Google DeepMind put a humanoid robot to work on a real car factory line — and it sorted parts correctly 96% of the time. That is the headline from Gemini Robotics 2.5, announced on November 10, 2026, and it is the first time one of these models has been tested on an active production line instead of in a lab.
How I checked this: I read Google DeepMind's official announcement, "Gemini Robotics 2.5: Accelerating Physical Intelligence," line by line, then cross-checked the key numbers against independent same-day coverage from TechTimes and other outlets. Every figure in this article is DeepMind's own claim unless I say otherwise — no outside lab has tested this model yet, and the one number that matters most comes with a catch I will get to.
One model, three leaps
Gemini Robotics 2.5 is a vision-language-action model — a VLA. In plain terms, it takes what the robot's cameras see plus a spoken or written instruction, and turns that into arm and hand movements. DeepMind calls it its most capable VLA model to date, and credits Gemini 4 Argon with speeding up the research behind it: Argon acted as a simulation engine, generating thousands of practice runs from a real camera scene and building digital twins, so the robot model could learn new skills faster.
DeepMind says it tested the model across more than 200 distinct tasks — serving food with tongs, tying a bow on a gift box, pipetting liquid, handling delicate and flexible objects. Compared with previous baselines, success rates rose 30% to 100% depending on the task type, and tasks finished 20% to 30% faster. The model runs on more than one body: Apptronik's Apollo 2 humanoid, with its five-finger, 22-degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered parallel grippers.
That multi-body point matters more than it sounds. A model that only works on one robot is a demo; a model that moves between a humanoid and a dual-arm rig is a platform. It is the same week Google shipped four TPU systems in a single year — our breakdown of Google Launches TPU 8tx and TPU 8i Max: Four AI Chips in One Year — No Prices Yet covers the hardware side of that push — and DeepMind is clearly trying to make its robotics stack just as portable.
The Mercedes-Benz test
The part of the announcement I kept coming back to is the factory pilot, because it is not a lab demo. DeepMind partnered with Apptronik and Mercedes-Benz to run Apollo 2 humanoids on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The job was automotive kitting: sorting individual metal brackets from bulk bins into trays, in the exact order each vehicle on the line needs. It is repetitive, physically taxing work — exactly the kind of job factories want to hand off so technicians can focus on complex builds.
Apollo 2 did it autonomously, using Gemini Robotics ER to track the workflow and Gemini Robotics 2.5 to control its hands and arms, correcting for gaps as it went. The result DeepMind reports: a 96% success rate.
Now the honest catch. DeepMind's announcement does not say how many trials that 96% covers, how it defines a success, or how often a human stepped in. As TechTimes pointed out the same day, 24 successes in 25 attempts and 960 in 1,000 both round to 96%, but they mean very different things. Mercedes has also not publicly confirmed the figure itself. So treat 96% as a promising pilot number from the company that built the model — not as a production guarantee. If you want to see how the industry is learning to measure robot claims properly, our piece on AGIBOT's Worken: A New Way to Measure What Robots Actually Do is worth a read alongside this one.
Image: Google DeepMind
Safety got a real upgrade too
The safety work in this release deserves as much attention as the 96%. Gemini Robotics 2.5 uses inference-time reasoning to steer around physical hazards in real time — DeepMind says that boosted safe task completion by 33% around obstacles and 43% around nearby people. And the company open-sourced a new benchmark, ASIMOV-Jailbreak, with hundreds of prompts that disguise unsafe requests across dozens of scenes, to test whether robot AI can be tricked into dangerous actions. On it, Gemini 4 Argon resisted 99.5% of jailbreak attempts, against 82% for the previous baseline, Gemini Robotics ER 2.
That matters because robots that share space with people need a different kind of safety from chatbots. A wrong sentence is embarrassing; a wrong movement near a person is an injury. The open benchmark is the most checkable part of this whole announcement — anyone can run it — which is more than can be said for the 96%.
Why this matters
I tested nothing hands-on here — nobody outside DeepMind's partner group has access to this model yet, and I won't pretend otherwise. What I checked is the shape of the claim, and the shape is new: this is the first DeepMind robotics release whose headline result comes from a working factory floor rather than a controlled lab. The dexterity numbers (200+ tasks, up to 100% better success) are lab figures; the Mercedes pilot is the real-world anchor, and DeepMind is careful to call it a proof of concept, not a deployment.
The limits are just as clear. There is no public API, no price, and no general-availability date — Gemini Robotics 2.5 is rolling out to DeepMind's robotics partners only, including Agile Robots, Apptronik and Boston Dynamics. If you build robots, you cannot call this model today. And the safety benchmark, while open, is DeepMind's own test, not an independent one.
Still, the direction is hard to miss. Google's AI answered from orbit last week — see Google's AI Just Answered From Space: What Project Suncatcher's First Orbit Reply Means — and this week its robots are sorting parts on a German car line. The same month also brought robots a sense of touch, covered in Enact Intelligence Raises $30M to Give Robots a Sense of Touch. The frontier is moving from chat windows into physical work, and the companies that can prove reliability on a factory floor — with trial counts published, not just percentages — will be the ones that matter.
My verdict: if you follow robotics, this is the most important launch of the week — a humanoid doing real kitting on a live line is a genuine milestone. If you were hoping to use the model, wait: partners-only means the rest of us watch for now. Watch for two things next — an independent replication of the Mercedes numbers with a trial count attached, and the first partner deployments beyond the pilot. Those will tell us whether 96% was a demo or a turning point.
Would you trust a humanoid robot working beside you on a factory floor? Tell us in the comments — and if you work in manufacturing, would 96% be good enough for your line?
Sources
- Google DeepMind, "Gemini Robotics 2.5: Accelerating Physical Intelligence," November 10, 2026 — https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/
- TechTimes, "Gemini Robotics 2.5 Reports 96% Success Sorting Mercedes Car Parts but Publishes No Trial Count," November 11, 2026 — https://www.techtimes.com/articles/330432/20261111/gemini-robotics-25-reports-96-success-sorting-mercedes-car-parts-publishes-no-trial-count.htm
Related reading on our blog: Google Unveils Gemini 4 Carbon: Its Strongest AI Yet, Built for Cyber Defense


Comments
Post a Comment