Google's new AI agent will play video games like humans
Google's artificial intelligence division, DeepMind, has revealed its latest project called SIMA. Short for Scalable, Instructable, Multiworld Agent, SIMA is currently in the research phase and aims to understand 3D worlds and image recognition while following natural language instructions. Unlike traditional gaming AI bots that are designed to win, SIMA is being trained to play video games in a manner similar to human players.
SIMA's training and learning process
Google DeepMind is collaborating with eight game developers, including Hello Games and Tuxedo Labs, to train SIMA. The AI agent has been tested in games like No Man's Sky and Teardown, among others, to learn the basics of gameplay. To enhance its prediction skills, Google has recorded human players interacting with the games while providing instructions for SIMA. This approach allows the AI agent to learn from a variety of gaming environments and setups.
SIMA's current capabilities and future potential
Currently, SIMA has mastered around 600 fundamental gaming skills, such as turning left or opening the game menu. However, complex tasks like 'find resources and build a camp' remain challenging for the AI agent. As per Google DeepMind researcher Tim Harley, with further advancements in AI models, SIMA could eventually perform more intricate functions within games. It would be able to adapt new functions that it hasn't performed before.
SIMA's performance and future research directions
In terms of performance, SIMA has shown promising results. An agent trained on eight games performed better than the one trained on a single game. When introduced to a new game, the agent trained on eight games showed nearly the same level of performance as an agent specifically trained for that game. This indicates SIMA's potential to perform beyond its training, although more research is needed for it to perform at par with human levels in familiar and unfamiliar games.
The role of language in SIMA's functioning
Language input is crucial for SIMA's successful operation. In tests where language training or instructions were not provided, the AI agent tended to perform common actions rather than following specific directives. DeepMind researchers note that this early-stage research shows potential for developing an entire new genre of generalist, language-driven AI agents. As SIMA is subjected to more training modules, it is expected to become more versatile and capable of carrying out complex tasks.
