Google DeepMind announces Gemini Robotics 2.5

Google DeepMind presents Gemini Robotics 2.5 as a vision-language-action model that expands the range of skills robots can learn and the speed at which they acquire them.[1] The company says those improvements are intended to let robots operate in less structured real-world environments and assist people with more tasks.[1] The model is also described as natively working with many robot types, including Apptronik’s Apollo 2.[1] That compatibility claim does not establish that every robot platform is supported or that the model is available for independent deployment.

Google DeepMind says its evaluated tasks included serving food with tongs, tying a bow on a gift box and pipetting liquid.[1] The company describes these as combining advanced tool use, manipulation of deformable objects and delicate handling.[1] Across the tasks, it reports success rates 30% to 100% higher than previous baselines, depending on the task, and completion times 20% to 30% faster.[1] The results are presented as task-dependent ranges, not as one overall success rate. Professionals assessing the claims should keep the specific evaluation tasks and baseline comparisons in view.

Google DeepMind reported a real-world proof of concept on Mercedes-Benz production lines.[1] In that demonstration, Gemini Robotics 2.5 helped Apptronik’s Apollo 2 humanoid pick and pack automotive parts, with a reported 96% success rate.[1] The figure relates to that particular proof of concept; the evidence does not establish performance across other production-line work or facilities. The demonstration offers manufacturing teams a concrete task to compare with their own needs, but it is not evidence of broad deployment.

Google DeepMind says Gemini Robotics 2.5 is rolling out to its robotics partners.[1] Ars Technica reports that it is not generally available, is in the hands of partners including Agile Robots, Apptronik and Boston Dynamics, and that the current release runs on cloud-based servers.[2] Google DeepMind says it improved physical-safety steerability and introduced a benchmark for robotics-specific adversarial jailbreaks.[1] Ars reports a 98.8% jailbreak-safety score when Argon was included in the loop, while Gemini Robotics ER 2 alone scored 73.5% on Google’s safety benchmark.[2] Ars also reports an open-source database of 1,500 prompts designed to trick robotic AI into unsafe behavior and says the full ASIMOV dataset is available on Hugging Face for embodied-AI researchers.[2]

DeepMind director Kanishka Rao said closing the last 4% gap would be a steep road.[2] The report also notes the concern that becoming very good at a particular application can come at the cost of generalization.[2] These points qualify the reported task results: strong performance on specific evaluations does not, by itself, establish equally strong performance in unfamiliar settings. Teams considering a trial can use the announced task examples and limitations to define what they would need to test in their own environment.