A humanoid that finally uses its legs
On 30 July Google DeepMind published Gemini Robotics 2, and the demonstration that carries the release shows an Apptronik Apollo 2 walking, crouching, stretching and clearing a cluttered room. The technical claim underneath is narrower and more interesting than the footage: for the first time the system drives an entire humanoid, legs and torso and arms and multi-fingered hands, through one unified policy rather than a locomotion controller bolted onto a separate manipulation stack.
That distinction is the engineering. A robot that has to decide whether to step closer or reach further is solving one problem, not two, and the previous generation could not make that trade because the halves of it were governed separately. Gemini Robotics 2 can also run tasks that last several minutes across hundreds of decision points while coordinating with a second robot.
Three models, three different doors
The release is not one thing. It is Gemini Robotics 2, a vision-language-action model that controls full humanoids and two-armed robots; Gemini Robotics ER 2, an embodied reasoning model that acts as the planning brain, sequences multi-step work and orchestrates several machines; and Gemini Robotics On-Device 2, a version optimised to run locally with no network connection.
The access tiers do not match the billing. Gemini Robotics ER 2 is available now through Google AI Studio to any developer, with a private preview on the Gemini Enterprise Agent Platform. Both models that actually produce robot actions are restricted to registered early-access partners. The model you can call today reasons about a physical scene and hands back a plan. The models that move a machine are behind a gate.
Read the whole table, not the best row
Google published per-task success rates, which is more disclosure than these releases usually carry, and the numbers deserve to be read in full. On an Apollo fitted with Inspire hands, general whole-body manipulation scored 76.3 percent picking from a shelf, 68.4 percent from a table and 45.7 percent from the floor. On a Franka Duo with grippers the results are stronger: 74.2 percent on general pick and place, 78.9 percent on diverse tool kitting and 89.6 percent on precise insertion.
The multi-finger results are where the honesty lives. Unscrewing a bulb reached 92 percent. The other multi-finger tasks landed between 32 and 44 percent. A sixty-point spread inside one capability is not a rounding difference; it is the signal that dexterity remains a research problem with a few solved corners. Any vendor quoting you the 92 without the 32 is selling the demonstration.
The number that changes the economics is 200
Buried under the humanoid footage is the figure with commercial consequences. DeepMind reports that a new two-armed robot embodiment can be adapted in a few hours using fewer than 200 examples, despite differences in shape, sensors and degrees of freedom, and it names Dexmate, SO101 and Trossen platforms alongside the Apollo and Franka hardware.
Per-robot data collection has been the structural cost of industrial robotics. Every new arm, gripper or chassis has historically meant its own teleoperation campaign, and that expense is what kept capable models locked to the hardware they were trained on. If a few hundred examples genuinely ports a policy across embodiments, the barrier moves from data collection to integration, and integration is something a mid-sized European engineering firm already knows how to sell.
What a European operator can actually do with this in 2026
The usable product this quarter is the reasoning model, and it is not nothing. ER 2 accepts video, understands a physical scene, orchestrates multi-step tasks and coordinates more than one machine, which maps onto inspection routing, warehouse task sequencing and planning layers above robots you already own. It is reachable from Europe through Google AI Studio today, and it is the right place to run a small, contained pilot.
What it does not support is a capital plan. Google frames this as early-stage research with no near-term consumer deployment, every demonstrated action required dedicated training through teleoperation, video and simulation, and Engadget noted the sector's history of humanoid demonstrations that quietly relied on hidden operators. DeepMind also introduced ASIMOV-Agentic, a benchmark for whether a commanded action could produce a harmful outcome, and its robotics lead Carolina Parada put the position plainly: "There's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply." Budget for a pilot of the planner, not for a humanoid on your line.
Read next: 87 Years Fell to an Answer Anyone Can Check | Google's Flagship Model Is Late With No New Date
