Gemini Robotics 2 Moves Full-Body Control From Lab Demo to Multi-Robot Coordination

2. Gemini Robotics 2 Moves Full-Body Control From Lab Demo to Multi-Robot Coordination

Google DeepMind announced Gemini Robotics 2 on July 30, 2026, expanding its robotics model from upper-body control to full whole-body motion covering feet to fingertips. The update runs on Apptronik's Apollo 2 humanoid, enabling crouching, walking, stretching, and fine manipulation tasks like sealing a Ziploc bag, tying a trash bag, and unscrewing a lightbulb using five-fingered hands. DeepMind also updated Gemini Robotics ER 2, its embodied-reasoning vision-language layer, which now tracks when tasks begin and end and coordinates multiple robots of different types working together. A companion on-device model adapts faster to new hardware embodiments, including robots with "drastically different shapes, sensors and degrees of freedom," without requiring an internet connection.

The scope jump matters competitively. Figure AI, Physical Intelligence, and Boston Dynamics are all racing to prove that a single model stack can generalize across robot morphologies and real-world task complexity. Gemini Robotics 2 answers that race on two fronts simultaneously: whole-body motor control and cross-robot coordination. The multi-robot tasking demo, where Apollo 2 directs Google's dual-arm robot to sort tools, signals that DeepMind is positioning Gemini Robotics ER 2 as an orchestration layer, not just a single-arm controller. That architectural bet, if it holds in production, shifts the competitive frame from "which model moves a robot arm best" to "which platform can coordinate a fleet."

DeepMind's own caveat is worth tracking: the announcement notes that robots "have more to advance in movement speed." Speed is the remaining gap between controlled demos and warehouse-grade deployment. Watch whether Physical Intelligence's pi0 or Figure's next model iteration closes that speed gap before DeepMind does, because whoever solves speed at whole-body scale first sets the commercial timeline for humanoid robotics.

Source: Google DeepMind's new AI model can control a robot's entire body