What Google's Gemini Robotics 2 means for you

Google DeepMind has built a set of AI models that act as the "brain" for robots, letting them plan tasks, coordinate with other robots, and control their whole bodies, including hands and fingers, rather than just a single arm.

On 30 July 2026, Google DeepMind announced Gemini Robotics 2, a family of three models built for physical robots rather than chatbots or apps.

The first, Gemini Robotics 2, is a "vision-language-action" model, technology that turns what a robot sees and hears into actual movement. It controls full humanoid robots and two-armed machines, including detailed hand movements.

The second, Gemini Robotics ER 2 ("embodied reasoning"), handles higher-level thinking: understanding instructions, talking with people, planning multi-step jobs, and now coordinating several robots working together on the same task.

The third, Gemini Robotics On-Device 2, is a smaller, efficient model that runs directly on the robot rather than over the internet. Google said it can adapt to a completely new robot body within hours, using fewer than 200 examples.

The models already work with several robot platforms, including Apptronik's Apollo 2, Franka Duo, Dexmate, SO101 and Trossen machines. A single version of the software can run across these different robot bodies and plan actions over several minutes.

In testing, general whole-body tasks succeeded between 45.7% and 76.3% of the time. Simpler gripper tasks succeeded 74.2% to 89.6% of the time. Tasks needing fine finger control were far less reliable, ranging from 32% to 92%. Google DeepMind acknowledged this gap directly, saying "the multi-finger dexterous manipulation remains challenging".

You won't be using these models yourself any time soon. Gemini Robotics ER 2 is available in Google AI Studio and through a private preview, while the two more advanced models are going only to early-access partners who build or operate robots.

The bigger picture: most humanoid robots shipped today come from Chinese manufacturers. Google is trying to become the software layer that powers robots made by many different companies, in the way its Android system powers phones from many different manufacturers.

Google has not said when these models will reach a wider group of developers or the public. For now, progress depends on how well Gemini Robotics 2 performs with the early-access partners testing it on real robot hardware.