Google DeepMind partners with game studios to prototype AI gameplay agents

Google DeepMind described SIMA 2 as a Scalable Instructable Multiworld Agent designed to follow natural language instructions, observe on-screen content, and act through standard keyboard and mouse inputs—without requiring game-specific APIs or code access. The agent is powered by the company’s Gemini models and is intended to function as an interactive companion capable of real-time reasoning and conversation.

The research team reported that SIMA 2 achieved human-like play across complex 3D research environments and commercial video games, including No Man's Sky, Valheim, and Hydroneer. These environments were selected to test the agent’s ability to generalize across different game mechanics, visual styles, and interaction paradigms.

The announcement highlighted a multi-year collaboration with game developers to prototype new gameplay experiences using AI agents. Specific partnerships mentioned include Fenris Creations, Hello Games, Coffee Stain Studios, and Foulball Hangover. The collaborations aim to explore how AI can enhance game design, player interaction, and narrative experiences without altering existing game codebases.

The work builds on 15 years of AI research in games, from the Deep Q-Network (DQN) that learned to play Atari games directly from pixels to AlphaGo’s defeat of world champion Lee Sae Dol in 2016. The team noted that AlphaGo’s Move 37 and AlphaZero’s novel strategies influenced both AI research and human play, demonstrating how AI breakthroughs can inspire new directions in both technology and culture.

The post emphasized that while earlier systems optimized for high scores, SIMA 2 focuses on understanding and interacting with game worlds in a human-like manner. This shift reflects a broader research goal: developing AI that can operate in unstructured, real-world-like environments where objectives and rules are not predefined.