Gemini Robotics 2.5 runs humanoid kitting pilot on Mercedes lines

Gemini Robotics 2.5 runs humanoid kitting pilot on Mercedes lines

10 November 2026: Google DeepMind has released Gemini Robotics 2.5, a model that turns camera input and spoken or written instructions into robot movements, and says it drove Apptronik's Apollo 2 humanoid through a parts-kitting pilot on active Mercedes-Benz production lines with a 96% success rate. DeepMind says the model is rolling out to its robotics partners.

Kitting on two live assembly lines

The proof of concept ran in Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany. Kitting is the repetitive preparation job of taking individual parts from bulk bins and placing them in trays in the exact order each vehicle on the line needs. In the pilot, Apollo 2 reached into supply bins, picked up metal brackets and placed them in sequencing containers, correcting for gaps as it went.

Two models shared the work. Gemini Robotics ER, DeepMind's embodied reasoning model, kept track of the workflow, while Gemini Robotics 2.5 controlled the robot's hands and arms. Apptronik's chief executive Jeff Cardenas said the models improved within a few weeks of post-training and carried over to Mercedes production with little loss of performance, The Decoder reports.

DeepMind does not say how many shifts or parts the pilot covered, and the 96% figure is its own.

What DeepMind claims for the model

DeepMind says it evaluated Gemini Robotics 2.5 on 200 distinct tasks, from serving food with tongs to tying a bow on a gift box and pipetting liquid. Against its own previous baselines it reports a success rate 30% to 100% higher depending on the type of task, with tasks completed 20% to 30% faster. These are the company's own comparisons, not an independent benchmark.

The model works across robot types, including Apollo 2 with its five-finger SharpaWave hand of 22 degrees of freedom, and the Franka Duo platform with two-finger parallel grippers. DeepMind also lists Gemini Robotics ER 2 and a smaller On-Device 2 model built to run locally on the robot.

Safety near people

For a plant, the safety claims matter as much as the dexterity. DeepMind says the model's real-time control lets the robot steer around hazards during a task, and that in its own tests safe task completion rose 33% near obstacles and 43% near people. It has also published ASIMOV-Jailbreak, an open benchmark on Hugging Face of prompts that disguise unsafe requests, to test whether robotics models can be talked into dangerous actions.

For a manufacturer weighing humanoids, the pilot is a data point on one task, kitting, which is repetitive, well bounded and already done in a fixed sequence. Our catalogue lists 63 robot manufacturers and 25 robotics integrators, and Mercedes-Benz among our car makers.

Sources

• Google DeepMind: Gemini Robotics 2.5 accelerates real-world physical intelligence, 10 November 2026, 10 November 2026
• The Decoder: Gemini Robotics 2.5 scores 96 percent sorting metal parts in a Mercedes-Benz assembly pilot, 10 November 2026, 10 November 2026