Gemini Robotics 2.5 kits car parts at 96% in Mercedes pilot

Google DeepMind on November 10 introduced Gemini Robotics 2.5, a vision-language-action model that, in a proof-of-concept on Mercedes-Benz production lines, had Apptronik’s Apollo 2 humanoid pick metal brackets out of bulk bins and place them into sequencing trays with a 96% success rate. That 96% comes from a working factory floor and covers one kitting task. The lab gains and safety numbers come from DeepMind’s own evaluations, and the model is rolling out to its robotics partners.

A vision-language-action model looks at a scene through cameras, reads an instruction and decides how the robot’s arms and hands should move. In the Google DeepMind announcement, Gemini Robotics 2.5 is that motion layer. Planning is handled by a separate model, Gemini Robotics ER, which tracks a multi-step job, such as which part goes into which tray slot next. In the factory pilot the two worked as a pair.

What the Mercedes pilot measured

The pilot ran on active lines in Kecskemét, Hungary, and Sindelfingen, Germany. The task was automotive kitting: sorting individual parts from bulk bins into organized trays in the exact order each vehicle on the line needs them. Apollo 2 reached into supply bins, picked up metal brackets and placed them into sequencing containers, correcting for any gaps along the way, with Gemini Robotics ER keeping track of the workflow and Gemini Robotics 2.5 driving the hands and arms.

DeepMind still calls the result a proof-of-concept and describes it as “promising progress toward widespread real-world industrial applications.” The company pitches the use case as offloading monotonous prep work so technicians can focus on complex build tasks.

The lab numbers and Argon’s part

In the lab, DeepMind evaluated the model on 200 distinct tasks, including serving food with tongs, tying a bow on a gift box and pipetting a liquid. They mix tool use, handling of soft and deformable objects, and delicate grips. Against previous baselines, Gemini Robotics 2.5 reaches a 30% to 100% higher success rate depending on the type of task, and finishes tasks 20% to 30% faster.

The model runs on different robot bodies, as earlier Gemini Robotics models did. DeepMind names two: Apollo 2 with the five-finger, 22 degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered parallel grippers.

Much of the speed-up in development is credited to Gemini 4 Argon. DeepMind used Argon as a simulation engine: from a real-world camera scene it builds a digital twin and generates thousands of varied simulated practice runs, which, combined with real training demonstrations, help the robot generalize to new scenes and objects faster. DeepMind says Argon can also help researchers hypothesize from video footage why task failures happen and test the best improvements both in simulation and directly on real robots.

Who scored 99.5% on jailbreaks?

On physical safety, Gemini Robotics 2.5 uses reasoning at inference time to steer around hazards as it moves. In DeepMind’s evaluations this raised safe task completion by 33% around obstacles and by 43% around nearby people.

The 99.5% jailbreak resistance figure belongs to Gemini 4 Argon, against 82% for the previous baseline, Gemini Robotics ER 2. DeepMind says Argon currently leads its agentic robotics evaluations on refusing unsafe tasks, resolving uncertainty and working close to humans. The test behind the jailbreak score, ASIMOV-Jailbreak, is being released as an open-source benchmark: hundreds of prompts that disguise unsafe requests across dozens of scenes.