Google's robot hit 96% in factory sorting, with the last 4% still unresolved

Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model for several robot types. The company evaluated it on more than 200 tasks and used it with the Apollo 2 humanoid in a pilot on active Mercedes-Benz production lines in Hungary and Germany.

Apollo 2 sorted automotive parts with 96% success

The robot picked metal parts from bulk bins and arranged them into kits in production order. Gemini Robotics ER tracked the workflow while Gemini Robotics 2.5 controlled the hands and arms. Google reports a 96% success rate.

Across 200 internal tasks, the new model delivered success rates 30% to 100% above previous baselines depending on the task and completed them 20% to 30% faster. Tasks included using tongs, tying a ribbon, and pipetting liquid.

Value is shifting from movement to learning speed

The important change is how quickly a robot can learn another activity and transfer it across bodies. The model works with Apollo 2's five-finger hand with 22 degrees of freedom and Franka Duo's two-finger grippers. Gemini 4 Argon also turns camera scenes into digital twins and thousands of simulated practice runs.

That could reduce expensive tuning for each work cell. Buyers still need to measure integration time, human interventions, and stability after a change in parts or lighting, rather than counting demonstrations alone.

Four failures in a hundred remain costly on a line

The 96% figure comes from a proof of concept and does not reveal the pilot duration or the cost of failures. On an active production line, one misplaced bracket can stop downstream work. DeepMind itself says closing the last 4% will be a steep road.

Safety results also come from the vendor. Inference guidance increased safe completion by 33% around obstacles and 43% near people. ASIMOV-Jailbreak opens testing of disguised unsafe requests, but it does not yet create a shared industry standard.

Hours without human intervention will decide deployment

The useful signals will be sustained partner operation, stop counts, and performance after environmental changes. Gemini Robotics 2.5 is rolling out only to robotics partners, not as a generally available product. If pilots preserve performance without frequent supervision and retraining, robotics moves closer to an ordinary capital purchase rather than a research project.

Lilith's verdict

On video, 96% looks almost finished. On a production line, four misplaced parts in a hundred are four small levers that can stop an entire shift.