Glossary
November 10, 2026

Google's AI Program for Robots Sorts Car Parts on Real Mercedes-Benz Factory Lines in a Trial

Google DeepMind, Google's AI research lab, released a new version of the AI program it uses to control robots. In a trial on working Mercedes-Benz car factory lines, a human-shaped robot running it sorted metal parts correctly 96 percent of the time, the company says. It was a short trial, the numbers come from Google, and the program is going only to a few partner companies for now.

Google DeepMind, Google's AI research lab, today released Gemini Robotics 2.5, the newest version of the AI program it uses to control robots. The program takes in what a robot's cameras see and a plain-language instruction, such as "put the brackets in the tray," and turns them into the movements of the robot's arms and hands.

The company tested it on real factory work. Working with Mercedes-Benz, the carmaker, and Apptronik, a company that builds human-shaped robots, it put Apptronik's Apollo 2 robot on active production lines at Mercedes plants in Kecskemét, Hungary, and Sindelfingen, Germany. The robot's job was a common, repetitive one: pick metal brackets out of big bins of loose parts and place them in trays in the order each car on the line will need them. A second Google program kept track of where the robot was in the job, while Gemini Robotics 2.5 moved its hands. Google says the robot got the task right 96 percent of the time.

The company also says it tested the program on more than 200 different tasks, from serving food with tongs to tying a bow on a gift box, and that it succeeded 30 to 100 percent more often than Google's own earlier versions, depending on the task, while working 20 to 30 percent faster. Google says the robots are also better at finishing a job safely when obstacles or people are nearby.

Why it matters: most factory robots today repeat one exact motion in a fenced-off space. Sorting jumbled parts by hand is the kind of job that has been hard to automate, because no two bins look alike. A robot that can do it reliably could take over tiring, repetitive work, and could change the kinds of jobs people do on a factory floor.

There are real limits. Every number comes from Google itself, and the comparisons are against its own older programs, not rivals. Google calls the Mercedes work a proof of concept, and it has not said how many attempts the trial included, how long it ran, or how often people had to step in. And 96 percent, while impressive, still means about one mistake in every 25 tries, and a factory needs far better than that before a robot can work a line on its own. The program is going only to a handful of partner companies for now, not to the public.

Sources

  1. Google DeepMind — Gemini Robotics 2.5 announcementdeepmind.google · primary
  2. The Decoder — Gemini Robotics 2.5 scores 96 percent in a Mercedes-Benz assembly pilotthe-decoder.com
  3. Interesting Engineering — report on the factory trialinterestingengineering.com
Follow the timeline
Get an email when new entries are added.
© 2026 Sugarpine