An open-weight robot foundation model is attacking the robot-data bottleneck

Perceptron AI released Isaac 0.5, a 36B-parameter embodied foundation model spanning video understanding, reasoning and robot control. It was trained on 3 trillion multimodal tokens, about 1 million hours of general video and 100,000 hours of robotics-oriented experience across 35+ robot systems. The company reports 97.2% on LIBERO and, more interestingly, says increasing general-video pretraining from 1,000 hours to 1 million reduced the teleoperation data required to reach the same action loss from roughly 5,900 hours to 28. Those figures are company-reported and the benchmark margins are tiny, but if the data-scaling result survives independent testing, it points toward a much cheaper path to cross-embodiment learning. HPCwire

The deeper signal is video → physical prior → small amount of embodiment-specific training, rather than collecting mountains of expensive robot demonstrations independently for every body.

Physical-AI components just got their own public-market vehicle

Mech-Mind Robotics listed in Hong Kong on September 1, raising about HK$2.35 billion / US$300 million and entering the market at roughly HK$12.7 billion capitalization. Its stack explicitly combines 3D vision, deep-learning perception, motion planning, its Mech-GPT multimodal model and dexterous-hand hardware, essentially an industrial eye → brain → hand architecture. Retail demand was reported at more than 3,800× oversubscription. O'Melveny

This one matters for the investment map. Capital is no longer flowing only toward glamorous humanoid OEMs. Perception, robot brains, dexterity and integration layers are becoming independently investable businesses.

Germany and Ukraine are moving from drone procurement toward a shared autonomous-defense stack

Ukraine said on September 2 that it is finalizing a deal with Germany covering the joint development of drones, missiles and related software. It builds on Germany-backed procurement of roughly 15,000 Strila interceptor drones, a German-Ukrainian interceptor-drone joint venture, and financing for 50,000 Ukrainian Shrike attack drones. Reuters

That progression is important: buy hardware → co-manufacture → co-develop autonomy/software. Ukraine’s battlefield feedback loop is increasingly becoming part of Europe’s defense-industrial infrastructure rather than remaining a wartime exception.

Defense investors are now funding the coordination layer between autonomous machines

Aitan emerged from stealth on September 2 with a $41 million round, bringing total equity funding to $54 million. The company describes its product as infrastructure for rapidly building unmanned forces, combining command-and-control networks, edge AI and robotic weapon systems, and says its technology accumulated hundreds of thousands of operating hours alongside the IDF. Axios

The interesting bit is what investors are buying. Not another drone. The nervous system connecting heterogeneous autonomous platforms. That fits the wider pattern we have been watching in NATO, Ukraine and U.S. programs: value is migrating upward from individual bodies toward fleet coordination and resilient C2.

Wayve’s learned driving stack reached paying passengers in London

On September 3, Uber began offering Wayve-powered autonomous rides in London using Ford Mustang Mach-Es. Fewer than 20 vehicles are initially operating, with a licensed operator onboard while regulatory approval for fully driverless service continues. Wayve’s core claim is that its learned AI Driver can adapt across roads, weather, vehicles and cities rather than relying primarily on city-specific rule stacks. Reuters

It is a useful embodied-agent milestone because the robot body here is not humanoid. A generalizing learned policy is being attached to an existing mass-produced physical platform and inserted directly into a commercial service.

One smaller signal I would keep pinned

Wetour Robotics demonstrated a human-motion capture system combining 8-channel surface EMG with first-person vision, intended to recover information cameras miss during occlusion and add force-related information to robot demonstrations. The company’s own tests found vision-only hand tracking absent in 21.8% of frames in one task, with a 4.32-second continuous dropout. Crucially, however, its force-estimation pipeline has so far been validated end-to-end on synthetic data and still needs live validation, so I would classify this as promising infrastructure research, not a proven breakthrough yet. GlobeNewswire

That direction remains worth watching: human muscle activation + vision + motion → richer robot-training trajectories. Cameras tell the model where the hand went. Muscle signals begin to tell it something about how the contact happened.

Signal read

🔥 Robot-foundation models: open competition is getting serious.
🟢 Embodied-AI capital markets: broadening from robot brands into core perception/manipulation infrastructure.
🟢 Autonomous-defense integration: shifting toward software, shared C2 and multinational manufacturing loops.
🟢 Learned autonomy in public deployment: meaningful commercial milestone in London.
🟢 Human sensorimotor data capture: becoming richer than video-only demonstrations.
🟡 Neural interfaces / bionics / biohybrids / autonomous materials: I found no September 1–3 development strong enough to justify padding this cycle.

Core takeaway: the connective tissue is winning this round. The interesting developments are increasingly not isolated robot bodies but the things that let intelligence cross the membrane into matter reliably: foundation models that transfer between embodiments, eye-brain-hand components, muscle-derived training signals, fleet-level command networks and learned policies dropped into existing machines.

The body race continues. But the real contest is starting to look like a race for the nervous system. 🦾⚡

[

Reuters

Uber and AI-firm Wayve launch London's first robotaxis

5 days ago

](https://www.reuters.com/business/uber…REDACTED/?utm_source=chatgpt.com)[

Reuters

Ukraine is finalising drone deal with Germany, Zelenskiy says

5 days ago

](https://www.reuters.com/business/aerospace-defense/ukra…REDACTED/?utm_source=chatgpt.com)[

Axios

Exclusive: Aitan, battle-tested by IDF, raises $41 million

6 days ago

](https://www.axios.com/2026/09/02/aita…REDACTED?utm_source=chatgpt.com)[

GlobeNewswire

Wetour Robotics Demonstrates sEMG-Vision System Targeting

7 days ago

](https://www.globenewswire.com/news-release/2026/09/01/3354055/0/en/weto…REDACTED.html?utm_source=chatgpt.com)

Sunday 10:02 AM

Scan physical AI research