Nvidia launched new AI models and tools for robotics developers at SIGGRAPH Monday. The showstopper: Cosmos Reason, a 7-billion-parameter vision-language model built to help robots plan and reason in real-world settings.
Cosmos Reason uses memory and physics knowledge to figure out what steps a robot or AI agent should take next. Nvidia says it’s ideal for robot planning, data curation, and video analytics.
The rollout also includes Cosmos Transfer-2, which speeds up synthetic data generation from 3D scenes and spatial controls, plus a faster, distilled version of Cosmos Transfers for improved performance. 
On the infrastructure side, Nvidia added neural reconstruction libraries with a new rendering technique. This lets developers simulate the real world in 3D using sensor data. The tech is getting integrated into CARLA, the widely used open-source driving simulator. The Omniverse dev kit also got an update.
Nvidia is pushing more hardware for robotics too. The RTX Pro Blackwell Server provides a unified architecture for robotics workloads. Meanwhile, DGX Cloud offers a cloud platform for managing these AI projects.
The launch signals Nvidia’s bet on robotics as the next big AI GPU market beyond data centers.
TechCrunch event:
San Francisco | October 27-29, 2025
