Physical AIRoboticsAI01 sources

Gemini Robotics 2 takes Google's VLA to whole-body control

A full-size black humanoid robot standing in a frontal pose against a plain background, with torso, arms, legs and head visible.

Cadop, via Wikimedia Commons (CC BY-SA 4.0)CC BY-SA

Google DeepMind released three models on 30 July 2026 under the Gemini Robotics 2 name: a vision-language-action model, an embodied-reasoning model called Gemini Robotics ER 2, and an on-device variant. The change of scope is the substance — where earlier versions drove arms, this generation covers whole-body motion, described as walking, crouching, stretching and manipulating objects within one continuous task.

DeepMind describes a single checkpoint driving Apptronik’s Apollo 2 humanoid with two different hand types, and the same model running on a Franka Duo with a Robotiq gripper and on Dexmate, SO101 and Trossen platforms. Multi-finger dexterity is shown on a 22-degree-of-freedom SharpaWave hand. The release also introduces multi-robot collaboration, letting different types of robot communicate and work together on one task.

Availability is uneven. Gemini Robotics ER 2 is on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, while the vision-language-action and on-device models go to early-access partners only. A separate document, the Gemini Robotics 2: Safety Technical Report, was published alongside.

Sources

  1. [1]Gemini Robotics 2 brings whole body intelligence to robotsGoogle DeepMind·