Google DeepMind Unveils Gemini Robotics 2 for Whole-Body Humanoid Control
The new system replaces fragmented software stacks with a single policy to coordinate humanoid movement and multi-robot collaboration.
Google DeepMind released Gemini Robotics 2 on July 30, 2026, introducing a system capable of whole-body humanoid control. The release marks a shift from limited tabletop manipulation to a unified framework that coordinates a robot's entire physical form.
The new system enables "whole-body control," allowing a humanoid to coordinate walking, crouching, and reaching under one continuous policy. This replaces the previous industry standard of using separate, stitched-together software modules for locomotion and manipulation. The architecture is split into three specialized models: Gemini Robotics 2 (VLA) for motor control, Gemini Robotics-ER 2 for embodied reasoning and planning, and Gemini Robotics On-Device 2 for local, offline execution.
By unifying these functions, Google addresses a major bottleneck where separate software stacks often caused instability during complex movements, such as reaching while crouching. This transition allows the AI to manage the robot's balance and limb coordination simultaneously, moving beyond the constraints of previous iterations that were primarily limited to tabletop tasks like picking up a cup.
DeepMind demonstrated the VLA model on the Franka Duo bi-arm platform and Apptronik’s Apollo 2 humanoid, including the 22-DOF SharpaWave hand. While not entirely hardware-agnostic, the system utilizes a "motion-transfer" approach to adapt to new robot designs using approximately 200 examples or a few hours of data. Additionally, the release introduces multi-robot collaboration, demonstrated through real-time task handoffs between an Apollo 2 humanoid and a Franka F3 Duo platform.
This unification of control and collaboration suggests a path toward general-purpose robotic agents for industrial use. By reducing the need for exhaustive per-robot retraining, Google is moving closer to a universal "robot brain" deployable across various humanoid forms. Currently, Gemini Robotics-ER 2 is available in public preview via the Gemini API and Google AI Studio, while the VLA and On-Device models remain restricted to early-access partners and the Trusted Tester Program.