Physical AI Radar · 2026-09-25
A daily radar on the Physical AI frontier: eight lanes, trilingual digests, traceable evidence.
- Generated:
2026-09-25 01:40 UTC - Window:
papers 2026-09-22 → 2026-09-25 · posts 2026-08-26 → 2026-09-25 · no repeats within 7 days (UTC) - Evidence tags: O = first-party official · R = paper/preprint · M = media/secondary
Today’s picks
1. DreamStream: Towards Policy-Oriented Generative Simulation for End-to-End Driving
- Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence:
[R]｜ Source: arXiv 2609.26792 ｜2026-09-24 - Signals:
closed-loop success reported·real-robot experiments·open code or weights - Why it matters
AI draft: If you lean on generative simulators for closed-loop evaluation, this preprint argues photo-realism metrics can misrank fidelity and offers a policy-feature metric plus code, though the gains are self-reported on driving benchmarks only.
Faithfully evaluating end-to-end driving policies in simulation requires observations that are not merely photo-realistic, but preserve the scene features a policy relies on to make decisions.
2. VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models
- Lane: Simulation & evaluation (sim2real / closed-loop) ｜ Evidence:
[R]｜ Source: arXiv 2609.25376 ｜2026-09-24 - Signals:
closed-loop success reported·open code or weights - Why it matters
AI draft: Before shipping a quantized VLA, treat precision recipes as model-specific: this preprint reports that which layers you protect swings closed-loop success sharply and that one model’s calibration fix degrades another, mostly in simulation.
Post-training quantization reduces the memory requirements of vision-language-action (VLA) models, but precision selection must account for the interaction between layer scope, numerical format, and calibration.
3. MATE: Multi-Agent Virtual Teleoperation Platform for Humanoid Collaboration Data Collection
- Lane: Data engine (collection / curation / scaling) ｜ Evidence:
[R]｜ Source: arXiv 2609.26520 ｜2026-09-24 - Key numbers:
24.1 hours(“multi-humanoid collaboration dataset comprising 24.1 hours of coordinated behavior”) - Signals:
quantitative results·real-robot experiments - Why it matters
AI draft: Collecting multi-humanoid collaboration data no longer requires owning several robots: distributed operators in a shared physics sim produced 24.1 hours of coordinated demos, and the authors report zero-shot transfer to one physical humanoid.
Humanoid robots require diverse embodied experiences to acquire complex loco-manipulation and collaborative skills.
4. Accelerating a ROS 2 Node with an AI Agent and NVIDIA Isaac ROS
- Lane: Systems & orchestration (long-horizon / multi-policy) ｜ Evidence:
[O]｜ Source: NVIDIA Technical Blog (Robotics) ｜2026-09-22 - Why it matters
AI draft: NVIDIA’s own walkthrough of having an AI agent port a ROS 2 node to Isaac ROS is a reminder that message transport, not the CUDA kernel, usually bounds graph latency; it shows the workflow but gives no timing figures.
GPU acceleration can speed up compute-intensive robotics workloads, but a fast CUDA kernel alone does not guarantee a fast ROS 2 graph.
5. Isaac ROS 5.0 brings AI agents to robotics development, says NVIDIA
- Lane: Systems & orchestration (long-horizon / multi-policy) ｜ Evidence:
[M]｜ Source: The Robot Report ｜2026-09-22 - Why it matters
AI draft: A trade-press summary of the Isaac ROS 5.0 release: agentic workflows and new platform support are the headline, so check NVIDIA’s release notes for which hardware and ROS distributions are covered before scheduling a migration.
The release brings new agentic workflows and platform support to NVIDIA Isaac ROS to help developers deploy robotics applications faster.
6. Bringing a Frontier World Model to the Convenience Store: Inside Telexistence’s DreamZero Experiment on AWS
- Lane: Foundation models (VLA / WAM) ｜ Evidence:
[O]｜ Source: AWS Physical AI Blog ｜2026-09-09 - Why it matters
AI draft: A first-party account of Telexistence running a world model on AWS for convenience-store tasks; it signals cloud-hosted world models reaching retail pilots, but the post is framed as an experiment and gives no success or latency figures.
Introduction Imagine walking into a convenience store in Tokyo and watching a robot scan your items, bag your groceries, and restock the shelves — all without missing a beat.
7. ANYbotics opens the door for inspections with ANYmal robots
- Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence:
[M]｜ Source: The Robot Report ｜2026-09-24 - Why it matters
AI draft: Per trade-press reporting, ANYmal can now badge through access-controlled doors via dormakaba and LEGIC integration, so the practical blocker for unattended inspection shifts from locomotion to facility credentials and security policy.
ANYbotics partnered with dormakaba and LEGIC Identsystems to enable the ANYmal quadruped to gain access like a badged employee.
8. Epson introduces AX6 cobot with compact design, no-code programming
- Lane: Embodiment & supply chain (hardware / cost) ｜ Evidence:
[M]｜ Source: The Robot Report ｜2026-09-24 - Why it matters
AI draft: Trade-press coverage of Epson entering the power-and-force-limited cobot segment with a compact six-axis arm and no-code programming; the summary here carries no payload, reach or price, so compare against the spec sheet before shortlisting.
Epson Robots has expanded its six-axis robot line with the AX6, a lightweight force- and power-limited arm.
Frontier baseline (curated)
1. 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.
2. 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.
3. Regulation (EU) 2023/1230 becomes fully applicable
- Lane: Safety, permissions & compliance ｜ Evidence:
[O]｜ Source: EUR-Lex / TUV SUD ｜2027-01-20 - Key numbers:
fully applicable 2027-01-20·replaces Directive 2006/42/EC - Why it matters: A hard date you can copy into a project plan: after it, non-compliant machinery cannot be placed on the EU market.
4. GR00T N1.7: open reasoning VLA model for humanoid robots
- Lane: Foundation models (VLA / WAM) ｜ Evidence:
[O]｜ Source: NVIDIA / Hugging Face ｜2026 - Key numbers:
3B parameters·open weights - Why it matters: A 3B open reasoning VLA gives in-house teams a fine-tunable, deployable starting point instead of building a foundation model.
5. 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.
6. 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’.
Lane distribution
| Lane | # |
|---|---|
| Foundation models (VLA / WAM) | 4 |
| Data engine (collection / curation / scaling) | 6 |
| Training & self-improvement (RL / post-training) | 4 |
| Simulation & evaluation (sim2real / closed-loop) | 6 |
| Edge & real-time (on-device inference / control rate) | 5 |
| Safety, permissions & compliance | 5 |
| Systems & orchestration (long-horizon / multi-policy) | 3 |
| Embodiment & supply chain (hardware / cost) | 4 |
Recent cadence
- Source mix:
[O]16 ·[R]16 ·[M]5 - Picks per recent run: 09-19 8 · 09-20 8 · 09-21 8 · 09-22 8 · 09-23 6 · 09-24 2 · 09-25 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.