Gemini Robotics 2.5 got 96% on a Mercedes line. DeepMind's own robotics lead calls the last 4% a steep road.
Google DeepMind's new robot model sorted car brackets at two Mercedes-Benz plants. The safety numbers in its blog post and in Ars Technica's write-up don't match.
Ninety-six percent sounds like an A. On a car assembly line, Google DeepMind’s own robotics team says it isn’t, and it said so in the same November 10 announcement.
Google DeepMind announced Gemini Robotics 2.5 on November 10. It’s the company’s newest vision-language-action model, which means it takes camera images plus a plain-language instruction and turns them into arm and hand movements. The headline number in DeepMind’s post is a 96% success rate in a proof of concept on active Mercedes-Benz production lines.
What the robot actually did
DeepMind worked with Apptronik and Mercedes-Benz. Apptronik’s Apollo 2 humanoid ran at plants in Kecskemét, Hungary, and Sindelfingen, Germany, doing a job called kitting. It reached into bulk bins, picked out metal brackets, and placed them into trays in the order each car on the line needs them, filling gaps in the sequence as it went.
Two models split the work. According to DeepMind, Gemini Robotics ER tracked which step came next, and Gemini Robotics 2.5 controlled the arms and hands. The company calls the run a pilot, and the wattba.eu write-up says plainly that it was a trial, not a permanent deployment.
Apptronik’s CEO Jeff Cardenas is quoted in DeepMind’s post saying the models improved “dramatically in just a few weeks of post-training” and carried over to Mercedes production “with very little drop-off.”
The part Ars Technica pushed on
Ars Technica’s Ryan Whitwam wrote that 96% is impressive next to earlier robot AI, but probably not enough for Mercedes, because a person doing a repetitive task like this gets it right virtually every time. He quotes DeepMind director Kanishka Rao calling the last 4% a “steep road.”
Carolina Parada, who heads DeepMind’s robotics work, gave Ars the trade-off in one sentence: “We can be very good at a particular application at the cost of generalization. Or we can just push on the generalization axis and maybe not get to 99.9 in every application.”
That’s an unusually candid thing for a lab to say next to a launch.
The other numbers, and who measured them
Across more than 200 lab tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid, DeepMind reports success rates 30% to 100% higher than its previous baselines, depending on the task, and runs 20% to 30% faster. Those are Google’s own comparisons against its own earlier models, and the range is wide enough that no single figure describes it. A “100% higher” result on one task can sit next to a 30% gain on another.
Ars adds two caveats from its reading. The model isn’t generally available; it’s with robotics partners including Agile Robots, Apptronik and Boston Dynamics. And the current release runs on cloud servers, so the accuracy gains come with latency.
| Claim | DeepMind’s blog post | Ars Technica’s write-up |
|---|---|---|
| Jailbreak resistance, Gemini 4 Argon in the loop | 99.5% | 98.8% |
| Previous baseline, Gemini Robotics ER 2 | 82% | 73.5% |
| Benchmark size | “hundreds of prompts” | 1,500 prompts |
I can’t tell from the sources which pair is right. The benchmark is called ASIMOV-Jailbreak and Google says it’s open source; Ars links the broader ASIMOV dataset on Hugging Face. Anyone who wants to cite a safety score should check the dataset page directly, not either article.
Where Argon fits
DeepMind says Gemini 4 Argon helped build the model by generating thousands of simulated practice runs from real camera scenes and by helping researchers guess why tasks failed. That ties this release to Google’s general model line, which I covered when Gemini 4 Argon launched. The company says the robot’s hazard-avoidance reasoning raised safe task completion by 33% around obstacles and 43% around nearby people, again measured by DeepMind.
My read
My read: the interesting thing here isn’t the 96%. It’s that a lab put “we may not get to 99.9” in the same news cycle as a product launch. Factory work has no use for a robot that fails one pick in 25 unless a person is checking behind it, and nothing in the sources says whether anyone was checking behind it in these trials. Neither DeepMind nor Ars gives the number of picks behind the percentage, either.
What I’d watch is whether Mercedes or Apptronik names a production date, because so far the only commitment on the page is a rollout to partners. For the money side of the same bet, see our earlier piece on a startup selling synthetic data to robot makers.
Note on sources
The 96%, the 30% to 100% range and the 33% and 43% safety figures come from DeepMind’s own blog post and are company-reported, not independently tested. The Ars Technica quotes from Kanishka Rao and Carolina Parada are as Ars reported them. WATTBA is a secondary write-up, used only to confirm the trial status. I read the blog post and Ars in full.