Physical AI Radar · 2026-10-01

Language: ZH · EN · JA

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

• Generated: 2026-09-30 23:40 UTC
• Window: papers 2026-09-27 → 2026-10-01 · posts 2026-09-01 → 2026-10-01 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. WorldLine: Action-Driven Visual Simulation for Robotic Manipulation

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2609.38059 ｜ 2026-09-29
• Key numbers: 10,000 hours (“dynamics from more than 10,000 hours of action-free robot”) · 2,000 hours (“grounds them using over 2,000 hours of action trajectories”) · 74% (“predicts trajectory success with 74% mean accuracy across”)
• Signals: quantitative results · closed-loop success reported · real-robot experiments
• Why it matters AI draft: If real-robot time is your bottleneck for screening candidate behaviors, the authors’ self-reported 74% mean accuracy at predicting trajectory success hints at a pre-filter, but as a preprint claim it is a triage aid, not a substitute for real evaluation.

Real-world robot learning is constrained by the cost of collecting experience and evaluating candidate behaviors.

2. RoboFin3D: A Sim-to-Real Platform for Robotic Surface Finishing

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2609.37560 ｜ 2026-09-29
• Key numbers: 77.15% (“of unsanded regions from 77.15% to 84.47%”) · 84.47% (“regions from 77.15% to 84.47%”) · 97.41% (“and real training reaches 97.41%”)
• Signals: quantitative results · real-robot experiments
• Why it matters AI draft: Where every grinding or sanding trial consumes a workpiece, this preprint quotes a figure on unsanded regions going from 77.15% to 84.47% and real training reaching 97.41%; check what each measures before moving trials into simulation.

Grinding and sanding are fundamental processes in industrial robotic surface finishing. However, physical trials are expensive and consume workpieces, making reproducible experiments difficult.

3. Cooperative Multi-Agent Vision-Language-Action Models via Reinforced Fine Tuning

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2609.36588 ｜ 2026-09-29
• Key numbers: 11 tasks (“and $π_{0.5}$ backbones across 11 tasks in RoboTwin”)
• Signals: closed-loop success reported · real-robot experiments
• Why it matters AI draft: Teams planning several cooperating VLA agents should note this preprint treats single-agent pretraining as the obstacle and tests RL fine-tuning across 11 RoboTwin tasks; the page quotes no success rate, so check the size of any gain in the paper.

We study reinforcement learning (RL) methods for cooperative multi-agent Vision-Language-Action (VLA) models. This problem is challenging because VLAs are pretrained on large-scale single-agent data and therefore lack the fine-grained…

4. Urgent Actions Go First: Urgency-Aware Denoising for Real-Time VLA Control

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2609.37772 ｜ 2026-09-29
• Key numbers: 1.89x (“achieves up to a 1.89x speedup in average”)
• Signals: closed-loop success reported · real-robot experiments · latency or control-rate numbers
• Why it matters AI draft: For diffusion or flow-matching VLA policies whose iterative denoising eats the control budget, the authors claim up to a 1.89x average speedup; the quote names no hardware, so benchmark on your own controller before relaxing latency budgets.

Diffusion and flow-matching Vision-Language-Action (VLA) policies generate action chunks through iterative denoising, incurring substantial inference latency that severely limits real-time robotic control.

5. Innodata opens motion-capture lab to help humanoids move more like people

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-09-30
• Signals: real-robot experiments
• Why it matters AI draft: Humanoid teams buying motion data have another outside supplier: per this secondary report, Innodata records movement with Vicon cameras at sub-millimeter accuracy as robot training data; ask how it maps onto your robot’s body before signing.

Innodata uses Vicon cameras to measure movements with sub-millimeter accuracy and to collect real-world training data for robots.

6. Tackling construction labor shortages: ASI and SoftBank partner on autonomous fleets

• Lane: Systems & orchestration (long-horizon / multi-policy) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-09-30
• Why it matters AI draft: Contractors running heavy equipment from several brands should note the reported $225M push for brand-agnostic autonomy on infrastructure projects; this is a secondary report of a partnership and shows nothing yet about field reliability.

ASI and SoftBank join forces with $225M to scale brand-agnostic autonomous heavy equipment across infrastructure projects.

7. 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

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [M] ｜ Source: MONOist (ITmedia) ｜ 2026-09-30
• Why it matters AI draft: Plants built around FANUC arms may soon be offered Hitachi AI as a jointly backed package, but the excerpt describes a partnership with joint demonstrations and gives no results, so wait for validation figures before planning around it.

日立製作所とファナックはフィジカルAI（人工知能）実装の共同事業化に向けた戦略的パートナーシップを締結した。日立のAI技術とファナックの産業用ロボットを組み合わせたフィジカルAIの共同実証と共同展開を推進する。

8. 从海拉鲁到现实世界：视频模型如何理解「变化」

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [M] ｜ Source: 雷峰网 Leiphone ｜ 2026-09-30
• Why it matters AI draft: If you are weighing video models as world models, this secondary piece shows game-scene tests such as fire spreading across grass in Zelda; that is qualitative plausibility, and the excerpt says nothing about robot actions, so it should not move a data budget.

HiDream-O1-Video 的三场世界回应测试。在《塞尔达传说：旷野之息》中，玩家点燃一片草地，火焰会沿着草叶蔓延，热空气随之升起。

Frontier baseline (curated)

1. The reliability gap: industrial buyers want 95-99% uptime; most humanoids run 30-90 minutes

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: robozaps / industry analyses ｜ 2026
• Key numbers: 95-99% expected uptime · 30-90 min continuous operation · MTBI and autonomy rate are the metrics that matter
• Signals: quantitative results
• Why it matters: The 2027 rate limiter is unlikely to be model capability; it is mean time between interventions plus the certification gap.

2. GR00T N1.7: open reasoning VLA model for humanoid robots

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [O] ｜ Source: NVIDIA / Hugging Face ｜ 2026
• Key numbers: 3B parameters · open weights
• Why it matters: A 3B open reasoning VLA gives in-house teams a fine-tunable, deployable starting point instead of building a foundation model.

3. 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.

4. Bringing a frontier world model to the convenience store: Telexistence’s DreamZero experiment on AWS

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [O] ｜ Source: AWS Physical AI Blog ｜ 2026-09-09
• Key numbers: 14B world action model · best closed-loop 45% · real-data-trained run 0% in sim · MSE -55.6% from 1k to 2k steps · ~583 ms/step on 8xB200 · curation flagged 39.3% vs human 49.7%
• Signals: quantitative results
• Why it matters: Turns ‘is a frontier model right for my robot’ into measured numbers, and honestly records the 0% when real data was evaluated in sim - the domain-gap diagnosis is itself the deliverable.

5. 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.

6. 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.

Lane distribution (picks, last 7 days)

Recent cadence

• Source mix (picks, last 7 days): [O] 3 · [R] 30 · [M] 19
• Picks per recent run: 09-25 8 · 09-26 8 · 09-27 8 · 09-28 4 · 09-29 8 · 09-30 8 · 10-01 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