Google DeepMind Unveils Gemini Robotics 2, Full-Body Control for Humanoid Robots

On July 30, 2026, Google DeepMind announced the launch of Gemini Robotics 2, a new generation of AI models designed to control humanoid robots. Unlike previous versions, which focused mainly on hand dexterity, the new system controls the robot's entire body, from feet to fingers, allowing machines to walk, crouch and perform precise movements within a single, unified decision-making chain.

Three models instead of one

Gemini Robotics 2 is essentially a set of three interconnected systems. Gemini Robotics 2 itself is a vision-language-action (VLA) model that turns camera images and natural-language instructions into concrete motor movements. Gemini Robotics ER 2 serves as the higher-level planning layer, breaking a complex task into stages, tracking progress from a continuous video stream, and calling external tools, such as Google Search, when the robot needs additional information. The third component, Gemini Robotics On-Device 2, runs locally on the device itself, without a connection to cloud servers, which matters in locations with poor connectivity or for tasks requiring an immediate response.

ER 2's authors, Steven Hansen and Peng Xu of Google, note in the technical description that the system acts as a high-level brain for the robot, communicating with people, understanding the physical context of its surroundings, and coordinating the work of multiple machines operating in the same space. That last capability has practical significance on factory floors, where several robots work side by side and must avoid collisions and duplicated tasks.

Dexterity still far from human level

The figures published by DeepMind show the system performs well, but not flawlessly. Object-picking success reaches 68.4 percent when working on a tabletop and 76.3 percent when reaching for a shelf, while tasks requiring precise insertion of parts into sockets rise to 89.6 percent. That is still a noticeable margin of error for tasks a person performs almost automatically, such as tying a knot, zipping up a jacket, or screwing in a lightbulb.

Bloomberg, which first reported the launch, pointed to exactly this gap between the announcements and the robots' actual manual dexterity, reflected in the headline of its piece on Gemini AI for robots struggling with precise movements. Competitors, including companies developing their own physical AI models, face the same problem: grasping and manipulating small objects remains one of the hardest tasks in robotics, despite progress in these same systems' language reasoning.

Hardware partners

Gemini Robotics 2 does not operate in a vacuum. DeepMind has spent months building a network of hardware makers testing successive versions of the model on their own machines. Apptronik is integrating the system into its Apollo 2 humanoid, Boston Dynamics is using it in Atlas and the quadrupedal Spot, and Agile Robots and Franka, maker of the F3 Duo platform, are testing it in industrial tasks. Earlier announcements mentioned plans to test Gemini-powered robots at Hyundai factories, where the machines are meant to attempt assembly tasks.

Safety and availability

Alongside the new models, DeepMind also published a Safety Technical Report and the ASIMOV-Agentic benchmark, designed to assess the safety of robotic agents operating in close contact with people. In tests of adherence to safety instructions and maintaining a safe distance from humans, ER 2 consistently outperforms the previous ER 1.6 version. The models are already available through the Gemini API and Google AI Studio, and for select customers also in private preview on the Gemini Enterprise Agent platform.

For the Polish market, the launch currently has mostly indirect significance. None of the hardware partners mentioned has announced deployments in Poland yet, and humanoid robotics remains in the industrial testing phase even in the United States and South Korea. Still, the pace at which Google DeepMind is updating successive generations of the model, from the first Gemini Robotics in March 2025, through the On-Device version in June of that same year, to today's launch, shows that the company treats physical AI as one of its priorities alongside language models themselves.