Google DeepMind Unlocks Full‑Body Control with Gemini Robotics 2

Google DeepMind’s Gemini Robotics 2 now moves from feet to fingertips under one learned policy, a first for the family that previously focused on upper‑body control alone. The announcement, posted on gcn.com, highlights three distinct models: a vision‑language‑action (VLA) core, an embodied reasoning version (ER 2) for multi‑step planning, and an on‑device edge variant that runs locally on robot hardware. This shift marks a leap toward what DeepMind calls physical AI, where robots can coordinate complex movements across the entire body.

The system has been demonstrated on Apptronik’s Apollo 2, tested with Sharpa and Inspire five‑fingered hands, and shown driving a two‑armed Franka Duo equipped with a Robotiq gripper. Additional trials involved Dexmate, Trossen and SO101 research platforms, covering a wide range of dexterous manipulation tasks. Apptronik’s Robot Park in Austin, a 90,000‑square‑foot data collection hub opened in June 2026, continuously feeds high‑quality training data back to DeepMind, refining the models in real‑world conditions.

Gemini Robotics ER 2 serves as the high‑level brain, a vision‑language model that communicates with humans, understands the physical world and plans tasks lasting several minutes. It also measures a robot’s ability to refuse unsafe actions, adding a safety layer to longer sequences that involve hundreds of decisions. By integrating reasoning with whole‑body control, the platform aims to make humanoid robots more reliable in real‑world environments.

The release underscores a broader trend: physical AI is moving from lab experiments to practical deployment. Companies like Apptronik are building dedicated facilities to generate the massive datasets needed for such models, while edge versions promise lower latency for on‑site operations. Analysts note that the ability to coordinate legs, torso, arms and fingers under a single policy could reduce the engineering overhead of designing separate controllers for each body part.

Historical attempts at humanoid control often required separate policies for upper and lower bodies, leading to latency and synchronization issues. Gemini Robotics 2’s unified approach could set a new benchmark, similar to how earlier vision‑language models consolidated perception and language understanding. The move also signals DeepMind’s intent to compete in the growing robotics market, where physical AI is seen as the next frontier after language and multimodal generation.

Looking ahead, the question is whether this unified control will translate into reliable, safe robots in everyday settings. As more companies adopt similar data‑driven pipelines, the industry may see a rapid acceleration of humanoid capabilities, but also heightened scrutiny over safety and ethical deployment.

FAQ

What makes Gemini Robotics 2 different from earlier versions?
Gemini Robotics 2 introduces a single learned policy that controls legs, torso, arms and fingers together, whereas previous models focused only on upper‑body manipulation.

Which robots have been used to test the new system?
The platform has been demonstrated on Apptronik’s Apollo 2, Sharpa and Inspire five‑fingered hands, a two‑armed Franka Duo with a Robotiq gripper, and research bots from Dexmate, Trossen and SO101.

How does the ER 2 reasoning layer improve robot performance?
ER 2 acts as a high‑level brain, planning multi‑step tasks lasting several minutes, communicating with humans and evaluating safety by refusing unsafe actions.

What is the significance of Apptronik’s Robot Park for this development?
Robot Park provides a continuous stream of high‑quality training data, enabling DeepMind to refine Gemini Robotics 2 in real‑world conditions and accelerate improvements in whole‑body control.