NVIDIA introduces physical AI agent skills at CVPR 2026 designed to accelerate autonomous vehicle, robotics, and vision AI development

On June 3, NVIDIA introduced new physical AI agent skills at CVPR 2026 that integrate with NVIDIA Cosmos 3 to unify scene reconstruction, data generation, policy training, and evaluation into a single, scalable pipeline.

For autonomous vehicles, the tools leverage the 32-billion-parameter Alpamayo 2 Super VLA model (a vision-language-action model that converts fleet data into 3D scenes and runs reinforcement learning in simulations). New Metropolis skills and Isaac frameworks similarly automate synthetic scenario creation for vision AI and robotics, reducing manual training work.

NVIDIA is offering these tools openly via GitHub and providing preconfigured environments called "Physical AI Launchables" on NVIDIA Brev to enable research institutions to move from model capabilities to real-world autonomous systems.