Google DeepMind
Robotics · Artificial Intelligence · Manufacturing
Gemini Robotics 2.5 hits 96% success in Mercedes-Benz factory pilot
November 10, 2026
DeepMind's robot AI jumped from lab demos to live assembly lines, a gap most humanoid robotics efforts haven't crossed.
- Google DeepMind launched Gemini Robotics 2.5, a vision-language-action model, on November 10, 2026, calling it its most capable robot control model to date.
- The model turns camera input and language instructions into robot arm and hand movements, and is designed to work across robot bodies including Apptronik's Apollo 2 humanoid and the Franka Duo two-armed platform.
- In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 humanoids used the model to sort metal automotive brackets into trays on live production lines in Kecskemét, Hungary, and Sindelfingen, Germany, hitting a 96% success rate.
- Across more than 200 lab-tested tasks, DeepMind reports 30% to 100% higher success rates than its prior baseline, depending on task type, and 20% to 30% faster task completion.
- DeepMind also released ASIMOV-Jailbreak, an open safety benchmark testing robotics models against adversarial attempts to bypass physical safety behavior.
- DeepMind has not disclosed the trial count, run duration, failure definition, or human intervention rate behind the 96% figure, leaving the pilot's real-world rigor unverified.
- Factory pilots are the toughest proving ground for humanoid robots because unstructured, safety-critical environments expose brittleness that lab demos don't, making this a meaningful signal for industrial automation even without independent verification.