Google's Gemini Robotics 2.5 Runs a Real Mercedes Production Line: Apptronik's Apollo 2 Sorted Car Parts With a 96% Success Rate

Google DeepMind announced Gemini Robotics 2.5 on Tuesday, calling it its most capable vision-language-action model outside the lab, and the demonstration that carries the most weight is not a benchmark chart: on active Mercedes-Benz production lines in Kecskemét, Hungary and Sindelfingen, Germany, an Apptronik Apollo 2 humanoid running the model picked metal brackets from bulk bins and sequenced them into kits for each vehicle, with a company-reported 96% success rate.

The pilot task is automotive kitting, a repetitive, physically demanding job with little tolerance for error: parts must come out of unordered bins and be arranged in the exact sequence each car on the line requires. DeepMind's Gemini Robotics ER model tracks the workflow while the Robotics 2.5 action model controls the arms and hands, correcting assembly gaps along the way. Getting a VLA model from lab demos onto a line with takt-time pressure is the step most robotics announcements skip, which is why automakers are watching this one closely.

On evaluations, Google reports that the model was tested across more than 200 activities spanning dexterity and tool use, including serving food with tongs, pipetting liquids and tying a bow, a canonical nightmare task because ribbon deforms unpredictably. Compared with previous baselines, the company claims 30% to 100% higher success rates depending on task type and 20% to 30% faster completion. These are all company-reported figures and have not been independently verified.

The model is deliberately hardware-agnostic. It runs natively on the Apollo 2 humanoid with its five-finger, 22-degree-of-freedom SharpaWave hand, and on the Franka Duo platform with simple two-finger parallel grippers, so the same software spans a wide range of embodiments.

Safety got its own release track. Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards, which Google says improves safe task completion around obstacles by 33% and around nearby people by 43%. The company also released ASIMOV-Jailbreak, an open benchmark of hundreds of disguised dangerous requests across dozens of scenarios, and reports that Gemini 4 Argon resists 99.5% of them versus an 82% baseline for Gemini Robotics ER 2.

Training efficiency leans on Gemini 4 Argon as a simulation engine: it generates thousands of diverse practice episodes from real camera scenes to build high-fidelity digital twins, and can reason about task failures from video clips before testing fixes in simulation and on real robots. DeepMind says the model is being rolled out progressively to robotics partners. The 4% gap between the pilot's success rate and production-grade reliability is the honest part of the story, and closing it is what will decide whether humanoids stay on automotive lines after the pilots end.