Google’s SIMA 2 Agent Uses Gemini to Reason and Act in Virtual Worlds

Why It Matters

By combining advanced language reasoning with embodied interaction, SIMA 2 showcases a viable path to AI systems that can learn and operate in real‑world settings with minimal human supervision, accelerating progress toward autonomous robotics and general AI.

Summary

Google DeepMind unveiled SIMA 2, an embodied AI agent that leverages the Gemini 2.5 flash‑lite model to reason, understand language, and act in virtual environments. Compared with its predecessor, SIMA 2 doubled task‑completion rates, demonstrating complex, commonsense behavior in games like No Man’s Sky and photorealistic worlds generated by DeepMind’s Genie. The agent can self‑improve by generating its own tasks and reward signals, reducing reliance on human‑labeled data. DeepMind positions SIMA 2 as a stepping stone toward more general‑purpose robots and broader AGI applications.