Game-Playing AI: Nvidia NitroGen Introduced

Nvidia has introduced its new open-source AI model, NitroGen. Trained on over 1,000 games, the model can convert game footage directly into action. Here are the details…
Nvidia has announced NitroGen, a new model that brings artificial intelligence to the gaming world. Developed as open source, this model not only guides players but can also play games directly on its own. NitroGen is based on the G-Assist approach previously introduced by Nvidia and will support over 1,000 games.
AI Control for Over 1,000 Games At the core of NitroGen lies a “vision-to-action” approach that converts visual data directly into control inputs. The model was trained using over 40,000 hours of publicly available game videos, covering more than 1,000 games. During the training process, gamepad layouts displayed on-screen by streamers were analyzed to isolate joystick movements and button presses. Thus, the artificial intelligence learned which actions were being performed from in-game images.

On the technical side, NitroGen relies on a policy model based on GROOT, which Nvidia developed for its robotics work. However, instead of robot motor commands, it generates direct gamepad inputs. Nvidia states that this allows the model to perceive in-game situations and provide human-like reactions. In tests conducted, NitroGen offers up to a 52% increase in success compared to models trained from scratch, even in games it has never encountered before.
NitroGen‘s scope is currently limited mainly to games played with a controller. Since the model is designed to predict gamepad actions, it is not directly used in keyboard-mouse oriented games. However, Nvidia points out that these limits can be expanded by developers thanks to its open-source structure.

NitroGen is also intended to be used in robotics and simulation fields outside of games. Reflexes and decision-making mechanisms learned from different game genres serve as a valuable data source for systems operating in the physical world. Nvidia sharing the entire model and datasets as open source could accelerate research in this area.










