NVIDIA NitroGen plays over 1000 games on its own

NVIDIA announced NitroGen, a new artificial intelligence model developed in collaboration with institutions such as Stanford and Caltech. NitroGen is described as an open-source foundational model capable of playing over 1000 games. However, this technology is not limited to the gaming world; it also holds great potential in simulations and robotics.

NVIDIA NitroGen: “GPT for Action”

NVIDIA Research Director Jim Fan described NitroGen as “an open-source foundational model capable of playing over 1000 games.” The research team views this work as “GPT for Action.” The application of large-scale training technology used in language and visual fields, this time to motor control and game mechanics, is considered a significant milestone in artificial intelligence research.

NitroGen is based on the GROOT N1.5 architecture, originally designed for robotics. Tested in games, this architecture has proven successful in hundreds of games with varying mechanics and physics. This approach is said to potentially contribute to robots operating more flexibly in unpredictable environments in the future.

Researchers used over 40,000 hours of publicly available game videos to train NitroGen. Streams showing real-time gamepad interactions were particularly critical to the model’s learning process.

NitroGen was tested across a wide range of games, from RPGs and platform games to battle royale and racing games. The initial version focuses particularly on fast motor control, or “player instinct.” According to the research, the model achieved a 52% higher success rate in procedurally generated worlds and games it had never seen before, compared to models trained from scratch.

All research outputs of NitroGen have been shared as open source. Model weights, datasets, and code are available for use by researchers working in the fields of gaming, robotics, and LLM. Therefore, anyone can adapt or improve the model for their own projects.

Read More: GTA 6 Hype Could Cause Price Increases