Physical AI 前沿雷达 · 每周汇总 · 2026-W39
2026-09-21 → 2026-09-27 共发布 48 条，按主线分组。标题链接到原始来源，日期链接到当天的雷达。
- 证据标注：O = 一手官方 · R = 论文/预印本 · M = 媒体或二手
[O]21 ·[R]14 ·[M]13
基座模型（VLA / WAM） (2)
[O]Bringing a Frontier World Model to the Convenience Store: Inside Telexistence’s DreamZero Experiment on AWS — AWS Physical AI Blog ｜ 2026-09-25[R]Decoupled Early Exits for Task-Dependent Compute Allocation in Flow-Matching VLAs — arXiv 2609.29382 ｜ 2026-09-27
数据引擎（采集 / 清洗 / scaling） (4)
[O]How to Train a Cross-Embodiment Robot Navigation Policy with AI Agents — NVIDIA Technical Blog (Robotics) ｜ 2026-09-21[O]Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video — NVIDIA Blog ｜ 2026-09-22[R]MATE: Multi-Agent Virtual Teleoperation Platform for Humanoid Collaboration Data Collection — arXiv 2609.26520 ｜24.1 hours｜ 2026-09-25[R]Difference-Aware Retrieval Policies for Imitation Learning — Toyota Research Institute ｜46%｜ 2026-09-26
训练与自我改进（RL / 后训练） (7)
[O]Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps — Hugging Face Blog ｜ 2026-09-21[M]Video Friday: Humanoid Robot Takes On Monkey Bars — IEEE Spectrum Robotics ｜ 2026-09-22[O]Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL — Hugging Face Blog ｜ 2026-09-22[O]Jun Kim, oMLX creator and maintainer, joins Hugging Face to support the MLX community — Hugging Face Blog ｜ 2026-09-23[O]At AI Day Singapore, NVIDIA and Partners Showcase AI Advancements Across Southeast Asia — NVIDIA Blog ｜ 2026-09-24[R]Uncertainty-Gated Exploration Noise Suppresses Task Collapse in Online RL Fine-Tuning of a Flow-Matching Vision-Language-Action Policy — arXiv 2609.28838 ｜96.02%｜ 2026-09-27[R]Learning from Mixed-Quality Deployment Experience for Robot Manipulation — arXiv 2609.29000 ｜ 2026-09-27
仿真与评测（sim2real / 闭环） (11)
[O]Piloting the world’s first double-blind AI evaluations — Google DeepMind ｜ 2026-09-21[O]Boots on the Ground: ‘WARDOGS’ Goes All Out on GeForce NOW at Early-Access Launch — NVIDIA Blog ｜ 2026-09-22[O]BenchMIRT: What are LLM benchmarks actually measuring? — Hugging Face Blog ｜ 2026-09-22[O]How UK AISI and EvalEval Are Making Benchmark Results Reproducible — Hugging Face Blog ｜ 2026-09-23[O]How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows — Hugging Face Blog ｜ 2026-09-24[R]DreamStream: Towards Policy-Oriented Generative Simulation for End-to-End Driving — arXiv 2609.26792 ｜ 2026-09-25[R]VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models — arXiv 2609.25376 ｜ 2026-09-25[R]RoboRecover: Benchmarking Robot Policy Recovery under Execution Deviations — arXiv 2609.28952 ｜ 2026-09-26[R]Rolling-WAM: World Action Models with Rolling Imagination — arXiv 2609.30247 ｜4.5x｜ 2026-09-26[R]AnchorDream: Repurposing Video Diffusion for Embodiment-Aware Robot Data Synthesis — Toyota Research Institute ｜36.4%｜ 2026-09-27[M]SeeAct AI穆尧：具身智能终局，一定是从“被训练”走向“自我进化”｜物理AI50人 — 雷峰网 Leiphone ｜ 2026-09-27
边缘与实时性（端侧推理 / 控制频率） (8)
[O]NVIDIA Brings Real-Time AI to Broadcast, Sports and Global Streaming at IBC — NVIDIA Blog ｜ 2026-09-21[O]Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original — Hugging Face Blog ｜ 2026-09-21[M]Cyborg Roaches Can Stab You With Needles — IEEE Spectrum Robotics ｜ 2026-09-22[M]Barbara Mazzolai Wants to Build a New Field of Robotics — IEEE Spectrum Robotics ｜ 2026-09-23[M]This Robot Will Draw Your Blood Now — IEEE Spectrum Robotics ｜ 2026-09-23[R]Albireo: Adaptive, Energy-Efficient Inference Framework for Video Object Detection on the Edge — arXiv 2609.29648 ｜17.6%·14.4%·24.9%｜ 2026-09-26[R]Continuous Online Fault Detection for Mobile Robots via Adaptive Edge Models — arXiv 2609.29194 ｜4.30 ms｜ 2026-09-26[R]Exploiting answer-invariant redundancies in satellite imagery for efficient VLM inference on edge — arXiv 2609.29029 ｜78%·69%·45%｜ 2026-09-27
安全、权限与合规 (1)
[O]Why Deploying Physical AI at Scale Demands Safety at Every Layer — NVIDIA Blog ｜ 2026-09-22
系统与编排（长程 / 多策略） (7)
[O]Physical AI Takes the Wheel: How the World’s Robotaxi Leaders Are Building With NVIDIA Technologies — NVIDIA Blog ｜ 2026-09-21[O]Introducing agentic video understanding with Gemini — Google DeepMind ｜ 2026-09-21[O]Give Your Coding Agents a Memory You Own — Hugging Face Blog ｜ 2026-09-22[O]NVIDIA Isaac ROS 5.0 Advances Agentic, Open Source Robotics Development — NVIDIA Blog ｜ 2026-09-23[O]Accelerating a ROS 2 Node with an AI Agent and NVIDIA Isaac ROS — NVIDIA Technical Blog (Robotics) ｜ 2026-09-25[M]Isaac ROS 5.0 brings AI agents to robotics development, says NVIDIA — The Robot Report ｜ 2026-09-25[M]Qualcomm to acquire PickNik Robotics and keep MoveIt open-source — The Robot Report ｜ 2026-09-26
本体与供应链（硬件 / 成本） (8)
[M]The Best Way to Explore Lunar Craters Is a Giant Robot Ball — IEEE Spectrum Robotics ｜ 2026-09-21[O]NeoMME: an efficient Multimodal-native and Multilingual Encoder — Hugging Face Blog ｜ 2026-09-23[M]ANYbotics opens the door for inspections with ANYmal robots — The Robot Report ｜ 2026-09-25[M]Epson introduces AX6 cobot with compact design, no-code programming — The Robot Report ｜ 2026-09-25[M]Video Friday: Life’s Better With a Little Robot Goose — IEEE Spectrum Robotics ｜ 2026-09-26[M]Agility Robotics, maker of Digit humanoid, exploring wheeled robots — The Robot Report ｜ 2026-09-26[R]DAWN: Noise-Robust Quadruped Parkour via Depth-Denoising World Models — arXiv 2609.29092 ｜ 2026-09-27[M]这次云栖，斑马智能亮出了从汽车到具身智能的入场券 — 雷峰网 Leiphone ｜ 2026-09-27
本仓库为技术观察，不构成投资建议、产品承诺或安全认证结论。论文结论以作者自报为准。