Google DeepMind Unveils Whole-Body AI Model for Humanoid Robots

On July 30, 2026, Google DeepMind released Gemini Robotics 2, a vision-language-action model that for the first time coordinates the movement of a humanoid robot's entire skeleton, not just its upper body as in previous generations. The company demonstrated how it works in practice on Apptronik's Apollo 2 robot.

From upper body to full skeleton

Earlier versions of the Gemini Robotics models could only control a robot's arms and hands, assuming a stable, stationary base. Gemini Robotics 2 changes that approach: the model translates camera images and natural-language commands into motor commands covering walking, crouching, bending at the waist, and precise hand movements with multiple degrees of freedom.

In demonstration footage, Apptronik's Apollo 2 humanoid picks up objects lying on the floor, scans store shelves for a specific product on voice command, and carries out multi-step tasks lasting several minutes. The model also drives hands with 22 degrees of freedom, enabling actions such as unscrewing a light bulb or tying up a trash bag.

Three models in one system

Gemini Robotics 2 is the main model controlling movement, but DeepMind introduced two complementary tools alongside it. Gemini Robotics ER 2 handles planning for multi-step tasks and coordinating the work of several robots at once, while Gemini Robotics On-Device 2 runs locally on the robot's hardware without a cloud connection, which matters in places without stable internet, such as warehouses or factories.

ER 2 is already available in Google AI Studio and, in private preview, in the Gemini Enterprise Agent Platform. Access to the full VLA model and the On-Device version remains limited to trusted test partners, including Apptronik, Sharpa Robotics, and integrators working with the Nvidia Isaac GR00T platform.

Numbers that reveal the limits

DeepMind published specific performance figures showing the system is still far from human-level reliability. On whole-body tasks with the Apollo 2 robot fitted with Inspire hands, success rates are 68.4 percent for picking up from a table, 76.3 percent from a shelf, and just 45.7 percent when reaching for objects on the floor. In multi-finger manipulation with the SharpaWave hand, unscrewing a light bulb succeeds in 92 percent of attempts, but other tasks, such as screwing in a light bulb, tying up a bag, or arranging tools, fall in the 32-44 percent range.

On the Franka Duo gripper, a classic legless industrial arm, results are markedly higher: 74.2 percent for general picking and placing, 89.6 percent for precise part insertion, and 78.9 percent for sorting tools. The spread shows that the more complex and whole-body the coordination required, the harder the task is for the current generation of the model.

Safety for agentic robots

Alongside the models, DeepMind introduced a new benchmark, ASIMOV-Agentic, meant to measure the safety of autonomously operating robots, including their ability to refuse a dangerous command and detect proximity to humans during operation. This responds to growing concerns that robots controlled by large language models could take risky actions without adequate oversight.

To solve the hardest problems at scale, robots of every shape and size need AI models that give them the ability to think, act, and interact intelligently in order to perform tasks safely - Carolina Parada, head of robotics at Google DeepMind

What this means for the industry

Gemini Robotics 2 fits into the race among major tech companies for dominance in humanoid robotics, where besides Google DeepMind, competitors include Nvidia with its Isaac GR00T platform and robot makers themselves, such as Apptronik and Boston Dynamics. A model designed to run on many different mechanical builds after brief fine-tuning could lower the cost of deploying robots in warehouses and factories, since it removes the need to build control software from scratch for every new piece of hardware.

For now, access remains limited to a narrow group of test partners, so mass deployment at Polish logistics or manufacturing companies will have to wait. Still, the very architecture, in which a single AI model is meant to operate robots from different manufacturers, signals the direction physical-labor automation is heading in the coming years.