Google DeepMind is working on SIMA, an artificial intelligence designed to learn gaming skills in a surprisingly human-like way. This system, an acronym for Scalable, Instructable, Multiworld Agent, represents a qualitative leap in the interaction between human players and artificial intelligence in video games. SIMA is currently the subject of research and is not yet ready for the general public, however its potential suggests a revolution in the way we will interact with virtual worlds. Unlike traditional AI agents, often programmed to win at all costs, SIMA focuses on performing assigned tasks, learning to play in ways that harmonize with the group dynamics of human players.
Tim Harley, Google DeepMind researcher and co-leader of the SIMA project, highlighted SIMA's peculiarity during a meeting with journalists: “It is not trained to win, but to actively participate in the game, following the instructions received”. This feature clearly distinguishes it from other AI agents, offering a more natural and engaging shared gaming experience. SIMA was formed in collaboration with eight game developers, including Hello Games, Embracer, Tuxedo Labs and Coffee Stain. Through integration into titles such as No Man's Sky, Teardown, Valheim and Goat Simulator 3, SIMA has mastered the fundamentals of gaming, demonstrating exceptional adaptability without the need for custom APIs or access to source codes.
The approach taken to train SIMA includes learning through natural language instruction, understanding three-dimensional worlds and image recognition. Initially, the team created a new environment in the Unity engine, where agents had to build sculptures to test their ability to manipulate objects. Next, pairs of human players – one commanding the game and the other providing instructions – were recorded to capture the verbal instructions and motivations behind the actions taken, providing SIMA with a rich learning database. SIMA currently has around 600 basic abilities, including turning left, climbing stairs and opening the menu to use a map. Over time, it is expected to be able to perform more complex functions within the games, such as searching for resources and building a base camp.
The SIMA project aims to create a unique gaming experience, where the AI agent acts as a teammate rather than a simple AI-driven NPC (Non-Player Character). Frederic Besse, co-leader of the SIMA project, points out that it is still premature to outline all the possible applications of AI agents like SIMA in the gaming sector. However, the goal is clear: to make SIMA the perfect AI gaming companion, capable not only of performing complex actions but also, in the future, of communicating effectively with human players, leading them to victory.
#Google39s #gaming #isn39t #built #win