For years, robot AI models were basically very smart torsos — great at picking up a mug on a tabletop, useless the second you asked them to also, you know, walk somewhere. Google DeepMind just fixed that, and the fix comes with fingers that can tie knots.
One Brain, the Whole Body
On July 30, Google DeepMind released Gemini Robotics 2, a family of three models that for the first time controls a humanoid's legs, torso, arms, and multi-fingered hands under a single learned policy. Previous versions of Gemini Robotics were strictly upper-body specialists, fine for tabletop chores but stuck if the object they needed was on the floor or a shelf.
The lineup splits the labor: the core VLA model turns vision and language into motor commands, Gemini Robotics ER 2 (built on Gemini 3.5 Flash) handles high-level planning, and an On-Device 2 model lets a robot adapt to a new body in hours rather than months. DeepMind showed it off on Apptronik's Apollo 2 humanoid — complete with a 22-degree-of-freedom hand — tying knots and sealing a ziplock bag, plus cameos from a Franka Duo arm and Boston Dynamics' Spot.
The Numbers Aren't Perfect, and That's the Point
Success rates ranged from 45.7% picking objects off the floor to 76.3% off a shelf, with dexterity tasks landing anywhere from 32% to 92%. That's not a highlight reel of flawless robots — it's a snapshot of a genuinely hard problem getting incrementally, measurably less hard.
The sleeper feature here is multi-robot collaboration: different robot types passing off subtasks to each other through shared semantic understanding, which sounds a lot like the early scaffolding for robots that coordinate instead of just coexist. Access is staggered — ER 2 is in public preview, the VLA model is early-access-partners-only, and On-Device 2 is locked to trusted testers, so most of us are watching from the bleachers for now.
Robots walking and grabbing at the same time shouldn't feel like a milestone in 2026, but here we are, applauding a machine for successfully sealing a sandwich bag. Progress is weird like that.
Source: Google DeepMind