ADاشتہار / AdvertisementGoogle Ad
Gemini Robotics 2.5 Hits 96% on a Real Mercedes LineTechnology
Google DeepMind’s Gemini Robotics 2.5 put Apptronik’s Apollo 2 humanoids to work on live Mercedes-Benz lines, sorting car parts with a 96% success rate.
Share:
Google DeepMind has unveiled Gemini Robotics 2.5, its most capable model yet for controlling robots, and the headline test did not take place in a laboratory. Apptronik’s Apollo 2 humanoid robots, running the new model, worked on live Mercedes-Benz production lines in Hungary and Germany, picking metal car parts from bulk bins and packing them into sequencing containers in the order each vehicle needed them. DeepMind reported a 96% success rate for the work.
The pilot ran at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany. A second DeepMind model, Gemini Robotics ER, kept track of the workflow while Gemini Robotics 2.5 controlled the robot’s hands and arms, correcting for gaps as parts were moved. The job is known in factories as kitting: gathering the right parts, in the right order, before assembly begins.
ADاشتہار / AdvertisementGoogle Ad
What the new model changes
DeepMind describes Gemini Robotics 2.5 as a vision-language-action model, which means it connects what a robot sees and what it is told to the movements it makes. The company says it tested the model on more than 200 tasks, with success rates 30% to 100% higher than earlier baselines depending on the task, and tasks finished 20% to 30% faster. It also says the model can control whole humanoid bodies, including stepping, squatting and bending to keep balance.
The 96% figure deserves a note of care. DeepMind has not published how many trials the number covers, how it defines a successful pick, or how often a person stepped in, and no independent check of the figure has been published. A proof of concept on one job, however real the factory floor, is not the same as a robot working a full shift unsupervised, and Mercedes-Benz has not said it plans to buy fleets of the robots.
The model is rolling out only to DeepMind’s robotics partners, including Agile Robots, Apptronik and Boston Dynamics, along with more than 100 trusted testers. There is no public version, no price and no date for wider availability. Apptronik said the models improved sharply within a few weeks of post-training and carried over to the Mercedes-Benz lines with very little drop-off.
Alongside the model, DeepMind released an open safety benchmark called ASIMOV-Jailbreak, which tests whether a robot can be talked into unsafe actions. For now, the Mercedes-Benz pilot is best read as a sign of how far humanoid robots have moved from demonstrations towards real factory work, and how much testing still lies between a promising trial and a production line that depends on a robot every day.
Related Coverage
