Google DeepMind announced new research partnerships with video game developers to test generalist artificial intelligence agents inside persistent commercial titles, extending an initiative that began with Atari systems after the lab's founding in 2010.
The research builds on SIMA, short for Scalable Instructable Multiworld Agent, and its successor SIMA 2. Powered by Google DeepMind's Gemini models, the agent processes on-screen visuals, follows natural language instructions, and inputs actions using standard keyboard and mouse controls without requiring source code or direct API integration. DeepMind has tested the system across commercial 3D environments, including Hello Games' No Man's Sky, Coffee Stain's Valheim, and Foulball Hangover's Hydroneer.
Evolution of Game Agents
DeepMind's early work focused on mastering specific titles under closed rule sets. In 2015, the lab published research on Deep Q-Networks that learned to play 49 Atari 2600 games directly from raw pixels. The lab subsequently developed AlphaGo, which defeated world champion Go player Lee Sae Dol in 2016, followed by AlphaGo Zero, AlphaZero, and MuZero. In 2019, its AlphaStar system reached Grandmaster status in StarCraft II. The underlying reinforcement learning methods also supported AlphaFold's protein structure predictions, which received the 2024 Nobel Prize in Chemistry.
Rather than optimizing for raw point totals, the newer SIMA models act as interactive companions intended to evaluate real-time reasoning and assist developers with automated quality assurance testing across frequent code updates.
The EVE Research Roadmap
A central component of DeepMind's current program is a partnership with Fenris Creations, the studio behind the persistent space simulation EVE Online. The collaboration has already deployed the Aura Guidance system, which uses Gemini to answer new player questions using historical Rookie Help data.
DeepMind's experimental plan begins inside an offline sandbox instance of EVE Online, keeping experimental models separated from the live player base. Research will next move to EVE Frontier, an environment built with programmable "Smart Assemblies" and mutable rules, before the team evaluates deploying mature capabilities into live servers of EVE Online and the first-person tactical game EVE Vanguard.
The lab is also conducting research alongside Keen Software House on Space Engineers, RubberbandGames on Wobbly Life, Strange Loop Games on Eco, Thunderful Games on ASKA and Steamworld Build, Digixart on Road 96, and Tuxedo Labs with Saber Interactive on Teardown.
