Google DeepMind's Gemini Robotics 2 now controls humanoid robots' entire bodies

Google DeepMind's progression from controlling a robot's upper body to its entire body represents a significant step in making humanoid robots capable of handling real-world work. The announcement on Thursday describes Gemini Robotics 2 as supporting "whole-body motions" — a shift that opens up tasks requiring coordination across the robot's limbs and fingers. The company acknowledges that robots "have more to advance in movement speed," indicating that performance still lags behind a fully capable system, yet this update positions the technology toward practical deployment in environments where multiple robots or complex multi-step actions are needed.

The companion update to Gemini Robotics ER 2 — the vision-language model that interprets instructions and analyzes surroundings — adds a layer of operational maturity. By understanding "when tasks begin and end" and detecting human proximity for safety, the model moves beyond isolated motion control into a more complete robotic system. Google's emphasis on safety (branding it the "safest robotics model to date") signals that deployment concerns around human-robot interaction are part of the development roadmap. The On-Device Model improvement, allowing faster adaptation to robots with different physical configurations, suggests Google is working toward flexibility across hardware partners like Apptronik, whose Apollo 2 robot appears in the company's demonstration videos.