Embodied AI Radar · Weekly 2026-W44 · 具身智能雷达周报
2026-10-26 → 2026-11-01 · 436 new arXiv papers · 43 with code
本周共收录 436 篇 arXiv 新论文，其中 43 篇附代码。
This week in one paragraph. Two threads stood out. First, pretraining on human and egocentric data at scale: HuGe and RynnVLA-Latent both report results from thousands of hours of mixed human and robot video. Second, measuring before believing: S², RoboSeed, CuraEval, RoboJudge and the paired π0.5 test on BEHAVIOR-1K all question how we score policies, and the paired test finds that earlier tricks do not transfer uniformly. CoRL 2026 opens on Nov 9, and its venue page now lists 170+ accepted papers, filterable by topic, code and Oral/Spotlight.
一段话看本周。 两条主线最明显：一是用人类与第一视角数据做大规模预训练，HuGe 和 RynnVLA-Latent 都报告了基于数千小时人类与机器人混合视频的结果；二是先测量、再相信，S²、RoboSeed、CuraEval、RoboJudge 以及 BEHAVIOR-1K 上的 π0.5 配对测试都在追问如何评估策略，配对测试发现已有技巧的效果并不能一致迁移。CoRL 2026 将于 11 月 9 日开幕，会议页面现已收录 170+ 篇录用论文，可按主题、代码和 Oral/Spotlight 筛选。
Editor's picks · 本周精选
Top 8 of 116 new picks, ranked by community signals (Hugging Face upvotes, GitHub stars). 本周新增 116 条精选，按社区信号（Hugging Face 点赞、GitHub 星标）列出前 8 条。
- PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations — 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%。 · details · 详情
- Rollback: Experience Backtracking in World Models for Policy Improvement — 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。 · details · 详情
- ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning — 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。 · details · 详情
- RynnUnify: Towards Unified Embodied Brain — 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. / 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%。 · details · 详情
- Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success — S² is a training-free offline metric of how well a model's representation aligns with physical motion. Reports it tracks robot task success across model families and can rank VLM backbones before VLA fine-tuning. / S² 是一种无需训练的离线指标，衡量模型视觉表征与物理运动几何的对齐程度（把柏拉图表征假说扩展到机器人）；报告它与多类模型的真实任务成功率正相关，可在微调成 VLA 之前筛选 VLM 骨干。 · details · 详情
- HuGe: Scaling Humanoid Generalist Vision-Language-Action Models with Egocentric Experience — Humanoid VLA family pre-trained on 2,000+ h of egocentric human and robot data, predicting latent motion tokens for a whole-body controller; uses pseudo-proprioception and arm re-rendering to bridge the embodiment gap. / 人形机器人通用 VLA 系列：在 2,000 多小时第一视角人类与机器人数据上预训练，输出潜在动作 token 驱动预训练的全身控制器并附手部信号；通过仿真回放构造伪本体感知、重渲染手臂外观缩小人-机差异，报告 HumanoidArena 最优，并在 Unitree G1 上验证人类数据增益。 · details · 详情
- DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning — 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%。 · details · 详情
- RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions — 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%。 · 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: AGQ · Coupling Execution Horizon and Sampling Depth in Flow-Matching VLA Policies · Echo · FastHarness · FocusVLA · From Reasoning to Action · Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates · Multi-View Latent Action Pretraining for Cross-Embodiment Robot Learning from Videos · OmniBrain-0 · PaceVLA · Plan First, Act Later · Profiling a Frontier Agent with EBench · Quantize by Doing · Retrieve to Act · RoboSeed · Robot-Aligned 3D Gaussian Splatting for Precise Reaching with Wrist-Camera VLA Policies · Vision-Language-Action or World-Action? Just Action Models · Vision-Language-Delta-Action · VISTA · What Actually Matters? A Paired Evaluation of Inference Settings for π0.5 on BEHAVIOR-1K
- Manipulation & Dexterous Hands · 操作与灵巧手: AdaDexTrack · AdaPolicy · CARD · ContactBench 1.0 · COWBOY · cuArms · FORCA · Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions · Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well · Real2Sem2Real · Reconstruction-Augmented Visuomotor Pre-Training of 3D-Grounded Bimanual Dexterity from RGB-Only Human Videos · RoboLongR · RoboSTD · Scaling Dexterous Manipulation via Action Representation Pretraining · ScheduleGen
- Tactile & Visuotactile Sensing · 触觉与视触觉感知: ACTS · Tactile-Based Robotic Manipulation · VecTile
- Locomotion & Humanoids · 运动控制与人形机器人: Action-Aligned Latent Planning for Versatile Humanoid Loco-Manipulation · Differentiable Contact Dynamics for Legged Locomotion · ELEVATE · GRAFT · HumanoidNDM · HuMiT · Learning Force-Reactive Humanoid Loco-Manipulation · METATRON · REACH · SMR · Solving a Giant Rubik's Cube with a Quadruped Robot · SONIC-K1 · SWAP · Test-Time Motion Steering for Perceptive Humanoid Fall Recovery · ViT-H · Whole-Body Humanoid Control via Stochastic Motion Tokenization · Zero-HLM
- World Models · 世界模型: Adapting Video Models for Robot Control via Motion-guided Asymmetric Prediction · Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention · DeXplicit · GAWM · LaWAM · OSWAM · Physical Canvas · PhysTrace · PICA-WAM · RynnWorld-Latent · TAPESTRY · Trace-WAM · Training Controllable World Models via Action Influence Maximization · TrajWAM · Video World Models Need Predictable Latents · WorldFlow
- Navigation & Embodied Agents · 导航与具身智能体: Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation · EmbodiedWorld-200K · Gaussian Episodic Memory for Long-Term Spatiotemporal Reasoning · HAVANA · InJePA · MOON · One Scan is Enough · See What I Need, Dream What I Want
- Planning, Control & Safety · 规划、控制与安全: A generalist policy for muscle-driven control across individual musculoskeletal models · Keeping Contact
- Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习: CuraEval · EgoForesight · Flexible Language-Guided Representation Editing for Adaptive Robot Behavior · HIDRA · Progress-Guided Preference Generation for Aligning Generative Control Policies · Training in Deployment
- Sim-to-Real & Simulators · 仿真与 Sim-to-Real: DexSuite · Gym-Anything for Robotics · MARE · SceneBloom
- Datasets & Benchmarks · 数据集与基准: AstroBench · Behavior Spatial Benchmark · BiDex-100 · EgoContact · Embodied Playground · EmbodiedLab · FLORA · RoboJudge · RobotAgentBench · SO-101 Bench
- Hardware & Open-Source Robots · 硬件与开源机器人: DexPort · EgoThumb · LocoPilot · OpenJaws · PHD-Hand · SwapBot · UMI-Live for In-The-Wild Bimanual Robotics Data Collection
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
- CVPR 2026 · AdaDexTrack: Dynamic Modulation for Adaptive and Generalizable Dexterous Manipulation Tracking · project
- CoRL 2026 · LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies · project
- CoRL 2026 · MAST-VLA: Skill-Conditioned Adaptation with Trajectory-Affordance Priors for Multi-Skill Robot Manipulation
- CoRL 2026 · HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model
- CoRL 2026 · TAPESTRY: A Single Backbone for Video Generation and Robot Control · project
- Humanoids 2026 · Whole-Body Humanoid Control via Stochastic Motion Tokenization
- Humanoids 2026 · REACH: An RL-based Hierarchical Humanoid Crawling Framework for Adaptable Confined Space Traversal · project
- Humanoids 2026 · FORGE: Feasible Online Reduced-Order Reference Generation and Whole-Body Execution for Agile Humanoid Locomotion
- IJRR 2026 · Skeleton-Guided Geometry-Aware Auto-Labeling for Robotic Grasping: From 4-DoF Learning to 6-DoF Execution
- IROS · EgoThumb: A Vision-Driven Supernumerary Robotic Thumb for Autonomous Grasp and Holding Assistance · project
- IROS 2026 · Revisiting Direct Control: A Field Evaluation of Sensor Fusion Efficacy in Teleoperating Full-Scale Autonomous Vehicles via Bird's Eye View
- IROS 2026 · SVIn+: Multi-Modal Framework for Robust Underwater State Estimation using 3D Sonar, Visual-Inertial, Water-Pressure, and DVL · project
- IROS 2026 · A Multi-Layer Navigation System for Smart Wheelchairs with a Limited Input Interface
- IROS 2026 · AtomTree: A Hierarchical Framework for State-Aware Embodied Instruction Following with LLMs · project
- NeurIPS 2026 · One Scan is Enough: Demonstration-Free Adaptation for Language-Guided Navigation
- NeurIPS 2026 · Multi-Objective Reinforcement Learning Using Routed Ensembles and Trajectory Attribution
- NeurIPS 2026 · HIDRA: Hierarchical Dual-Routing Attention for Replay-Free Lifelong Imitation Learning · code
- NeurIPS 2026 · CSO: Refining Robotic Policies via Skill Distribution Alignment and Skill-Grained Optimization
- NeurIPS 2026 · Self-Supervised Doppler-Guided RF Odometry
- NeurIPS 2026 · Coupled Guidance for Flow Matching
- RA-L · RAIWAY: A Large-Wheeled Bipedal Robot with Contact-Informed Blind Locomotion
- RA-L · PushAround: Collaborative Path Clearing via Physics-Informed Hybrid Search
- RA-L · Keeping Contact: Task-Adaptive Friction Margins for Bimanual Non-Prehensile Manipulation · project
- RA-L 2026 · SEVI: Safe Evidential Real-World Reinforcement Learning · project
- RA-L 2026 · Input Bias Correction in Gaussian Process Regression
- RA-L 2026 · DSGN: A Dual-Stream Gated Network for Scalable Simultaneous Surveillance and Containment · code
- WACV 2027 · Wavelet Phase Diffusion for Structurally and Semantically Consistent Sim-to-Real Translation · project
- WACV 2027 · Closed-Loop Aware Planning via Imitation-Free Speed Scoring for End-to-End Autonomous Driving · project
- WACV 2027 · Decoding Geometry, Not Tokens: Efficient VLMs for Robot Affordance and Trajectory Prediction
- WACV 2027 · SymAnt: Grounding Long-Horizon Action Anticipation with Adaptable Symbolic Priors · 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。
- PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations · ▲ 154
- Rollback: Experience Backtracking in World Models for Policy Improvement · ▲ 82
- ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning · ▲ 66
- HiSTEVE 1.0: Harnessing Human-Interface VLA Agents for Open-World Games · ▲ 59
- Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success · ▲ 56
- RynnUnify: Towards Unified Embodied Brain · ▲ 56
- HuGe: Scaling Humanoid Generalist Vision-Language-Action Models with Egocentric Experience · ▲ 44
- DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning · ▲ 32
Rising on GitHub · GitHub 升温榜
Stars gained in the 7 days to 2026-11-01. 截至 2026-11-01 的 7 天内新增星标。
- huggingface/lerobot +166 · LeRobot
- google-deepmind/mujoco +145 · MuJoCo
- LambdaHumanoid/openlambda +109 · λ0: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation
- Ropedia/ropedia-w0 +102 · Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience
- Physical-Intelligence/openpi +101 · π0.5: a Vision-Language-Action Model with Open-World Generalization
- ros-navigation/navigation2 +95 · The Marathon 2: A Navigation System
- HexaFuture/HexaShell +95 · HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
- TheRobotStudio/SO-ARM100 +83 · SO-ARM100 / SO-101
By topic · 分主题统计
- VLA & Robot Foundation Models · VLA 与机器人基础模型: 54
- Manipulation & Dexterous Hands · 操作与灵巧手: 84
- Tactile & Visuotactile Sensing · 触觉与视触觉感知: 7
- Locomotion & Humanoids · 运动控制与人形机器人: 43
- World Models · 世界模型: 32
- Navigation & Embodied Agents · 导航与具身智能体: 44
- Perception, SLAM & 3D Mapping · 感知、SLAM 与三维建图: 32
- Planning, Control & Safety · 规划、控制与安全: 31
- Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习: 52
- Sim-to-Real & Simulators · 仿真与 Sim-to-Real: 14
- Datasets & Benchmarks · 数据集与基准: 22
- Hardware & Open-Source Robots · 硬件与开源机器人: 6
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: 28
- Dexterous hand · 灵巧手: 19
- Humanoid · 人形: 32
- Quadruped / legged · 四足 / 足式: 18
- Mobile robot · 移动机器人: 15
- Drone / aerial · 无人机: 29
- Autonomous driving · 自动驾驶: 16
- Underwater / marine · 水下 / 海洋: 2
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
- RoboSeed: Multidimensional Evaluation of Vision-Language-Action and World-Action Models: Capability, Resource Affordability, and Research Evidence · code
- OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation · code · project
- Quantize by Doing: Task-Guided Mixed-Precision Quantization of Vision-Language-Action Models · code
- Vision-Language-Action or World-Action? Just Action Models · code · project
- RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions · code · project
- JevVLA: Jev-Inspired Energy-Based Adaptation for Vision-Language-Action Policies · code
Manipulation & Dexterous Hands · 操作与灵巧手
- A Tutorial on Shape Servoing of Deformable Objects-Part I: Fundamentals · code · project
- GAWM: A Compact Latent World-Action Model for Multi-Task Robot Manipulation · code
- SO-101 Bench: Accessible Spatial Manipulation · code · project
- HIDRA: Hierarchical Dual-Routing Attention for Replay-Free Lifelong Imitation Learning · code
- JevVLA: Jev-Inspired Energy-Based Adaptation for Vision-Language-Action Policies · code
- RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions · code · project
Locomotion & Humanoids · 运动控制与人形机器人
- Is It the Curriculum or Just the Easy Start? Benchmarking Unsupervised Environment Design on Quadrupeds · code
- TACOS: Terrain-Aware COntact Supervision for Visual Footstep Planning · code
- HuGe: Scaling Humanoid Generalist Vision-Language-Action Models with Egocentric Experience · code · project
World Models · 世界模型
- RoboSeed: Multidimensional Evaluation of Vision-Language-Action and World-Action Models: Capability, Resource Affordability, and Research Evidence · code
- Vision-Language-Action or World-Action? Just Action Models · code · project
- GAWM: A Compact Latent World-Action Model for Multi-Task Robot Manipulation · code
- RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions · code · project
- Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention · code
- Video World Models Need Predictable Latents · code
Navigation & Embodied Agents · 导航与具身智能体
- OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation · code · project
- Rendering-Free Event Frame Simulation for Fast Reinforcement Learning of Autonomous Robot Navigation · code
- EmbodiedLab: Benchmarking Embodied Agents That Act to Know · code · project
- VISTA: Active Exploration for Self-Improvement in Embodied Agents · code · project
- Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation · code · project
- MOON: Learning When to Forget and How to Move On in Vision-and-Language Navigation · code
Perception, SLAM & 3D Mapping · 感知、SLAM 与三维建图
- RESPLE2: Inertial-Informed Recursive Non-Uniform Spline Estimation for Multi-Sensor Odometry with Global Refinement · code
- VPR-Space: A Benchmark for Visual Place Recognition in Space Analog Domains · code
- Bracketing Shows the Way: Adaptive Multi-Exposure Policy as Input for Robust Visual SLAM · code
Planning, Control & Safety · 规划、控制与安全
- Fundamental Limits of Sampling-Based Reachability: Geometry and Flow Distortion · code
- ISB-NMPC: Interception-Oriented Spherical-Bearing Nonlinear Model Predictive Control for Quadrotors · code
- Multimodal MPC via Bayesian Set Optimization for Dynamic Manipulation of Deformable Linear Objects · code
Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习
- Rendering-Free Event Frame Simulation for Fast Reinforcement Learning of Autonomous Robot Navigation · code
- GAWM: A Compact Latent World-Action Model for Multi-Task Robot Manipulation · code
- HIDRA: Hierarchical Dual-Routing Attention for Replay-Free Lifelong Imitation Learning · code
- DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning · code · project
Sim-to-Real & Simulators · 仿真与 Sim-to-Real
- Rendering-Free Event Frame Simulation for Fast Reinforcement Learning of Autonomous Robot Navigation · code
- SO-101 Bench: Accessible Spatial Manipulation · code · project
Datasets & Benchmarks · 数据集与基准
- Is It the Curriculum or Just the Easy Start? Benchmarking Unsupervised Environment Design on Quadrupeds · code
- SO-101 Bench: Accessible Spatial Manipulation · code · project
- VPR-Space: A Benchmark for Visual Place Recognition in Space Analog Domains · code
- EmbodiedLab: Benchmarking Embodied Agents That Act to Know · code · project
- Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation · code · project
- FLORA: An Open Model Suite for Scaling Robotics Data Engines · code · project
---
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。