Google’s SIMA 2 agent uses Gemini to reason and act in virtual worlds

Google DeepMind has introduced SIMA 2, an advanced generalist AI agent that leverages Gemini, Google’s large language model, to enhance its reasoning and interaction capabilities within virtual environments. Unlike its predecessor SIMA 1, which had limited success (31%) in completing complex tasks, SIMA 2 doubles that performance and can handle complex tasks in previously unseen 3D game environments. Trained on extensive video game data, SIMA 2 integrates Gemini’s language and reasoning skills with embodied agent capabilities—meaning it perceives and acts within virtual worlds similarly to how humans or robots would. This integration enables SIMA 2 to understand instructions contextually, reason internally (e.g., associating a “ripe tomato” with the color red), and even follow emoji-based commands. A key innovation of SIMA 2 is its self-improving ability. While SIMA 1 relied solely on human gameplay data, SIMA 2 starts with that baseline but then autonomously generates