Google’s advanced artificial intelligence lab DeepMind has trained a new model capable of playing high-end games like No Man’s Sky, Valheim and even Goat Simulator without requiring special instructions.
Named SIMA (Scalable Inscrutable Multiworld Agent) it is able to follow natural language instructions from a player and carry out tasks in a range of game settings.
This builds on years of research teaching AI how to play games, from IBM Watson playing chess in the 80s to neural networks able to build games from scratch.
DeepMind researcher Tim Harley says SIMA isn't built to win at games, rather to do what its told by a player. Essentially it would be like having your own Gemini AI assistant that can interact with the virtual world.
It opens up the potential of having virtual teammate not out for victory at any cost, that won't double cross you and that you can make do the boring jobs.
What is SIMA and how was it trained?
Introducing SIMA: the first generalist AI agent to follow natural-language instructions in a broad range of 3D virtual environments and video games. 🕹️It can complete tasks similar to a human, and outperforms an agent trained in just one setting. 🧵 https://t.co/qz3IxzUpto pic.twitter.com/02Q6AkW4uqMarch 13, 2024
The big breakthrough with SIMA was independence. Teaching an AI to play a single game is a significant achievement, but the DeepMind team have managed to build a generalist AI agent.
It can work across a range of games, navigating a variety of environments and perform tasks from simple natural language instructions.
“This research marks the first time an agent has demonstrated it can understand a broad range of gaming worlds, and follow natural-language instructions to carry out tasks within them, as a human might,” the team explained in the research paper published with the new model.
Its training involved playing video games and being exposed to as many environments as possible. They worked with developers such as Hello Games and Tuxedo Labs, letting SIMA lose to help it understand navigation, menu use, flying and crafting.
“By learning from different gaming worlds, SIMA captures how language ties in with game-play behavior,” the researchers explained.
What can SIMA do that you can’t do with in game AI?
SIMA can do anything a human can do in game, but it hasn’t been trained to win — which sounds a lot like what happens when I try to play games.
What really makes it impressive is SIMA can exist across games without requiring special access to data not normally available to bots or players.
SIMA can do anything a human can do in game, but it hasn’t been trained to win — which sounds a lot like what happens when I try to play games.
DeepMind built an AI that doesn't need fine-tuning for a specific game, doesn't need access to a games source code to achieve an instructed goal and can learn in response to game play.
Version one was trained on 600 basic skills including navigation, object interaction and menu use.It can complete any tasks within 10 seconds and usually much faster.
"We want our future agents to tackle tasks that require high-level strategic planning and multiple sub-tasks to complete," the team explained.
This could include the human player asking its SIMA-powered teammate to go away and find the resources to build a camp, or to walk the perimeter of an existing camp and sound an alert if enemies are incoming.
A new type of gameplay on the horizon
Bots in games are nothing new, and neither is AI as a companion, guide or even a way to simplify the game play in easy mode — but this takes the concept to a new level.
In the future you could have a personal assistant, built on top of the SIMA technology that sits alongside your Steam, Xbox or Playstation account and can join you in supported games.
Personally I’d welcome something to fend off the attackers when I join a multiplayer game, I might last more than about two minutes in Rust before being shot.
