Gemini Robotics 2: Whole-Body Control, and the Failure Rates

Everyone Quoted the 92 Percent. The Same Table Says 36 and 32.

Gemini Robotics 2 is the first model to drive a humanoid's legs, torso, arms and fingers under a single policy. DeepMind published the failure rates alongside it.

Google DeepMind released Gemini Robotics 2 on Thursday: three models that, for the first time, control a full humanoid — legs, torso, arms and multi-fingered hands — under one policy rather than a stitched-together locomotion and manipulation stack. It also published a scorecard that contradicts its own headlines.

What the table says

On Apptronik's Apollo with Inspire hands: pick from a shelf 76.3%, from a table 68.4%, from the floor 45.7%. With Sharpa hands: unscrew a bulb 92%, tie a trash bag 44%, close a ziplock 40%, screw a bulb in 36%, sweep a dustpan 32%. On the two-arm Franka Duo gripper setup the numbers are much higher — precise insertion 89.6%, tool kitting 78.9%.

The number that travelled was the best cell in the table

Nearly every write-up led with the light bulb at 92%. The reverse operation — putting it back — is 36%. Reported as a dexterity breakthrough, the scorecard describes a system that fails roughly two thirds of fine-motor tasks. DeepMind publishing those figures at all is unusual and worth crediting; quoting only the flattering one is not.

"Available" is doing heavy lifting

Three models shipped, and they are not equally shipped. ER 2, the embodied-reasoning model, is in public preview through the Gemini API and AI Studio — and it does not move a robot. The VLA model that generates motor commands is early-access partners only. On-Device 2 is restricted to Trusted Testers. Nobody outside Google's partner list can run the thing that produced these results.

Autonomous, but narrowly

Google states the footage is real-time and fully autonomous, a pointed contrast with the teleoperated humanoid demos that dominate this category — and that claim holds. But the robots were trained per demonstrated task, through human teleoperation, video and simulation, and DeepMind concedes the system is "limited in generalizing to out-of-distribution tasks". This is autonomous execution of specifically trained tasks, not general capability. One other correction worth making: the Boston Dynamics tie-up being attached to this news was announced at CES on 5 January 2026. The demo hardware here is Apptronik's.

Why It Matters

Whole-body control under a single policy is the real technical gate for humanoids doing useful work — an upper-body model cannot crouch to pick something off a floor. The published numbers show that gate being opened and how far past it the field still has to go.