Nvidia and Research Partners Develop Versatile AI Model NitroGen, Trained on 40,000 Hours of Gameplay

Nvidia, in collaboration with Stanford University, the California Institute of Technology, and other research institutions, has developed an artificial intelligence system named NitroGen. This model is capable of learning and performing tasks across a wide variety of virtual environments and video games.

The system was trained on 40,000 hours of gameplay from over 1,000 different video games. Training utilized public gameplay recordings, including videos where player inputs from controllers were displayed in real time. NitroGen successfully operates across diverse genres such as RPGs, platformers, battle royales, and racing games, in both 2D and 3D formats.

NitroGen is based on the GROOT N1.5 architecture, originally created for robotics. Its application in gaming demonstrates the potential for developing universal algorithms capable of acting in unpredictable settings. Researchers note that such developments could be useful not only in gaming but also in simulating real-world physical processes and in training robots.

The model is focused on rapidly acquiring motor skills and shows a 52% higher success rate in task completion compared to systems trained from scratch. It can adapt to procedurally generated worlds and previously unseen games.

All research materials, including code, datasets, and model parameters, are publicly available for further study and development.