Nvidia launches open physical AI stack for robotics and autonomous systems
Toolkit spans simulation, orchestration and deployment across industrial, medical and consumer robotics
Nvidia has introduced a suite of open physical AI models and development frameworks aimed at accelerating work on humanoids, autonomous vehicles and other embodied systems.
The company said the stack supports the entire robotics development lifecycle, from high-fidelity simulation and synthetic data generation to orchestration in the cloud and deployment at the edge.
It is built around OpenUSD and Nvidia Omniverse libraries, which serve as ground-truth simulation sources.
Included in the release are Nvidia Cosmos world models, Isaac Lab-Arena for policy evaluation, the Alpamayo portfolio of autonomous vehicle models and datasets, and the OSMO orchestration framework.
According to the company, these components are designed to interoperate across research and production environments.
Caterpillar is using Nemotron open models running on Nvidia’s Jetson Thor platform to power its Cat AI Assistant. The system enables operators to ask natural-language queries such as “Hey Cat” and receive step-by-step instructions from inside equipment cabs.
Omniverse libraries are also being used to create digital twins of factories and job sites, feeding simulation insights back into equipment and fleets.
Developers and manufacturers have adopted the stack across industrial, domestic and medical robotics. LEM Surgical used Jetson AGX Thor, Holoscan and Isaac technologies in its FDA-cleared Dynamis Robotic Surgical System, generating synthetic training data with Cosmos Transfer and Isaac Sim.
NEURA Robotics trained humanoid and service robots in OpenUSD environments, followed by post-training using the Isaac GR00T-Mimic framework.
Related reading
- NVIDIA calls on Congress to reauthorise National Quantum Initiative
- GeForce NOW adds flight control support and expands cloud gaming line-up
- Microsoft unveils UniRG to improve accuracy in AI-generated radiology reports
Other robotics developers, including AgiBot, Intbot and ROBOTIS, demonstrated combined workflows using Cosmos Predict and Reason models with Isaac Sim and variants of GR00T to enhance sim-to-real transfer and social reasoning in deployed robots.
Nvidia has published technical documentation, tutorials and a robotics learning path to support adoption, and is inviting developers to join community initiatives including the Cosmos Cookoff physical AI challenge, which tasks teams with applying Cosmos Reason in robotics and autonomous platforms.
The Recap
- NVIDIA unveiled a suite of open physical AI models and frameworks.
- Tools support humanoids, autonomous vehicles and other embodiments.
- Developers can participate in the Cosmos Cookoff physical AI challenge.
