Particle.news
Download on the App Store

DeepMind Debuts SIMA 2, a Gemini-Powered Agent That Learns in Virtual Worlds

The preview emphasizes self-directed training plus stronger instruction following, with real‑world robotics still unscheduled.

Overview

  • SIMA 2 runs on Gemini 2.5 flash‑lite, enabling the agent to reason about goals, explain its steps, and follow instructions via text, voice, sketches, or emojis.
  • DeepMind reports roughly doubled performance over SIMA 1, citing 65% task completion versus 31% on complex tasks in tests.
  • Trained on human gameplay across eight commercial titles including No Man’s Sky and Goat Simulator 3, the agent generalizes to unfamiliar 3D environments.
  • Genie 3 generates new photorealistic worlds and tasks, and a reward model scores attempts so the agent learns new behaviors through trial and error with minimal new human data.
  • Researchers note limits with long multi‑step missions, a shortened memory window, and precise mouse/keyboard control, and they are offering only a limited research preview with no timeline for robotics.