Google DeepMind has released Gemini Robotics 2, and the short version is that its robots can now use their whole bodies. Announced on July 30, the model is the first from the company to control a humanoid's legs, torso, arms and multi-finger hands under a single learned policy. The 2025 version handled the upper body only.
That difference is bigger than it sounds. Whole-body control means a machine can twist, lean, crouch and reach at the same time, coordinating its balance with whatever its hands are doing. As The Robot Report describes it, the AI plans and adjusts movement from the torso down to the legs rather than treating the arms as a bolt-on.
DeepMind is shipping three models rather than one, according to MarkTechPost: a vision-language-action model that drives the robot, an embodied reasoning model that handles planning, and a smaller on-device version for tasks that cannot wait on a network round trip. One checkpoint reportedly runs Apptronik's Apollo 2 humanoid with two different hands as well as a Franka research arm, which suggests the skills carry across robot bodies instead of being tuned to one machine.
The demos lean on fiddly, real-world jobs rather than staged spectacle. In testing, the system unscrewed a light bulb with a 92% success rate, a task that needs steady grip, careful force and constant balance adjustments. Dexterity at that level has been one of the hardest problems in robotics for years.
The third piece is multi-robot collaboration, where two or more machines coordinate on a shared task. That is the part most relevant to warehouses, factories and any setting where a single robot is not enough. It also raises the familiar questions about the jobs those machines are meant to do, which is a debate worth having honestly rather than dismissing.
None of this means household robots are about to arrive. Lab success rates are not the same as reliability in a messy kitchen, and humanoids remain expensive and power-hungry. What Gemini Robotics 2 shows is that the intelligence layer, the software that decides how a body should move, is improving faster than the hardware around it. Google has been busy on the software side more broadly, having recently shipped a trio of cheaper Gemini Flash models and teased its next flagship.
Commentarii · 0