A humanoid robot sorted metal parts for Mercedes-Benz assembly lines in a pilot involving Google DeepMind and Apptronik. Google DeepMind reports that the robot completed the kitting task with a 96 percent success rate, placing parts into trays in the order required for each vehicle.
Two models shared the factory task
The pilot ran on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apptronik’s Apollo 2 robot picked metal parts from bulk bins and arranged them in trays for the assembly process.
Two AI models handled different parts of the work. Gemini Robotics ER monitored the workflow, while Gemini Robotics 2.5 controlled the robot’s hands and arms. That division gave the system one model to oversee the task and another to carry out physical movements.
Google DeepMind describes Gemini Robotics 2.5 as a vision-language-action model, or VLA. It processes what a robot sees, interprets language instructions, and turns them into movements. The model is intended to work across different kinds of robots, including Apollo 2 and the Franka Duo platform, which uses two-finger parallel grippers.
Performance claims span more than factory work
Google DeepMind says it evaluated the model on more than 200 tasks, including serving food with tongs and pipetting liquids. Against the company’s own previous baselines, it reported success rates that were 30 to 100 percent higher, depending on the task type. The company also says tasks were completed 20 to 30 percent faster.
Those figures cover a range of tasks and comparisons, while the 96 percent result refers to the Mercedes-Benz pilot. They show how the company is presenting progress across both general robot skills and a specific production use case. The source does not provide further detail on how many sorting attempts made up the pilot result.
Apptronik CEO Jeff Cardenas said the models improved significantly within a few weeks of post-training and moved into Mercedes production with minimal performance loss. The account points to adaptation after training as part of the pilot, though it does not describe the post-training process in detail.
Safety is part of the evaluation
Google DeepMind says Gemini Robotics 2.5 uses real-time control to help robots avoid physical hazards while carrying out tasks. In the company’s evaluations, safe task completion improved by 33 percent near obstacles and by 43 percent near people.
The company also published ASIMOV-Jailbreak, an open-source benchmark on Hugging Face. It contains hundreds of prompts that disguise unsafe requests across dozens of scenes. The benchmark is designed to check whether robotics models can be manipulated into dangerous actions.
Google DeepMind reports that its Gemini 4 Argon model scored 99.5 percent for jailbreak resistance on the benchmark, compared with 82 percent for Gemini Robotics ER 2. Those are the company’s reported evaluation results; the article does not provide additional information about how the scores were measured.
Simulation helps create training examples
Gemini 4 Argon also served as a simulation engine during development of Gemini Robotics 2.5. Google DeepMind says Argon can use a real camera scene to generate thousands of simulated training runs and create a digital twin. These simulated examples are combined with real-world demonstrations to help the robot adapt to unfamiliar scenes and objects.
The same setup can be used to investigate mistakes. Argon can propose possible causes based on video recordings, then test potential improvements in simulation and on the physical robot. This connects training, error analysis, and real-world trials in the development process described by the company.
Google DeepMind says Gemini Robotics 2.5 is being rolled out to partners Agile, Apptronik, and Mercedes-Benz. The production pilot offers a concrete example of the model directing a robot through a factory task, while the broader claims cover speed, success across other tasks, and safety evaluations.