The game AI has reached a milestone breakthrough. NVIDIA and Stanford University recently jointly released a new general game agent - NitroGen, which was trained on over 1000 different types of games, with a total of 40,000 hours of high-quality game data, demonstrating unprecedented cross-game generalization capabilities. More notably, the research team announced that they will fully open-source the training dataset and model weights, providing a powerful infrastructure for the global AI and game research community.

 A general intelligent agent that can play any game

The core goal of NitroGen is to break through the limitations of current game AI, which is typically "one game, one model". Traditional reinforcement learning models usually require training from scratch for a single game, while NitroGen learns in a massive and diverse game environment (covering platform jumping, strategy, shooting, puzzle, and simulation genres), acquiring general perception, decision-making, and operation capabilities. Experiments show that the model can quickly adapt and reach human-playable levels in new games it has never seen before.

image.png

 Project address: https://nitrogen.minedojo.org/

40,000 hours of data open-sourced, promoting the democratization of game AI

The research team emphasized that NitroGen's success lies not only in its model architecture but also in the high-quality, large-scale training data. To this end, they also released a dataset called GameVerse-1K, including:

- Complete interaction trajectories of more than 1,000 commercial and open-source games;

- 40,000 hours of human and AI gameplay recordings;

- Accurate alignment of each frame, operation commands, reward signals, and state metadata.

All data and model weights will be open-sourced via GitHub and Hugging Face, supporting academic research and non-commercial applications.

 Technical highlights: End-to-end visual input + General action space

NitroGen uses pure visual input (raw pixels) without requiring internal game APIs or state access, truly playing games like humans. At the same time, it designed a unified action abstraction layer, mapping the complex controls of different games (such as keyboards, controllers, touchscreens) to a standardized action space, allowing the model to generalize across platforms.

 Industry significance: Beyond games, it is also a testing ground for general intelligence

AIbase believes that the value of NitroGen goes far beyond entertainment scenarios. As a complex, dynamic, and high-dimensional simulation environment, games are an ideal testing ground for training general artificial intelligence (AGI). The paradigm of "large-scale multi-game pre-training + fast migration" verified by NitroGen could be applied to robot control, autonomous driving, industrial simulation, and other fields in the future.

By choosing to fully open-source this time, NVIDIA and Stanford not only accelerate scientific research iterations but also send a clear message to the industry: open collaboration is the fastest path to general intelligence.

Currently, the NitroGen code and GameVerse-1K dataset are available on the official repository, and developers can immediately download and experience them. This intelligent revolution ignited by games is accelerating into the real world.