Google DeepMind Unveils AI Capable of Executing Commands in Unfamiliar 3D Games

Gray Frame Corner

Google DeepMind introduces AI capable of performing various tasks in unfamiliar 3D games.

The AI can understand natural-language instructions and navigate through different gaming environments.

Collaboration with game studios helped train the Scalable Instructable Multiworld Agent (SIMA) on nine games.

SIMA learns from diverse environments, including those created in Unity, without needing game source code.

It operates with only on-screen visuals and user instructions, using keyboard and mouse inputs like a human.

SIMA undergoes evaluation on hundreds of basic skills across categories like navigation and object interaction.

Performance evaluations show SIMA trained on multiple games outperforms those trained on single titles.

SIMA demonstrates the ability to generalize its skills to new environments beyond its training.

Language input is essential for SIMA's successful performance, as observed in tests without language training.

DeepMind emphasizes ongoing research to enhance SIMA's capabilities and develop more general AI agents for various tasks.

White Frame Corner