Overview
- SIMA 2 integrates Google’s Gemini model to reason about scenes and act in virtual environments, with DeepMind reporting roughly double the performance of SIMA 1.
- Training drew on human gameplay from eight commercial titles, including No Man’s Sky and Goat Simulator 3, plus three company-built worlds to map pixels to mouse‑keyboard actions.
- DeepMind used its Genie 3 world model to generate novel photorealistic environments, where SIMA 2 navigated and interacted with previously unseen objects.
- The agent self-improves by attempting Gemini-generated tasks, receiving AI feedback and reward-model scores, and learning via repeated trial and error.
- DeepMind stresses the system remains experimental, struggling with long multi-step tasks and fine motor control, with limited memory and no timeline for robotics deployment or wider release.