Google DeepMind on Tuesday released Gemini Robotics 2.5, a vision-language-action model the lab said it is rolling out to robotics partners. In a blog post, DeepMind said a proof of concept with Apptronik’s Apollo 2 humanoid on Mercedes-Benz production lines in Kecskemet, Hungary, and Sindelfingen, Germany, completed automotive kitting at a 96% success rate. Ars Technica also reported the release and said the model is not generally available.
DeepMind said it evaluated the model across 200 distinct tasks, including tool use and handling deformable objects, and reported a 30% to 100% higher success rate than previous baselines, depending on the task, and that it completed tasks 20% to 30% faster. Ars noted that those comparisons were based on laboratory testing.
For the factory trial, DeepMind said Apollo 2 used the lab’s embodied reasoning model to track the workflow and Gemini Robotics 2.5 to control its hands and arms as it sorted automotive parts from bulk bins into sequenced trays. The lab described the work as a proof of concept and a pilot. Kanishka Rao, a DeepMind director, told Ars that closing the remaining 4% gap will be a “steep road.” Carolina Parada, DeepMind’s robotics head, told Ars the lab can optimize for a single application “at the cost of generalization,” or push generalization and “maybe not get to 99.9 in every application.”
DeepMind said Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards, and that in its evaluations this raised safe task completion by 33% around obstacles and by 43% around nearby people. The lab said it is releasing ASIMOV-Jailbreak, an open-source benchmark with hundreds of prompts that disguise unsafe requests. Ars reported that the dataset is available on Hugging Face. DeepMind said Gemini 4 Argon, which Google unveiled in September, scored 99.5% jailbreak resistance on the benchmark, compared with 82% for the prior Gemini Robotics ER 2 baseline.
The lab said Argon also served as a simulation engine that generated thousands of practice runs from real-world camera scenes to help train Gemini Robotics 2.5. Ars reported that the 2.5 release still runs on cloud servers, rather than fully on the robot, and that the resulting latency remains a limitation. DeepMind listed Agile, Apptronik, and Mercedes-Benz among the partners supporting the release.