Google’s AI research company DeepMind has revealed a new AI system called SIMA (Scalable, Instructable, Multiworld Agent) that is being trained to play video games more like a human player.
Unlike previous game AI that aims to solve games optimally, SIMA learns gaming skills in a way that meshes better with how people actually play.
SIMA represents a fundamentally different approach to game AI. While not intended to replace existing non-player characters, SIMA could eventually serve as another player in your party that understands natural language instructions and can navigate 3D environments using computer vision.
“SIMA isn’t trained to win a game; it’s trained to run it and do what it’s told,” explained DeepMind researcher Tim Harley.
Currently, SIMA is only a research project, but the goal is for it to ultimately learn how to play any video game, even open-world titles with no defined critical path. To train the system, DeepMind worked with game developers like Hello Games, Embracer, Tuxedo Labs, and Coffee Stain Studios. Researchers plugged SIMA into games like No Man’s Sky, Teardown, Valheim, and Goat Simulator 3 to teach it basic skills like turning, climbing, using maps and menus.
So far, SIMA has learned about 600 foundational gaming skills. While it can handle basic tasks, more complex instructions like “find resources and build a camp” are still very challenging for current AI agents.
However, the team ultimately wants SIMA to understand and execute higher-level directives versus just low-level actions.
SIMA’s novel approach doesn’t require access to game source code or a custom API to play. This allowed DeepMind to easily integrate SIMA into commercial games during training. As the AI system develops further, it could pave the way for more human-like companion AI in video games.
More info on Google DeepMind: https://deepmind.google/discover/blog/sima-generalist-ai-agent-for-3d-virtual-environments/
