Google DeepMind on Wednesday unveiled a unique artificial intelligence (AI) model that can play 3D video games like a human. The AI ​​model, called the Scalable Instruction Multi-World Agent (SIMA), is learning how to interact with various game environments and perform different tasks. The model is currently under research and is still being trained to build expertise in performing more complex actions. Google says that once perfected, the AI ​​model could have a wide range of use cases both online and in the real world.

in a blog post, DeepMind’s SIMA team explained that the AI ​​model was not developed to be a super-intelligent gamer that can beat any game. Instead, the goal is to teach it how to move around a 3D game in an open-world situation and use natural language instructions to understand how humans react and interact with it. Google emphasizes that this is a very difficult task and should be a priority.

“This is an important goal for artificial intelligence in general, because while large language models have given rise to powerful systems that can capture knowledge about the world and generate plans, they currently lack the ability to act on our behalf,” the SIMA team said. .”

Tasks performed by the AI ​​model SIMA
Image source: Google DeepMind

To create a learning environment for the AI ​​model, Google DeepMind worked with eight game studios and trained SIMA on nine different video games. These include Hello Games’ No Man’s Sky, Tuxedo Labs’ Teardown, Coffee Stain Studios’ Goat Simulator 3 and Valhalla, among others. AI models are exposed to new interactive worlds with each game and must learn to navigate the world, interact with objects, use menus, and more.

The company has also created four research environments, including a new environment called Build Lab built using Unity. In this special environment, the AI ​​model was tasked with constructing sculptures out of building blocks to test its object manipulation and intuitive understanding of the physical world.

Google emphasized that the current version of SIMA assesses 600 basic skills, including navigation such as turning left and driving a car, climbing ladders and interacting with objects such as making a helmet. The task is basically simple and can be completed in 10 seconds.

The company says training AI models on a variety of 3D video games to follow human instructions and act accordingly could have a significant impact in the future. The company is now focused on training SIMA with more complex instructions “that require high-level strategic planning and multiple sub-tasks to complete, such as finding resources and setting up camp.” Google hints that these tasks can translate to the real world, the company Efforts are being made to find ways to help humanity.