Google DeepMind Releases Gemini Robotics 2.5, Achieving 96% Success Rate in Mercedes-Benz Pilot
Google DeepMind launches its latest Vision-Language-Action model, Gemini Robotics 2.5, which achieved a 96% success rate in sorting parts during a pilot at Mercedes-Benz production lines and introduces a new robot safety benchmark.
On November 10, 2026, Google DeepMind officially released Gemini Robotics 2.5, its most capable Vision-Language-Action (VLA) model to date. The model aims to enhance robot dexterity, reliability, and safety, enabling them to operate effectively outside the lab in more unstructured, real-world environments.
According to official data, Gemini Robotics 2.5 demonstrated significant improvements across over 200 distinct tasks. Compared to previous baselines, the new model achieved 30% to 100% higher success rates and completed tasks 20% to 30% faster. These tasks ranged from using everyday tools to handling delicate, flexible objects.
To validate reliability in active industrial settings, DeepMind partnered with Apptronik and Mercedes-Benz for a proof-of-concept on production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 humanoid robots equipped with Gemini Robotics 2.5 autonomously performed automotive kitting—sorting individual parts from bulk bins into organized trays. This pilot achieved a 96% success rate, suggesting robots could take on monotonous work, freeing technicians for complex build tasks.
In terms of safety, DeepMind introduced inference-time reasoning guidance to help robots navigate securely around obstacles and people, boosting safe task completion by 33% and 43%, respectively. Additionally, the company released ASIMOV-Jailbreak, a new open-source benchmark testing adversarial jailbreaks specific to robotics. In this test, Gemini 4 Argon achieved a 99.5% jailbreak resistance, a significant improvement over the 82% rate of the previous baseline.
It is important to note that these performance metrics are based on internal evaluations by Google DeepMind and specific pilot scenarios, and have not yet been widely reproduced by independent third parties. Gemini Robotics 2.5 is currently rolling out to partners.