Google DeepMind has presented its new AI model for robots. In a pilot at Mercedes-Benz plants, humanoid robots used it to sort vehicle parts, with a success rate of 96 percent according to DeepMind.
Google DeepMind has presented Gemini Robotics 2.5. The company calls it its most capable vision-language-action model to date. It is meant to teach robots more skills and let them learn these faster. According to DeepMind it was tested on over 200 different tasks, among them serving food with tongs, tying a bow on a gift box and pipetting a liquid.
Compared with its predecessor, the model achieves a success rate that is 30 to 100 percent higher depending on the task and completes tasks 20 to 30 percent faster, according to DeepMind. It works with different types of robot, among them the humanoid Apollo 2 from Apptronik and the Franka Duo platform with two-fingered grippers.
Pilot in Sindelfingen and Kecskemét
For the field test, DeepMind worked with Apptronik and Mercedes-Benz. At the plants in Sindelfingen and in Kecskemét in Hungary, Apollo robots took metal brackets out of supply bins and placed them in the order in which they are needed on the line for each vehicle. According to DeepMind the pilot reached a success rate of 96 percent. Robotics director Kanishka Rao conceded, according to heise, that the road from there to near-human reliability is steep.
New safety features come on top. In DeepMind's own tests the share of safely completed tasks rose by 33 percent around obstacles and by 43 percent around people. DeepMind is also releasing ASIMOV-Jailbreak, an open dataset of instructions that disguise unsafe tasks as harmless ones. It can be used to check whether a robot can be led into unsafe actions. The model is going to DeepMind's robotics partners first.

