Google rolls out Gemini Robotics 2.5 to partners, with factory reliability still under test
Google rolls out Gemini Robotics 2.5 to partners, with factory reliability still under test
Google DeepMind introduced Gemini Robotics 2.5 on November 10 and began rolling it out to robotics partners. The vision-language-action model turns visual input and instructions into physical movements. Google reports a 96% success rate for automotive parts sorting in a proof-of-concept using Apptronik’s Apollo 2 at Mercedes-Benz production lines in Hungary and Germany. That figure is a company result, not an independently reproduced reliability test. Google DeepMind
The factory work combines separate systems: Gemini Robotics ER tracks the workflow, while Robotics 2.5 controls the hands and arms. DeepMind also evaluated the new model across more than 200 tasks, including manipulating flexible objects and using tools. The announced rollout remains limited to partners; the release does not establish public model weights or a general-use API. Google DeepMind
Independent reporting adds an important operational limit. Ars Technica reports that Robotics 2.5 currently runs on cloud servers and that latency remains visible. Its interviews with DeepMind also describe the difficulty of closing the remaining reliability gap in factory work. The report corroborates the announcement, rather than independently validating its success rates. Ars Technica
The accompanying ASIMOV-Jailbreak safety benchmark deserves a separate access check. Its Hugging Face page is visible, but downloading the dataset requires accepting conditions and sharing account information. The page’s metadata lists CC BY 4.0, while its README calls the project MIT-licensed. Researchers should verify which terms govern the files they intend to use rather than assume everything is an unrestricted download. Google’s dataset repository
The practical advance is a broader system being tested inside an operating workplace. What factories need next is evidence about sustained operation: recovery from mistakes, the amount of human supervision, cycle times and the cost of interruptions. A promising pilot moves that discussion forward, but it leaves those deployment questions open.
Sources rechecked November 15, 2026. Original announcement dates retained.
Sources and citations
Revision history
Published .