Google DeepMind unveils Gemini Robotics 2.5; 96% success rate picking auto parts

Google DeepMind on November 9, 2026, unveiled Gemini Robotics 2.5, a vision-language-action model. In a proof of concept on a Mercedes-Benz production line, Apptronik’s Apollo 2 humanoid robot reportedly achieved a 96% success rate picking and packing auto parts.

According to Google DeepMind, Apptronik’s Apollo 2 picked and packed auto parts on a Mercedes-Benz production line and recorded a 96% success rate in a proof of concept. When integrating robots into manufacturing processes, the results from combining an actual robot with a specific task can be assessed alongside model task evaluations. However, the figure should be understood as applying only to the task in the proof of concept.

What stands out in this announcement is that it includes both task evaluations of the model and a proof of concept on a factory line. The company’s statement that it evaluated capabilities across more than 200 different tasks indicates the scope of its assessment of generality. By contrast, the 96% figure on the Mercedes-Benz production line is the result of a proof of concept involving a specific task performed by Apptronik’s Apollo 2. The former should not be recast as evidence of the latter’s operational track record, nor should the 96% be interpreted as the automation rate for an entire factory. The available information does not specify the denominator for the success rate, trial conditions or duration of the task. In a related article, we reported on November 9, 2026, that humanoid.guide published a report on the use of robot models, including a report on the use of Gemini Robotics 2.5 on an auto production line. This announcement includes the model provider’s account and figures from a proof of concept using a robot, showing a link between model capability assessments and real-world tasks. However, there is no basis for treating the two pieces of information as independent verification under the same task conditions. Assessing use in manufacturing requires looking not only at the model’s skills but also at the robot’s movements, the shape and placement of parts, and procedures for human intervention when a task fails. The demonstration described here involved Apollo 2 handling parts; it does not substantiate transferability to other robots or long-term operation on a mass-production line. For manufacturers, including Mercedes-Benz, the focus of evaluation will include the limited scope of the task, safety checks and the division of roles between people and robots, not just the success-rate figure. The company also cited faster skill acquisition as an area of improvement, but the information provided in the announcement includes no figures for the time required to learn, the amount of training data or the maintenance of performance after updates. Key details to verify include the conditions used to calculate the 96% figure, the number of trials, recovery procedures after failures, and the range of parts and processes covered by the demonstration. Without this information, productivity and deployment costs cannot be assessed from the figure. See the original source for details.