DeepMind announces partnership with Fenris Creations to develop generalist gaming agents in the EVE Universe

TL;DR

DeepMind announced a research partnership with Fenris Creations to develop general‑purpose AI agents that can operate in the persistent EVE Online universe, targeting capabilities such as continual learning, long‑term memory, and multi‑agent planning to unlock novel gameplay and inform broader AI research.

Gaming as a catalyst for AI breakthroughs

DeepMind’s AI research has historically used games as testbeds. Starting with the Deep Q‑Network (DQN) that learned 49 Atari 2600 titles from raw pixels (Nature, 2015), the lab progressed to AlphaGo (2016), AlphaGo Zero, AlphaZero, MuZero, and AlphaStar, each demonstrating increasingly general reinforcement‑learning methods. Successes in games directly fed into non‑gaming domains, most notably AlphaFold, which leveraged game‑derived techniques to solve protein‑structure prediction and earned a 2024 Nobel Prize in Chemistry.

From mastering games to understanding them: the SIMA agenda

DeepMind’s next research question is whether an AI can understand any game world like a human player, without access to internal APIs or source code. The Scalable Instructable Multiworld Agent (SIMA) series addresses this by:

  • Observing the game screen exactly as a player would.
  • Interpreting natural‑language instructions.
  • Controlling the game via standard keyboard and mouse inputs.

SIMA 2, built on the Gemini family of frontier models, demonstrates real‑time reasoning, conversation, and human‑level play across 3D environments such as No Man’s Sky, Valheim, and Hydroneer. A truly general gaming agent would enable AI companions that comprehend the game world and NPCs that adapt beyond scripted behavior, as well as robust QA testing that tolerates continuous code changes.

The Fenris Creations partnership: a living laboratory

Fenris Creations, the independent studio behind the EVE Universe, provides a unique, evolving sandbox for testing these capabilities. EVE Online’s persistent, player‑driven economy and multi‑scale conflict present four core challenges that align with DeepMind’s frontier AI goals:

  1. Continual learning – acquiring new skills while retaining prior knowledge in a world that changes daily.
  2. Long‑term memory – storing and retrieving information across timescales far beyond current model context windows.
  3. Long‑horizon planning – making decisions that span weeks, months, or years.
  4. Complex multi‑agent dynamics – handling cooperation, competition, negotiation, and emergent social behavior at massive scale.

The partnership spans three EVE environments:

  • EVE Online – a large‑scale, single‑shard persistent universe for studying galaxy‑wide strategy.
  • EVE Vanguard – a first‑person, fast‑paced tactical layer that bridges ground‑level action with the broader universe.
  • EVE Frontier – an open, programmable sandbox with “Smart Assemblies” that let the rules of the world evolve, testing agents’ ability to adapt to novel mechanics.

Early成果 and roadmap

A concrete outcome of the collaboration is the Aura Guidance system, which uses Gemini to surface community‑generated help content to new pilots in EVE Online. The research roadmap proceeds in stages:

  1. Offline sandbox – a safe copy of EVE Online for initial agent training and safety testing.
  2. EVE Frontier experiments – co‑development of agents that coexist with humans in an open‑ended setting.
  3. Live integration – once agents demonstrate mature capabilities, they may be introduced to EVE Online and EVE Vanguard to enrich player experiences.

Implications for AI and gaming

If successful, the partnership could:

  • Deliver AI companions that understand natural language, remember past interactions, and plan over extended horizons.
  • Enable dynamic, adaptive NPCs that react to emergent player behavior without manual scripting.
  • Provide a testbed for continual‑learning algorithms that scale to real‑world complexity.
  • Transfer insights from persistent virtual economies to real‑world domains such as finance, logistics, and collaborative robotics.

Conclusion

DeepMind’s collaboration with Fenris Creations marks a strategic shift from AI that masters games to AI that understands and co‑evolves with complex, persistent virtual worlds. By tackling continual learning, long‑term memory, and multi‑agent dynamics within the EVE Universe, the partnership aims to create breakthrough gameplay experiences and generate transferable AI knowledge for broader scientific challenges.

Sources

Related