Gemini Robotics 2 brings whole body intelligence to robots

Details

Google DeepMind's Gemini Robotics 2 is a major step in applied robotics AI, introducing a three-model suite that addresses the key gaps in current robotic systems: whole-body coordination, fine manipulation, and cross-embodiment transfer.\n\nGemini Robotics 2 is the primary vision-language-action (VLA) model, converting vision and language input directly into motor control. It handles full humanoid robots — feet to fingertips — and bi-arm robots, with a new emphasis on dexterous manipulation across both hands and grippers. The same model checkpoint controls multiple embodiments including the Apptronik Apollo 2 robot with different hand types and the Franka Duo with a Robotiq gripper.\n\nGemini Robotics ER 2 is the embodied reasoning (VLM) layer, acting as a planning agent: it communicates with humans, understands the physical world, and orchestrates multi-step tasks lasting several minutes. It adds a new capability: coordinating multiple robots working as a team — enabling faster task completion through collaboration.\n\nGemini Robotics On-Device 2 is an efficient VLA model optimized to run locally on robotic hardware, with the key advance that it can adapt to completely new robot embodiments in a few hours of data — critical for real-world deployment where hardware varies. This combination of intelligence, dexterity, and fast embodiment transfer represents a significant advance toward robots that can be deployed and retrained outside controlled lab environments.