A humanoid robot tries out on a Mercedes assembly line: Google's new robot 'brain' sorts parts at a 96% success rate in a pilot

The backstory: factory robots are capable, but rigid

Picture a factory robot and you probably see a big arm swinging around a car plant. Those machines are fast and precise, with one big weakness: every move is scripted in advance by engineers. Shift a part slightly, or ask for a different job, and the robot often has to be stopped, reprogrammed, and retested.

So the messy jobs that require looking, thinking, and then doing, like fishing the right parts out of a jumbled bin and lining them up in order, are still often done by people. For years, tech companies have tried to give robots a more general-purpose "brain" that can see, understand, and act, instead of just replaying a routine.

What happened

On November 10, Google DeepMind introduced Gemini Robotics 2.5, which it calls its most capable "vision-language-action" model yet. The name is a mouthful, but the idea is simple: see through cameras, understand instructions in language, then move the arms and hands to get the job done. DeepMind said the model is starting to roll out to its robotics partners.

The company said it tested the model on more than 200 tasks, including serving food with tongs, tying a bow on a gift box, and transferring liquid with a pipette. Why are those hard? Tongs are a tool, ribbon is floppy and changes shape, and pipetting demands a steady hand, all tricky for robots. Compared with earlier versions, DeepMind said success rates rose by 30% to 100% depending on the type of task, and the work got done 20% to 30% faster.

The same brain can run in different robot bodies. One is Apptronik's Apollo 2 humanoid, fitted with a SharpaWave hand that has five fingers and 22 "degrees of freedom," which you can think of as the number of independent ways the joints can move; more means a more nimble hand. Another is the Franka Duo platform, which has simple two-fingered grippers that open and close like pliers.

The headline: a tryout on a real factory floor

Working with Apptronik and Mercedes-Benz, DeepMind put Apollo 2 on active car production lines in Kecskemet, Hungary, and Sindelfingen, Germany. Its job was "kitting": pulling metal brackets out of bulk bins, placing them into sequencing containers in the order each vehicle needs, and filling any empty slots. In plain terms, it was laying out the right parts for each car ahead of the workers who install them.

One model, Gemini Robotics ER, tracked the overall workflow, while Gemini Robotics 2.5 controlled the arms and hands. DeepMind said the trial hit a 96% success rate. Apptronik co-founder and CEO Jeff Cardenas said the models improved sharply after a few weeks of extra training and lost little performance when moved onto the Mercedes lines.

What 96% really means

96% sounds great, but flip it around: roughly 4 out of every 100 attempts didn't go right. That's strong for a lab. On a car assembly line, though, a single misplaced part can mean stopping to fix it.

This was also a proof of concept, more like a job tryout than a permanent hire. It shows what the robot could do on one defined task. It does not mean Mercedes has switched its lines over to humanoids, and it doesn't show the same rate would hold across continuous shifts.

Safety: could it hurt someone?

Once robots work near people, safety matters more than speed. DeepMind said the new model reasons about hazards as it works and steers around them. In the company's tests, safe task completion rose 33% near obstacles and 43% with people nearby.

There's also the risk of "jailbreaking": tricking a robot into something dangerous by disguising the request as something normal. The team open-sourced a test set called ASIMOV-Jailbreak with hundreds of these disguised prompts. DeepMind said Gemini 4 Argon resisted 99.5% of them, up from 82% for the earlier Gemini Robotics ER 2. Again, these safety numbers come from the company's own evaluations.

DeepMind added that during development it used Gemini 4 Argon to generate lots of varied practice scenarios from real camera footage and to build "digital twins," virtual copies of real scenes, so robots could practice more in simulation and train faster. The company framed the release as a step toward what it calls physical AGI, meaning AI that can do general work in the real world.

What it could change

If general-purpose robot brains like this mature, changing a production line or switching to a new task might mean a few weeks of teaching rather than a big reprogramming effort. That's a strong draw for car, electronics, and logistics companies that still rely on lots of fiddly manual work.

But the road from "96% in a pilot" to "reliable on every shift, every day" is long. Reliability, cost, maintenance, and safety alongside human workers all need much more real-world evidence. What can be said with confidence today is narrow: a new model is going out to partners, and the 96% figure comes from a controlled proof of concept.

What it means for you

You're unlikely to bump into a humanoid robot anytime soon. But if you work in a factory or warehouse, technology like this could first take over repetitive sorting and parts-prep jobs, and the work may shift toward supervising, tuning, and handling exceptions.

Longer term, if robots become more flexible and cheaper to deploy, the cost of making cars and other goods could come down. And as robots move into spaces shared with people, "can it be tricked, and can it hurt someone?" becomes everyone's question, which is why safety takes up so much of this announcement.

Original source published: Google DeepMind, Nov 10, 2026, 23:00 Beijing time (UTC+8)