SIMA 2: An Agent that Plays, Reasons, and Learns With You in Virtual 3D Worlds

https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/

Gemini core; reasons about goals, explains actions, converses, answers questions about environment; follows sketches, emoji, multiple languages; unseen games ASKA and MineDojo and Genie 3 worlds; self-improvement from Gemini-proposed tasks and rewards. Limited research preview for academics and developers.

Agent-based automated playtesting and balancing, Generalist instructable game agents (SIMA), LLM explanation of simulation state, Open-ended co-generation of environments and agents