Google DeepMind has created a new AI testing platform using video games. This platform is called SIMA, which stands for Scalable Instructable Multiworld Agent. SIMA is not designed to win games. Its purpose is different. SIMA learns general skills useful in virtual worlds. This research aims to develop AI assistants that perform tasks in complex environments following natural language instructions.
(Google DeepMind develops game AI testing platform)
SIMA works across many different games. It learns skills by watching humans play. It observes keyboard and mouse actions paired with language descriptions. This method teaches SIMA how to act on natural language directions within a game. The AI learns to understand and navigate 3D worlds. It figures out object locations and how to interact with them. SIMA learns basic but essential skills like turning, climbing ladders, and using in-game menus.
The team trained SIMA on nine different video games. These included popular titles like No Man’s Sky and Teardown. SIMA also trained in four research environments built with tools like Unity and Unreal Engine. This variety is crucial. It tests SIMA’s ability to apply learned skills to completely new situations. SIMA must perform tasks in games it has never seen before. Its instructions come only as simple, natural language commands. For example, “Turn left” or “Climb that ladder.”
DeepMind believes gaming offers a perfect testing ground. Games simulate complex worlds with clear goals. They provide rich, controllable environments for AI research. Developing AI that follows instructions reliably in these settings is a major step. It paves the way for more helpful real-world AI applications. Imagine an AI helper in complex software or a detailed simulation. SIMA represents progress towards that future capability.
(Google DeepMind develops game AI testing platform)
Currently, SIMA can perform over 600 basic skills. These range from simple movement to interacting with objects. Its performance improves significantly with training on more games. DeepMind is actively expanding SIMA’s capabilities. They are adding more games and partners to the project. The research team continues refining SIMA’s ability to understand and execute complex instructions within diverse virtual spaces. This work focuses on creating more adaptable and useful AI agents.