Google DeepMind Says Gemini Robotics 2 Enables Full Body Control

Google DeepMind says its latest model, Gemini Robotics 2, enables robots to reason through every movement, unlocking a broad range of tasks, the Robot Report reported on August 2. The update is a notable step toward truly general-purpose robots, with control over the entire body rather than just hands or arms.

What the Model Does

Previous versions of Gemini Robotics could handle specific tasks like grasping, manipulation, or navigation. Gemini Robotics 2 adds "full body control" - meaning the model reasons about the robot's posture, balance, arm trajectory, hand grip, and locomotion as a unified problem. This unlocks tasks that require coordinated whole-body motion, such as climbing stairs while carrying objects, opening doors, or working in tight spaces.

"Gemini Robotics 2 enables robots to reason through every movement, unlocking a broad range of tasks, DeepMind said," the Robot Report wrote.

Why It's Hard

Full body control has been a long-standing challenge in robotics because the state space is enormous - a humanoid robot can have dozens of joints, each with multiple degrees of freedom. Earlier approaches required specialized controllers for each task. Foundation-model approaches, like Gemini Robotics 2, instead train a single large model on diverse robot data and let it generalize.

What's Next

DeepMind is publishing Gemini Robotics 2 to academic and commercial partners. The model has been tested on multiple robot platforms, including the Apptronik Apollo and several Boston Dynamics models. Early benchmarks show meaningful improvements over the previous version on a wide range of tasks. The full release is planned for late 2026, with broader commercial availability in 2027.