Physical AI Radar · Weekly roundup · 2026-W40
52 picks published 2026-09-28 → 2026-10-04, 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]2 ·[R]32 ·[M]18
Foundation models (VLA / WAM) (9)
[O]NVIDIA Cosmos 3 on AWS: Omnimodal World Models for Physical AI — AWS Physical AI Blog ｜ 2026-09-29[R]Fast Plans, Faithful Actions: Closing the Planning-Execution Gap in Hierarchical Vision-Language-Action Models — arXiv 2609.30833 ｜91.0%·96.2%·95.85%｜ 2026-09-29[R]ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning — arXiv 2609.34982 ｜9.8%·6.1%·11.4%｜ 2026-09-30[R]Rethinking Causal Action Tokenization with Conditional Annealing in Flow Matching — arXiv 2609.35469 ｜ 2026-09-30[M]从海拉鲁到现实世界：视频模型如何理解「变化」 — 雷峰网 Leiphone ｜ 2026-10-01[R]DSDyn-VLA: A Dual-Stream Dynamic Manipulation Framework with Motion Perception, Future Awareness, and Realtime Correction — arXiv 2609.39198 ｜ 2026-10-02[R]Toward Real-Time VLAs: Stage-Aware Two-Step Flow Denoising and System-Level Evaluation — arXiv 2609.39822 ｜61.557 ms·21.956 ms｜ 2026-10-02[M]Runway introduces Praxis-1 world action model for robotics — The Robot Report ｜ 2026-10-03[R]ChunkVLA-AM: Parallel Action Chunking for Vision-Language-Action Robot Control in Additive Manufacturing — arXiv 2610.01856 ｜92.9%｜ 2026-10-04
Data engine (collection / curation / scaling) (3)
[R]Using Non-Expert Data to Robustify Imitation Learning via Offline Reinforcement Learning — Toyota Research Institute ｜ 2026-09-28[R]Praxis: Distilling Physical Interaction Priors from Egocentric Videos for Generalizable Whole-Body Manipulation — arXiv 2609.30735 ｜ 2026-09-29[R]IronMind: Scaling Humanoid Dexterous Manipulation via Camera-Space Ego-Centric Pretraining — arXiv 2609.39403 ｜10,000 hours·55.0%·11.7%｜ 2026-10-04
Training & self-improvement (RL / post-training) (4)
[R]VLaRL: Augmenting Vision-Language-Action Models with Simulation-Trained Latent-Conditioned Residual RL — arXiv 2609.30868 ｜ 2026-09-29[R]Cooperative Multi-Agent Vision-Language-Action Models via Reinforced Fine Tuning — arXiv 2609.36588 ｜11 tasks｜ 2026-10-01[R]DexPolicy: Scheduled Exploration for Trajectory-Guided Dexterous Manipulation — arXiv 2610.00360 ｜49.4%·68.1%·14.1%｜ 2026-10-03[R]GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed — arXiv 2609.39600 ｜72.42%·71.35%·61.75%｜ 2026-10-04
Simulation & evaluation (sim2real / closed-loop) (12)
[R]Beyond Binary Success: Sample-Efficient and Statistically Rigorous Robot Policy Comparison — Toyota Research Institute ｜70%·50%｜ 2026-09-28[R]TACTIC: Tactile and Vision Conditioned Contact-Centric Control for Whole-Arm Manipulation — Toyota Research Institute ｜ 2026-09-28[R]Capturing Visual Environment Structure Correlates with Control Performance — Toyota Research Institute ｜ 2026-09-29[R]F4R: Failure-Driven Recognition, Reconstruction, Refinement, and Redeployment for Continual Robot Self-Improvement — arXiv 2609.35575 ｜93.75%·90.0%｜ 2026-09-30-
[M]**[动易科技刘晨澔：人形机器人羽毛球对拉 40 拍背后的分层革命IROS 2026](https://www.leiphone.com/category/private/zoJ7chGiDi7IPHvN.html)** — 雷峰网 Leiphone ｜ 2026-09-30 [R]WorldLine: Action-Driven Visual Simulation for Robotic Manipulation — arXiv 2609.38059 ｜10,000 hours·2,000 hours·74%｜ 2026-10-01[R]RoboFin3D: A Sim-to-Real Platform for Robotic Surface Finishing — arXiv 2609.37560 ｜77.15%·84.47%·97.41%｜ 2026-10-01[R]Faster and Better? Benchmark Bugs and Design Limitations Distort the Evaluation of Vision-Language-Action Acceleration — arXiv 2609.37771 ｜ 2026-10-02[R]Token-World: World Modeling in Vision-Language Model Token Space for Robot Manipulation — arXiv 2610.00575 ｜ 2026-10-03[O]Isaac Lab on AWS: From Simulation to Registered Policy — AWS Physical AI Blog ｜ 2026-10-03[R]ReCo: Response-Consistent Locomotion with Policy-Aware MPC for Legged Manipulation — arXiv 2610.01612 ｜28.7%·27.4%｜ 2026-10-04[R]EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation — arXiv 2609.38905 ｜50.4%·83.5%·83.1%｜ 2026-10-04
Edge & real-time (on-device inference / control rate) (8)
[R]EdgeVLN: Runtime-Aware Deployment Ready Quantized Vision Language Navigation Model — arXiv 2609.35570 ｜16 GB·11.35 GB｜ 2026-09-30[R]RAVEL: Asynchronous Rolling Inference for Flow-Based Vision-Language-Action Models — arXiv 2609.34170 ｜ 2026-09-30[R]Urgent Actions Go First: Urgency-Aware Denoising for Real-Time VLA Control — arXiv 2609.37772 ｜1.89x｜ 2026-10-01[M]日立とファナックがフィジカルAI協業 日立エッジAI×ファナックロボ有効性も検証 — MONOist (ITmedia) ｜ 2026-10-01[R]DiffWAM: A Fast and Efficient Navigation World Action Model — arXiv 2609.39763 ｜74.40%｜ 2026-10-02[R]Kinematic MeanFlow: One-Step Action Generation Policy for Robotic Foundation Models — arXiv 2610.00864 ｜67.5%·74.4%·30.3%｜ 2026-10-03[R]IndoorBEV: A Lightweight Real-Time LiDAR BEV Perception System for Indoor Mobile Robots — arXiv 2610.00355 ｜0.6M parameters·169.6 ms·200 ms｜ 2026-10-03[R]Towards Agile Vision-Based Multi-UAV Flight: Revisiting State Estimation — arXiv 2609.39611 ｜40%·57%·300 ms｜ 2026-10-04
Safety, permissions & compliance (4)
[M]Gecko Robotics works with NVIDIA to add AI agent security and control — The Robot Report ｜ 2026-09-29[M]Your Robot’s Safety Functions Already Work. What If the Input Lies? — The Robot Report ｜ 2026-10-02[M]Inside Omron’s next-generation LD mobile robots — The Robot Report ｜ 2026-10-03[R]DODGER: Safety-Guided Reinforcement Learning for Robot Navigation Among Dynamic Obstacles — arXiv 2609.38873 ｜ 2026-10-04
Systems & orchestration (long-horizon / multi-policy) (3)
[M]Tackling construction labor shortages: ASI and SoftBank partner on autonomous fleets — The Robot Report ｜ 2026-10-01[M]ANYbotics launches Shift to streamline robot fleet operations, scale inspections — The Robot Report ｜ 2026-10-02[R]DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication — arXiv 2610.02161 ｜ 2026-10-04
Embodiment & supply chain (hardware / cost) (9)
[M]ダイフクが物流倉庫の完全自動化へヒューマノイドを開発、2030年に実証実験 — MONOist (ITmedia) ｜ 2026-09-28[M]三箭齐发！Sharpa三大新品IROS全球首秀，三位一体再次刷新灵巧操作上限 — 雷峰网 Leiphone ｜ 2026-09-29[M]本田重申Momenta合作，中国智驾全球布局提速 — 雷峰网 Leiphone ｜ 2026-09-29[M]ForceN to give a crash course on force and torque sensing for humanoids at RoboBusiness — The Robot Report ｜ 2026-09-30[M]Meet Flourish One, the Raspberry Pi-powered humanoid built for busy parents — The Robot Report ｜ 2026-09-30[M]Innodata opens motion-capture lab to help humanoids move more like people — The Robot Report ｜ 2026-10-01[M]Atlas Robot’s New Hand May Outperform Humanlike Designs — IEEE Spectrum Robotics ｜ 2026-10-02[M]Boston Dynamics drops pinkie on new humanoid hand — The Robot Report ｜ 2026-10-02[M]Video Friday: Albatross Falls, Spins, Self-Rights, and Sails Away — IEEE Spectrum Robotics ｜ 2026-10-03
This repository is technical observation. It is not investment advice, a product commitment, or a safety certification conclusion. Paper claims are author-reported.