Google DeepMind releases Gemini Robotics 2, giving humanoid robots whole-body AI control

Summary

Google Deepmind released Gemini Robotics 2 on July 30, a family of AI models designed to give humanoid and other robots whole-body coordination, fine-grained manipulation, and the ability to coordinate tasks across multiple robots. The system was demonstrated with Apptronik's humanoid robot performing tasks including unscrewing a light bulb (92% success rate), clearing trash, and picking up objects. Gemini Robotics 2 differs from earlier robotics AI by simultaneously controlling head, torso, arms, and legs rather than treating each limb independently. Google said the model series targets a range of robot form factors.

Why it matters

Whole-body coordination is the main unsolved bottleneck in humanoid robotics: robots that cannot balance and manipulate at the same time fail on tasks that require both. A 92% success rate on a screw-in task with simultaneous whole-body movement, if reproducible outside a lab, represents a practical advance. Google DeepMind is competing with Tesla (Optimus), Figure, Boston Dynamics, and Agility Robotics for leadership in the humanoid AI stack, and a robotics-focused Gemini model family signals Alphabet treating robotics as a product line rather than a research project.

What to watch

• Whether Gemini Robotics 2 is made available to third-party robot manufacturers beyond Apptronik, and on what commercial terms
• Competing releases from Tesla, Figure, and Boston Dynamics, all of which have announced comparable whole-body coordination projects for 2026
• China's parallel push, with Beijing's robot export restriction rules from July 30 signalling intent to control the robotics supply chain