Physical AI Radar · Weekly roundup · 2026-W41
56 picks published 2026-10-05 → 2026-10-11, grouped by lane. Titles link to the source; dates link to that day’s radar.
- Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary
[O]0 ·[R]39 ·[M]17
Foundation models (VLA / WAM) (12)
[R]NarrativeFlow: Flow-Based Vision-Language-Action Model Using Robot Velocity Fields — arXiv 2610.00981 ｜ 2026-10-05[R]Divide-and-Remember: Recursive Action-Relevant Memory for Long-Horizon VLA Policies — arXiv 2610.00982 ｜ 2026-10-05[R]FastOPD: On-Policy Distillation for Lightweight VLA Deployment — arXiv 2610.02832 ｜84%·78.1%｜ 2026-10-06[R]SimpleTouch: Can Vision-Language-Action Models Master Contact-Rich Manipulation Without Tactile Policy Pretraining? — arXiv 2610.02784 ｜50 demonstrations·77.5%·45.2%｜ 2026-10-06[R]ESP: Energy-Score Policy for One-Step Multimodal Action Generation — arXiv 2610.07696 ｜ 2026-10-07[R]Adapting Vision-Language-Action Models to Unknown Visual Disruptions During Execution — arXiv 2610.07946 ｜43.9%·53.2%·58.7%｜ 2026-10-07[M]IROS 2026大奖揭晓：长期记忆拿下最佳论文，人形机器人打网球、端托盘双双获奖 — 雷峰网 Leiphone ｜ 2026-10-08[R]Tell Robot What Not to Do: A Negation Understanding Perspective — arXiv 2610.11952 ｜2.60%·88.45%·12.4%｜ 2026-10-09[R]Humanoid World Action Model With Joint State–Action Generation — arXiv 2610.12026 ｜70.6%·43.3%｜ 2026-10-10[M]再裁员5000人！保时捷明确回归燃油车，高端车型计划涨价20%；腾讯拟发行50亿美元离岸债券，或筹钱加码AI；OpenAI全面上线GPT-6 — 雷峰网 Leiphone ｜ 2026-10-10[R]ARC: A Reasoning Recipe for Robot Foundation Models — arXiv 2610.12386 ｜ 2026-10-11[R]Rewiring Semantics, Dynamics, and Control: A Simple yet Effective Action-Centric Tri-Stream Transformer — arXiv 2610.11416 ｜ 2026-10-11
Data engine (collection / curation / scaling) (5)
[R]UniWAM: Unified World-Action Model — arXiv 2610.02054 ｜ 2026-10-05[M]对话如祺出行COO韩锋：世界模型越强，真实数据才是真壁垒 — 雷峰网 Leiphone ｜ 2026-10-08[R]Leveraging Human-In-The-Loop Demonstrations in Reinforcement Learning for Digital Twin-Driven Robot Flexibility — arXiv 2610.12140 ｜100%·17%｜ 2026-10-09[M]这届 IROS 2026 上，具身学术与产业边界正在消融 — 雷峰网 Leiphone ｜ 2026-10-10[R]SkillWeave: Weaving Heterogeneous Demonstrations into Long-Horizon Manipulation Skills — arXiv 2610.12046 ｜27%·65%·87%｜ 2026-10-11
Training & self-improvement (RL / post-training) (4)
[R]PROMO: Preference-conditioned Multi-Objective Reinforcement Learning for Quadrupedal Robots — arXiv 2610.01260 ｜30.4%·38.7%·59.0%｜ 2026-10-05[R]UniIntervene++: An Adaptive Intervention Agent for Efficient Real-World Reinforcement Learning — arXiv 2610.03620 ｜89.67%·0.77%·94.6%｜ 2026-10-06[R]PEARS: Physical-Prior-Guided Efficient Adaptation via Failure Reasoning and Diffusion Steering for Tactile Manipulation — arXiv 2610.08784 ｜53.2%·95%·90%｜ 2026-10-08-
[M]**[独家清华 AIR 首届博士李健雄创立本溯智能：五源资本领投，押注动作原生具身模型](https://www.leiphone.com/category/ai/WbD0Tf6P1oA2WDdq.html)** — 雷峰网 Leiphone ｜ 2026-10-09
Simulation & evaluation (sim2real / closed-loop) (14)
[R]Is Success All You Need? Investigating the Impact of Input Perturbations on VLA Behaviour in Tabletop Manipulation Tasks — arXiv 2610.01351 ｜ 2026-10-05[R]WBAG: A Whole-Body and Attached-Geometry Safety Framework for Vision-Language-Action Manipulation — arXiv 2610.01083 ｜ 2026-10-05[R]World Action Learning via Interaction-Centric Spectral Latent Guidance — arXiv 2610.03607 ｜99.20%·93.80%·57.7%｜ 2026-10-06[M]川崎重工とNVIDIAが目指す「次世代デジタルシップヤード構想」とは何か — MONOist (ITmedia) ｜ 2026-10-06[R]Demo: Closed-Loop Sionna-Isaac Sim Co-Simulation Framework for Wireless-Aware Robot Navigation over ROS 2 — arXiv 2610.08618 ｜16ms·60 Hz·7.4%｜ 2026-10-07[R]StairVLA: Stage-Aware Hierarchical Action Generation for Vision-Language-Action Models — arXiv 2610.07756 ｜96.5%·97.8%·115.0 ms｜ 2026-10-07[R]Agentic RSR: Real-to-Sim-to-Real through Scene Reconstruction and Execution-Grounded Robot Policies — arXiv 2610.10479 ｜80%｜ 2026-10-08[R]RealtimeWAM: How Fast Can I Run My World Action Model? — arXiv 2610.10079 ｜63.09 ms·82.75%·87.41%｜ 2026-10-08[R]WARP-VLA: Wrist-Camera Adaptation for View-Robust Policy Execution in Vision-Language-Action Models — arXiv 2610.11508 ｜39.2%·78.3%｜ 2026-10-09[R]REACT: Rolling Denoising and Dual Decoupling for Reactive Robot Control with VLA Models — arXiv 2610.12007 ｜ 2026-10-09[R]RoboRender: Robot-Oriented Video Generation for Visual Sim-to-Real Transfer — arXiv 2610.09254 ｜71%·7.1x·3.6x｜ 2026-10-10[R]USDCraft: Geometrically Grounded Programmatic Modeling of Articulated 3D Assets for Simulation — arXiv 2610.11322 ｜ 2026-10-10[R]RESETTLE: Robotic Recovery through Disagreement-Triggered Retrieval and Efficient Corrective Control — arXiv 2610.12185 ｜8.70%·6.28%·6.83%｜ 2026-10-11[R]Residual Modeling Closes the Regression and Generative Policy Gap in Robot Learning — arXiv 2610.12231 ｜ 2026-10-11
Edge & real-time (on-device inference / control rate) (6)
[R]TOAST: Stochastic Robot Action Tokenization for Autoregressive Vision-Language-Action Models — arXiv 2610.00899 ｜ 2026-10-05[R]HexVIO: Towards All-Day Stereo-Inertial Tracking Through Commodity DSPs — arXiv 2610.03283 ｜67%·86%·30 fps｜ 2026-10-06[R]LightVLN: Efficient Aerial Vision-and-Language Navigation with Compact Memory and History-Guided Local Aggregation — arXiv 2610.05024 ｜50.93%·36.14%·25.83%｜ 2026-10-07[R]ActTune: Action-Aware Precision and GPU Operating-Point Adaptation for Energy-Efficient Vision-Language-Action Inference — arXiv 2610.08444 ｜ 2026-10-07[R]RFPO: Rectified Flow Policy Optimization for Embodied Control — arXiv 2610.10453 ｜2.4%·98.5%·4.39 ms｜ 2026-10-08[M]AWS launches open-source Physical AI Toolchain for robotics — The Robot Report ｜ 2026-10-09
Safety, permissions & compliance (2)
[M]FCC robot restrictions could accelerate shift to local AI — The Robot Report ｜ 2026-10-06[R]FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems — arXiv 2610.12432 ｜99.95%·97%｜ 2026-10-11
Systems & orchestration (long-horizon / multi-policy) (3)
[R]InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation — arXiv 2610.02196 ｜ 2026-10-05[R]MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile Manipulation — arXiv 2610.03476 ｜7.50%·27.50%·32.5%｜ 2026-10-06[R]ManiUnit: A Manipulation Skill Dataset and Benchmark for Long-Horizon Tasks — arXiv 2610.12089 ｜56%·78.7%·49.3%｜ 2026-10-10
Embodiment & supply chain (hardware / cost) (10)
[M]ルネサスが低耐圧GaNデバイス市場に参入、業界トップクラスの性能指数を実現 — MONOist (ITmedia) ｜ 2026-10-07[M]Boston Dynamics appoints former Amazon exec Rohit Prasad as new CEO — The Robot Report ｜ 2026-10-07[M]研究了1933篇 IROS 论文，我们看到了机器人学的六项新变化 — 雷峰网 Leiphone ｜ 2026-10-08[M]Harmonic Drive to share how integrated actuators can reduce engineering at RoboBusiness — The Robot Report ｜ 2026-10-08[M]This Disembodied Hand Is All the Robot You Need — IEEE Spectrum Robotics ｜ 2026-10-09[M]Jabil discusses the pace of humanoid robot development and production — The Robot Report ｜ 2026-10-09[M]Video Friday: Robot Decommissioning Takes a Fun Turn — IEEE Spectrum Robotics ｜ 2026-10-10-
[M]**[让触觉像视觉一样可规模化，哥大李昀烛团队 FlexiTac 的开源实践IROS 2026](https://www.leiphone.com/category/robot/Pz34omTBnW9Y0LhO.html)** — 雷峰网 Leiphone ｜ 2026-10-10 [R]Being-M0.7: A Latent World-Action Model for Humanoid Robots — arXiv 2610.11283 ｜10,000 hours｜ 2026-10-11[M]Why humanoid robot demos still fail the generalization test — The Robot Report ｜ 2026-10-11
This repository is technical observation. It is not investment advice, a product commitment, or a safety certification conclusion. Paper claims are author-reported.