Google develops new AI that can learn and master every video game

a Terminator next to a No Man's Sky explorer
Credit: Skydance Media/Hello Games

a Terminator next to a No Man's Sky explorer
Credit: Skydance Media/Hello Games

Google Deepmind, the Google subsidiary that focuses on artificial intelligence research has revealed the Scalable Instructable Multiworld Agent, or SIMA for short, an AI capable of following commands on different games. In short, it's an AI that can play video games.

Unlike its predecessors, such as the StarCraft II-focused AlphaStar, SIMA isn't confined to mastering individual games but has been developed to excel across various titles, all while comprehending and executing natural-language instructions, making it the first of its kind in the AI world.

In a lengthy blog post on the DeepMind website, the creators revealed that SIMA was trained in collaboration with eight prominent game studios on nine distinct video games, including renowned titles like No Man’s Sky, Goat Simulator 3, and Valheim.

The genres are vastly different, giving SIMA experience in unique virtual landscapes, from navigating terrains to piloting spacecraft. Beyond learning the intricacies of various games DeepMind crafted specialised research environments like the Construction Lab, utilising Unity, to assess SIMA's adaptability and problem-solving skills in varied scenarios.

A No Man's Sky explorer looking at the horizon
expand image
Credit: Hello Games
SIMA was trained in No Man's Sky among other games

What sets SIMA apart is its simplicity. The AI doesn't have the need for game-specific APIs or giving it access to source code as it operates with just two inputs: on-screen visuals and simple commands inputted by the user.

According to DeepMind, SIMA is trained in over 600 unique skills across all nine games they have trained it in it. Some examples include the ability to drive a car in Goat Simulator 3, chopping down trees in Valheim, and shooting Asteroids in No Man's Sky.

As basic as they are, DeepMind researchers are looking to eventually train SIMA to perform more involved tasks such as "find resources and build a camp."

The applications could pose an interesting question for developers and whether they'd be able to find implementations for the SIMA Agent in the near future. Would they be able to program Agent to do certain tasks while finding bugs during the QA process? Only time will tell.

This Article's Topics

Explore new topics and discover content that's right for you!