Google DeepMind introduced SIMA, an AI agent being trained to develop gaming skills to mimic human gameplay rather than acting like a dominating AI. SIMA, short for Scalable, Instructable, Multiworld Agent, is still in the research phase.
SIMA aims to acquire the ability to play various video games. Its including those without a straightforward path to completion and open-world games.
It is not meant to replace current game AI but rather to act as a cooperative player within your group. SIMA combines natural language instructions with comprehension of 3D environments and image recognition.
“SIMA isn’t trained to win a game; it’s trained to run it and do what it’s told,” explained Tim Harley, a Google DeepMind researcher and co-lead of SIMA.
Google’s SIMA: Advancing AI in Gaming
Google collaborated with eight game developers, including Hello Games, Embracer, Tuxedo Labs, Coffee Stain, and others, to train and evaluate SIMA. They connected SIMA to games such as No Man’s Sky, Teardown, Valheim, and Goat Simulator 3 to teach it the fundamentals of gameplay. Google mentioned in a blog post that SIMA doesn’t require a custom API or access to source codes to play these games.
Harley explained that the team selected games focused on open play rather than narrative to assist SIMA in acquiring general gaming skills. If you’re familiar with Goat Simulator, you understand that the game revolves around performing random, spontaneous actions, which is what the team wanted SIMA to learn.
To achieve this goal, the team initially constructed a new environment in the Unity engine where agents had to create sculptures to test their grasp of object manipulation. Next, Google captured pairs of human players — one playing the game and the other providing instructions — to gather language instructions.
Then, players played independently to demonstrate what prompted their actions in the game. All of this data was used to train the SIMA agents to predict future events on the screen.
SIMA: Gaming AI Now and Tomorrow
SIMA currently possesses around 600 basic skills, such as turning left, climbing ladders, and accessing menus to use maps. In the future, Harley mentioned that SIMA could be trained to perform more intricate tasks within games. However, tasks like “finding resources and constructing a camp” remain challenging because AI agents cannot execute actions on behalf of humans.
SIMA is not designed to be an AI-driven non-playable character (NPC) like those developed by Nvidia and Convai. Instead, it is intended to be another player in a game that influences the outcome. Frederic Besse, co-lead of the SIMA project, mentioned that it’s too early to determine the potential uses of AI agents like SIMA in gaming beyond the research domain.
Similar to AI NPCs, SIMA might eventually learn to communicate verbally, but it’s currently far from achieving that capability. SIMA is still in the process of learning how to play games and adapt to unfamiliar ones. Google noted that with more advanced AI models, SIMA might eventually handle more complex tasks and become an ideal AI team member to guide players to victory.
Leave a Reply