Google DeepMind Laboratory presented SIMA - an artificial intelligence model that will play games for you. It is designed to play more like a human, rather than a superhuman artificial intelligence. SIMA, which stands for Scalable Instructable Multiworld Agent, is currently just a research experiment.
Over time, SIMA should learn to play any game, even games with non-linear progression and open-world games. Although it is not intended to replace existing game artificial intelligence - rather as another player who is well suited for team play. It combines natural language communication learning with understanding of 3D worlds and image recognition.
"SIMA is not trained to win the game; it is trained to play and perform what it is instructed to do," said Google DeepMind researcher and SIMA co-lead Tim Harley.
Google worked with eight game developers, including Hello Games, Embracer, Tuxedo Labs, Coffee Stain, and others, to train and test SIMA. Researchers connected SIMA to games like No Man's Sky, Teardown, Valheim, and Goat Simulator 3 to teach the AI the basics of gameplay. A Google blog post explains that SIMA does not require a special API to play games or access to the code.
Harley said that the team chose games that were more focused on open-ended action than story to help SIMA learn general gaming skills. If you have played or watched the gameplay of Goat Simulator, you know that the essence of the game is to do random, spontaneous things, and researchers hoped that SIMA would learn this spontaneity.
To achieve this, the team first created a new environment in the Unity engine where the player had to create sculptures to test object manipulation understanding. Then Google recorded pairs of human players - one played, and the other gave instructions on what to do next - to record speech instructions. After that, the players played on their own to show what led to their actions in the game. All of this was fed to SIMA to learn to predict what would happen next on the screen.
Currently, SIMA has about 600 basic skills, such as turning left, climbing stairs, and opening a menu to use a map. Ultimately, SIMA will be able to be assigned more complex functions in the game. Tasks such as finding resources and building a camp are still challenging for the model.
SIMA is not intended to replace AI-based NPC, but to be a player. SIMA project lead Frederick Bess said that it is still too early to talk about the benefits such artificial intelligence agents can bring to games outside the research sphere.
However, like AI-based NPCs, SIMA may eventually learn to talk, but not yet. SIMA is still learning to play games and adapt to ones it hasn't played before. Google says that thanks to advanced artificial intelligence models, SIMA will eventually be able to perform more complex tasks and become an ideal member of a team of players.
Comments (0)
There are no comments for now