Gemini Robotics 2.5 model achieved a 96% success rate in sorting parts in a pilot at Mercedes-Benz

Google DeepMind announced the results of a pilot deployment of the Gemini Robotics 2.5 model on active production lines at Mercedes-Benz facilities in Kecskemét (Hungary) and Sindelfingen (Germany). The humanoid robot Apollo 2 from Apptronik, performing a kitting task, sorted metal parts from crates and placed them into bins in the order needed for a specific vehicle on the line. According to Google DeepMind, this pilot achieved a 96% success rate, with the tasks running on a pair of models – Gemini Robotics ER monitored the progress of the work, while Gemini Robotics 2.5 controlled the robot's hands and arms. Apptronik CEO Jeff Cardenas said, according to the source, that the models improved significantly over several weeks of additional training and that the transition to production at Mercedes-Benz occurred with minimal loss of performance.

The Gemini Robotics 2.5 model is a vision-language-action model for controlling robots, building on an earlier version with full control of the robot's body. According to the company's own data, it was tested on more than 200 tasks, including serving food with tongs and pipetting liquids, and the company states that, compared to the previous version, it achieves a 30 to 100% higher success rate depending on the task type and performs tasks 20 to 30% faster. The model works with various types of robots, including the five-fingered humanoid robot Apollo 2 with 22 degrees of freedom and the Franka Duo platform with two-finger grippers. One of the demonstrated tasks was tying a bow, which, according to the source, shows improved manipulation of deformable objects.

To develop the model, the company used the Gemini 4 Argon language model as a simulation engine, which generates thousands of simulated practice runs from real video of a scene and creates a digital twin of the environment; combined with real training data, this is meant to allow robots to generalize more quickly to new scenes and objects. The company also published an open source safety benchmark called ASIMOV-Jailbreak on Hugging Face, which tests whether robotic models can be induced into dangerous actions through covertly formulated requests. According to the company, the Gemini 4 Argon model achieves 99.5% resistance to jailbreaking, compared to 82% for the previous version, Gemini Robotics ER 2; safe task completion, according to the company's own tests, improved by 33% near obstacles and by 43% near people.