SIMA 2

SIMA 2 is Google DeepMind’s second generation Scalable Instructable Multiworld Agent, built on Gemini to reason about goals and act inside complex 3D virtual environments. It takes visual input from games, combines it with language prompts, and outputs keyboard and mouse actions, so it can follow high level instructions such as exploring, gathering resources, or solving in game objectives. Compared with the original SIMA, SIMA 2 is designed to plan multi step strategies, explain what it is trying to do, and improve via self play across many different titles. DeepMind uses these rich simulated worlds as a testbed for more general embodied intelligence, with a long term view toward agents that can eventually transfer skills to real world robotics and interactive systems.