Physical AI Radar · 2026-09-27

Language: ZH · EN · JA

A daily radar on the Physical AI frontier: eight lanes, trilingual digests, traceable evidence.

• Generated: 2026-09-26 23:40 UTC
• Window: papers 2026-09-23 → 2026-09-27 · posts 2026-08-28 → 2026-09-27 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. Uncertainty-Gated Exploration Noise Suppresses Task Collapse in Online RL Fine-Tuning of a Flow-Matching Vision-Language-Action Policy

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2609.28838 ｜ 2026-09-26
• Key numbers: 96.02% (“under which 96.02% of the action”)
• Signals: quantitative results · open code or weights
• Why it matters AI draft: Anyone planning to keep fine-tuning a deployed flow-matching VLA online should monitor per-task competence, which the authors say continued updates often destroy; the page’s lone figure lacks its metric, but open code is flagged, so the self-reported claim can be checked.

Online reinforcement learning fine-tuning of pretrained flow-matching vision-language-action (VLA) policies promises robots that keep learning after deployment, but continued updates often destroy competence on individual tasks while the…

2. Exploiting answer-invariant redundancies in satellite imagery for efficient VLM inference on edge

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2609.29029 ｜ 2026-09-26
• Key numbers: 78% (“Rift reduces energy by 78% and latency by”) · 69% (“78% and latency by 69%”) · 45% (“while increasing accuracy from 45% to 73%”) · 73% (“accuracy from 45% to 73%”)
• Signals: quantitative results · latency or control-rate numbers
• Why it matters AI draft: Teams running VLMs on-device over high-resolution imagery may not need exhaustive tiled inference: the authors report 78% less energy and 69% lower latency with accuracy up from 45% to 73%, but on satellite queries, with nothing here on robot cameras or control loops.

Onboard vision-language models could enable satellites to answer queries directly, but exhaustive tiled inference over high-resolution imagery is slow and energy-intensive.

3. Decoupled Early Exits for Task-Dependent Compute Allocation in Flow-Matching VLAs

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2609.29382 ｜ 2026-09-26
• Signals: closed-loop success reported · latency or control-rate numbers
• Why it matters AI draft: For teams budgeting compute on a flow-matching VLA built from a VLM backbone plus action expert, per-task compute allocation may become a tuning knob; closed-loop and latency results are flagged, but with no figures on this page the trade-off cannot yet be sized.

Flow-matching Vision-Language-Action (VLA) models have emerged as a potential solution for generalist robot control, designed by combining a pretrained Vision-Language Model (VLM) backbone with an action expert that generates continuous…

4. Learning from Mixed-Quality Deployment Experience for Robot Manipulation

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2609.29000 ｜ 2026-09-26
• Signals: real-robot experiments
• Why it matters AI draft: If you log deployment runs, this preprint argues partial progress and failures are training material rather than data to filter out; real-robot experiments are flagged, but the page shows no figures, so the size of the gain remains unverified here.

Robot policies deployed in real environments naturally accumulate mixed-quality experience, including successful executions, partial progress, and failures.

5. AnchorDream: Repurposing Video Diffusion for Embodiment-Aware Robot Data Synthesis

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: Toyota Research Institute ｜ 2026-09-09
• Key numbers: 36.4% (“with relative gains of 36.4% in simulator benchmarks”)
• Signals: quantitative results · real-robot experiments
• Why it matters AI draft: Where demonstration collection is the bottleneck, video-diffusion data synthesis is worth a trial, but the only figure shown is a 36.4% relative gain in simulator benchmarks; real-robot experiments are flagged without numbers here, so don’t cut teleoperation budgets yet.

The collection of large-scale and diverse robot demonstrations remains a major bottleneck for imitation learning, as real-world data acquisition is costly and simulators offer limited diversity and fidelity with pronounced sim-to-real gaps.

6. DAWN: Noise-Robust Quadruped Parkour via Depth-Denoising World Models

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [R] ｜ Source: arXiv 2609.29092 ｜ 2026-09-26
• Why it matters AI draft: Quadruped teams whose vision-based gaits assume clean depth in training and hand-tuned filters at deployment should watch this; the page shows no figures or real-robot signal, so it is not yet grounds for replacing an existing filter pipeline.

Vision-based legged locomotion methods assume clean depth at training time and rely on hand-tuned post-processing filters at deployment.

7. 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

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [M] ｜ Source: 雷峰网 Leiphone ｜ 2026-09-24
• Why it matters AI draft: An interview whose core claim is that robots lack experience rather than data; useful as one framing when setting data strategy, but as secondary opinion with no figures it should not by itself shift collection spending.

机器人真正缺的，不是数据，而是经验。那是具身智能最荒芜的时候。

8. 这次云栖，斑马智能亮出了从汽车到具身智能的入场券

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: 雷峰网 Leiphone ｜ 2026-09-24
• Why it matters AI draft: A secondary report of Zebra positioning its automotive AI for embodied intelligence; the visible excerpt describes only an in-car assistant scene, with no robot product, price or delivery to plan sourcing around.

周五晚上六点，你坐进车里，什么也没说。 AI 系统已经知道，这周六是女儿的生日。

Frontier baseline (curated)

1. π*0.6 / Recap: a VLA that learns from experience via advantage-conditioned policies

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2511.14759 ｜ 2026
• Key numbers: RL from autonomous rollouts + human corrections · value function fine-tuned on online data
• Why it matters: The methodological basis for ‘deployment is training’: to use it, operations must first make intervention logs and success judgements first-class data.

2. Fine-tuning π0 for robotic manipulation on Amazon SageMaker HyperPod EKS

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [O] ｜ Source: AWS Physical AI Blog ｜ 2026-09-10
• Key numbers: MSE -89.7% (DROID) · MSE -88.9% (LIBERO) · 8xH100, 20k steps, 6h15m · 50-step chunk <200 ms at 1 ODE step · checkpoint ~45 GB
• Signals: quantitative results
• Why it matters: An earlier version of the same work reported ~96% because eval episodes overlapped training; enforcing the split moved it to ~89%. The best evaluation-discipline case of 2026.

3. Gemini Robotics 2 brings whole body intelligence to robots

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [O] ｜ Source: Google DeepMind ｜ 2026
• Key numbers: full humanoid control, feet to fingertips · bi-arm robots
• Why it matters: Whole-body control enters the generalist VLA feature list, so upper-body manipulation and lower-body balance stop living in two separate systems.

4. NVIDIA Cosmos 3: physical reasoning, world generation and action generation in one open model

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [O] ｜ Source: NVIDIA Technical Blog ｜ 2026-08
• Key numbers: single open model · Nano / Super / Edge variants
• Why it matters: Merging physical reasoning, world generation and action generation into one open model means ‘imagining’ and ‘acting’ stop being two stacks.

5. NVIDIA Jetson Thor: up to 2070 FP4 TFLOPS and 128 GB at 40-130 W

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [O] ｜ Source: NVIDIA ｜ 2026
• Key numbers: 7.5x AI compute vs Orin · 3.5x energy efficiency vs AGX Orin · Jetson T2000: 400 FP4 TFLOPS / 16 GB
• Signals: quantitative results · latency or control-rate numbers
• Why it matters: A generational jump in on-device compute creates room for world models and test-time compute, yet the real-time budget still has to be measured item by item.

6. Jetson-PI: onboard real-time control via foresight-aligned asynchronous inference

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2607.12659 ｜ 2026
• Key numbers: 8.66x control frequency vs naive PyTorch on Orin · +14.8% average success vs VLASH on LIBERO
• Signals: latency or control-rate numbers
• Why it matters: Nearly an order of magnitude in control frequency from systems engineering on the same hardware - deployment optimisation often has more headroom than swapping models.

Lane distribution (picks, last 7 days)

Recent cadence

• Source mix (picks, last 7 days): [O] 21 · [R] 14 · [M] 13
• Picks per recent run: 09-21 8 · 09-22 8 · 09-23 6 · 09-24 2 · 09-25 8 · 09-26 8 · 09-27 8

Standards & compliance watch

• ISO 25785-1 — Safety requirements for dynamically stable industrial mobile robots — Under development; committee-stage as of 2026-09.
• Regulation (EU) 2023/1230 (Machinery Regulation) — Fully applicable 2027-01-20.
• EU AI Act official timeline (Council of the EU) — High-risk deadlines 2027-12-02 / 2028-08-02 per Digital Omnibus.

Per-item analysis on this page was drafted by radar-analyst (claude-opus-5.5) and merged automatically after deterministic checks and an agent review, without human review; titles and numbers are untouched.

This repository is technical observation. It is not investment advice, a product commitment, or a safety certification conclusion. Paper claims are author-reported.

Methodology · Weekly roundup · Archive