Physical AI Radar · 2026-10-04

Language: ZH · EN · JA

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

• Generated: 2026-10-03 23:40 UTC
• Window: papers 2026-09-30 → 2026-10-04 · posts 2026-09-04 → 2026-10-04 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. ReCo: Response-Consistent Locomotion with Policy-Aware MPC for Legged Manipulation

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2610.01612 ｜ 2026-10-01
• Key numbers: 28.7% (“root-mean-square error (RMSE) by 28.7% and 27.4% relative”) · 27.4% (“(RMSE) by 28.7% and 27.4% relative to the”)
• Signals: quantitative results · closed-loop success reported · real-robot experiments
• Why it matters AI draft: If you pair an RL locomotion policy with MPC to track the end effector while the base walks, this self-reported preprint claims RMSE cuts of 28.7% and 27.4% with real-robot and closed-loop signals; the quoted text does not show the baseline, so check it first.

Continuous legged manipulation requires accurate end-effector tracking while the base keeps walking. Combining reinforcement learning (RL) with model predictive control (MPC) suits this task: the learned policy provides robust locomotion,…

2. DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication

• Lane: Systems & orchestration (long-horizon / multi-policy) ｜ Evidence: [R] ｜ Source: arXiv 2610.02161 ｜ 2026-10-01
• Signals: closed-loop success reported
• Why it matters AI draft: For teams taking VLM- or VLA-driven robots beyond single-robot settings, this preprint targets exactly that gap, but the page flags only closed-loop results with no figures or real-robot signal, so it is one to track rather than a design input.

Vision-language models (VLMs) and vision-language-action models (VLAs) have recently driven rapid progress in general-purpose robots, yet most progress has focused on single-robot settings.

3. ChunkVLA-AM: Parallel Action Chunking for Vision-Language-Action Robot Control in Additive Manufacturing

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2610.01856 ｜ 2026-10-01
• Key numbers: 92.9% (“succeeded in 39 ( 92.9% )”)
• Signals: quantitative results · closed-loop success reported · real-robot experiments
• Why it matters AI draft: Anyone weighing VLA control for additive-manufacturing cells gets a self-reported 92.9% success figure with real-robot and closed-loop signals, though the page does not show the task mix or comparison, which matter before scoping a pilot.

Vision-language-action (VLA) models unify visual perception, language understanding, and action generation, offering new opportunities for automation in additive manufacturing (AM).

4. IronMind: Scaling Humanoid Dexterous Manipulation via Camera-Space Ego-Centric Pretraining

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [R] ｜ Source: arXiv 2609.39403 ｜ 2026-09-30
• Key numbers: 10,000 hours (“budgets from 250 to 10,000 hours”) · 55.0% (“000-hour model achieves a 55.0% success rate”) · 11.7% (“compared with at most 11.7% for every pretraining”) · 5,000 hours (“pretraining budget up to 5,000 hours and 5.0% without”)
• Signals: quantitative results · closed-loop success reported
• Why it matters AI draft: If you are budgeting egocentric human-video collection for humanoid hands, this self-reported preprint suggests the payoff appears only at scale, 55.0% success against at most 11.7% for budgets up to 5,000 hours; the page flags no real-robot experiments.

Egocentric human video offers a scalable data source for dexterous manipulation, yet using it to train humanoid robots presents two challenges: (1) an embodiment gap, as human hands differ structurally from robot end-effectors and…

5. EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2609.38905 ｜ 2026-09-30
• Key numbers: 50.4% (“autonomous simulation success from 50.4% to 83.5% over”) · 83.5% (“success from 50.4% to 83.5% over two RSI”) · 83.1% (“EmbodiRSI achieves 83.1% scene-balanced autonomous real-world”) · 75.0% (“compared with 75.0% for adaptation using”)
• Signals: quantitative results · real-robot experiments
• Why it matters AI draft: When adapting manipulation policies, this self-reported preprint argues the next data round should follow the policy’s current failure modes, reporting simulation success rising from 50.4% to 83.5% over two RSI rounds and 83.1% real-world against 75.0% for another adaptation.

Adapting robot manipulation policies to new tasks and environments remains highly data-intensive, while the data needed for further improvement depends on the policy’s current capabilities and failure modes.

6. Towards Agile Vision-Based Multi-UAV Flight: Revisiting State Estimation

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2609.39611 ｜ 2026-09-30
• Key numbers: 40% (“acceleration estimation errors by 40% and 57% across”) · 57% (“errors by 40% and 57% across the three”) · 300 ms (“exhibit a constant ~ 300 ms delay in acceleration”)
• Signals: quantitative results · closed-loop success reported · real-robot experiments · latency or control-rate numbers
• Why it matters AI draft: If multi-UAV collision avoidance relies on onboard estimates of neighbours’ motion, this self-reported preprint points to a constant ~300 ms acceleration lag in some estimates and claims 40% and 57% lower acceleration error; check your estimator’s lag before tightening spacing.

Agile multi-UAV flight requires accurate and low-latency onboard estimation of the kinematic states of neighboring UAVs for collision avoidance, motion coordination, etc.

7. GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2609.39600 ｜ 2026-09-30
• Key numbers: 72.42% (“similarly sized models at 72.42%”) · 71.35% (“with GPT-6 Astra ( 71.35% )”) · 61.75% (“averaging 61.75% versus 53.32% for”) · 53.32% (“averaging 61.75% versus 53.32% for the fast”)
• Signals: quantitative results · real-robot experiments · latency or control-rate numbers
• Why it matters AI draft: For robot perception stacks held up by serial autoregressive grounding, this self-reported preprint targets that latency, but the visible evidence is mainly benchmark scores, 61.75% average against 53.32%, so measure end-to-end latency in your own pipeline before swapping.

Autoregressive (AR) grounding models serialize spatial predictions, introducing sequential latency and imposing a causal order on output tokens.

8. DODGER: Safety-Guided Reinforcement Learning for Robot Navigation Among Dynamic Obstacles

• Lane: Safety, permissions & compliance ｜ Evidence: [R] ｜ Source: arXiv 2609.38873 ｜ 2026-09-30
• Signals: real-robot experiments
• Why it matters AI draft: For mobile robots sharing space with people and moving obstacles, this preprint flags real-robot experiments but the page quotes no collision or success figures, so it cannot yet support a safety case or justify replacing an existing avoidance layer.

Robots operating in human-centered environments must safely navigate among multiple dynamic obstacles to avoid collisions with people and surrounding infrastructure.

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. 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.

3. Fleet-scale reinforcement learning for generalist robot policies

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2605.00416 ｜ 2026
• Key numbers: deploy -> shared experience -> improve -> redeploy loop · autonomous rollouts + human interventions across a fleet
• Why it matters: Fleet scale is treated as a training asset rather than an operations burden, which changes what fleet telemetry is designed for.

4. China approves 60 guiding technical documents incl. embodied-AI data quality (part 1: real data)

• Lane: Safety, permissions & compliance ｜ Evidence: [O] ｜ Source: SAMR / SAC announcement ｜ 2026-09
• Key numbers: 60 guiding technical documents (“China approves 60 guiding technical documents incl”) · embodied-AI data quality, part 1: real data
• Why it matters: Embodied-AI data quality enters the national standards system for the first time, turning data governance from team habit into an auditable specification.

5. 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.

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] 2 · [R] 32 · [M] 18
• Picks per recent run: 09-28 4 · 09-29 8 · 09-30 8 · 10-01 8 · 10-02 8 · 10-03 8 · 10-04 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