Google DeepMind on Nov. 10 introduced Gemini Robotics 2.5, its latest vision-language-action model for robots, the company said. In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 humanoid robots used the models to kit automotive parts on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany, reaching a 96% success rate. Ars Technica reported the same announcement.
Why it matters: DeepMind said the robots sorted parts from bulk bins into trays in the order needed for each vehicle. Ars Technica quoted DeepMind director Kanishka Rao as saying the remaining 4% gap to typical human reliability will be a “steep road.”
DeepMind said it evaluated Gemini Robotics 2.5 on more than 200 tasks, including using tools and handling flexible objects. Ars Technica said the company reported a 30% to 100% higher success rate than the previous version and 20% to 30% faster performance in lab tests, and that an Apollo 2 robot could tie a bow. DeepMind is also releasing ASIMOV-Jailbreak, an open benchmark of adversarial prompts; it said a related system, Argon, reached 99.5% jailbreak resistance, compared with 82% for an earlier baseline, Gemini Robotics ER 2.
The company said Gemini Robotics 2.5 is rolling out to robotics partners. Carolina Parada, who leads DeepMind’s robotics work, told Ars Technica that pushing for generality can come at the cost of 99.9% reliability on a single factory task.