A team of researchers from Nvidia, Stanford, and Caltech jointly developed NitroGen, an open-source base model trained to play over 1,000 games: RPGs, platformers, battle royale, racing, 2D, 3D, and more. However, the model’s potential is much greater — it opens up new possibilities for digital modeling and robotics.
A team of researchers from Nvidia, Stanford, and Caltech jointly developed NitroGen, an open-source base model trained to play over 1,000 games: RPGs, platformers, battle royale, racing, 2D, 3D, and more. However, the model’s potential is much greater — it opens up new possibilities for digital modeling and robotics.
NitroGen is based on the GROOT N1.5 architecture, which was originally designed to control robots. The fact that it has proven so successful in games means that this experience can be brought back to the real world. This will significantly help robots better navigate in unpredictable and complex environments, writes Tom's Hardware.
NitroGen was adapted for games with completely different mechanics. The researchers used over 40,000 hours of public gameplay videos from streamers. Particularly useful were videos in which players superimposed their real-time gamepad interactions on the broadcast.
The debut version of NitroGen focuses on quick reactions and motor skills. Research data shows that the new language model is extremely effective in various areas. It is able to operate in games with random level generation and completely new projects, outperforming conventional models in terms of success rate by 52%.
