Physical AI Radar · 2026-10-05

Language: ZH · EN · JA

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

• Generated: 2026-10-04 23:40 UTC
• Window: papers 2026-10-01 → 2026-10-05 · posts 2026-09-05 → 2026-10-05 · no repeats within 30 days (UTC)
• Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary

Today’s picks

1. Is Success All You Need? Investigating the Impact of Input Perturbations on VLA Behaviour in Tabletop Manipulation Tasks

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2610.01351 ｜ 2026-10-01
• Signals: closed-loop success reported
• Why it matters AI draft: If you pick VLA checkpoints on benchmark success rate alone, this self-reported preprint questions that habit; the quoted text only notes high benchmark success and shows no figures, so see what else it measures before changing your evaluation.

Vision-Language-Action (VLA) models have achieved high task success rates on robot manipulation task benchmarks.

2. WBAG: A Whole-Body and Attached-Geometry Safety Framework for Vision-Language-Action Manipulation

• Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence: [R] ｜ Source: arXiv 2610.01083 ｜ 2026-10-01
• Signals: closed-loop success reported
• Why it matters AI draft: For teams putting VLA manipulation into the real world, where collisions can involve different parts of the robot, this is a safety-layer candidate to evaluate; it is a self-reported preprint showing a closed-loop signal but no figures on the page.

Vision-language-action (VLA) policies have demonstrated impressive capabilities in generalizable robotic manipulation, but their deployment in the real world remains challenging due to potential collisions involving different parts of the…

3. NarrativeFlow: Flow-Based Vision-Language-Action Model Using Robot Velocity Fields

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2610.00981 ｜ 2026-10-01
• Signals: closed-loop success reported · real-robot experiments
• Why it matters AI draft: If you hold demonstrations from several robot platforms, robot velocity fields as an embodiment-agnostic, motion-centric representation may let you pool them; the self-reported preprint carries real-robot and closed-loop signals but no figures on the page.

We focus on language-conditioned flow-based manipulation, where robot flows (robot velocity fields) serve as embodiment-agnostic, motion-centric representations for leveraging data collected from multiple robot platforms.

4. Divide-and-Remember: Recursive Action-Relevant Memory for Long-Horizon VLA Policies

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [R] ｜ Source: arXiv 2610.00982 ｜ 2026-10-01
• Signals: closed-loop success reported · open code or weights
• Why it matters AI draft: If your tasks need history because the current observation alone does not determine the action, try this self-reported preprint, flagged with open code, before building your own memory module; the page shows a closed-loop signal but no figures.

Vision-language-action (VLA) models struggle on history-dependent manipulation tasks, where the current observation alone does not determine the action, and the policy needs a memory of the history.

5. PROMO: Preference-conditioned Multi-Objective Reinforcement Learning for Quadrupedal Robots

• Lane: Training & self-improvement (RL / post-training) ｜ Evidence: [R] ｜ Source: arXiv 2610.01260 ｜ 2026-10-01
• Key numbers: 30.4% (“energy by up to 30.4%”) · 38.7% (“position error by 38.7%”) · 59.0% (“peak body-attitude deviation by 59.0% relative to the”)
• Signals: quantitative results · open code or weights
• Why it matters AI draft: If your quadruped RL fixes the tracking, stability and energy trade-off in a scalar reward at training time, this self-reported preprint with open code reports up to 30.4% on energy, 38.7% on position error and 59.0% on peak body-attitude deviation; the baseline is cut off on the page.

Quadrupedal locomotion requires balancing conflicting objectives such as command tracking, stability, and energy efficiency, yet conventional reinforcement learning (RL) hardcodes these priorities into a fixed scalar reward at training…

6. TOAST: Stochastic Robot Action Tokenization for Autoregressive Vision-Language-Action Models

• Lane: Edge & real-time (on-device inference / control rate) ｜ Evidence: [R] ｜ Source: arXiv 2610.00899 ｜ 2026-10-01
• Signals: closed-loop success reported
• Why it matters AI draft: If you run an autoregressive VLA that turns continuous actions into discrete token sequences, the tokenizer is worth revisiting; this self-reported preprint shows only a closed-loop signal on the page, and despite its edge filing no latency or speed figures.

Autoregressive Vision-Language-Action models often represent continuous robot actions as discrete token sequences, enabling action prediction with standard next-token objectives.

7. UniWAM: Unified World-Action Model

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [R] ｜ Source: arXiv 2610.02054 ｜ 2026-10-01
• Why it matters AI draft: If your VLA is trained on action labels alone, this self-reported preprint argues that gives limited grounding in world dynamics; with no figures and no real-robot or closed-loop signal on the page, treat it as a design direction, not a result.

Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics.

8. InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation

• Lane: Systems & orchestration (long-horizon / multi-policy) ｜ Evidence: [R] ｜ Source: arXiv 2610.02196 ｜ 2026-10-01
• Why it matters AI draft: If adding a humanoid task today means retraining the controller, this self-reported preprint proposes reusing existing skills at test time and keeping what the robot learns from its own attempts; no figures, real-robot or closed-loop signal appear on the page yet.

We study test-time evolution for humanoid loco-manipulation: solving tasks that a controller was never trained for by repurposing its existing skills, improving from its own attempts, and retaining what it learns, without retraining.

Frontier baseline (curated)

1. 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.

2. 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.

3. NVIDIA Cosmos 3: physical reasoning, world generation and action generation in one open model

• Lane: Foundation models (VLA / WAM) ｜ Evidence: [O] ｜ Source: NVIDIA Technical Blog ｜ 2026-08
• Key numbers: single open model · Nano / Super / Edge variants
• Why it matters: Merging physical reasoning, world generation and action generation into one open model means ‘imagining’ and ‘acting’ stop being two stacks.

4. 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.

5. LeRobotDataset v3.0: a standardised, streamable format for robot learning data

• Lane: Data engine (collection / curation / scaling) ｜ Evidence: [O] ｜ Source: Hugging Face ｜ 2026
• Key numbers: Parquet + MP4 + relational metadata · multi-episode packing · native streaming for multi-TB datasets
• Why it matters: The format question is converging on LeRobot v3 for data and OpenUSD for assets; data you can convert in one step is worth far more than the same hours locked in a private format.

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] 2 · [R] 37 · [M] 17
• Picks per recent run: 09-29 8 · 09-30 8 · 10-01 8 · 10-02 8 · 10-03 8 · 10-04 8 · 10-05 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