Google DeepMind Launches Gemini Robotics 2: Full-Body Humanoid Control From Fingertips to Toes
Google DeepMind released Gemini Robotics 2 on July 30 — the previous generation could only control a robot's upper body; this one runs the whole thing from head to toe, and can adapt to an entirely new robot in a matter of hours.
Three Key Takeaways
From half-body to whole-body: "whole-body intelligence" for humanoids. Gemini Robotics 2 is a family of three models: a vision-language-action model (VLA) for full-body motion control, an embodied reasoning model for multi-step planning and multi-robot coordination, and an on-device lightweight variant that adapts to new robot hardware within hours. The previous generation was limited to arms and hands; this one brings in leg movement, balance, and walking.
Real-world numbers: usable, but still clumsy. DeepMind demonstrated the system on Apptronik's Apollo 2 humanoid: loading a tape into a boombox, screwing in a lightbulb, tying up a garbage bag, and placing a watering can on a specified shelf. Measured success rates — 68.4% picking objects from a table, 45.7% from the floor, 76.3% from a shelf. In other words: big items from high surfaces are fine; bending down to grab small things still fails about half the time.
Hours to adapt to a new body — that is the real headline. Programming a control system for a new robot used to take months of parameter tuning and data collection. Gemini Robotics 2's on-device model claims to complete adaptation within hours — hand it an unfamiliar robot, and it figures out how to move. If this holds up in production environments, robot companies would no longer need to train a control model from scratch for every new form factor.
WangDou's Take
A 68% table-pickup rate and 46% floor-pickup rate — translated into human terms, that means if you ask it to grab a mug from the kitchen, there is a one-in-three chance it drops it; ask it to pick up a sock from the floor, and it fails half the time. On a factory assembly line, that is a non-starter. But in the 2026 embodied-AI landscape, it is genuinely leading-edge. One thing Google got right: they did not hype the success rate — they just put the numbers on the table. The more interesting story is the "adapt to a new body in hours" capability. If that actually works, Google is selling a universal robot brain, not a controller for one specific machine. Every company building humanoid hardware will want to plug into that API — and Google just happens to need a new ecosystem foothold beyond software. The playbook is Android all over again, except this time the phones have legs.
Source: Bloomberg, Engadget, Google DeepMind Blog