Embodied AI Radar — agent picks / 精选

ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

2026-10-28T00:00:00Z

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. 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。

Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well

2026-10-28T00:00:00Z

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%. 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%。

PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations

2026-10-28T00:00:00Z

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. 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%。

Vision-Language-Action or World-Action? Just Action Models

2026-10-28T00:00:00Z

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%). 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%。

Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution

2026-10-28T00:00:00Z

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). 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%。

GAWM: A Compact Latent World-Action Model for Multi-Task Robot Manipulation

2026-10-28T00:00:00Z

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. 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。

Keeping Contact: Task-Adaptive Friction Margins for Bimanual Non-Prehensile Manipulation

2026-10-28T00:00:00Z

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. 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。

Trace-WAM: Learning to Act via 3D Motion Prediction for World Action Models

2026-10-28T00:00:00Z

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. 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）。

TAPESTRY: A Single Backbone for Video Generation and Robot Control

2026-10-28T00:00:00Z

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. 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。

RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions

2026-10-28T00:00:00Z

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. 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。

RoboSTD: Zero-Shot Single-to-Dual Transfer via Sagittal Mirroring for Bimanual Learning

2026-10-28T00:00:00Z

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. 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。

Adapting Video Models for Robot Control via Motion-guided Asymmetric Prediction

2026-10-28T00:00:00Z

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%). 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%。

OpenJaws: An Open Hardware UMI Framework with Stereo Localization for Generalist Manipulation

2026-10-28T00:00:00Z

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. 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。

REACH: An RL-based Hierarchical Humanoid Crawling Framework for Adaptable Confined Space Traversal

2026-10-28T00:00:00Z

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. 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。

CARD: Contact-Anchored Retargeting for Dexterous Manipulation

2026-10-28T00:00:00Z

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%. 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%。

A generalist policy for muscle-driven control across individual musculoskeletal models

2026-10-28T00:00:00Z

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. 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。

GRAFT: Geometry-Aware Residual Adaptation for Foundation Trackers

2026-10-28T00:00:00Z

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. 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。

Whole-Body Humanoid Control via Stochastic Motion Tokenization

2026-10-28T00:00:00Z

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. 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。

SO-101 Bench: Accessible Spatial Manipulation

2026-10-28T00:00:00Z

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%. 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%。

HIDRA: Hierarchical Dual-Routing Attention for Replay-Free Lifelong Imitation Learning

2026-10-28T00:00:00Z

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. 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。

cuArms: GPU-Accelerated Conflict-Based Planning for Multi-Arm Manipulation

2026-10-28T00:00:00Z

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. 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。

Rollback: Experience Backtracking in World Models for Policy Improvement

2026-10-28T00:00:00Z

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. 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。

OmniBrain-0: A Unified Agentic Model for Embodied Intelligence

2026-10-28T00:00:00Z

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. 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。

Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention

2026-10-28T00:00:00Z

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. 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%。

RoboLongR: Causally Consistent Rewards Teach Robots to Progressively Master Long-Horizon Manipulation

2026-10-28T00:00:00Z

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). 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%。

One Scan is Enough: Demonstration-Free Adaptation for Language-Guided Navigation

2026-10-28T00:00:00Z

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. 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%。

See What I Need, Dream What I Want: A Navigation Foundation Model on a Self-Constructed Representation Space

2026-10-28T00:00:00Z

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. 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。

FocusVLA: Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models

2026-10-28T00:00:00Z

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. 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%。

Test-Time Motion Steering for Perceptive Humanoid Fall Recovery

2026-10-28T00:00:00Z

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. 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。

RoboJudge: Multimodal Language Models as Judges for Embodied Video Generation

2026-10-28T00:00:00Z

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). 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）。

PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models

2026-10-28T00:00:00Z

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. 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。

Progress-Guided Preference Generation for Aligning Generative Control Policies

2026-10-28T00:00:00Z

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. 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。

Learning Force-Reactive Humanoid Loco-Manipulation

2026-10-27T00:00:00Z

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. 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%。

Differentiable Contact Dynamics for Legged Locomotion: Policy Learning and Contact-Implicit Trajectory Optimization

2026-10-27T00:00:00Z

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. 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。

OSWAM: An Effective World Action Model with One-Step Action Generation

2026-10-27T00:00:00Z

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. 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。

Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy

2026-10-27T00:00:00Z

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. 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。

Robot-Aligned 3D Gaussian Splatting for Precise Reaching with Wrist-Camera VLA Policies

2026-10-27T00:00:00Z

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. 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。

EmbodiedLab: Benchmarking Embodied Agents That Act to Know

2026-10-27T00:00:00Z

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. “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。

Zero-HLM: Generalizable Humanoid Loco-Manipulation without Real Robot Data

2026-10-27T00:00:00Z

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. 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%。

VISTA: Active Exploration for Self-Improvement in Embodied Agents

2026-10-27T00:00:00Z

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. 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。

Video World Models Need Predictable Latents

2026-10-27T00:00:00Z

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. 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。

RobotAgentBench: Towards Open-world Robot Autonomy

2026-10-27T00:00:00Z

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. 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。

Scaling Dexterous Manipulation via Action Representation Pretraining

2026-10-27T00:00:00Z

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. 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%。

RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions

2026-10-27T00:00:00Z

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. 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%。

AdaPolicy: Action-Driven Anticipatory Robot Policies via Next-Chunk Prediction

2026-10-27T00:00:00Z

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. 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。

HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model

2026-10-27T00:00:00Z

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. 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。

Behavior Spatial Benchmark: Evaluating Spatial Understanding for Robotics

2026-10-27T00:00:00Z

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. 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。

AdaDexTrack: Dynamic Modulation for Adaptive and Generalizable Dexterous Manipulation Tracking

2026-10-27T00:00:00Z

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. 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。

MOON: Learning When to Forget and How to Move On in Vision-and-Language Navigation

2026-10-27T00:00:00Z

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. 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。

AGQ: Action-Guided Ultra-Low-Bit Quantization for Vision-Language-Action Models

2026-10-27T00:00:00Z

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. 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%。

Coupling Execution Horizon and Sampling Depth in Flow-Matching VLA Policies

2026-10-27T00:00:00Z

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. 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。

DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning

2026-10-27T00:00:00Z

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. 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%。

Echo: Coupled Action and Task Status Signal Denoising for Vision Language Action Models

2026-10-27T00:00:00Z

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. 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。

DeXplicit: Making Physical State Explicit in Dexterous Manipulation World Models

2026-10-27T00:00:00Z

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. 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。

HuMiT: Low-Latency Whole-Body Humanoid Teleoperation via Compact Reference Tracking

2026-10-27T00:00:00Z

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. 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。

Real2Sem2Real: 3D Semantic Augmentations for Robot Learning from a Single Human RGB Demonstration

2026-10-27T00:00:00Z

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. 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。

Multi-View Latent Action Pretraining for Cross-Embodiment Robot Learning from Videos

2026-10-27T00:00:00Z

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. 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%。

Training Controllable World Models via Action Influence Maximization

2026-10-27T00:00:00Z

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. 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。

Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions

2026-10-27T00:00:00Z

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. 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）。

TrajWAM: An Efficient Trajectory-Guided World Action Model for Dynamic Robotic Manipulation

2026-10-27T00:00:00Z

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. 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。