Google Deepmind introduced new II-tools SIMA, aimed at mastering game skills to mimic human actions rather than following algorithms of super-powerful AIs. Sima (Scalable, Instructable, Multiworld Agent) is currently in the study phase.
Sima is learning to play various video games, even those without a linear path or open worlds. Its goal is not to replace existing game AI but to act as an additional player in the game, interacting effectively with teams using natural language instructions and 3D-world understanding and image recognition. “Sima is not trained to win; it is trained to play and follow instructions,” said Tim Harley, a project manager at Google DeepMind.
Google worked with 8 game developers, including Hello Games, Embracer, Tuxedo Labs, and Coffee Stain, to train and test Sima. Sima was connected to games like No Man’s Sky, Teardown, Valheim, and Goat Simulator 3 to learn basic gameplay. It’s worth noting that Sima doesn’t require a game-specific API or access to source codes.
Deepmind selected games with open gameplay rather than a strong narrative to help Sima develop general gaming skills. For instance, the spontaneous actions in Goat Simulator aimed to instill a sense of spontaneity in Sima.
Training involved creating a new Unity engine environment where agents built sculptures to test object manipulation abilities. Pairs of players controlled the game and provided instructions to develop language understanding. This data helped Sima agents predict in-game events.
Sima currently possesses around 600 basic skills like turning left, climbing stairs, or accessing the map menu. Harley believes Sima will eventually handle more complex tasks but acknowledges challenges like resource gathering and camp building that AI struggles with.
Sima was designed as a game-altering player rather than a game AI card like NVIDIA or Convai. DeepMind mentions it’s premature to discuss using such AI agents beyond research applications.
Although Sima may learn to communicate and talk in the future, it’s a work in progress. Google anticipates that as AI models advance, Sima will tackle more sophisticated tasks and become an ultimate AI collaborator towards game success.