SIMA 2: An Agent That Plays, Reasons, and Learns with You in Virtual 3D Worlds (deepmind.google)

DeepMind/Google announced SIMA 2, a large multi‑author project that builds an interactive agent that can play, reason, and learn alongside people inside richly featured 3D game worlds. The team partnered with many commercial game developers (Valheim, No Man’s Sky, Teardown, Space Engineers, and others) to train and test the agent across diverse virtual environments, and they used access to the Genie 3 model as part of the system’s language and reasoning stack. The paper and release emphasize human-in-the-loop development backed by extensive cross‑team support for safety, responsibility, and tooling. SIMA 2 is significant because it pushes embodied AI beyond narrow task bots toward a single agent that integrates multimodal perception, language understanding, long‑horizon planning, and online learning across heterogeneous, open-ended simulations. Technically, the project combines large language model reasoning (Genie 3), visual and action policy components, and large-scale interactive training on real commercial game environments — enabling richer human–agent collaboration, tool use and emergent behaviors in complex 3D settings. For the AI/ML community this offers a new benchmark and development platform for studying generalization, instruction following, continual learning, and safety practices in interactive, open-world domains.