Google DeepMind has entered a research partnership with Fenris Creations, the independent studio behind the massively multiplayer online game EVE Online. This collaboration aims to develop artificial intelligence systems that can perform long-horizon planning and continual learning within complex, dynamic environments. The research will utilize EVE Online's persistent virtual universe, which has been running continuously since 2003, as a testbed for AI development.
The partnership builds on DeepMind's 15-year history of using games to advance AI research. Early work included training AI agents to play Atari games from raw pixels, a feat documented in a 2015 Nature paper. This foundational research paved the way for more complex systems like AlphaGo, which defeated the world champion in the game of Go in 2016, and AlphaZero, which generalized across chess, shogi, and Go without human data. More recently, AlphaStar achieved Grandmaster level in StarCraft II in 2019. These game-playing AIs have also informed breakthroughs in other scientific fields, such as protein structure prediction with AlphaFold.
EVE Online offers a unique environment for this research due to its single-shard universe, which accommodates thousands of players interacting within a shared economy. Player-driven political structures and real supply-and-demand dynamics create a complex system that demands AI capable of sustained strategy and adaptation. The research program is structured in stages, beginning in an offline instance of EVE Online before progressing to more interactive environments. This approach allows for the study of AI capabilities such as continual learning without forgetting, memory extending beyond immediate context, long-horizon planning over weeks or months, and complex multi-agent dynamics involving cooperation, competition, and negotiation.
This collaboration is part of a broader trend of AI companies using sophisticated game environments to test and develop advanced AI models. Fenris Creations, formerly known as CCP Games, became independent earlier this year after a $120 million management buyout from Pearl Abyss. DeepMind's minority stake was integrated into this transaction. Hilmar Pétursson, CEO of Fenris Creations, stated that the partnership pushes into uncharted territory where AI must learn and adapt on timescales not demanded by other game environments, helping to understand human-AI coexistence.
DeepMind has also partnered with other game developers, including those using the Unity platform, to integrate AI research into game development tools. This includes creating more intelligent non-player characters (NPCs) and dynamic game environments. The studio's work on generalist agents, such as the SIMA model capable of following natural language instructions across various 3D environments, further demonstrates its commitment to AI that can understand and act on human direction.
The research in EVE Online is not focused on product development but on fundamental AI capabilities. The environment provides two decades of player interaction data, offering a dataset of emergent human decision-making in a complex, rules-light system. This sustained strategic planning capability could have applications beyond gaming, including areas like infrastructure planning and economic modeling. A prototype called Aura Guidance, powered by Gemini, is already being used within EVE Online to improve new player retention.