Gemini Robotics 2 controls a full humanoid with one learned model

A humanoid robot that walks, crouches and works is now driven by one model.

On July 30, 2026, Google DeepMind unveiled Gemini Robotics 2, its first AI that controls a full humanoid, legs, torso, arms and fingers, under one learned policy.

From upper body to whole body

DeepMind says its previous models controlled the humanoid’s upper body to achieve table-top tasks. Gemini Robotics 2 extends that to whole-body control: one vision-language-action model handles the whole body.

In the demo, Apptronik’s Apollo 2 robot, asked to put a watering can into a bin on the bottom shelf, walks to the table, picks up the can, takes a few steps to the shelves and places it. Locomotion, navigation and manipulation, all coming from the same model making the same kind of decisions.

The family and the numbers

The release is actually three models. Gemini Robotics 2 is the action model. ER 2 handles embodied reasoning and multi-step planning. On-Device 2 runs locally on the robot with no cloud dependency, which matters for latency and for anywhere connectivity is not guaranteed.

DeepMind’s reported numbers vary by task. With 22-degree-of-freedom hands, success reaches 92% for unscrewing a light bulb, while other multi-finger tasks land between 32% and 44%, including 44% for tying a trash bag. DeepMind also says On-Device 2 adapts to new bi-arm robot embodiments in a few hours, typically with fewer than 200 examples. That last figure is the strategic one. A model that transfers to new hardware that cheaply is not a robot product, it is an operating system candidate.

Availability splits by layer: ER 2 is already in Google AI Studio, while the action models go to early-access partners.

The platform question

The humanoid race just gained a serious operating-system contender, and it sharpens the industry’s central strategic question. One camp bets on a general brain that runs on anyone’s body, which is Google’s play here. The other bets on vertical integration, one company building brain and body together.

If a few hours and 200 examples is really all it takes to move this model onto new hardware, the brain-for-any-body camp just got its strongest evidence yet.