Nvidia introduces new AI models and tools to accelerate autonomous driving research

Nvidia has announced new open AI models and development tools as it continues building core technology for physical AI, including robots and self-driving vehicles that must understand and navigate the real world. The company unveiled Alpamayo-R1, a reasoning vision language model designed for autonomous driving research, during the NeurIPS conference in San Diego. Nvidia says it is the first model of its kind built specifically for driving applications, combining visual and text-based inputs to help vehicles interpret their environment and respond appropriately.

Alpamayo-R1 is built on Nvidia’s Cosmos-Reason model , part of a family of reasoning models first released in January 2025 and expanded later that year. These systems are designed to think through decisions step by step, which Nvidia believes is essential for reaching level 4 autonomy — a stage where vehicles can operate independently in defined conditions. The company says models like Alpamayo-R1 can give autonomous vehicles a form of “common sense,” allowing them to handle complex or subtle driving tasks more like humans. The new model is available now on GitHub and Hugging Face.

Nvidia also released the Cosmos Cookbook, a set of guides and resources to help developers train and fine-tune Cosmos models for their own needs. The material includes detailed instructions on data collection, synthetic data creation, inference tools, and evaluation methods, aiming to make it easier for teams to build and test advanced reasoning models.

These updates come as Nvidia pushes further into physical AI, a focus CEO Jensen Huang has emphasized as the next major wave of artificial intelligence. Bill Dally, Nvidia’s chief scientist, said earlier this year that robotics will play a central role in the future and that Nvidia aims to build the “brains” behind those systems. Developing models capable of powering autonomous vehicles and robots, he said, is a crucial step toward that goal.