Physical AI Radar · 2026-10-03

Language: ZH · EN · JA

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

• Generated: 2026-10-02 23:40 UTC
• Window: papers 2026-09-29 → 2026-10-03 · posts 2026-09-03 → 2026-10-03 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. Token-World: World Modeling in Vision-Language Model Token Space for Robot Manipulation

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2610.00575 ｜ 2026-09-30
• Signals: closed-loop success reported · latency or control-rate numbers
• Why it matters AI draft: If your world-model evaluator renders RGB frames only to re-encode them as policy inputs, this self-reported preprint argues that indirect interface is the weak point; the page flags closed-loop and latency results but quotes no figures, so it is not yet grounds to re-architect.

A common approach to world-model simulation for vision-language-action (VLA) systems is to predict future RGB observations and then re-encode them into policy inputs, introducing an indirect interface between simulation and downstream…

2. DexPolicy: Scheduled Exploration for Trajectory-Guided Dexterous Manipulation

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2610.00360 ｜ 2026-09-30
• Key numbers: 49.4% (“Target success rises from 49.4% to 68.1% (FPO)”) · 68.1% (“rises from 49.4% to 68.1% (FPO)”) · 14.1% (“14.1% to 45.4% (GRPO)”) · 45.4% (“14.1% to 45.4% (GRPO)”)
• Signals: quantitative results · closed-loop success reported · open code or weights
• Why it matters AI draft: For dexterous RL, the excerpt’s claim that exploration noise for finding contacts undermines precise object control is worth testing in your own setup; target success reportedly rises 49.4% to 68.1% (FPO) and 14.1% to 45.4% (GRPO), with no real-robot result shown.

Reinforcement learning (RL) for dexterous manipulation must discover finger-object contacts and then control the object precisely; the action noise that serves the first goal can interfere with the second.

3. Kinematic MeanFlow: One-Step Action Generation Policy for Robotic Foundation Models

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2610.00864 ｜ 2026-10-01
• Key numbers: 67.5% (“latency of GR00T-N1.6 by 67.5% ~74.4% across L40”) · 74.4% (“of GR00T-N1.6 by 67.5%~ 74.4% across L40 and”) · 30.3% (“end-to-end latency reductions of 30.3% ~54.9%”) · 54.9% (“latency reductions of 30.3%~ 54.9%”)
• Signals: quantitative results · open code or weights · latency or control-rate numbers
• Why it matters AI draft: Teams whose GR00T-N1.6 deployment is bottlenecked by multi-step flow matching may get latency back without new hardware: the authors self-report 67.5%–74.4% lower latency including on L40 and 30.3%–54.9% end-to-end, but no closed-loop success is shown, so verify task performance holds.

In this paper, we study how to achieve one-step action generation in Robotic Foundation Models (RFMs), aiming to overcome the high inference latency of multi-step flow matching.

4. IndoorBEV: A Lightweight Real-Time LiDAR BEV Perception System for Indoor Mobile Robots

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2610.00355 ｜ 2026-09-30
• Key numbers: 0.6M parameters (“IndoorBEV contains only 0.6M parameters and requires 2.3”) · 169.6 ms (“a mean latency of 169.6 ms under a 200”) · 200 ms (“169.6 ms under a 200 ms perception deadline”)
• Signals: quantitative results · real-robot experiments · open code or weights · latency or control-rate numbers
• Why it matters AI draft: For compute-constrained indoor mobile robots, a 0.6M-parameter LiDAR BEV stack averaging 169.6 ms against a 200 ms perception deadline is a candidate, but a mean that close to the deadline means measuring worst-case latency on your own hardware before adopting it.

Efficient indoor LiDAR perception is challenging because mobile robots must understand cluttered three-dimensional environments under strict latency and memory constraints.

5. Isaac Lab on AWS: From Simulation to Registered Policy

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [O] ｜ Source: AWS Physical AI Blog ｜ 2026-10-02
• Signals: real-robot experiments
• Why it matters AI draft: For teams training in Isaac Lab on AWS, this first-party post names a missing layer in Physical AI on AWS as systems move from research into production, and walks a path ending in a registered policy; it gives no figures, so read it as a workflow pattern, not a performance claim.

Introduction: The missing layer in Physical AI on AWS Physical AI systems, those that perceive, reason, and act in the physical world, are moving fast from research into production deployments.

6. Video Friday: Albatross Falls, Spins, Self-Rights, and Sails Away

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: IEEE Spectrum Robotics ｜ 2026-10-02
• Why it matters AI draft: Designers of sea-deployed platforms may note the quoted ALBATROSS concept: air-dropped, autorotating to the water without a parachute, passively self-righting, then reusing the same rigid wings as sails; it is a secondhand Video Friday item with no figures, so check the primary source.

We’ve developed ALBATROSS, a hybrid aerial–marine robot that can be released from the air, autorotate down to the water without a parachute, passively self-right after landing, and then reuse the same rigid wings as sails for autonomous…

7. Runway introduces Praxis-1 world action model for robotics

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-10-02
• Signals: open code or weights
• Why it matters AI draft: If you are choosing a world action model to build on, the report says Runway plans to ship Praxis-1 with open weights once it is publicly released; this is secondhand, and the page shows no release date or figures, so keep it on a watchlist rather than in a plan.

When Runway does release Praxis-1 publicly, it plans to ship it with open weights rather than as a closed model.

8. Inside Omron’s next-generation LD mobile robots

• Lane: Safety, permissions & compliance ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-10-02
• Why it matters AI draft: AMR buyers may want to weight the fleet-management and safety software as heavily as the vehicle, per the Omron project manager quoted here; that is one vendor’s view relayed secondhand, with no figures to back it.

In mobile robotics, software, safety, and fleet management are becoming just as important as the robot itself, according to an OMROM project manager.

Frontier baseline (curated)

1. Isaac Lab-Arena: reusable APIs to author, scale and analyse robot policy evaluations

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [O] ｜ Source: NVIDIA ｜ 2026
• Key numbers: policy evaluation at scale · solver-agnostic
• Why it matters: Evaluation moves from one-off scripts to a reusable pipeline - the engineering piece that turns ‘can demo’ into ‘can regress’.

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. 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.

4. RD-VLA: implicit test-time compute scaling via latent iterative reasoning

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2602.07845 ｜ 2026
• Key numbers: constant memory regardless of reasoning depth · up to 80x inference speedup vs token-based reasoning VLAs
• Signals: latency or control-rate numbers
• Why it matters: Robots can ‘think longer’ without paying token latency - the key to moving test-time compute inside the control loop.

5. RoboWorld: fast and reliable neural simulators for generalist robot policy evaluation

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2607.01060 ｜ 2026
• Key numbers: closed-loop rollouts in a video world model · VLM judge · rankings strongly correlated with real-world eval
• Why it matters: If rankings inside a world model align with real-robot rankings, the evaluation budget shifts from human hours to GPU hours - real-robot spot checks still cannot be dropped.

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] 2 · [R] 30 · [M] 20
• Picks per recent run: 09-27 8 · 09-28 4 · 09-29 8 · 09-30 8 · 10-01 8 · 10-02 8 · 10-03 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