Nvidia introduces Alpamayo to help autonomous vehicles reason like humans

Nvidia has unveiled Alpamayo, a new family of open source AI models, simulation tools, and datasets designed to improve how autonomous vehicles and robots understand and respond to the real world. The announcement was made at CES 2026 and marks Nvidia’s latest push to bring advanced reasoning into physical AI systems, particularly self-driving cars.

Nvidia CEO Jensen Huang described the launch as a turning point for physical AI, saying machines are beginning to understand, reason, and act in complex environments. According to Huang, Alpamayo gives autonomous vehicles the ability to think through rare and difficult driving situations, operate safely in crowded or unpredictable settings, and clearly explain the decisions they make on the road.

At the center of the new platform is Alpamayo 1, a 10-billion-parameter vision, language, and action model built around chain-of-thought reasoning. The model allows autonomous vehicles to break down problems into steps and consider multiple outcomes before choosing the safest action. This approach is meant to help vehicles handle edge cases they may not have encountered before, such as navigating a busy intersection during a traffic light outage.

Ali Kani, Nvidia’s vice president of automotive, explained that Alpamayo reasons through every possible option before selecting a path forward. Beyond controlling steering, braking, and acceleration, the system can describe what action it plans to take and why it chose that action, offering greater transparency into vehicle behavior.

The underlying code for Alpamayo 1 is available on Hugging Face, allowing developers to adapt the model to their needs. Companies can fine-tune it into smaller and faster versions, use it to train simpler driving systems, or build tools such as automated data labeling and decision evaluation software.

Nvidia is also releasing supporting tools and data alongside Alpamayo. This includes an open dataset with more than 1,700 hours of driving footage collected across different regions and conditions, with a focus on rare and complex scenarios. In addition, the company introduced AlpaSim, an open-source simulation framework available on GitHub that recreates real-world driving environments. Together, these tools are intended to help developers safely train, test, and validate autonomous driving systems at scale.