Clarity for AI buildersSponsored by AI Routing
adoption

Gemini Robotics 2.5 reportedly reaches 96% success in Mercedes-Benz parts-sorting pilot

Gemini Robotics 2.5 reportedly reaches 96% success in Mercedes-Benz parts-sorting pilot
Google DeepMind’s Gemini Robotics 2.5 reached 96% success in a Mercedes-Benz pilot, according to The Decoder. The report describes a test with Apptronik on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apptronik’s Apollo 2 humanoid robot performed kitting: picking metal parts from bulk bins and placing them into trays in the sequence needed for each vehicle. Two models shared the work, with Gemini Robotics ER monitoring the workflow and Gemini Robotics 2.5 controlling the robot’s hands and arms. The result concerns a specific assembly-support task, not autonomous vehicle manufacturing as a whole. The supplied report does not specify the number of attempts, the duration of the pilot or how success was defined. Those omissions limit what the reported 96% rate establishes about sustained production performance or the frequency of human intervention. According to The Decoder’s account of Google DeepMind’s announcement, Gemini Robotics 2.5 is a vision-language-action model that turns visual input and language instructions into robot movements. Google DeepMind says it evaluated the model on more than 200 tasks, including serving food with tongs and pipetting liquids. Against its own previous baselines, the company reports success rates 30% to 100% higher, depending on task type, and task completion 20% to 30% faster. Those broader evaluation figures are separate from the Mercedes-Benz pilot result. They describe company-reported comparisons across tasks, rather than an independently established improvement on the production lines. The model is designed for different hardware, including Apollo 2, whose five-fingered hand has 22 degrees of freedom, and the Franka Duo platform with two-finger parallel grippers. Supporting both configurations would give the model a wider hardware scope, but does not establish identical performance across them. The report also describes a training approach using Google DeepMind’s Gemini 4 Argon language model as a simulation engine. Starting from a real camera scene, Argon can reportedly create a digital twin and generate thousands of simulated training runs. Combined with real demonstrations, this is intended to improve generalization to unfamiliar scenes and objects. Argon can also propose explanations for errors seen in video and test potential improvements in simulation and on a physical robot. These are described as training and debugging capabilities, not separate measurements of factory reliability. On safety, Google DeepMind reports improvements in safe task completion of 33% near obstacles and 43% near people in its own evaluations. It also published ASIMOV-Jailbreak, an open-source benchmark containing hundreds of prompts that disguise unsafe requests across dozens of scenes. The reported jailbreak-resistance score is 99.5% for Gemini 4 Argon, compared with 82% for Gemini Robotics ER 2. That benchmark result addresses resistance to unsafe prompts; it should not be read as the robot’s physical safety rate. According to The Decoder, Google DeepMind is rolling Gemini Robotics 2.5 out to partners Agile, Apptronik and Mercedes-Benz. The supplied evidence describes a partner rollout and a production-line pilot, rather than general public availability or factory-wide deployment.