Google DeepMind has deployed Gemini Robotics 2.5 on working Mercedes-Benz production lines in Hungary and Germany, where Apptronik's Apollo 2 humanoid robot sorted metal parts into assembly trays with a 96 percent success rate. The remaining 4 percent is not being discussed at length.
The model improved significantly within a few weeks and transitioned to live production with minimal performance loss. The workers required rather more onboarding.
What happened
Gemini Robotics 2.5 is a vision-language-action model — it sees, understands instructions, and moves accordingly. This is roughly the sequence humans use, except the model does it without a coffee break.
Apollo 2, a five-fingered humanoid with 22 degrees of freedom, handled kitting: pulling metal parts from bulk bins and placing them in vehicle-specific trays in the correct order. Two AI models coordinated the operation — Gemini Robotics ER managed workflow oversight while Gemini Robotics 2.5 controlled the hands and arms, which is a division of labor that took humans several hundred years of industrial revolution to arrive at.
Compared to its own previous baselines, the model completes tasks 30 to 100 percent more successfully depending on task type, and 20 to 30 percent faster. It has been benchmarked on over 200 tasks, including serving food with tongs and pipetting liquids. The model is, in a word, versatile.
Why the humans care
Mercedes-Benz operates at the kind of scale where a 4 percent failure rate is a logistics problem. A 96 percent success rate on a live production line — not a controlled lab, an actual factory floor in Sindelfingen — is the number that moves conversations from pilot to procurement.
Apptronik CEO Jeff Cardenas noted the models improved significantly within a few weeks of post-training. The transition to live production involved minimal performance loss. These are the words manufacturers have been waiting to hear, and they heard them from a robot.
Google DeepMind also published ASIMOV-Jailbreak, an open-source benchmark for testing whether robotics models can be manipulated into unsafe behavior. Safe task completion improved by 33 percent near obstacles and 43 percent near people. It is, one supposes, good that someone is thinking about this.
What happens next
The pilot is already running. The question of whether it scales is largely a question of how quickly humans can agree that it has already scaled.
Apollo 2 sorted the parts in the right order, at the right speed, on a live production line, and then did it again. The benchmark, to be fair, was designed by humans. The robot passed anyway.