DeepMind announced Gemini Robotics 2.5 for robot control
Google DeepMind announced Gemini Robotics 2.5, a vision-language-action model it describes as its most capable to date for controlling robots. The company blog post is dated November 10, 2026, with a publication timestamp of 3:00 p.m. UTC, which is 10:00 AM ET. This article summarizes that announcement and later press coverage. It is not a hands-on test of the model on a robot, and the performance figures below are company claims.
Real, reported: based on the public sources cited below. Nothing added or dramatized. Sources checked November 12, 2026, at 7:18 AM ET (EST). Task counts, success rates, and factory-floor results come from DeepMind's post or from press accounts of that post. The opened sources do not present this release as a peer-reviewed paper or name a scientific journal. Those results should be read as a company claim, not as an independently verified benchmark.
What DeepMind says it is releasing
DeepMind wrote: "Today, we are introducing Gemini Robotics 2.5, our most capable vision-language-action (VLA) model to date." It said the model "increases both the breadth of skills a robot can learn and the speed at which it can acquire them," with the aim of helping robots operate in more unstructured, real-world environments. The post says the company "tested Gemini Robotics 2.5's capabilities in task breadth, performance, and complexity across over 200 distinct tasks, though this approach could scale to many more." DeepMind listed examples that include serving food with tongs, tying a bow on a gift box, and pipetting a liquid, describing them as combining tool use, handling of deformable objects, and delicate manipulation.
On factory work, DeepMind said: "During a real-world proof-of-concept on Mercedes-Benz production lines, our models helped Apptronik's Apollo 2 humanoid pick and pack automotive parts with a 96% success rate." The post locates that trial on "active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany." It describes the job as automotive kitting: sorting parts from bulk bins into organized trays in the order needed for each vehicle. DeepMind said the trial used an embodied reasoning model, Gemini Robotics ER, to keep track of the workflow, and Gemini Robotics 2.5 to control the robot's hands and arms as it picked up metal automotive brackets and placed them into sequencing containers, "all while correcting for any gaps."
In a later section, DeepMind wrote: "Our evaluations show significant increases in task precision, robustness, and speed. Compared to previous baselines, Gemini Robotics 2.5 achieves a 30% to 100% higher success rate depending on the type of task, while completing tasks 20% to 30% faster." It also said inference-time reasoning guidance "boosted safe task completion around obstacles by 33%" in its evaluations. Those percentages are DeepMind's own evaluation claims. We have not independently measured the tasks, the baselines, or the factory trial.
The same DeepMind post includes a comment from Jeff Cardenas, identified as co-founder and chief executive of Apptronik: "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." That comment appears on DeepMind's announcement page. It is Apptronik's characterization, not an independent factory audit.
The post also says the company has "improved physical safety steerability and introduced a new benchmark to test against adversarial jailbreaks specific to robotics." Later, DeepMind says it is releasing ASIMOV-Jailbreak, "a new open-source benchmark containing hundreds of prompts that disguise unsafe requests across dozens of scenes." In that section it reports: "Argon achieves a 99.5% jailbreak resistance — a significant improvement over the 82% rate of our previous baseline, Gemini Robotics ER 2." That 99.5% figure is DeepMind's reported score for Gemini 4 Argon, not for Gemini Robotics 2.5, and it is a company claim. DeepMind said Gemini Robotics 2.5 "is rolling out to our robotics partners" and that, like earlier versions, it works with multiple robot types, including Apptronik's Apollo 2 humanoid.
How SiliconANGLE covered the announcement
SiliconANGLE, in an article first posted at 11:30 AM ET on November 11 — the day after DeepMind's post — reported that the lab "debuted the next generation of Gemini Robotics, the company's most capable vision-language-action model to date." The outlet repeated DeepMind's description of a proof-of-concept in which the model "successfully packed automotive parts with a 96% success rate" at Mercedes-Benz. On the evaluation figures, SiliconANGLE wrote: "In evaluations, Robotics 2.5 now achieves a 30% to 100% success rate depending on the task and completes tasks 20% to 30% faster compared with the previous version’s baseline." That wording is SiliconANGLE's. DeepMind's own post says Gemini Robotics 2.5 achieves a "30% to 100% higher success rate depending on the type of task." DeepMind's acknowledgements on the same post say: "Finally, we’d like to thank our partners: Agile, Apptronik, and Mercedes-Benz teams for their support on this release." SiliconANGLE wrote that the model "is rolling out today via DeepMind’s robotics partners, including Agile Robots SE, Apptronik and Boston Dynamics Inc." Boston Dynamics appears in SiliconANGLE's account; it is not named in DeepMind's opened acknowledgements or in the trial passages cited here.
SiliconANGLE's article also discusses a separate DeepMind genomics model announced the following day. This article is only about the robotics release.
How to read the reported figures
A 96% success rate on a described factory kitting trial, and a 30% to 100% higher success rate against "previous baselines," cannot be compared with other robots unless the tasks, time limits, and baselines are the same. DeepMind did not, in the opened post, publish a full task-by-task table or an external audit of the Mercedes-Benz proof-of-concept. The new robotics jailbreak benchmark is also a company-introduced test. Those limits do not make the announcement unimportant; they do mean the numbers should be read as the lab's own reported results.
Our advice (general guidance)
If you work with robots or evaluate vendors, treat Gemini Robotics 2.5 as a newly announced company model. Ask for the task definitions behind any success-rate claim, and run the jobs you actually need rather than relying on a single factory example. Physical robots can injure people or damage equipment; follow the safety process your workplace already uses, and do not assume a vendor's jailbreak benchmark covers every way a system can be misused.
Be cautious of third-party pages or messages that claim to offer unofficial downloads, API keys, or "partner access" to the model. Open DeepMind's own product or blog pages on your own, and ignore anyone who asks you to disable safety controls or to install software from an unfamiliar link.
These are editorial precautions, not individualized engineering, procurement or safety advice.
Sources
- Google DeepMind: Gemini Robotics 2.5 accelerates real-world general dexterity (Nov 10, 2026, 10:00 AM ET)
- SiliconANGLE: Google DeepMind releases new advanced robotics AI model and next-generation gene prediction model (first posted Nov 11, 2026, 11:30 AM ET; page also shows a later update)
Published November 12, 2026 by FraudMeNot Team. Label: Real, reported.