imitation Learning with Low-Cost Exoskeletons", "type": "hardware", "arxiv": "2503.03081", "year": 2025, "venue": "CoRL 2025", "org": "Shanghai Jiao Tong University", "url": null, "code": "https://github.com/AirExo/AirExo-2", "project": null, "topics": [ "hardware", "manipulation", "learning" ], "tags": [ "exoskeleton", "in-the-wild", "imitation-learning" ], "robot_types": [], "highlight": false, "added": null, "takeaway_en": "Exoskeleton demos are converted into pseudo-robot demonstrations; the paper reports its RISE-2 policy trained on them alone, with no robot data, performs comparably to one trained on teleoperated robot data.", "takeaway_zh": "将外骨骼示教转换为“伪机器人”示教；论文报告仅用这些数据（无任何机器人数据）训练的 RISE-2 策略，效果与用遥操作机器人数据训练的策略相当。", "github_stars": 44, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2503.03081.html", "labs": [ "sjtu" ] }, { "id": "arxiv-2407.03162", "title": "Bunny-VisionPro: Real-Time Bimanual Dexterous Teleoperation for Imitation Learning", "type": "hardware", "arxiv": "2407.03162", "year": 2024, "venue": null, "org": "UC San Diego, HKU", "url": null, "code": "https://github.com/Dingry/BunnyVisionPro", "project": null, "topics": [ "hardware", "manipulation" ], "tags": [ "teleoperation", "vr", "hand", "bimanual", "haptics" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "Apple Vision Pro teleoperation of two dexterous robot hands with real-time retargeting, collision and singularity avoidance, and low-cost vibration haptics for the operator.", "takeaway_zh": "用 Apple Vision Pro 实时遥操作双灵巧手，带实时重定向、碰撞与奇异点规避，并为操作者提供低成本振动触觉反馈。", "github_stars": 357, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2407.03162.html", "labs": [ "ucsd", "hku" ] }, { "id": "arxiv-2403.07869", "title": "TeleMoMa: A Modular and Versatile Teleoperation System for Mobile Manipulation", "type": "hardware", "arxiv": "2403.07869", "year": 2024, "venue": null, "org": "UT Austin, Stanford, Sony AI", "url": null, "code": "https://github.com/UT-Austin-RobIn/telemoma", "project": "https://robin-lab.cs.utexas.edu/telemoma-web/", "topics": [ "hardware", "manipulation" ], "tags": [ "teleoperation", "mobile-manipulation", "whole-body" ], "robot_types": [ "arm", "mobile" ], "highlight": false, "added": null, "takeaway_en": "One modular teleop interface that mixes RGB-D body tracking, VR controllers, keyboard and joystick for whole-body mobile manipulators (Tiago++, HSR, Fetch), in sim and real.", "takeaway_zh": "模块化遥操作接口，可组合 RGB-D 人体追踪、VR 手柄、键盘与摇杆，控制 Tiago++、HSR、Fetch 等移动操作机器人做全身动作，仿真和真机都支持。", "github_stars": 86, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2403.07869.html", "labs": [ "stanford", "ut-austin" ] }, { "id": "arxiv-1910.00093", "title": "An Open Torque-Controlled Modular Robot Architecture for Legged Locomotion Research", "type": "hardware", "arxiv": "1910.00093", "year": 2019, "venue": null, "org": "MPI-IS, NYU (Open Dynamic Robot Initiative)", "url": null, "code": "https://github.com/open-dynamic-robot-initiative/open_robot_actuator_hardware", "project": null, "topics": [ "hardware", "locomotion" ], "tags": [ "quadruped", "open-source", "torque-control", "low-cost" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "The open-source Solo quadruped and its modular torque-controlled actuators, built from off-the-shelf and 3D-printed parts so labs can reproduce a dynamic legged robot.", "takeaway_zh": "开源四足机器人 Solo 及其模块化力矩控制关节，采用现成件和 3D 打印件，便于各实验室复现动态腿式机器人。", "github_stars": 1509, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-1910.00093.html", "labs": [ "nyu" ] }, { "id": "arxiv-2505.21864", "title": "DexUMI: Using Human Hand as the Universal Manipulation Interface for Dexterous Manipulation", "type": "paper", "arxiv": "2505.21864", "year": 2025, "venue": "CoRL 2025", "org": "Stanford, Columbia, CMU, NVIDIA", "url": null, "code": "https://github.com/real-stanford/DexUMI", "project": "https://dex-umi.github.io/", "topics": [ "hardware", "manipulation" ], "tags": [ "real", "exoskeleton", "data-collection", "dexterous-hand", "inpainting" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "DexUMI — an exoskeleton matched to the robot hand's fingertip workspace, plus inpainting that swaps the human hand for the robot's; reports 86% average success (Inspire, XHand) and ~3.2× faster collection than teleop.", "takeaway_zh": "DexUMI：按目标机器人手的指尖工作空间设计的可穿戴手部外骨骼，再用视频修补把画面中的人手换成机器人手；论文报告在 Inspire 与 XHand 上平均成功率 86%，采集效率约为遥操作的 3.2 倍。", "github_stars": 270, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2505.21864.html", "labs": [ "stanford", "nvidia", "cmu", "columbia" ] }, { "id": "arxiv-2408.11805", "title": "ACE: A Cross-Platform Visual-Exoskeletons System for Low-Cost Dexterous Teleoperation", "type": "paper", "arxiv": "2408.11805", "year": 2024, "venue": "CoRL 2024", "org": null, "url": null, "code": null, "project": "https://ace-teleop.github.io/", "topics": [ "hardware" ], "tags": [ "real", "teleoperation", "exoskeleton", "low-cost", "humanoid" ], "robot_types": [ "arm", "hand", "humanoid", "legged" ], "highlight": false, "added": null, "takeaway_en": "ACE — a 3D-printed bimanual exoskeleton for wrist pose plus hand-facing webcams for finger pose (about $600); one setup teleoperates arm-hand systems, humanoids and quadruped-mounted arms.", "takeaway_zh": "ACE：3D 打印的双臂外骨骼获取手腕位姿、朝向手部的摄像头估计手指姿态（约 600 美元）；同一套设备可遥操作机械臂+灵巧手、人形机器人以及四足搭载的机械臂。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2408.11805.html", "labs": [] }, { "id": "arxiv-2404.16823", "title": "Learning Visuotactile Skills with Two Multifingered Hands", "type": "hardware", "arxiv": "2404.16823", "year": 2024, "venue": "ICRA 2025", "org": "UC Berkeley", "url": null, "code": "https://github.com/ToruOwO/hato", "project": "https://toruowo.github.io/hato/", "topics": [ "hardware", "tactile", "manipulation" ], "tags": [ "bimanual", "dexterous-hand", "teleoperation", "touch", "low-cost" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "HATO — bimanual setup from two UR5e arms and repurposed prosthetic hands with fingertip touch sensors, VR teleop for data collection, and visuotactile policies for tasks like pouring and handovers.", "takeaway_zh": "HATO：用两台 UR5e 机械臂和改装的假肢手（指尖带触觉传感器）搭建双臂系统，用 VR 遥操作采集数据，并训练视觉-触觉策略完成倒酒、交接等任务。", "github_stars": 174, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2404.16823.html", "labs": [ "berkeley" ] }, { "id": "openarm", "title": "OpenArm", "type": "hardware", "arxiv": null, "year": 2025, "venue": null, "org": "Enactic", "url": null, "code": "https://github.com/enactic/openarm", "project": "https://openarm.dev", "topics": [ "hardware", "manipulation" ], "tags": [ "arm", "humanoid", "teleoperation", "force-feedback", "open-source" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "Fully open-source 7-DoF humanoid arm for contact-rich work — CAD, electronics, firmware, ROS 2 and MuJoCo/Genesis models, plus leader–follower force-feedback teleop and gravity compensation.", "takeaway_zh": "完全开源的 7 自由度人形机械臂，面向接触丰富的研究：CAD、电路、固件、ROS 2/MoveIt 2 以及 MuJoCo/Genesis 模型全部公开，支持主从力反馈遥操作与重力补偿（软件 Apache-2.0，硬件 CERN-OHL-S）。", "github_stars": 3653, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/openarm.html", "labs": [] }, { "id": "arxiv-2412.10447", "title": "TidyBot++: An Open-Source Holonomic Mobile Manipulator for Robot Learning", "type": "hardware", "arxiv": "2412.10447", "year": 2024, "venue": "CoRL 2024", "org": "Princeton, Stanford", "url": null, "code": "https://github.com/jimmyyhwu/tidybot2", "project": null, "topics": [ "hardware", "manipulation" ], "tags": [ "mobile-manipulator", "holonomic", "open-source", "teleoperation", "low-cost" ], "robot_types": [ "arm", "mobile" ], "highlight": false, "added": null, "takeaway_en": "An inexpensive open-source mobile base with powered casters, so it is fully holonomic; carries arbitrary arms and comes with phone teleoperation for collecting mobile-manipulation demos for imitation learning.", "takeaway_zh": "低成本开源移动底盘，采用主动脚轮实现全向（完整约束）运动，可搭载任意机械臂，并配有手机遥操作界面，方便为模仿学习采集移动操作示教数据。", "github_stars": 622, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2412.10447.html", "labs": [ "stanford", "princeton" ] }, { "id": "arxiv-2407.10353", "title": "UMI-on-Legs: Making Manipulation Policies Mobile with Manipulation-Centric Whole-body Controllers", "type": "paper", "arxiv": "2407.10353", "year": 2024, "venue": "CoRL 2024", "org": null, "url": null, "code": "https://github.com/real-stanford/umi-on-legs", "project": null, "topics": [ "hardware", "locomotion", "manipulation" ], "tags": [ "quadruped", "umi", "whole-body-control", "loco-manipulation", "sim", "real" ], "robot_types": [ "legged" ], "highlight": false, "added": null, "takeaway_en": "Robot-free UMI gripper demos train the manipulation policy; a sim-trained whole-body controller tracks its end-effector trajectories on a quadruped. Reports >70% success on all tasks and zero-shot reuse of an arm policy.", "takeaway_zh": "用手持 UMI 夹爪采集无需机器人的示教来训练操作策略，再由仿真训练的全身控制器在四足机器人上跟踪其输出的末端轨迹；报告所有任务成功率超过 70%，并能零样本复用原为固定机械臂训练的策略。", "github_stars": 561, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2407.10353.html", "labs": [] }, { "id": "arxiv-2502.07730", "title": "DOGlove: Dexterous Manipulation with a Low-Cost Open-Source Haptic Force Feedback Glove", "type": "hardware", "arxiv": "2502.07730", "year": 2025, "venue": "RSS 2025", "org": "Tsinghua University", "url": null, "code": "https://github.com/TEA-Lab/DOGlove", "project": "https://do-glove.github.io/", "topics": [ "hardware", "manipulation", "tactile" ], "tags": [ "glove", "teleoperation", "force-feedback", "haptics", "dexterous-hand", "open-source", "low-cost" ], "robot_types": [ "hand" ], "highlight": false, "added": null, "takeaway_en": "An open-source teleoperation glove buildable in hours for under $600, with 21-DoF hand capture, 5-DoF cable-driven force feedback and fingertip haptics; reports the feedback matters for contact-rich tasks without vision.", "takeaway_zh": "开源遥操作手套，几小时即可组装、成本低于 600 美元：21 自由度手部动捕、5 自由度绳驱力反馈和指尖振动触觉；报告在无视觉输入时力/触觉反馈对接触密集任务至关重要，采集数据可训练模仿学习策略。", "github_stars": 144, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2502.07730.html", "labs": [ "tsinghua" ] }, { "id": "arxiv-2504.13165", "title": "RUKA: Rethinking the Design of Humanoid Hands with Learning", "type": "hardware", "arxiv": "2504.13165", "year": 2025, "venue": "RSS 2025", "org": "NYU", "url": null, "code": "https://github.com/ruka-hand/RUKA", "project": "https://ruka-hand.github.io/", "topics": [ "hardware", "manipulation" ], "tags": [ "hand", "tendon-driven", "3d-printed", "open-source", "low-cost" ], "robot_types": [ "humanoid" ], "highlight": false, "added": null, "takeaway_en": "A human-sized, tendon-driven hand (5 fingers, 15 underactuated DoF) from 3D-printed and off-the-shelf parts; joint-to-actuator control is learned from MANUS-glove mocap data instead of hand-modelled.", "takeaway_zh": "人手大小的腱驱动灵巧手（5 指、15 个欠驱动自由度），由 3D 打印件和现成零件制成；关节到电机的控制映射用 MANUS 数据手套的动捕数据学习得到，而非手工建模。", "github_stars": 202, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2504.13165.html", "labs": [ "nyu" ] }, { "id": "arxiv-2411.13677", "title": "Bimanual Dexterity for Complex Tasks", "type": "hardware", "arxiv": "2411.13677", "year": 2024, "venue": "CoRL 2024", "org": "CMU", "url": null, "code": null, "project": "https://bidex-teleop.github.io/", "topics": [ "hardware", "manipulation" ], "tags": [ "teleoperation", "bimanual", "hand", "mocap-glove", "low-cost" ], "robot_types": [ "arm", "hand" ], "highlight": false, "added": null, "takeaway_en": "Bidex — a portable, low-latency bimanual dexterous teleop rig (mocap gloves + teacher arms) for 50+ DoF robots; reports better data, collected faster, than Vision Pro or SteamVR teleop.", "takeaway_zh": "Bidex：便携、低延迟的双臂灵巧手遥操作系统（动捕手套 + 示教臂），可控制 50+ 自由度的机器人；论文报告其采集的数据质量和速度均优于 Vision Pro 与 SteamVR 遥操作。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2411.13677.html", "labs": [ "cmu" ] }, { "id": "arxiv-2610.32969", "title": "DexPort: A Hand-Held Skill Interface for Dexterous Hand Data Collection", "type": "paper", "arxiv": "2610.32969", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://amberxie88.github.io/dexport/", "topics": [ "hardware", "manipulation", "datasets" ], "tags": [ "data-collection", "dexterous-hand", "portable", "real" ], "robot_types": [ "hand" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "A portable hand-held rig for collecting demonstrations with a high-DoF robot hand, using the same hand hardware for collection and deployment, so there is no retargeting from human hands and no full robot setup needed.", "takeaway_zh": "便携手持式数据采集装置，面向高自由度机器人灵巧手；采集与部署使用同一只机械手，无需从人手动作重定向，也无需搭建完整机器人系统。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.32969.html", "labs": [] }, { "id": "arxiv-2610.29030", "title": "PHD-Hand: A Printable Humanoid Dexterous Hand with Linkage-Driven 3-DOA Digits", "type": "paper", "arxiv": "2610.29030", "year": 2026, "venue": "Humanoids 2026", "org": null, "url": null, "code": null, "project": null, "topics": [ "hardware", "manipulation" ], "tags": [ "dexterous-hand", "linkage-driven", "3d-printable", "low-cost", "humanoid" ], "robot_types": [ "hand", "humanoid" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "Fully linkage-driven humanoid hand with 3 actuated DoF per finger; a flexure hinge and spring-steel drive rods replace many tiny parts, so it prints on a budget 3D printer at near human-hand size.", "takeaway_zh": "完全连杆驱动的人形灵巧手，每根手指 3 个主动自由度；用柔性铰链和弹簧钢驱动杆取代大量微型零件，可用入门级 3D 打印机制造，尺寸接近真人手。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.29030.html", "labs": [] }, { "id": "arxiv-2610.27998", "title": "LocoPilot: Single-Operator Bilateral Teleoperation for Loco-Manipulation Policy Learning", "type": "paper", "arxiv": "2610.27998", "year": 2026, "venue": null, "org": "IIT", "url": null, "code": "https://github.com/iit-DLSLab/LocoPilot", "project": null, "topics": [ "hardware", "manipulation", "locomotion" ], "tags": [ "teleoperation", "data-collection", "quadruped", "mobile-manipulation", "haptics", "ros2", "real" ], "robot_types": [ "legged" ], "highlight": false, "added": "2026-10-26", "takeaway_en": "One operator drives a Unitree Go2 + Piper arm with a leader arm and two handheld controllers, with sensorless force feedback; records a unified 11-D action reused at deployment (shown by fine-tuning MolmoAct2).", "takeaway_zh": "单人即可同时操控 Unitree Go2 四足与 Piper 机械臂：主端机械臂加两个手柄，无需力传感器即可力反馈；ROS 2 管线记录与部署共用的 11 维统一动作，并通过微调 MolmoAct2 验证从采集到部署的全流程。", "github_stars": 6, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.27998.html", "labs": [] }, { "id": "arxiv-2610.36598", "title": "OpenJaws: An Open Hardware UMI Framework with Stereo Localization for Generalist Manipulation", "type": "paper", "arxiv": "2610.36598", "year": 2026, "venue": null, "org": null, "url": null, "code": null, "project": "https://open-jaws.github.io/", "topics": [ "hardware", "manipulation", "datasets" ], "tags": [ "umi", "data-collection", "gripper", "stereo", "open-hardware", "3d-printed" ], "robot_types": [], "highlight": false, "added": "2026-10-28", "takeaway_en": "Fully open UMI-style handheld gripper with a low-cost stereo camera for metric, low-drift trajectories, plus UMI-Loc-Bench-v1 for localization accuracy; hardware, software and data under a permissive license.", "takeaway_zh": "完全开源的 UMI 式手持夹爪：3D 打印、标准 ISO 接口，用低成本双目相机获得公制尺度、低漂移的轨迹，并提出手持采集设备定位精度基准 UMI-Loc-Bench-v1；硬件、软件与数据以可商用许可证发布。", "github_stars": null, "page": "https://hu-zlatan.github.io/embodied-ai-radar/e/arxiv-2610.36598.html", "labs": [] } ] }