NVIDIA and Researchers Unveil NitroGen, an Open-Source AI That Can Play 1,000 Games

A team of researchers from NVIDIA, Stanford University, and Caltech has introduced NitroGen, an open-source AI foundation model designed to play more than 1,000 video games. The system was trained on over 40,000 hours of publicly available gameplay footage, including videos that display players’ real-time gamepad inputs.

NitroGen is a unified vision-to-action model that interprets raw video frames and outputs corresponding gamepad actions. It was trained through large-scale imitation learning rather than reward-based reinforcement methods. The model performs well across a wide range of genres, including RPGs, platformers, racing, and battle royale titles, and can adapt to unseen games and procedurally generated environments.

Built on the GROOT N1.5 architecture originally developed for robotics, NitroGen demonstrates capabilities that could extend beyond gaming. Its developers suggest that the same approach could help robots operate more effectively in unpredictable real-world settings. Early evaluations indicate that NitroGen achieves a 52% improvement in task success rates compared to models trained from scratch.

The model has been released under an open-source license, with its dataset, pretrained weights, and code available for research and experimentation. Potential applications include automated testing for video games, next-generation game AI, and advancing research in embodied AI and robotics.