Google's Gemini Robotics 2 can now drive a humanoid from feet to fingertips.

Google DeepMind released Gemini Robotics 2 on July 30, and the headline capability is one the whole industry has been circling: for the first time, a single Gemini model can control an entire humanoid, from feet to fingertips, not just the upper body. Where DeepMind's earlier robotics models drove a humanoid's arms for tabletop tasks, the new model handles whole-body motion — walking, crouching, stretching and balancing to work in the cluttered, human-scaled spaces the world is actually built for. In the demonstration DeepMind leads with, it tells Apptronik's Apollo 2 to 'put the watering can into the green bin in the bottom shelf,' and the robot processes the instruction, walks to a table, picks up the can, steps over to the shelves and places it on the bottom one. The release is really a family of three: the flagship Gemini Robotics 2 vision-language-action model, an embodied-reasoning model called Gemini Robotics ER 2 that acts as a high-level 'brain,' and an on-device version tuned to run locally on the robot with no network connection.

Two claims underneath the demo matter more to a buyer than the watering can. The first is portability: DeepMind says the same model checkpoint drove three different bodies — Apollo 2 with SharpaWave hands, the same Apollo with Inspire hands, and a Franka Duo with a two-finger gripper — and that the on-device model can adapt to an entirely new robot with drastically different shape, sensors and degrees of freedom in a few hours, typically with fewer than 200 examples. The second is orchestration: the ER 2 reasoning model can now run task sequences lasting several minutes and hundreds of decisions, watch a video feed to track its own progress and know when a step is actually finished, self-correct when one fails, and coordinate more than one machine — a wheeled rover and a humanoid, say — to split a job. DeepMind also leaned hard on safety, introducing a benchmark it calls ASIMOV-Agentic and reporting that ER 2 will halt a humanoid when a person steps close and resume only once the area is clear. The reasoning model is available now through the Gemini API and Google AI Studio; the action and on-device models are going to early-access partners.

The reason to treat this as a milestone rather than a finished product is that DeepMind published the failure rates next to the wins, which is rarer than it should be. On general whole-body manipulation with Apollo, the model picked an object off a table about 68% of the time, off a shelf 76%, and off the floor only 46%. Multi-finger dexterity is shakier still: it unscrewed a light bulb 92% of the time but screwed one in just 36%, managed a dustpan 32% of the time and a ziplock bag 40%. DeepMind says plainly that 'multi-finger dexterous manipulation remains challenging' and that its robots 'have more to advance in movement speed.' A two-finger gripper on the Franka arm did far better at industrial-style work — precise insertion around 90%, tool kitting near 79% — which quietly underlines where reliable robot manipulation actually lives today. This is a research release, shipped with model cards and a safety report, not a product you deploy.

For anyone weighing humanoids, the significance is about who owns the hard part. The industry splits between makers building their own end-to-end AI — Figure with Helix, Tesla, 1X — and those willing to run a foundation model from a hyperscaler, and Gemini Robotics 2 makes the second path materially stronger; the fact that it can be re-pointed at a new body in hours also lowers the moat around any single robot's software. It is telling that the bodies in the demos are Western: Apollo is the marquee platform here, and DeepMind thanks Apptronik, Boston Dynamics and Agile Robots as partners — the same kind of names now sheltered by Washington's fresh import ban on Chinese robots. The practical caution for a procurement team is unchanged by a good demo: ask any vendor touting a 'Gemini-powered' robot whether the model runs on-device or in the cloud, what its autonomous success rate is on your specific tasks with no human in the loop, and how fast it moves — because a 46%-off-the-floor pick rate is a pilot metric, not a production one.