Google DeepMind has unveiled Gemini Robotics 2.5, a new AI model for controlling robot hands and arms, and it is showing off the first results from a car factory rather than just a lab. Announced on Tuesday 10 November, the model helped a humanoid robot sort metal parts on live Mercedes-Benz production lines with a reported 96% success rate. It also does better than its predecessor at fiddly jobs such as tying a bow, and DeepMind has released a new open test for whether robot AI can be tricked into doing something unsafe.
What Gemini Robotics 2.5 is
Gemini Robotics 2.5 is what researchers call a vision-language-action model. It takes in camera images and a spoken or written instruction, and turns them into movements for a robot’s arms and fingers. In DeepMind’s set-up it works as a pair. A separate reasoning model, Gemini Robotics ER, keeps track of the overall job and plans the steps, while the 2.5 model controls the hands and arms. The previous generation, Gemini Robotics 2, arrived in July and focused on giving robots of many shapes whole-body control.
Like earlier versions, it is designed to run on different machines rather than one custom robot. DeepMind names Apptronik’s Apollo 2 humanoid, which has five-fingered hands with 22 degrees of freedom, and the Franka Duo platform, which has two-fingered grippers. It is not on general sale. DeepMind says it is rolling out to its robotics partners, and Ars Technica reports it is not generally available.
Better at delicate, fiddly work
DeepMind tested the model across more than 200 tasks, including serving food with tongs, pipetting liquid and tying a bow on a gift box. Against its own earlier baselines, it reports a 30% to 100% higher success rate depending on the type of task, and tasks completed 20% to 30% faster. Some of the individual jumps are large: according to DeepMind’s charts, packing pastries with tongs went from 5% to 100%, and closing the lid of a toolbox from 0% to 93%.
The harder items still fail often. Folding trousers rose from 0% to 40%, hanging a cardigan from 0% to 62% and untying a ribbon bow from 0% to 70%. That matters because soft, floppy objects such as cloth and ribbon are notoriously difficult for robots, and those scores show the technology is moving, not that it has arrived. These figures are DeepMind’s own, from its own tests, and have not been independently checked.
The Mercedes-Benz test
The headline claim is the factory pilot. Working with Apptronik and Mercedes-Benz, DeepMind put Apollo 2 robots on active production lines at Kecskémet in Hungary and Sindelfingen in Germany. The job was kitting: taking individual metal brackets out of bulk bins and placing them into trays in the exact order each vehicle on the line needs. The robot reached a 96% success rate, according to DeepMind. The Decoder and Ars Technica both report the same figure from the company.
It is worth reading that number carefully. Ars Technica points out that a human doing a repetitive task like this gets it right almost every time, so 96% is still probably not enough for a manufacturer. It quotes DeepMind director Kanishka Rao calling the last 4% a “steep road”, and DeepMind’s robotics head Carolina Parada describing the trade-off between being very good at one application and being general-purpose. This is a proof of concept, not a roll-out of humanoids across Mercedes plants.
Apptronik’s chief executive Jeff Cardenas said the models improved a great deal in just a few weeks of post-training and carried over to the Mercedes lines with very little drop in performance. If that holds, it is the sort of evidence factories want: a robot trained largely elsewhere that copes with a new site.
Safety: a new open benchmark
DeepMind also says the model has better safety behaviour. It uses reasoning at the moment of action to steer around hazards, and in the company’s own evaluations safe task completion improved by 33% around obstacles and by 43% around nearby people. Ars Technica adds that the robot will stop or wait if something is in its way.
To push the wider industry, DeepMind has released ASIMOV-Jailbreak, an open-source test set on Hugging Face. It holds prompts that disguise unsafe requests across many scenes, to see whether a robot’s AI can be talked into something it should refuse. DeepMind says its Gemini 4 Argon model resists 99.5% of these attempts, against 82% for the earlier Gemini Robotics ER 2. Ars Technica reports slightly different figures from the same project, 98.8% against 73.5%, so treat the exact numbers with caution. There is still no agreed way to test physical AI for safety, which is the point of releasing it openly.
How Gemini 4 Argon helped
A big bottleneck in robotics is data. Chatbots learned from the open internet, but there is far less recorded data of robots doing real tasks. DeepMind says it used Gemini 4 Argon as a simulation engine: from a real camera view of a scene, it can generate thousands of varied practice runs and build a digital twin of the setting. Combined with real demonstrations, that helps the robot cope with new scenes and objects, and Argon can also suggest why a task failed from video footage. We looked at the same model’s wider performance in our report on Gemini 4 Argon’s benchmark results.
One practical limit: Ars Technica says the current release runs on cloud servers, so the gains in accuracy and speed still come with some delay, and the demonstration video it shows is sped up. For a robot working beside people, response time is as important as success rate.
Why it matters
This is the second major robot-brain announcement this week, after Nvidia opened up its GR00T 2 model on Monday. Together they show the leading AI companies racing to supply the software that makes humanoid and arm robots useful, while hardware makers such as Apptronik concentrate on the machines. For readers in the UK, the practical question is when this reaches workplaces here. Nothing in the announcements gives a date or a British customer, so for now this is a signal of direction, not a product you can buy. Our explainer on what humanoid robots can and cannot do is a good place to start for the basics.
What to watch next
Three things are worth following. First, whether Mercedes-Benz or any other manufacturer moves from a pilot to paid deployment, and what success rate they demand. Second, whether outside researchers adopt ASIMOV-Jailbreak and report their own results. Third, whether DeepMind widens access beyond its current partners.