From practice to production lines.
A robot that can pick up one familiar object is useful. A robot that can adapt its grip, use tools and put unfamiliar parts in the right order has a much bigger job description.
On November 10, Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model: software that connects what a robot sees and understands to its physical movements. The company says it evaluated the model across more than 200 distinct tasks, including serving food with tongs, tying a bow and pipetting liquid.
The most tangible development is a real-world automotive trial with Apptronik and Mercedes-Benz. It takes the story beyond a neat laboratory demonstration and into the repetitive preparation work that keeps a production line supplied.
Putting parts in order.
In the proof-of-concept, Apptronik’s Apollo 2 humanoid robots worked on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Their assignment was automotive kitting: taking parts from bulk bins and placing them into organized trays in the sequence needed for each vehicle.
Google reports a 96% success rate for that task. Its embodied-reasoning model tracked the workflow, while Gemini Robotics 2.5 controlled the robot’s hands and arms. The example involved picking metal automotive brackets and placing them into sequencing containers while correcting gaps.
That figure describes this particular pilot—not every task, an independently audited factory-wide result or an American deployment. The useful prospect is straightforward: less repetitive sorting for technicians and more time for complex assembly work. Google describes the trial as progress toward wider industrial applications.
More than a firm grip.
Factory parts are only one test of dexterity. Flexible materials, delicate objects and tools each demand a different kind of control. Google says the model works across different robot types, including Apollo 2 humanoids and the Franka Duo platform with two-finger grippers.
The company also describes improved guidance for working around physical obstacles and nearby people. That matters because a useful assistant must adjust its movements to the space it shares, not simply complete a task in an empty test area. These are reported research improvements, not a blanket guarantee for every workplace.
Practice makes progress.
Google credits Gemini 4 Argon with accelerating development. It can generate varied simulated practice runs from camera scenes and help researchers investigate failures using video, then test proposed improvements in simulation and on robots. Real training demonstrations remain part of that process.
That builds on the simulation-assisted robotics research we covered in October. The new development is the next robotics model and its reported production-line trial, rather than the same simulation idea announced again.
Gemini Robotics 2.5 is rolling out to Google’s robotics partners. Our reading: the constructive advance is a closer connection between learning a skill and doing useful work. Better hands, adaptable practice and a concrete factory task make that connection easier to see.