Confirmed 10 Nov 2026 · Labs A one-minute read

Google's robot model sorted parts on live Mercedes-Benz lines, succeeding 96% of the time.

Google DeepMind released Gemini Robotics 2.5 on November 10 and published the factory result. Caveat: it is a proof of concept on one task, and every figure is Google's own.

Until now

Humanoid robots do demos in laboratories and warehouses, not on car lines.

Now

They run a real job on 2 active plants, at a published rate.

Why it matters

Kitting is the job a plant gives a new worker: pick the right parts, in the right order, for the next vehicle. A machine that does it 96 times in 100 is not yet a shift, but it is no longer a video. The remaining 4 is the whole question.

Established

Apptronik's Apollo 2 ran kitting on active lines in Hungary and Germany.
Google reports a 96% success rate on that task.
Google released ASIMOV-Jailbreak, an open benchmark for unsafe robot requests.

Not established

That robots replace line workers. This is a proof of concept.
That the rate holds over a full shift or across other tasks.
Independent verification. Every performance figure comes from Google itself.

The five-minute readThe record, the timeline, the weight and 3 sources

The record

Google DeepMind published Gemini Robotics 2.5 on November 10, 2026. We read the company's own post at deepmind.google. The model is described as a vision-language-action model, which maps camera views and language instructions onto robot arm and hand motion. DeepMind reports evaluation across more than 200 distinct tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid, with success rates 30% to 100% higher than previous baselines depending on task type, and completion 20% to 30% faster. The reliability demonstration ran on active automotive production lines at Mercedes-Benz plants in Kecskemet, Hungary and Sindelfingen, Germany. Apptronik's Apollo 2 humanoid performed automotive kitting, sorting parts from bulk bins into trays in the order each vehicle required, using a 5-finger SharpaWave hand with 22 degrees of freedom. A second model, Gemini Robotics ER, tracked the multi-step workflow while Gemini Robotics 2.5 controlled grasping and placing. DeepMind reports a 96% success rate on that task and describes it as a proof of concept. On safety, DeepMind reports that inference-time reasoning guidance improved safe task completion by 33% around obstacles and 43% around nearby people, and released ASIMOV-Jailbreak, an open benchmark of hundreds of prompts that disguise unsafe requests. On that benchmark Gemini 4 Argon resisted 99.5% of attacks against 82% for the previous baseline, Gemini Robotics ER 2. DeepMind also reports using Gemini 4 Argon as a simulation engine, building digital twins from real camera scenes to generate practice runs. Apptronik chief executive Jeff Cardenas is quoted: "This is what we built Apollo for: useful work alongside people in real industrial settings." The model is rolling out to robotics partners with an early-access waitlist, and named partners include Agile, Apptronik and Mercedes-Benz. Not published: how long the robots ran, how many units were kitted, what the 4% failures were, what a human does on the same task, whether the line ran at normal speed, any cost, and any independent verification of a single figure.

How it got here

Sep 3
A robot company that had raised $1.9 billion committed $3.5 billion to buy computing power, and the seller took a stake in the buyer. Edition 012.
Oct 28
Korea commits more than 5 trillion won to AI and robots in small factories, the first state programme aimed at the factory floor rather than the laboratory. Edition 065.
Nov 5
China bars SoftBank from buying ABB Robotics' China business, part of a $5.375 billion sale, treating industrial robotics as strategic. Edition 079.
Nov 10
Google DeepMind publishes Gemini Robotics 2.5 with a 96% kitting result from Mercedes-Benz lines in Hungary and Germany, and open-sources the ASIMOV-Jailbreak safety benchmark.

Who has to react

Mercedes-Benz's works council and IG Metall, since kitting is a job someone currently does; every other carmaker running the same process; Apptronik's rivals in humanoids, who now have a published number to beat; NVIDIA, whose GR00T robot models compete directly; industrial safety regulators in Germany and Hungary; and insurers who price a line with machines on it.

What to watch

Would strengthen itMercedes-Benz or another carmaker moves humanoid kitting from proof of concept to contracted production at a named plant, or publishes a rate measured over a full shift, by December 31, 2027.
Would weaken itThe Mercedes-Benz trial ends without a production agreement, or no third party reproduces a comparable rate on a live line, by December 31, 2027.
Would strengthen itAn independent evaluation on ASIMOV-Jailbreak or a live line reports materially lower numbers than Google published, by June 30, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityHigh

PersistenceYear

Cost is medium. Running robots on live production lines at 2 Mercedes-Benz plants costs real money and real line time, and publishing a specific rate exposes the company to being measured against it. Commitment is medium: a proof of concept with an early-access waitlist commits DeepMind, Apptronik and Mercedes-Benz to very little beyond the claim itself. Reversibility is high, because a pilot can simply end and usually does, and nothing here is contracted. Persistence is a year rather than structural: the useful fact is the published number, which will be either beaten or quietly dropped within a few product cycles. We scored this honestly rather than generously, because every figure is the vendor's own.

The words

Kitting: Gathering the exact parts one vehicle needs into a tray, in build order, so the line worker never searches. It is an entry-level job on most plants.
Vision-language-action model: One model that takes camera images and a plain instruction and outputs robot movement, instead of separate perception, planning and control systems.
Proof of concept: A limited trial showing something can work, run alongside normal operations. It is not a deployment and carries no production volume.
Degrees of freedom: The number of independent ways a hand or joint can move. 22 is close to a human hand and is what fine assembly needs.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.