Google DeepMind has identified video games as a significant testing ground for artificial intelligence (AI) research. The company aims to expand AI research that can be applied to the real world by combining AI agents that understand language instructions and respond to user actions within complex 3D environments.
Alexandre Mufarech, Director of Google DeepMind, presented a session titled "How Google DeepMind Advances AI Research Through Video Games" at the Nexon Developer Conference 2026 (NDC) held on June 17 at the Pangyo Creative Economy Innovation Center in Seongnam.
Mufarech stated, "Games provide a minimal model for learning more complex dynamics," emphasizing that they are a core foundation of DeepMind's research program. He noted that through game-based research, Google DeepMind has expanded AI's reasoning, planning, and adaptability capabilities, building on the legacy of projects like AlphaGo, AlphaZero, and AlphaStar. The company validated AI's planning abilities through AlphaGo.
Currently, DeepMind is focusing on the AI agent SIMA and the world model Genie. SIMA is an embodied AI agent that understands and acts on natural language instructions in a 3D environment. It operates by recognizing only the pixels on the user's screen, manipulating characters through keyboard and mouse inputs without relying on game data or separate application programming interfaces (APIs).
Mufarech emphasized that SIMA is not designed to excel at specific games but aims to function as a general agent across various environments. He explained, "The core goal of SIMA research is to generalize to environments it has never encountered before," noting that complex and rich game sets serve as the foundation for this research.
The world model Genie generates the environment in which agents operate, simulating the dynamics of the world to respond to user actions. DeepMind introduced a short interactive environment based on a 2D platform game with Genie1 in February 2024, later expanding to a 3D fantasy world with Genie2. The recently unveiled Genie3 supports 24 frames per second generation, real-time interaction, and rendering up to 720p.
Genie3 features world consistency and "promptable world events," allowing the model to remember changes made by users and insert new events as prompts even after generation begins. Mufarech described this as similar to scripted events in video games.
DeepMind is integrating SIMA and Genie to create a cyclical research structure where AI agents act within generated virtual worlds and use the outcomes to improve the model.
However, Mufarech drew a line against the notion that AI could replace game creators. He stated, "Just generating a realistic 3D world does not make it a game," adding that games require story, gameplay, tension, and the creator's vision. He explained that AI could serve as a tool to express visions in new ways rather than replace creators.
Collaboration with the gaming industry is also emphasized. DeepMind trains AI agents based on the consent of game developers and is currently working with the development team of the massive multiplayer online game (MMO) EVE Online. Mufarech noted that EVE Online, with its persistent world and player-driven economy, provides a suitable environment for studying long-term planning, memory, and multiplayer interactions.
Mufarech concluded, "We must explore what AI research can do for games and confirm through iterative experiments whether we can create new player experiences and game experiences that would be impossible without AI."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.