Google DeepMind Expands Game AI Research With EVE Online Partnership

Google DeepMind is marking 15 years of game-based AI research with a shift in emphasis: from systems that optimize for victory to agents that can understand and operate inside open-ended game worlds.

The company’s earlier milestones include the Atari-playing Deep Q-Network, AlphaGo, AlphaZero, MuZero, and AlphaStar. These projects helped advance reinforcement learning, self-play, planning, and decision-making under imperfect information. DeepMind says the same research direction also contributed to work such as AlphaFold’s protein-structure predictions.

From high scores to general-purpose agents

That newer focus is represented by SIMA, a generalist agent that interprets what appears on screen, follows natural-language instructions, and uses standard keyboard and mouse inputs without requiring access to a game’s source code or APIs. SIMA 2, powered by Gemini, is designed to reason, converse, and act in real time across 3D environments and games including No Man’s Sky, Valheim, and Hydroneer.

For developers, the potential applications include companions that understand the game world, more responsive NPCs, and AI-driven quality assurance that can keep up as builds and content change. DeepMind also frames these systems as a way to explore gameplay concepts that would be difficult or impossible to create with conventional scripting.

The company’s latest research partnership is with Fenris Creations, the independent studio behind the EVE Universe. EVE Online offers a persistent, player-shaped setting with a complex economy, extensive trade networks, and social dynamics formed through alliances, conflict, and diplomacy. DeepMind says this kind of evolving environment provides a demanding test bed for frontier AI, while the collaboration will rely on developers and researchers building playable prototypes together and evaluating whether they are genuinely fun.

Source: Google DeepMind Blog