Google DeepMind’s new AI model can control a robot’s entire body

Google DeepMind has extended its Gemini Robotics model from upper-body manipulation to whole-body control, a quiet but consequential expansion. The previous version could move arms and grippers; Gemini Robotics 2 now coordinates feet, legs, torso, and fingertips in a single policy. The demonstration partner, Apptronik's Apollo 2 humanoid, was shown bending to retrieve a watering can and selecting specific items from a shelf. These are not party tricks. Whole-body coordination is the prerequisite for any humanoid robot to operate in spaces designed for humans, where reaching low, stepping over obstacles, and stabilizing while carrying loads are routine requirements.

The strategic logic is straightforward. Foundation models won the text and image races by absorbing broad, general capability from internet-scale data. The next frontier is physical capability, and the limiting factor is not compute but embodied data. Humanoid robots are scarce, expensive, and slow to operate, which makes high-quality motion data the scarcest resource in AI. By positioning Gemini as the shared "brain" that multiple hardware vendors can adopt, DeepMind is attempting to replicate the platform playbook that made Android dominant in mobile. Apptronik supplies the body; DeepMind supplies the mind; the value capture migrates upward to the model layer.

The competitive picture is tightening. Figure, Tesla Optimus, 1X, Agility, and a growing cohort of Chinese entrants are all racing to prove that their hardware can be the reference platform for general-purpose humanoid labor. A model that ports cleanly across bodies, as Gemini Robotics 2 is designed to do, weakens any single hardware maker's lock-in. It also raises the bar for everyone: the contest is no longer about who can build the slickest demo, but whose robots can run the most capable general policy at the lowest inference cost.

For the labor market, the implications remain distant but directional. Whole-body control does not put humanoid robots on factory floors next quarter. It does, however, close one of the two technical gaps, dexterity and locomotion, that have kept general-purpose humanoids confined to promotional videos. The remaining bottleneck is reliability over long horizons in unstructured environments. When that falls, the substitution question stops being theoretical and becomes a procurement decision. Until then, this is a capability milestone worth marking, not a deployment milestone worth fearing.