Google DeepMind has unveiled a significant advancement in artificial intelligence with the introduction of SIMA, which stands for Scalable, Instructable, Multiworld Agent. Unlike previous game-playing AIs that mastered a single game, SIMA is a generalist agent designed to learn and execute tasks across a wide array of 3D video games, a crucial step toward building more versatile and adaptable AI systems.
Announced this week through the company’s research blog, the project involved partnerships with several game developers, including Hello Games (*No Man’s Sky*) and Coffee Stain Studios (*Goat Simulator 3*). SIMA was trained in these rich, interactive environments not by accessing the game’s source code, but by observing human players and responding to natural language instructions. The AI processes visual data from the screen and user commands—such as “climb the ladder” or “find resources to build a shelter”—to learn how to perform complex, multi-step tasks.
This approach marks a departure from specialized AI like AlphaGo, which was trained on the rules of a single, structured game. SIMA’s strength lies in its ability to generalize its learned skills across different virtual worlds. Researchers at DeepMind emphasized that the goal isn’t to create an AI that can beat human players, but one that can understand and follow instructions to collaborate with them.
The underlying technology relies on advanced video and language models that enable the agent to connect linguistic commands with on-screen actions. While currently confined to virtual worlds, the principles behind SIMA have profound implications for the future of AI. The ability to understand and act within any given environment based on simple instructions could pave the way for more capable digital assistants, smarter robotics, and AI that can safely navigate and assist in real-world scenarios. The research demonstrates a powerful new paradigm for training AI to be more helpful and aligned with human intent.
English 
Deutsch