Google has revealed its AI agent SIMA, which can perform a range of tasks in three-dimensional games like never before, training to acquire playing skills to be able to play like humans instead of operating like traditional AI that follows specific commands.
The company has long been testing AI models that can win games such as chess and even learn games without knowing their rules.
According to Google, the AI agent has shown its ability to understand a wide range of game worlds and execute tasks within them based on natural language instructions.
The artificial intelligence is still in the research phase, with SIMA known as a multi-world agent that can be developed and trained.
SIMA learns how to play all kinds of video games, including those with no specific end and open-world games.
SIMA is not designed to replace AI in the game, but rather considered another partner that integrates well with your team, combining natural language learning, understanding of three-dimensional worlds, and image recognition.
Google stated that they are not training SIMA to win the game, but to execute the required commands.
Google collaborated with a group of game developers such as Hello Games, Embracer, Tuxedo Labs, Coffee Stain, and others to train and test SIMA.
Google AI researchers added SIMA to games like No Man’s Sky, Teardown, Valheim, and Goat Simulator 3 to teach the educational agent the basics of game practice.
The company affirmed that SIMA does not require a specific application programming interface to run games or access code instructions.
The researchers chose games focusing on free play rather than on the story to help SIMA acquire general gaming skills.
The team used four different research environments, including an environment built in the Unity engine, where the agent is directed to carve sculptures using building blocks.
SIMA was provided with a variety of environments and settings for learning, offering a diverse range of graphics and landscapes to choose from.
Each game opens up a new interactive world for SIMA, containing a variety of skills to learn, from simple navigation and menu usage to resource utilization, flying a spaceship, or crafting a helmet.
Following guidelines in tasks like these in video game worlds could lead to more useful AI agents in any setting they are placed in.
Currently, SIMA possesses about 600 basic skills, such as turning left, climbing ladders, and opening menus to use the map.
SIMA can be used to perform more complex tasks within the game. Actions like resource gathering and camp building remain challenging, as AI agents cannot perform actions on behalf of humans.
The aim of SIMA is not to be an AI-supported non-playable character, but an additional player in the game that influences the outcome.
Google explained that they have not yet determined the various uses that AI agents could bring to games beyond the research field. They pointed out that SIMA is still in the learning phase of playing games and adapting to games it has not played before.
Google said, “As AI models advance, SIMA might be capable of executing complex tasks and be an ideal member in the AI group that leads you towards success.”
