Humanoid robots running Google DeepMind’s new Gemini Robotics 2.5 model sorted metal brackets into trays on working Mercedes-Benz production lines with a 96% success rate, the lab said on 10 November. The proof of concept used Apptronik’s Apollo 2 robots at plants in Kecskemét, Hungary, and Sindelfingen, Germany, on a job called kitting: picking parts from bulk bins and laying them out in the order each car needs.
Gemini Robotics 2.5 is a vision-language-action model, meaning it turns what a robot sees and is told into arm and hand movements. In DeepMind’s own tests across more than 200 tasks, it succeeded 30% to 100% more often than its predecessor and finished tasks 20% to 30% faster. DeepMind also reports safer behaviour around obstacles and nearby people, and has released an open benchmark of disguised unsafe requests for robot models.
Kitting is repetitive factory work that has been hard to automate because parts arrive jumbled. A humanoid doing it on a real line, not a lab bench, is a step towards robots that fit into existing factories. But this was a proof of concept: DeepMind has not said how many attempts the 96% covers or who measured it, and the model is going only to its robotics partners for now.
Source: Google DeepMind