Google DeepMind did not just show a robot arm moving objects on a table. On July 30, it introduced Gemini Robotics 2 with demos of Apptronik's Apollo 2 humanoid bending, walking to a shelf, picking up a watering can and placing it in a target spot.
Three models form the stack
The release combines three parts: Gemini Robotics 2 turns vision and language into motor control; Gemini Robotics ER 2 plans and tracks multi-step tasks; Gemini Robotics On-Device 2 runs locally on robot hardware. Google says ER 2 can handle tasks lasting several minutes and involving hundreds of decisions.
Carolina Parada told WIRED: “It's another milestone in our path towards really getting towards what we call like physical AGI, which means we get a robot to do anything that a human can.” The measurable pieces are narrower: whole-body control, two-hand manipulation, multi-robot collaboration, local execution and safety stops.
The hand numbers are uneven
Google's charts show Apollo 2 with Inspire hands reaching 68.4% success picking from a table, 45.7% from the floor and 76.3% from a shelf. With the five-finger SharpaWave hand, unscrewing a bulb reached 92%, while screwing a bulb was 36%, tying a trash bag 44%, dustpan work 32% and sealing a ziplock 40%.
Those figures say the same thing as the videos, with less gloss: the system has moved beyond tabletop manipulation, but finger-level dexterity remains fragile.
On-device adaptation is the practical test
Google says On-Device 2 can adapt to new bi-arm embodiments in a few hours, typically with fewer than 200 examples. Its model card reports SO101 success rising from 6.7% to 53.3% with GRODv2, while Dexmate rose from 24.4% to 75.6%. These are model-card evaluation figures, not a universal claim for every robot.
The next checks are whether third-party hardware reproduces those adaptation numbers, whether multi-finger tasks move above the 32%-44% range, and whether safety stops work reliably around people in messy environments.
Sources: Google DeepMind release, Gemini Robotics ER 2 and On-Device 2 model cards, Google AI for Developers, CocoLoop, WIRED, The Verge, Sina Finance; verified model scope, three-model split, Apollo 2/SharpaWave/Inspire hands evaluation scope, 128k/64K settings, fewer-than-200-example adaptation and availability channels.