Physical AI Radar · 2026-10-02

Language: ZH · EN · JA

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

• Generated: 2026-10-01 23:40 UTC
• Window: papers 2026-09-28 → 2026-10-02 · posts 2026-09-02 → 2026-10-02 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. DSDyn-VLA: A Dual-Stream Dynamic Manipulation Framework with Motion Perception, Future Awareness, and Realtime Correction

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2609.39198 ｜ 2026-09-30
• Signals: closed-loop success reported · real-robot experiments · open code or weights · latency or control-rate numbers
• Why it matters AI draft: Teams handling moving objects, such as conveyor-belt picking, should note this self-reported preprint targets the dynamic case where the excerpt says VLAs struggle; the page flags real-robot and closed-loop work but quotes no figures, so treat it as a lead.

While Vision-Language-Action (VLA) models excel in static tasks, they struggle in dynamic environments where objects are in motion (e.g., conveyor belt manipulation).

2. DiffWAM: A Fast and Efficient Navigation World Action Model

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2609.39763 ｜ 2026-09-30
• Key numbers: 74.40% (“endpoint success rate of 74.40% on the 1”)
• Signals: quantitative results · closed-loop success reported · real-robot experiments · latency or control-rate numbers
• Why it matters AI draft: For UAV navigation built on video foundation models, the excerpt names costly future-video synthesis and geometric reconstruction as the bottleneck; the 74.40% endpoint success rate is author-reported and the page does not show which benchmark, so reproduce before adopting.

Pretrained video foundation models encode rich semantic and spatiotemporal priors for embodied navigation, yet converting these priors into UAV motion typically requires expensive future-video synthesis and geometric reconstruction.

3. Faster and Better? Benchmark Bugs and Design Limitations Distort the Evaluation of Vision-Language-Action Acceleration

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2609.37771 ｜ 2026-09-29
• Signals: closed-loop success reported · open code or weights · latency or control-rate numbers
• Why it matters AI draft: Anyone picking a VLA acceleration method from simulated manipulation benchmark results should hold off: this preprint argues flaws in those benchmarks skew the comparison, though the page shows no figures to size the distortion.

Simulated manipulation benchmarks are the standard tool for evaluating vision-language-action (VLA) policies and the acceleration methods that reduce their inference latency for on-robot deployment.

4. Toward Real-Time VLAs: Stage-Aware Two-Step Flow Denoising and System-Level Evaluation

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2609.39822 ｜ 2026-09-30
• Key numbers: 61.557 ms (“and model-inference time from 61.557 ms to 21.956 ms”) · 21.956 ms (“from 61.557 ms to 21.956 ms”)
• Signals: quantitative results · closed-loop success reported · latency or control-rate numbers
• Why it matters AI draft: Where inference latency keeps a VLA from matching high-rate execution, the author-reported cut in model-inference time from 61.557 ms to 21.956 ms is worth reproducing; the page shows no real-robot result, so on-hardware gains remain unverified.

Vision-language-action (VLA) models face a timing gap between low-rate inference and high-rate robot execution.

5. Atlas Robot’s New Hand May Outperform Humanlike Designs

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: IEEE Spectrum Robotics ｜ 2026-10-01
• Why it matters AI draft: For humanoid buyers, the signal is that Boston Dynamics redesigned the Atlas hand because the three-finger predecessor was never meant for mass production; this secondary report gives no figures, so price and volume remain open.

Today, Boston Dynamics is announcing a redesigned hand for its Atlas robot . The previous generation could do some crazy superhuman things with its three fingers, but it was never intended for mass production in the tens or hundreds of…

6. Boston Dynamics drops pinkie on new humanoid hand

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-10-01
• Why it matters AI draft: Hand designers can borrow the requirements method reported here, engineers taping pinkie and ring finger together for a day to find what they could not do; the secondary piece gives no task data, so the cost of fewer fingers is unquantified.

Boston Dynamics engineers taped their pinkies and ring fingers together for a day to learn what they weren’t able to do.

7. Your Robot’s Safety Functions Already Work. What If the Input Lies?

• Lane: Safety, permissions & compliance ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-10-01
• Why it matters AI draft: Safety teams should bring input integrity into the safety case: the piece argues a robot can obey its safety rules yet act on manipulated information, and urges stronger testing and evidence; it is secondary commentary without figures.

A robot can follow its safety rules and still act on manipulated information. See what this means for robot safety and how teams can strengthen their testing and evidence.

8. ANYbotics launches Shift to streamline robot fleet operations, scale inspections

• Lane: Systems & orchestration (long-horizon / multi-policy) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-10-01
• Why it matters AI draft: For plants running inspection robots, the evaluation question moves from collecting data to whether detected equipment failures become work orders for maintenance teams; this launch report has no figures, so fleet-scale reliability is not yet shown.

ANYbotics designed Shift to help maintenance teams run industrial inspections, detect failing equipment, and turn data into work orders.

Frontier baseline (curated)

1. RoboTTT: context scaling for robot policies via test-time training

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2607.15275 (NVIDIA) ｜ 2026-07-16
• Key numbers: 8K timestep visuomotor context · 3 orders of magnitude beyond prior policies · no inference-latency growth
• Signals: latency or control-rate numbers
• Why it matters: Long context enters the policy without added latency, so one-shot in-context imitation and on-the-fly improvement can live in the same inference pass.

2. 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.

3. 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.

4. Newton: open-source GPU physics engine for robotics (NVIDIA, Google DeepMind, Disney Research)

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [O] ｜ Source: NVIDIA Developer ｜ 2026
• Key numbers: built on Warp + OpenUSD · MuJoCo Warp up to 152x (locomotion) / 313x (manipulation) vs MJX on RTX 4090
• Signals: open code or weights
• Why it matters: A differentiable new simulation base plus Isaac Lab’s multi-backend switching lets training and evaluation reuse one actuator configuration.

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. Z-1: efficient RL post-training for flow-based VLA models

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2606.31846 ｜ 2026
• Key numbers: 80.6% average success on 24 RoboCasa tasks · +13.2 points over SFT init · public demos only
• Why it matters: Systematic GRPO post-training delivers double-digit gains without extra private demonstrations, breaking the default belief that only more data helps.

Lane distribution (picks, last 7 days)

Recent cadence

• Source mix (picks, last 7 days): [O] 1 · [R] 31 · [M] 20
• Picks per recent run: 09-26 8 · 09-27 8 · 09-28 4 · 09-29 8 · 09-30 8 · 10-01 8 · 10-02 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