A humanoid robot worked a real Mercedes line, and the 4 percent it missed is the whole story

Picture a Mercedes factory line in Hungary. A humanoid robot stands at a bench, reaches into a bin full of metal brackets, picks one up, and lays it in a tray in the exact order the next car on the line needs. Then it does it again. And again. Google DeepMind says it got 96 out of 100 right.

The robot is Apptronik's Apollo 2, and the brain is Gemini Robotics 2.5, announced on Tuesday, November 10. It is a vision-language-action model, which means it takes what its cameras see and plain language instructions and turns them into arm and hand movements. Two models split the job. One, called Gemini Robotics ER, keeps track of the workflow, which part goes in which slot next. The other, Gemini Robotics 2.5, moves the hands and arms. The hand has five fingers and 22 degrees of freedom, which is why it can pick up a metal bracket instead of just pushing boxes around.

DeepMind tested the model on more than 200 tasks before the factory run. Serving food with tongs. Tying a bow on a gift box. Pipetting liquid. Folding clothes. It says the new model succeeds 30 to 100 percent more often than the last one, depending on the task, and finishes 20 to 30 percent faster. Those are DeepMind's own numbers, not an independent test, so take them as the company's claim. DeepMind also used its Gemini 4 Argon model to build digital twins of real scenes and run thousands of practice rounds in simulation before the robot touched a real line.

The factory test ran on live lines in Kecskemet, Hungary and Sindelfingen, Germany. That is the part that matters to me. A lab demo is one thing. A robot working while real cars move down a real line is another. Kitting is boring, repetitive work, and it is physically hard on people. But one wrong part in a tray can hold up the car behind it, so boring does not mean easy.

"Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off."

Jeff Cardenas, Apptronik co-founder and CEO

Here is the other side of 96 percent. On a factory line, 4 misses in 100 is still a problem. DeepMind itself calls this a proof of concept, not a deployment, and the announcement gives no trial count and no word on how often a person stepped in. Mercedes has not said the robots are staying on the lines.

DeepMind also released an open safety test with the model, called ASIMOV-Jailbreak. It hides unsafe requests inside ordinary-sounding prompts across dozens of scenes, to see if a robot can be talked into doing something dangerous. DeepMind reports its Gemini 4 Argon refused 99.5 percent of those tricks, against 82 percent for the earlier model. A robot that works next to people needs that test as much as it needs good hands.

For now, the model goes to DeepMind's robotics partners only. No public API, no price, no date for everyone else. So the question is not whether a robot can sort car parts. It can, 96 times in 100. The question is what the last 4 look like, and who fixes them when they happen.

Cover image: AI-generated