Google DeepMind introduces SIMA, a versatile AI agent capable of performing tasks similar to a human player

Ever considered the idea of gaming with an AI system? It may sound like something out of a Sci-Fi movie, but Google DeepMind is pushing the boundaries of AI technology with the introduction of SIMA. Short for Scalable, Instructable, Multiworld Agent, SIMA is an AI agent that can learn and adapt gaming skills from various environments. This concept is currently being researched, with the AI agent able to play games like humans through natural language instruction and image recognition. Learn more about Google DeepMind’s groundbreaking SIMA technology.

What is SIMA?

According to Google’s blog post, SIMA is an artificial intelligence agent with the ability to learn game skills through natural language teaching and image recognition. It knows how to analyze different environments to prepare for action according to the instructed objective. This AI agent is trained to learn game skills to play just like humans. That’s why it can work as a video game companion, but it doesn’t work like regular game bots that do their own thing. Currently, SIMA has learned about 600 basic skills, including navigation, object interaction, driving a car, etc. Google said: “This work is not about achieving high game scores. Learning to play even one video game is a technical achievement for an AI system, but learning to follow instructions in different game settings can unlocking more useful AI agents for any environment.

How Google DeepMind trains SIMA

Google is working with game developers to provide SIMA with a platform to practice different games and settings. So far, the company has partnered with eight game studios to test SIMA’s capabilities. The AI agent was tested in nine different games, including Hello Games’ No Man’s Sky and Tuxedo Labs’ Teardown. Playing different games exposes the AI agent to different game setups and lots of learning. With each interactive world, it gains basic skills such as finding resources, navigating, flying a spaceship, etc.

To help SIMA learn skills and act spontaneously, the company worked with four research environments, including a new environment in the Unity engine where the agent was tasked with developing a sculpture to test understanding. Google also tested SIMA with two human players, with one playing the game and the other giving instructions to SIMA. This was done to improve the agent’s prediction skills.

Google further emphasized that with advanced AI models, SIMA would be able to perform complex game tasks that could act like humans and help ensure victory in games.