Google DeepMind has introduced Gemini Robotics 2.5, its most capable model yet for turning instructions into physical robot movements, after testing it on live Mercedes-Benz production lines where humanoid robots sorted car parts with a 96 per cent success rate.
DeepMind announced the new vision-language-action (VLA) model on Tuesday, 10 November, describing it as the most capable version of its robot AI to date. A VLA model connects visual and language understanding to robot actions, allowing a machine to see a task, understand an instruction and carry it out with its arms and hands.
The headline test took place far from the laboratory. DeepMind partnered with Apptronik and Mercedes-Benz to put Apollo 2 humanoid robots on active production lines at Mercedes-Benz plants in Kecskemet, Hungary, and Sindelfingen, Germany, where the robots performed automotive kitting: the repetitive, physically taxing job of sorting individual metal parts from bulk bins into trays arranged in the exact sequence each vehicle on the line needs.
In the proof-of-concept, the robots completed the task with a 96 per cent success rate. Two models split the work: Gemini Robotics ER, an embodied reasoning model, tracked the workflow and what remained to be done, while Gemini Robotics 2.5 itself controlled the humanoid's hands and arms, reaching into supply bins, picking up metal automotive brackets and placing them into sequencing containers, correcting for any gaps along the way.
DeepMind says the new model is a leap forward in three areas: the breadth of skills a robot can learn, the reliability with which it performs them, and the dexterity of its movements. Across more than 200 distinct tasks — from serving food with tongs and tying bows on gift boxes to transferring liquids with pipettes — the company reports success rates running 30 per cent to 100 per cent higher than its previous baselines, with the same jobs finishing 20 to 30 per cent faster. Those figures are DeepMind's own results, and there is no independent evaluation yet.
The model is designed to work across different kinds of robots without a full rebuild. DeepMind demonstrated it on Apptronik's Apollo 2 humanoid, fitted with the five-finger SharpaWave hand, which has 22 degrees of freedom, and on the Franka Duo, a two-armed platform with simple two-fingered grippers. That flexibility is part of the pitch: one model that can move from tabletop arms to complete humanoid bodies, including dynamic balance for stepping, squatting and bending.
Gemini Robotics 2.5 is rolling out now to DeepMind's robotics partners, including Apptronik, Agile Robots and Boston Dynamics, with more than 100 trusted testers, from industrial automation firms to robotics start-ups, working with the models. There is no public API, no price and no general release: outside that circle, the model is not available anywhere yet, and it still runs on cloud servers rather than on the robot itself, so some delay remains between seeing and acting.
Apptronik's co-founder and chief executive, Jeff Cardenas, said the pairing had improved dramatically in just a few weeks of post-training and carried over to the Mercedes-Benz production lines with very little drop-off. DeepMind calls the factory result a proof of concept, and neither company has said whether the humanoids will stay on the lines.
The launch is the latest step in a fast-moving race to put capable robots into real workplaces. Gemini Robotics began in March 2025 and picked up whole-body control earlier this year, and Figure's humanoid robots are already doing related work at BMW. What bears watching is whether the 96 per cent factory figure holds at scale: a pilot on two production lines is a promise the next deployment has to keep, and the gap between a demonstration and a deployment is where most robotics claims are tested.
For now, the message from DeepMind is that the robots are no longer confined to the lab. They are sorting parts on car lines in Hungary and Germany, learning new tasks faster than before, and the question is no longer whether a humanoid can do the work, but how soon it can do it reliably enough, and cheaply enough, to stay there.
