Google DeepMind is introducing Gemini Robotics 2.5, its latest vision-language-action model, and says Apptronik's Apollo 2 humanoid used it to do automotive kitting on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, at a 96% success rate. In its own evaluations, DeepMind says the model had a 30% to 100% higher success rate than previous baselines, depending on the task, and completed tasks 20% to 30% faster. Ars Technica reports that the factory video is sped up 2x and that the current 2.5 release still runs on cloud servers, not on the robot, so those speed gains still come with latency.
What DeepMind says the pilot shows
In a November 10 post, DeepMind robotics lead Carolina Parada describes a proof-of-concept on active Mercedes-Benz lines. Apollo 2, using Gemini Robotics ER to track the workflow and Gemini Robotics 2.5 to control its hands and arms, sorts metal brackets from bulk bins into sequencing containers in the order needed for each vehicle. DeepMind calls this a real-world reliability test, not a lab demo, and says the 96% figure is for that kitting task.
The 30% to 100% success-rate range and the 20% to 30% speed claim are from DeepMind's evaluations across more than 200 distinct tasks, such as serving food with tongs, tying a bow, and pipetting liquid. Ars Technica notes that those gains are based on laboratory testing, and that the factory clip is sped up "to make the robot look more human-like." The same story says Google has experimented with embedded robotics models, but this release is still cloud-based. We are not treating the 96% factory result and the lab speed-up as the same measurement.
What to watch
DeepMind says Gemini Robotics 2.5 is rolling out to its robotics partners, not as a general product. The remaining gap to a nearly error-free kitting line is its own next test: 96% is the number it published, and it is not a claim that the cell is ready to replace a person on every shift. Codema's Hardware & Infra coverage of devices, factories and the physical AI stack follows that side of the story, and the AI Models hub tracks the Gemini releases this work sits on, including Gemini 4 Argon, which DeepMind says helped accelerate the new robotics model.
FAQ
What is Gemini Robotics 2.5?
It is Google DeepMind's latest vision-language-action model. DeepMind says it is rolling out to its robotics partners, not as a general product, and Ars Technica reports the current release still runs on cloud servers, not on the robot.
What did the Mercedes-Benz pilot show?
DeepMind says Apptronik's Apollo 2 sorted metal brackets from bulk bins into sequencing containers in the order needed for each vehicle, on lines in Kecskemét, Hungary, and Sindelfingen, Germany. It reports a 96% success rate and calls the pilot a proof-of-concept reliability test on active lines.
Where do the 30% to 100% and 20% to 30% figures come from?
From DeepMind's own evaluations across more than 200 distinct tasks, such as serving food with tongs, tying a bow and pipetting liquid. Ars Technica notes those gains are based on laboratory testing, and Codema does not treat them as the same measurement as the 96% factory result.
Is the factory video shown at real speed?
No. Ars Technica reports the factory video is sped up 2x. It also says the model still runs on cloud servers, so the speed gains come with latency.
Does 96% mean the robot is ready to replace factory workers?
The 96% figure is for the kitting task only. Codema notes it is not a claim that the cell is ready to replace a person on every shift.
Sources
Topics
Kelly B.AI desk
Kelly B. is a Codema.ai house byline: a pen name used by our editorial desk, not one individual journalist. It covers Hardware & Infra: GPUs and the memory inside them (HBM, GDDR7, DDR5, NAND), hardware for running AI locally, GPU rental prices, chipmaking, and the data centres and power behind AI. Pieces under this name are written with AI assistance and reviewed by a human editor before they're published.
All stories by Kelly B.
