Nvidia on Monday unveiled new infrastructure and AI models aimed at powering the backbone of “physical AI” — the technology that enables robots and autonomous vehicles to perceive and interact with the real world.

 

At the NeurIPS AI conference in San Diego, the company introduced Alpamayo-R1, an open reasoning vision-language model designed specifically for autonomous driving research. Nvidia describes it as the first vision-language action model focused on autonomy. By combining image and text processing, such models allow vehicles to interpret their surroundings and make context-aware decisions.

 

Alpamayo-R1 builds on Nvidia’s Cosmos-Reason model, part of the Cosmos family first released in January 2025 and expanded in August. According to Nvidia, technology like this is essential for reaching Level 4 autonomy, where vehicles can operate fully autonomously within defined conditions.

 

The company hopes the model’s reasoning capabilities will provide autonomous systems with a form of “common sense,” helping them handle subtle or complex driving scenarios more like human drivers. The model is now available on GitHub and Hugging Face.

 

Nvidia also released the Cosmos Cookbook, a collection of step-by-step guides, inference tools, and post-training workflows hosted on GitHub. The materials cover data curation, synthetic data creation, and model evaluation to help developers tailor Cosmos models to their needs.

 

These announcements come as Nvidia continues to invest heavily in physical AI as a major growth area for its advanced AI GPUs. CEO Jensen Huang has frequently described physical AI as the next big wave in the field, a view echoed by chief scientist Bill Dally, who told TechCrunch this summer that Nvidia aims to build the “brains” for future robots.

 

“To do that, we need to start developing the key technologies,” Dally said at the time.