Google DeepMind uses EVE Online to test generalist game agents

Google DeepMind is expanding its game-based AI research through partnerships with game studios, including Fenris Creations and the EVE Universe. Its work builds on a lineage that began with Atari 2600 systems and the Deep Q-Network, which learned to play 49 different games without game-specific engineering, followed by AlphaGo, AlphaZero, MuZero, AlphaStar and AlphaFold.

The current focus is SIMA, the Scalable Instructable Multiworld Agent, and SIMA 2, a Gemini-powered companion designed to interpret on-screen environments, follow natural language instructions and operate through standard keyboard and mouse controls. Google DeepMind says general gaming agents could support adaptive NPCs, in-game companions, QA testing and post-launch content changes without requiring source code access.

The EVE partnership gives researchers a persistent world for studying continual learning, memory, long-horizon planning and complex multi-agent dynamics. Work will start in an offline EVE Online instance before moving to EVE Frontier experiments on human-agent coexistence, with any integration into EVE Online or EVE Vanguard considered only when capabilities are mature.