A Brain That Can Swap Bodies: Google DeepMind Aims To Be The ‘Android’ Of Robotics

Google DeepMind unveiled Gemini Robotics 2 on Wednesday, a suite of three AI models that gives humanoid and other robots whole-body control, sharper manipulation, and the ability to work together. The system can run locally on the robot itself and adapt to an entirely new machine body with a few hours of training data. DeepMind is pitching it as the “intelligence layer” for robots – the foundational operating system that every hardware maker builds on.

Rather than a single monolithic system, the company released three models across three access tiers:

• Gemini Robotics 2: A vision-language-action model that turns what a robot sees and is told into motor commands. It can drive a full humanoid from feet to fingertips, as well as standard two-armed systems.
• Gemini Robotics ER 2: An embodied-reasoning model that plans multi-step jobs over several minutes, recovers from failures, and lets multiple robots split a workflow between them.
• Gemini Robotics On-Device 2: A lightweight version that runs directly on the hardware, for industrial settings where cloud connectivity is unreliable or unavailable.

DeepMind demonstrated a single model checkpoint controlling three physically distinct robots: an Apptronik Apollo 2 fitted with two different sets of hands, and a separate two-armed rig with a gripper. One model ran hardware it was never designed for.

Unlike its 2025 system, which handled only a robot’s upper body, this version controls movement from the torso down through the legs, letting a humanoid walk, crouch, bend, and reach at once. In one demonstration, an Apollo 2 walked to a watering can and placed it on a lower shelf.

Success has been hit-or-miss, wit robots hitting roughly 68% accuracy picking objects from a table and 76% from a shelf, but only 46% from the floor. Multi-finger dexterity is still an unsolved problem, and simple grippers continue to outperform elaborate robotic hands. Movement needs to be faster and more reliable. This is a research release with a waitlist; nobody is putting it to work on a warehouse floor next quarter.

DeepMind says Gemini ER 2 its safest robotics model yet, based on benchmarks for adhering to safety constraints and avoiding humans, and it included an evaluation of whether a robot will refuse an unsafe instruction and defer to a human for help. The vision-language-action and on-device models are gated to more than 100 trusted testers and a handful of early partners; the reasoning model is in a more open preview.

The Bigger Bet

Alphabet, Nvidia, and a wave of Chinese tech firms are all racing to own the “physical AI” stack rather than any single robot – the same bet that made mobile operating systems so valuable. If humanoid machines ever become as common as laptops, whoever supplies the universal brain that transfers cleanly from one body to the next will own a toll booth on the entire industry.

As for DeepMind – so far it hasn’t kicked any Chinese kids in the head, so that’s a start. 

A Unitree G1 humanoid robot just roundhouse kicked a child in China’s Xinjiang region, sparking urgent conversations about public safety.

The robot, which was executing a pre-programmed, 360-degree spinning roundhouse kick, delivered enough force to double the child over.

While…

— Massimo (@Rainmaker1973) June 9, 2026