Embodied AI Radar · Weekly 2026-W43 · 具身智能雷达周报
2026-10-19 → 2026-10-25 · 485 new arXiv papers · 67 with code
本周共收录 485 篇 arXiv 新论文，其中 67 篇附代码。
> This week in one paragraph. Two threads stood out. First, *world action models*: several papers (GreenWAM, Agentic WAMs / AW-0, UniFlowWAM, CW-WAM, FRED, SkeLAM) make a video world model double as the policy or the policy evaluator. Second, *harnesses around frozen models*: instead of retraining the VLA, CarryOn, SERA and HexaShell wrap it (or a frontier VLM) in memory, progress tracking and self-written tools. CoRL 2026 acceptances keep arriving ahead of the conference (Nov 9–12).
>
> 一段话看本周。 两条主线最明显：一是*世界动作模型*，GreenWAM、AW-0、UniFlowWAM、CW-WAM、FRED、SkeLAM 等工作让视频世界模型同时充当策略或策略评估器；二是*给冻结模型加“外壳”*，CarryOn、SERA、HexaShell 不再重训 VLA，而是为其（或前沿 VLM）加上记忆、进度跟踪和自写工具。CoRL 2026（11 月 9–12 日）的录用论文也在陆续出现。
Editor's picks · 本周精选
Top 8 of 63 new picks, ranked by community signals (Hugging Face upvotes, GitHub stars). 本周新增 63 条精选，按社区信号（Hugging Face 点赞、GitHub 星标）列出前 8 条。
- Has GPT-6-Astra Achieved Embodied Intelligence? — 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). / 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”。 · details · 详情
- DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations — 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. / 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%。 · details · 详情
- Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience — 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. / 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。 · details · 详情
- Learning from Fleet Experience: Collective Experience Learning Across 96 Robots — 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%. / 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%。 · details · 详情
- HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents — 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. / 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。 · details · 详情
- GreenWAM: A World Foundation Model for Robotics — 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. / 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。 · details · 详情
- ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation — 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. / 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。 · details · 详情
- Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds — 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. / 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。 · details · 详情
All other picks by topic · 其余精选（按主题）
- VLA & Robot Foundation Models · VLA 与机器人基础模型: CarryOn · CondenseVLA · FastVLA · LTM · MambaAct · MorphAct · Multimodal Embodied Agents · NestedVLA · Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies · REKI · Towards Agentic Robots · VLA-Δ · VLA-Rolling · WAVE-0 · When One Step Is Enough
- Manipulation & Dexterous Hands · 操作与灵巧手: Beyond the End-Effector · Blind Precise Assembly · Can a Frontier VLM Serve as a Robot Policy for Memory-Dependent Manipulation? · From VLM Planners to Robotic Agents · Learn the Task, Not the Schedule · MoDex
- Tactile & Visuotactile Sensing · 触觉与视触觉感知: GelSlim 5.0 · OpenTac
- Locomotion & Humanoids · 运动控制与人形机器人: λ0 · SuperMimic · Watch Your Step
- World Models · 世界模型: Agentic World Action Models · CW-WAM · Robotic Video World Models · Self-Distilling Future Video Representations for World Action Models · Skeleton-Grounded Latent Action Model for Robot World Models · UniFlowWAM
- Navigation & Embodied Agents · 导航与具身智能体: Cracking Mobile Manipulation · Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance · TARSphere
- Perception, SLAM & 3D Mapping · 感知、SLAM 与三维建图: ACE-Brain-0.6 · SpaRV · Vision as Proprioception · Which Depth Estimator for Jetson AGX Orin? A 2026 Benchmark of Monocular and Stereo Methods
- Planning, Control & Safety · 规划、控制与安全: CusADi2 · PrismPlan · Sub-Millisecond Reflexes for Continuous Physical Human-Robot Interaction
- Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习: ACCRUE · AgenticRoboRL · TILDE · Toward Recursive Self-Improvement in Physical AI · Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy
- Sim-to-Real & Simulators · 仿真与 Sim-to-Real: Active Generative Simulation · HyperDrone
- Datasets & Benchmarks · 数据集与基准: EgoExplorer · Embodied Skill Bench · HandsOn · LIBERO-Industry · RoboCasa-Shift · RoboValue
Accepted at venues · 已被会议/期刊接收
From the authors' arXiv comments. 根据作者在 arXiv 上的备注。 All papers per conference · 按会议汇总
- CoRL 2026 · Structured Perception Representations for Off-road Dynamics Learning
- CoRL 2026 · MoDex: Towards Universal Multi-Object Dexterous Grasping · code
- CoRL 2026 · Noise-Invariant Spatial Encoding for 3D Manipulation Policies under Depth Corruption
- CoRL 2026 · Progress-Guided Preference Generation for Aligning Generative Control Policies · project
- CoRL 2026 · Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions · project
- CoRL 2026 · FR-VLA: Implicit Force Reasoning VLA for Contact-Rich Manipulation
- CoRL 2026 · CoopUQ: Robust Multi-Robot Social Navigation via Uncertainty-Aware Cooperative Forecasting · project
- CoRL 2026 · Training Controllable World Models via Action Influence Maximization
- CoRL 2026 · AWLS: Advantage-Weighted Latent Steering for One-Step Generative Robotic Manipulation Policies · project
- CoRL 2026 · SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model · project
- CoRL 2026 · Flexible Language-Guided Representation Editing for Adaptive Robot Behavior · project
- CoRL 2026 · HyperDrone: Simulation-to-Reality Transfer at Millions of Frames per Second
- CoRL 2026 · DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving · code
- CoRL 2026 · When One Step Is Enough: An Empirical Study of Euler Integration in Flow-Matching Policies
- CoRL 2026 · To Route or to Adapt? Multi-Expert Diffusion Planners in Unseen Regions · code
- CoRL 2026 · FocusVLA: Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models
- ICRA 2026 · CMAR-search: Commonsense and Memory Augmented Reasoning for Object Search in Dynamic Interactive Environments
- IROS · Pedicel-Targeted Strawberry Harvesting via Imitation Learning on a Low-Cost Teleoperated Robotic Arm · code
- IROS 2026 · Scalable, Intuitive, Mixed Reality-Based Semi-Autonomous Robotic Telemanipulation for the Execution of a Fine-Grained Bimanual Task
- IROS 2026 · Energy-Score Stabilized Attentive Mixture-Density Policy: A Lightweight Approach to Multimodal Robot Motion
- IROS 2026 · Bio-inspired Thrust-Assisted Trotting on a legged-Aerial Robot via Real-Time QP-MPC
- IROS 2026 · Distance Field Driven Reinforcement Learning for Reactive and Robust Collision Avoidance
- IROS 2026 · Joint-Angle Invariance of Jacobian-Based Performance Indices for Serial Robotic Manipulators
- IROS 2026 · SCOPE: Scene-graph Construction Optimized by Plan Error-correction for Robot Long-term Planning Datasets · project
- NeurIPS 2026 · NestedVLA: Learning to Consolidate and Generate Skills for Vision-Language-Action Models · code
- NeurIPS 2026 · Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing
- NeurIPS 2026 · Think Densely, Act Sparsely: Latent Expert Cognitive Chains for Vision-Language-Action Autonomous Driving
- NeurIPS 2026 · CondenseVLA: Learnable History Condensation for Efficient Multi-Frame VLA · code
- NeurIPS 2026 · EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models · project
- RA-L · TransNav: Transformer-Based End-to-End Navigation Method for Static and Dynamic Quadrotor Obstacle Avoidance
- WACV 2027 · Expand-and-Prune: Test-Time Candidate-Space Adaptation for End-to-End Autonomous Driving · code
- WACV 2027 · CLARO: Clutter Aware Robot Orchestrator for Open-Vocabulary Grasping using VLM Agents
- WACV 2027 · BodyBridge: Grounded Body Layers for Sim-to-Real Translation of Vulnerable Road Users
- WACV 2027 · Goal Images Are Control Targets: Spatially-Grounded Goal-Image Synthesis for Robotic Manipulation · project
Upvoted on Hugging Face · Hugging Face 热议
Papers from our feed listed on Hugging Face Daily Papers this week, by upvotes. 本周登上 Hugging Face Daily Papers 的收录论文，按点赞数排序。
- Has GPT-6-Astra Achieved Embodied Intelligence? · ▲ 478
- DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations · ▲ 161
- Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience · ▲ 151
- Learning from Fleet Experience: Collective Experience Learning Across 96 Robots · ▲ 137
- HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents · ▲ 123
- ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation · ▲ 91
- Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds · ▲ 78
- VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse · ▲ 64
By topic · 分主题统计
- VLA & Robot Foundation Models · VLA 与机器人基础模型: 67
- Manipulation & Dexterous Hands · 操作与灵巧手: 114
- Tactile & Visuotactile Sensing · 触觉与视触觉感知: 20
- Locomotion & Humanoids · 运动控制与人形机器人: 60
- World Models · 世界模型: 42
- Navigation & Embodied Agents · 导航与具身智能体: 61
- Perception, SLAM & 3D Mapping · 感知、SLAM 与三维建图: 39
- Planning, Control & Safety · 规划、控制与安全: 30
- Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习: 43
- Sim-to-Real & Simulators · 仿真与 Sim-to-Real: 9
- Datasets & Benchmarks · 数据集与基准: 24
- Hardware & Open-Source Robots · 硬件与开源机器人: 7
By robot type · 分机器人类型统计
Keyword-matched on title and abstract; a paper can count for several types. 按标题与摘要关键词匹配，一篇论文可计入多个类型。
- Robot arm · 机械臂: 27
- Dexterous hand · 灵巧手: 24
- Humanoid · 人形: 41
- Quadruped / legged · 四足 / 足式: 20
- Mobile robot · 移动机器人: 17
- Drone / aerial · 无人机: 15
- Autonomous driving · 自动驾驶: 18
- Underwater / marine · 水下 / 海洋: 3
New papers with code · 本周附代码的新论文
VLA & Robot Foundation Models · VLA 与机器人基础模型
- Echo: Coupled Action and Task Status Signal Denoising for Vision Language Action Models · code · project
- FastHarness: Experience-Driven Review Scheduling for Vision-Language-Action Policies · code
- Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies · code
- From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation · code · project
- AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics · code · project
- RoboCasa-Shift: Stage-Aware Robustness Evaluation of Vision-Language-Action Policies in Long-Horizon Household Manipulation · code
Manipulation & Dexterous Hands · 操作与灵巧手
- RGraph-Eval: Task-Grounded Scene Graphs for Evaluating Robot Video Generation · code
- AstroBench: Benchmarking Free-Floating Capture and Space Robotic Manipulation · code · project
- Alpha: Systematic Fault Injection and Robustness Evaluation for LLM-Driven Embodied Manipulation · code
- From Reasoning to Action: Benchmarking Frontier AI Models on LIBERO-10 Robot Manipulation Tasks · code · project
- DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations · code · project
- From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation · code · project
Locomotion & Humanoids · 运动控制与人形机器人
- Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience · code · project
- Vision as Proprioception: Estimating Humanoid Robot Pose from Video · code
World Models · 世界模型
- Self-Distilling Future Video Representations for World Action Models · code · project
- Latent Projection for Latent Planning · code
- Robotic Video World Models: A Survey of Applications, Research Challenges, Future Directions · code
- Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience · code · project
- PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models · code
- Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation · code
Navigation & Embodied Agents · 导航与具身智能体
- Externalise the Memory: A Pre-Registered Study of Where an Embodied Agent's Experience Should Live · code
- Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use · code · project
- TARSphere: A Robot-Oriented Full-Stack Framework for Long-Horizon Embodied Agents in Evolving Physical and Social Environments · code · project
- Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds · code · project
- TESS: LLM-guided Translation, Exploration, Simulation, and Search for Embodied Task Planning · code
- ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation · code · project
Perception, SLAM & 3D Mapping · 感知、SLAM 与三维建图
- RGraph-Eval: Task-Grounded Scene Graphs for Evaluating Robot Video Generation · code
- AOV-GS: Active Open-Vocabulary Semantic Mapping with 3D Gaussian Splatting · code
- PanoGraphGS: Unifying Panoptic Segmentation and Hierarchical 3D Scene Graphs with Gaussian Splatting · code
- Foundation Model Driven Optical Flow · code
- AnchorFlow-VIO: Fixed-Template Monocular Visual-Inertial Odometry on LaMAria · code
- SpaRV: Sparse Radiance Voxels for Scalable Geometrically Consistent LiDAR-Visual Mapping and Rendering · code
Planning, Control & Safety · 规划、控制与安全
- PrismPlan: A Shared GPU Substrate for Motion Planning Across Paradigms · code
- Pedicel-Targeted Strawberry Harvesting via Imitation Learning on a Low-Cost Teleoperated Robotic Arm · code
- Online Constraint Optimization-Based Indirect-Adaptive Robust Control for Autonomous Vehicles with Unknown Friction · code
Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习
- FastHarness: Experience-Driven Review Scheduling for Vision-Language-Action Policies · code
- DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving · code
- AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics · code · project
- LocoPilot: Single-Operator Bilateral Teleoperation for Loco-Manipulation Policy Learning · code
- Pedicel-Targeted Strawberry Harvesting via Imitation Learning on a Low-Cost Teleoperated Robotic Arm · code
Sim-to-Real & Simulators · 仿真与 Sim-to-Real
- NeuralSRender: G-Buffer Guided Video Diffusion for Sim-to-Real Rendering · code
- MARE: An Open-Source Lunar Rover Simulator with Shadow Ground Truth for Perception and Autonomy · code
- K-AoI: A Spatio-Temporal Synchronization Metric for Digital Twin-based Teleoperation in Dynamic Network Environments · code
Datasets & Benchmarks · 数据集与基准
- AstroBench: Benchmarking Free-Floating Capture and Space Robotic Manipulation · code · project
- From Reasoning to Action: Benchmarking Frontier AI Models on LIBERO-10 Robot Manipulation Tasks · code · project
- Which Depth Estimator for Jetson AGX Orin? A 2026 Benchmark of Monocular and Stereo Methods · code
- HandsOn: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation · code · project
- RoboValue: A Fine-Grained Sim-and-Real Benchmark for Unified Evaluation of Robotic Value Models · code · project
- TriSAR-3D and PaintFusion: A Novel Trimodal LiDAR-RGB-Thermal Dataset and Method for 3D Person Detection in Search and Rescue · code
Hardware & Open-Source Robots · 硬件与开源机器人
- Pedicel-Targeted Strawberry Harvesting via Imitation Learning on a Low-Cost Teleoperated Robotic Arm · code
---
Auto-tagged by keyword rules; picks are reviewed by the maintainer agent. Website · GitHub · New to the field? Reading path
主题为关键词规则自动标注；精选由维护智能体审阅。刚入门？看入门路线。