Google Labs Developers DeepMind has unveiled a new AI model with an unusual specialization. An algorithm called the Scalable Instructable Multiworld Agent (SIMA) has the ability to learn gamer skills — its creators told why this is necessary.
To create Google’s new AI collaborated with eight game developers, including Hello Games, Embracer, Tuxedo Labs and Coffee Stain. SIMA has been connected to No Man’s Sky, Teardown, Valheim and Goat Simulator 3 to teach basic gaming skills. According to the developers, AI does not need a special API to run games, as well as access to their source code.
For SIMA training, the created a new environment on the Unity engine, and then recorded the behavior of human players—one He gave instructions on how to pass, the second one carried them out. Such a set of training materials allowed the neural network to learn how to predict the future actions of the game character.
Currently, SIMA has About 600 basic skills (turning, climbing stairs, opening menus to use the map, and so on). Over time, the developers plan to Assign more complex functions to the AI agent. As a result, SIMA will be able to pass Any game, including open-world titles.
The developers note that they want to create a full-fledged companion for gamers – for example, for cooperative walkthrough. The timing of the release version of SIMA, ready for implementation in games, has not yet been announced.