Google DeepMind Ships Three Physical AI Models For Whole-Body Control, Dexterity And Multi-Robot Collaboration

Google DeepMind has released Gemini Robotics 2, designed to control humanoid and bi-arm robots.

The stack includes a vision-language-action model, an embodied reasoning model, and an on-device version for multi-step planning and dexterous manipulation.

DeepMind also introduced ASIMOV-Agentic, a new safety benchmark for agentic robotics.

Only the embodied reasoning model is publicly previewable, while the other two remain access-restricted.