The VLA model that intelligently controls any robot
Gemini Robotics 2.5 is Google DeepMind's most advanced Vision-Language-Action (VLA) model for robotics. It is designed to intelligently control virtually any type of robot — from tabletop robot arms to full humanoid bodies — enabling robots to reason about tasks and then physically interact with the world around them.
Controls robots of any shape or size, scaling its intelligence from simple robot arms to complex humanoid bodies.
Controls entire humanoid bodies from feet to fingertips, dynamically balancing the robot's centre of gravity so it can safely step, squat, and bend to navigate cluttered, real-world spaces rather than reaching from a static position.
Supports human-level dexterity with complex, 22-degree-of-freedom hands and grippers.
Uses reasoning to adapt its behavior to different situations and to complete unexpected, changeable, and complex tasks.
The model enables a robot to think about a task and then interact with the physical world around it, using reasoning to adapt behavior to the situation at hand.
For humanoids, the model coordinates the entire body from feet to fingertips, dynamically managing the robot's centre of gravity so it can step, squat, and bend while working in cluttered real-world spaces.
It is Google DeepMind's most advanced Vision-Language-Action (VLA) model for robotics, capable of intelligently controlling any type of robot, from tabletop robots to full humanoids.
It can operate robotic bodies of all sizes and shapes, scaling its intelligence from simple arms to complex humanoid bodies, including dexterous manipulation of hands and grippers and full control of a humanoid body from feet to fingertips.
It provides intelligent whole-body control, dynamically balancing a robot's centre of gravity so the robot can safely step, squat, and bend to navigate cluttered, real-world spaces instead of only reaching from a static position.
Yes. It unlocks human-level dexterity and supports complex, 22-degree-of-freedom hands and grippers.
Access is currently offered via a waitlist for early access, announced on the Gemini Robotics 2.5 page by Google DeepMind.
Yes. The model uses reasoning to adapt its behavior to different situations and to complete unexpected, changeable, and complex tasks.
The page does not state that the model is open source; access is provided through DeepMind's early-access waitlist.