Physical AI Radar · 2026-09-30

Language: ZH · EN · JA

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

• Generated: 2026-09-29 23:40 UTC
• Window: papers 2026-09-26 → 2026-09-30 · posts 2026-08-31 → 2026-09-30 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. EdgeVLN: Runtime-Aware Deployment Ready Quantized Vision Language Navigation Model

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2609.35570 ｜ 2026-09-28
• Key numbers: 16 GB (“NVIDIA Jetson Orin NX 16 GB”) · 11.35 GB (“using 11.35 GB resident memory while”)
• Signals: quantitative results · closed-loop success reported · real-robot experiments · latency or control-rate numbers
• Why it matters AI draft: Teams sizing vision-language navigation for a memory- and power-constrained module like the Jetson Orin NX 16 GB get a self-reported 11.35 GB resident-memory figure to budget against; it is a preprint claim, so measure on your own stack before choosing hardware.

Vision-language navigation (VLN) models perform well but target compute-rich platforms, limiting deployment on memory- and power-constrained robotic edge devices.

2. ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2609.34982 ｜ 2026-09-28
• Key numbers: 9.8% (“success rates by absolute 9.8%”) · 6.1% · 11.4% (“and 11.4%”)
• Signals: quantitative results · closed-loop success reported · real-robot experiments · open code or weights
• Why it matters AI draft: If you want a sturdier VLA manipulation policy through fine-tuning rather than retraining, this preprint reports success rates moving by an absolute 9.8% and the page flags an open code or weights release, so a cheap reproduction on your own tasks comes before adoption.

Vision-Language-Action (VLA) models have shown great promise for robotic manipulation by mapping multi-modal semantics to physical actions.

3. F4R: Failure-Driven Recognition, Reconstruction, Refinement, and Redeployment for Continual Robot Self-Improvement

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2609.35575 ｜ 2026-09-28
• Key numbers: 93.75% (“show that F4R achieves 93.75% In-Distribution and 90.0%”) · 90.0% (“achieves 93.75% In-Distribution and 90.0% Out-of-Distribution (OOD) success”)
• Signals: quantitative results · closed-loop success reported · real-robot experiments
• Why it matters AI draft: When expert demonstrations cover too little, a loop that learns from failures and redeploys is an alternative to collecting more of them; the authors self-report 93.75% in-distribution and 90.0% out-of-distribution success, a narrow gap worth verifying before planning around it.

The real-world performance of current vision-language-action models is fundamentally constrained by the limited coverage of expert demonstrations and their insufficient understanding of physical interactions.

4. Rethinking Causal Action Tokenization with Conditional Annealing in Flow Matching

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2609.35469 ｜ 2026-09-28
• Signals: closed-loop success reported · real-robot experiments
• Why it matters AI draft: Teams on autoregressive VLAs should treat the action tokenizer as a design choice rather than a compression step, as the authors argue current ones are semantically misaligned; with no figures on the page, it is not yet grounds to swap tokenizers.

Autoregressive Vision-Language-Action (VLA) models offer a scalable path to robot learning, yet existing action tokenizers treat tokenization as a compression problem, producing representations that are semantically misaligned with the…

5. RAVEL: Asynchronous Rolling Inference for Flow-Based Vision-Language-Action Models

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2609.34170 ｜ 2026-09-28
• Signals: closed-loop success reported · real-robot experiments · latency or control-rate numbers
• Why it matters AI draft: If costly VLM encoding and multi-step action generation are what slow your flow-based VLA, asynchronous inference is a direction to track, but the page lists no concrete figures, so there is nothing yet to size a control-loop budget against.

Flow-based vision-language-action (VLA) models are highly effective for generalist robot manipulation, yet their reliance on computationally expensive VLM encoding and multi-step iterative action generation imposes a significant latency…

6. ForceN to give a crash course on force and torque sensing for humanoids at RoboBusiness

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-09-29
• Why it matters AI draft: For humanoid builders, where force and torque sensors go and how to keep them from breaking are design-stage questions; this is a secondary preview of a conference session with no product data, so treat it as a checklist prompt, not evidence.

Attendees will learn about the many ways they can use force and torque sensors in humanoids and how to avoid breaking them while doing it.

7. Meet Flourish One, the Raspberry Pi-powered humanoid built for busy parents

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-09-29
• Why it matters AI draft: A home robot priced at $3,555 that relies on phone-based and cloud-trained chore automation makes connectivity part of its reliability story for buyers; this is secondary coverage of an announced release, so shipping status still needs a primary source.

Flourish is releasing a $3,555 phone-trainable home robot, targeting busy parents with personalized, cloud-trained chore automation.

8. 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

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [M] ｜ Source: 雷峰网 Leiphone ｜ 2026-09-29
• Why it matters AI draft: High-dynamic humanoid motion is described as reaching industrial-grade progress, but this is secondary conference coverage with no method detail or evaluation data on the page, so wait for primary material before it informs a motion-control choice.

顶会上的人形机器人高动态运动场景的最新工业级进展。羽毛球大概是所有球类里最“不讲道理”的项目。

Frontier baseline (curated)

1. 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.

2. Egocentric human video can outperform real-robot data for embodied pretraining

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [R] ｜ Source: arXiv 2606.20521 ｜ 2026
• Key numbers: -24% validation loss at equal data · +52.5% in-distribution success · +90% out-of-distribution success
• Why it matters: Teams without a fleet can still build a data advantage; teams with a fleet but no curation and judging pipeline hold data that is only a storage cost.

3. EgoScale: scaling dexterous manipulation with diverse egocentric human data

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [R] ｜ Source: arXiv 2602.16710 ｜ 2026
• Key numbers: 44k hours of egocentric video · +54% average success vs no-pretraining on a 22-DoF hand · transfers to lower-DoF hands
• Why it matters: Large-scale human motion is shown to be a reusable, embodiment-agnostic motor prior, so dexterous hands do not start from zero.

4. Post-train NVIDIA Cosmos 3 Edge for on-device robot control

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [O] ｜ Source: NVIDIA Technical Blog ｜ 2026-08
• Key numbers: 4B omni-model · 2B Nemotron reasoner · runs on Jetson Thor · receding-horizon control loop
• Signals: closed-loop success reported · open code or weights
• Why it matters: The first reproducible path for a world model on-device: post-train, serve on Thor, then evaluate closed-loop, with code and checkpoint released.

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. SHIELD: safety on humanoids via control barrier functions in expectation on learned dynamics

• Lane: Safety, permissions & compliance ｜ Evidence: [R] ｜ Source: arXiv 2505.11494 ｜ 2026
• Key numbers: runtime constraint specification · no retraining when constraints change
• Why it matters: Makes ‘change the constraint without retraining’ possible - the academic counterpart of a local authority gate.

Lane distribution (picks, last 7 days)

Recent cadence

• Source mix (picks, last 7 days): [O] 5 · [R] 26 · [M] 15
• Picks per recent run: 09-24 2 · 09-25 8 · 09-26 8 · 09-27 8 · 09-28 4 · 09-29 8 · 09-30 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