Google DeepMind releases Gemini Robotics 2 for humanoids

What happened: Google DeepMind unveiled Gemini Robotics 2 on 30 July 2026, a family of three models built to give humanoid and two-armed robots more capable, coordinated movement.

The details: Gemini Robotics 2 handles vision-language-action control, turning what a robot sees and hears into motor commands, including dexterous hand movements. Gemini Robotics ER 2 handles embodied reasoning: planning, talking to people, and coordinating multi-step tasks across several robots at once. Gemini Robotics On-Device 2 runs locally on the robot and adapts to a new robot body within hours using fewer than 200 examples.

The numbers: In DeepMind's tests, general whole-body manipulation succeeded 45.7-76.3% of the time. Gripper tasks scored 74.2-89.6%. Multi-finger dexterity ranged from 32-92%.

Who's affected: The models already support Apptronik Apollo 2, Franka Duo, Dexmate, SO101 and Trossen robot platforms. A single checkpoint can control different robot bodies and plan actions over several minutes.

The catch: DeepMind acknowledged the models still struggle with fine motor skills, saying "the multi-finger dexterous manipulation remains challenging."

Why it matters: Chinese manufacturers currently dominate humanoid robot shipments. Google is betting that its models, rather than any single robot maker's hardware, become the common intelligence layer running across the industry's machines.

What's next: Gemini Robotics ER 2 is available now in Google AI Studio and via private preview. The vision-language-action and on-device models are rolling out to early-access partners.

Details from Google DeepMind's announcement.