Physical AI Radar · 2026-09-26

Language: ZH · EN · JA

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

• Generated: 2026-09-26 01:40 UTC
• Window: papers 2026-09-22 → 2026-09-26 · posts 2026-08-27 → 2026-09-26 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. RoboRecover: Benchmarking Robot Policy Recovery under Execution Deviations

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2609.28952 ｜ 2026-09-25
• Signals: closed-loop success reported
• Why it matters AI draft: If your evaluation suite only starts policies from clean predefined states, as the authors say most benchmarks do, it says nothing about recovery after a mid-task deviation; the preprint reports closed-loop results, but no figures appear on this page.

Robot-policy benchmarks increasingly cover diverse tasks and preset out-of-distribution conditions, but typically evaluate complete trajectories from predefined initial states.

2. Albireo: Adaptive, Energy-Efficient Inference Framework for Video Object Detection on the Edge

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2609.29648 ｜ 2026-09-25
• Key numbers: 17.6% (“while reducing energy by 17.6% (Thor) and 14.4%”) · 14.4% (“by 17.6% (Thor) and 14.4% (Orin) and per-frame”) · 24.9% (“per-frame energy-delay product by 24.9% and 26.1%”) · 26.1% (“product by 24.9% and 26.1%”)
• Signals: quantitative results · open code or weights · latency or control-rate numbers
• Why it matters AI draft: For a perception stack that runs a detector on every frame of a long stream, the authors report 17.6% (Thor) and 14.4% (Orin) less energy with code released; nothing on this page ties the result to a robot task or control loop.

Video object detection on edge devices runs computationally expensive detectors over long frame streams, causing high energy consumption and sustained GPU utilization.

3. Rolling-WAM: World Action Models with Rolling Imagination

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2609.30247 ｜ 2026-09-25
• Key numbers: 4.5x (“it delivers a 4.5x steady-state replanning speedup”)
• Signals: closed-loop success reported · real-robot experiments · latency or control-rate numbers
• Why it matters AI draft: If you shelved world action models because joint video-action denoising made each replanning cycle too slow, the authors’ self-reported 4.5x steady-state replanning speedup with real-robot and closed-loop results reopens the question; success figures are not on this page.

World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action…

4. Continuous Online Fault Detection for Mobile Robots via Adaptive Edge Models

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2609.29194 ｜ 2026-09-25
• Key numbers: 4.30 ms (“the Student achieves a 4.30 ms CPU inference latency”)
• Signals: quantitative results · real-robot experiments · latency or control-rate numbers
• Why it matters AI draft: A fault detector you can keep adapting online at a self-reported 4.30 ms per inference on CPU fits beside a control loop without a GPU; real-robot results are reported, but the page shows no closed-loop test of acting on a detection.

Mobile robots require robust, real-time fault detection capable of continuous adaptation on constrained edge hardware.

5. Video Friday: Life’s Better With a Little Robot Goose

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: IEEE Spectrum Robotics ｜ 2026-09-25
• Why it matters AI draft: This is a video roundup, and the only text on the page names a small robot goose from The Robot Works; treat it as a place to spot hardware leads to verify elsewhere, not as evidence for any spec, price or shipment claim.

Life’s better with a little Goose. [ The Robot Works ] Thanks, Dave!

6. Difference-Aware Retrieval Policies for Imitation Learning

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [R] ｜ Source: Toyota Research Institute ｜ 2026-09-09
• Key numbers: 46% (“consistent performance improvements of 15-46% over standard behavior”)
• Signals: quantitative results
• Why it matters AI draft: If your behavior-cloned policy drifts off-distribution and compounds errors, the authors report a retrieval-based alternative gaining 15-46% over standard behavior cloning; the page gives no real-robot or closed-loop result, so verify on your own setup before switching.

Parametric imitation learning via behavior cloning can suffer from poor generalization to out-of-distribution states due to compounding errors during deployment.

7. Agility Robotics, maker of Digit humanoid, exploring wheeled robots

• Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-09-25
• Why it matters AI draft: A single trade report that the Digit maker is exploring wheeled form factors for different environments suggests even humanoid-first vendors are not betting on legs alone; if you are specifying a fleet, ask vendors which form factor is actually shipping.

Agility Robotics is exploring a range of robot designs and form factors, including wheels, to address different operating environments and customer needs.

8. Qualcomm to acquire PickNik Robotics and keep MoveIt open-source

• Lane: Systems & orchestration (long-horizon / multi-policy) ｜ Evidence: [M] ｜ Source: The Robot Report ｜ 2026-09-23
• Signals: open code or weights
• Why it matters AI draft: If MoveIt is in your manipulation stack, the report that Qualcomm will keep it open-source while integrating its Dragonwing and Arduino platforms means the near-term risk is roadmap tilt toward one silicon vendor, not a license change; confirm with the primary announcement.

Qualcomm plans to integrate the Dragonwing and Arduino platforms with the MoveIt open-source manipulation framework.

Frontier baseline (curated)

1. RoboHarness: memory-driven orchestration of heterogeneous robot policies

• Lane: Systems & orchestration (long-horizon / multi-policy) ｜ Evidence: [R] ｜ Source: arXiv 2607.18060 (Huawei Noah’s Ark Lab) ｜ 2026
• Key numbers: VLA + RL + TAMP as agentic skills · capability-aware routing · Memory Bridge for policy handoff without retraining
• Why it matters: Long-horizon systems are an orchestrator plus a policy pool; the operated object becomes a set of declarable capability boundaries rather than one model.

2. 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.

3. 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.

4. 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.

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. ISO 25785-1 (under development): safety for dynamically stable industrial mobile robots

• Lane: Safety, permissions & compliance ｜ Evidence: [O] ｜ Source: ISO ｜ 2026
• Key numbers: committee stage as of 2026-09 · covers legged / self-balancing robots
• Why it matters: Humanoids break the fixed-base premise of ISO 10218; until the dedicated standard lands, safety rests on architecture and runtime guarantees.

Lane distribution (picks, last 7 days)

Recent cadence

• Source mix (picks, last 7 days): [O] 24 · [R] 8 · [M] 15
• Picks per recent run: 09-20 8 · 09-21 8 · 09-22 8 · 09-23 6 · 09-24 2 · 09-25 8 · 09-26 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-fable-5.1) 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