Google DeepMind says Gemini Robotics 2.5 sorted parts at Mercedes-Benz with 96% success

Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action model that turns what a robot sees and is told into movements, and is rolling it out to Google's robotics partners.

In a proof of concept on Mercedes-Benz lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 humanoids using the model sorted car parts from bulk bins into trays in assembly order with a 96 percent success rate, Google said. Its Gemini Robotics ER model followed the workflow while Gemini Robotics 2.5 moved the robots' arms and hands.

Google said it tested the model on more than 200 tasks, such as serving food with tongs and tying bows, with success rates 30 to 100 percent above earlier baselines and work done 20 to 30 percent faster. In Google's own tests, reasoning at run time made robots complete tasks safely 33 percent more often around obstacles and 43 percent more often near people.

Google also released ASIMOV-Jailbreak, an open benchmark that checks whether robots resist disguised unsafe instructions. The factory result comes from a proof of concept, not a commercial rollout.