DeepMind ships Gemini Robotics 2 for humanoid whole-body control

A post today from Google DeepMind puts a specific number on something the humanoid industry has been asking about for a while: how fast a general-purpose model can be retargeted onto a new robot body. The claim, in DeepMind's own announcement written by Carolina Parada, is that Gemini Robotics On-Device 2 adapts to new embodiments with typically fewer than 200 examples and several hours of training.

The rest of the release is a three-model suite. Gemini Robotics 2 is the vision-language-action model for whole-body control, demoed on the Apptronik Apollo 2 humanoid walking, crouching, and placing objects on shelves in cluttered environments. Gemini Robotics ER 2 is the embodied reasoning model, planning multi-step tasks that run for several minutes and coordinating multi-robot teams. The on-device variant is the third leg. DeepMind reports numbers including 89.6% accuracy on Franka Duo precision insertion tasks and 78.9% on diverse tool assembly, plus multi-finger work on 22-degree-of-freedom SharpaWave hands including knot-tying and ziplock sealing.

Why this matters if you are not building a robot: the humanoid conversation for the last year has been dominated by hardware, with Apptronik, Boston Dynamics, Agile Robots and Franka Robotics all showing off platforms. If the same foundation stack can drive all four, and if new-body adaptation really is a few hundred demonstrations rather than a research project, the moat shifts from mechanical engineering to model access. Google starts to look like the operating system layer for a lot of humanoid work, and the robot vendors look like distribution.

The honest caveat is that these are DeepMind's own numbers, on DeepMind's own benchmark stack, including a new safety evaluation called ASIMOV-Agentic that measures whether the system refuses unsafe commands and asks for human help when uncertain. Success rates in the high eighties inside a curated demo are not the same as reliability on a factory floor, and the post does not disclose latency, compute footprint, pricing, or when the VLA and on-device pieces leave early access. What the reporting also does not define is what an 'example' actually is: a teleop episode, a trajectory, or an annotated demonstration.

What is worth watching from here is which humanoid maker declines to sign up. The vendors who ship on top of Gemini Robotics 2 get a working brain for free. The ones who insist on their own foundation model will need to justify that call to buyers who now have a reference point for how fast retargeting should be.