Industry
Gemini Robotics 2.5 helps Apollo 2 kit Mercedes parts at a 96% success rate, DeepMind says
Google DeepMind introduced Gemini Robotics 2.5 on 10 November, calling it the company's most capable vision-language-action model to date. In a blog post by Carolina Parada, the lab said the model increases both the range of skills a robot can learn and the speed at which it can acquire them, with gains in dexterity, reliability and safety.
The headline industrial result is a proof of concept with Apptronik and Mercedes-Benz. DeepMind said Apollo 2 humanoid robots used the new model on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany, to sort automotive parts from bulk bins into sequenced trays. The lab reported a 96% success rate on that kitting task. The embodied reasoning model, Gemini Robotics ER, tracked the workflow, while Gemini Robotics 2.5 controlled the hands and arms as the robot picked metal brackets and corrected for gaps, DeepMind said.
Apptronik marked the release on X the same day: "Tying a bow, sorting car parts... it's called range." DeepMind's post lists tying a bow on a gift box among more than 200 evaluation tasks, alongside serving food with tongs and pipetting liquid, as examples of tool use and handling deformable or delicate objects. The lab says Gemini Robotics 2.5 achieved a 30% to 100% higher success rate than previous baselines, depending on the task, and completed tasks 20% to 30% faster. Those figures are DeepMind's own evaluations.
Sundar Pichai, Google's chief executive, reposted Parada's announcement on X, quoting her description of Gemini Robotics 2.5 "for real-world general dexterity."
Jeff Cardenas, Apptronik's co-founder and chief executive, is quoted in the DeepMind post: "This is what we built Apollo for: useful work alongside people in real industrial settings. Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off. This is the foundation for bringing humanoid robots to industrial and commercial environments at scale."
The model works on more than one robot type, DeepMind said, including Apollo 2 with the five-finger, 22 degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered grippers. On safety, the lab said inference-time reasoning guidance improved safe task completion around obstacles by 33% and around nearby people by 43% in its evaluations. It is also releasing ASIMOV-Jailbreak, an open benchmark of adversarial robotics prompts. DeepMind said Gemini 4 Argon scores 99.5% jailbreak resistance on that test, compared with 82% for the previous Gemini Robotics ER 2 baseline.
Gemini Robotics 2.5 is rolling out to DeepMind's robotics partners, the post says. The Mercedes-Benz work is described as a proof of concept, not a claim that the robots are in full production use across the automaker's plants.