Google DeepMind Unveils Gemini Robotics 2 With Whole-Body Intelligence for Robots

In July 2026, Google DeepMind unveiled Gemini Robotics 2, an updated robot control system with whole-body intelligence. For the first time, the model gives a robot synchronized control of its entire body (legs, torso and arms) within a single motion-planning loop.

What is whole-body intelligence

Whole-body intelligence means the robot coordinates its movements as a single kinematic system rather than controlling each link in sequence. The first version of Gemini Robotics, announced by Google DeepMind at CES in January 2025, focused on dexterous manipulation: precise grasping, handling fragile objects and tasks that demand fine hand motor skills. Gemini Robotics 2 extends that principle to the entire locomotor apparatus. The goal is a robot that moves naturally instead of executing pre-programmed sequences.

The system plans the movements of the ankles, hips, spine and arms simultaneously, the way a person does when carrying a heavy box across uneven floor or bending down to pick something up off the ground. According to the Google DeepMind blog, the model accounts for the physical constraints of the environment in advance and redistributes load across the joints without waiting for a balance error.

Traditional control systems optimize each degree of freedom separately. Gemini Robotics 2 instead treats the body as a single system with a shared balance of forces and momentum, which allows it to perform movements without explicit programming of each joint.

How is the system trained?

Gemini Robotics 2 is built on top of the multimodal Gemini architecture: the robot receives visual and text input, reasons about the assigned task and translates that reasoning into motor commands in real time. For training, DeepMind combined teleoperation, in which a specialist controls the robot remotely and leaves a trace of correct decisions, with synthetic motion data. This substantially reduced the number of real-world demonstrations needed to learn a new task.

Key characteristics of the system, according to Google DeepMind's announcement:

• Architectural foundation: the multimodal Gemini model
• A single planning loop covers legs, torso and arms
• Support for humanoid platforms and conventional manipulators
• Training through teleoperation and synthetic motion data
• Target scenarios: unstructured and variable environments

According to the official

Google DeepMind blog, Gemini Robotics 2 delivers "whole-body intelligence," allowing a robot to plan and execute movements the way a living organism does, using all the motor resources available to it.

Where will the new system be used?

Whole-body intelligence opens up a class of tasks that narrow manipulators cannot handle. Warehouse logistics on uneven surfaces, cleaning in homes with changing layouts, assisting people with disabilities and servicing factory floors all share the same key bottleneck: coordinating the whole body. According to DeepMind, the system was tested in conditions as close to real life as possible: on slippery surfaces, in rooms with obstacles and on two-handed tasks involving asymmetric objects.

Google DeepMind is working with robotics partners to test the system on physical humanoid platforms. The pace of development has picked up: about a year and a half passed between the first and second versions of Gemini Robotics, while earlier research systems were updated much less often.

What this means

Gemini Robotics 2 moves physical AI from lab demonstrations toward functional systems for the real world. In whole-body robotics, Google DeepMind competes with Figure AI, Boston Dynamics and Tesla Optimus. But the integration with Gemini gives its robots a level of language and visual reasoning that purely motor-focused systems do not yet have. If the system proves reliable outside the lab, it could set a new standard for the industry.