Physical AI Radar · 2026-09-29

Language: ZH · EN · JA

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

• Generated: 2026-09-28 23:40 UTC
• Window: papers 2026-09-25 → 2026-09-29 · posts 2026-08-30 → 2026-09-29 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. NVIDIA Cosmos 3 on AWS: Omnimodal World Models for Physical AI

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [O] ｜ Source: AWS Physical AI Blog ｜ 2026-09-28
• Signals: real-robot experiments
• Why it matters AI draft: Teams whose perception, prediction or control models are starved for training data get a first-party follow-up to AWS’s earlier Cosmos-for-data-scarcity post; the page shows no figures, so hold any change to data budgets until measured results appear.

Introduction In our previous blog, we demonstrated how NVIDIA Cosmos world foundation models on AWS address the data scarcity problem in physical AI: the challenge of training robust perception, prediction, and control models when…

2. Fast Plans, Faithful Actions: Closing the Planning-Execution Gap in Hierarchical Vision-Language-Action Models

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2609.30833 ｜ 2026-09-25
• Key numbers: 91.0% (“on LIBERO-Long increases from 91.0% to 96.2%”) · 96.2% (“increases from 91.0% to 96.2%”) · 95.85% (“four suites increases from 95.85% to 98.45%”) · 98.45% (“increases from 95.85% to 98.45%”)
• Signals: quantitative results · closed-loop success reported · latency or control-rate numbers
• Why it matters AI draft: For anyone running a hierarchical VLA with a vision-language planner over an action expert, the authors report LIBERO-Long rising from 91.0% to 96.2%, but the page shows only benchmark deltas and no real-robot results, so reproduce before re-architecting.

Hierarchical vision-language-action (VLA) systems consist of a high-level vision-language planner and a low-level action expert that generates continuous actions.

3. 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

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: 雷峰网 Leiphone ｜ 2026-09-28
• Signals: quantitative results · latency or control-rate numbers
• Why it matters AI draft: Sharpa launching a tactile manipulation robot (D01), a hand (W02) and an exoskeleton data glove (AE01) together means hand hardware and tactile data capture could come from one vendor; this is a secondary report, so confirm specs and availability directly before shortlisting.

美国匹兹堡，9月28日——美国匹兹堡智能机器人与系统国际会议（IROS）现场，Sharpa正式发布三款重磅全自研新品：首个一体式触觉感知灵巧操作机器人D01、新一代全触觉超紧凑、轻量化灵巧手W02，以及高保真外骨骼触感数据手套AE01。

4. VLaRL: Augmenting Vision-Language-Action Models with Simulation-Trained Latent-Conditioned Residual RL

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2609.30868 ｜ 2026-09-25
• Signals: real-robot experiments
• Why it matters AI draft: If your VLA is imprecise in contact-rich manipulation, this preprint proposes a simulation-trained residual RL layer on top and flags real-robot experiments, but the page shows no figures, so it is not yet enough to justify adding that post-training stage.

Vision-language-action (VLA) models provide broad, instruction-conditioned manipulation behaviors, but their physical execution can remain imprecise during contact-rich interaction.

5. Praxis: Distilling Physical Interaction Priors from Egocentric Videos for Generalizable Whole-Body Manipulation

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [R] ｜ Source: arXiv 2609.30735 ｜ 2026-09-25
• Signals: closed-loop success reported
• Why it matters AI draft: Mobile-humanoid teams juggling workspace reach against hand-object contact that shifts with object pose may be able to draw interaction priors from egocentric video; the preprint flags closed-loop success but the page gives no figures, so treat it as a lead.

Mobile humanoid manipulation requires both reaching a usable workspace and preserving precise hand-object interactions as object poses and contact conditions change.

6. 本田重申Momenta合作，中国智驾全球布局提速

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: 雷峰网 Leiphone ｜ 2026-09-28
• Why it matters AI draft: Honda reportedly confirmed Momenta’s advanced driver-assistance stack for models launching from 2027, a timing marker for anyone judging whether Chinese autonomy suppliers win foreign production programmes; it is secondhand, so rely on Honda’s own statement before planning around it.

本田与Momenta此前达成的智能驾驶合作，在近期媒体沟通会上再次受到关注。 除了再次明确采用Momenta高阶智驾方案外，本田还给出了具体的量产时间表，称相关技术将从2027年上市的车型开始陆续导入。

7. Gecko Robotics works with NVIDIA to add AI agent security and control

• Lane: Safety, permissions & compliance ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-09-28
• Why it matters AI draft: Letting AI agents drive robot systems autonomously now has an off-the-shelf control layer to evaluate, with Gecko reportedly adopting NVIDIA’s new Open Agent Safety Platform; the report is secondhand with no figures, so its weight in a safety case is unproven.

Gecko Robotics is using the new NVIDIA Open Agent Safety Platform to ensure the secure autonomous control of systems.

8. Capturing Visual Environment Structure Correlates with Control Performance

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: Toyota Research Institute ｜ 2026-09-09
• Why it matters AI draft: Because policy rollouts are expensive even in simulation, TRI’s work suggests screening visual representations by how well they capture environment structure before committing to rollouts; the page gives no figures, so check how strong that correlation is before relying on it.

The choice of visual representation is key to scaling generalist robot policies. However, direct evaluation via policy rollouts is expensive, even in simulation.

Frontier baseline (curated)

1. 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.

2. XPENG commissions its humanoid production lines; IRON walks off the line autonomously

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [O] ｜ Source: PR Newswire ｜ 2026-09-08
• Key numbers: >80% of core processes automated (reported) · target 1,000+ units/month by year-end (reported)
• Why it matters: Line commissioning is a verifiable fact; the monthly target is reported second-hand. Separating the two is the easiest citation mistake to get caught on this cycle.

3. Dyna Robotics announces DYNA-2 world action model trained on 1M+ hours of human data (vendor claim)

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [M] ｜ Source: PR Newswire ｜ 2026-08-10
• Key numbers: 1M+ hours (~170 years of waking experience) · claimed first human-to-robot scaling law · claimed 20% -> 80-90% on high-precision manufacturing from pretraining scale alone
• Why it matters: Vendor-reported, yet directionally consistent with academic results: if a human-data scaling law holds, the data moat shifts from fleet size to collection and curation pipelines.

4. HumanEgo: zero-shot robot learning from minutes of human egocentric video

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [R] ｜ Source: arXiv 2605.24934 ｜ 2026
• Key numbers: 30 min/task -> 92.5% average success on 4 real tasks · 15 min -> 75% · +41% over matched-time teleoperation
• Why it matters: It pushes the cost per new task down to minutes of data - the most direct yardstick for whether your data strategy is behind.

5. 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.

6. 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.

Lane distribution (picks, last 7 days)

Recent cadence

• Source mix (picks, last 7 days): [O] 9 · [R] 21 · [M] 14
• Picks per recent run: 09-23 6 · 09-24 2 · 09-25 8 · 09-26 8 · 09-27 8 · 09-28 4 · 09-29 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