Google DeepMind has introduced Gemini Robotics 2, a suite of AI models designed to advance general-purpose physical intelligence through whole-body control, fine dexterity, and multi-robot collaboration. The system comprises three core models: Gemini Robotics 2, a vision-language-action (VLA) model that translates multimodal inputs into precise motor control for humanoids and bi-arm robots; Gemini Robotics ER 2, an embodied reasoning (ER) model that acts as a high-level agent to plan multi-step tasks lasting several minutes; and Gemini Robotics On-Device 2, an efficient VLA optimized for local execution. Technically, the platform enables sophisticated manipulation, such as controlling the 22-degree-of-freedom SharpaWave hand on the Apollo 2 humanoid for delicate tasks like tying knots. Furthermore, the on-device model utilizes "motion transfer" techniques to adapt to entirely new robotic embodiments—including the Dexmate and SO101 platforms—using less than 200 examples and only a few hours of data. To ensure safety in human-centric environments, the researchers introduced ASIMOV, a new benchmark for agentic safety orchestration, which evaluates a model's ability to refuse unsafe commands and proactively request human intervention during uncertain scenarios.