Google DeepMind Tests Gemini Robotics 2.5 on Mercedes-Benz Production Lines With 96 percent Success Rate

Google DeepMind announced on November 10th the launch of Gemini Robotics 2.5, the most capable vision-language-action model the company has developed to date. Apptronik’s Apollo 2 humanoid robot is being used by Google’s robotics partners to complete a variety of tasks, including automotive parts handling on active Mercedes-Benz production lines, with a 96 percent success rate. 

The Mercedes-Benz proof-of-concept for Gemini Robotics 2.5 took place at production facilities in Kecskemét, Hungary and Sindelfingen, Germany. In these environments, the Apptronik Apollo 2 humanoid robot performed automotive kitting tasks in which the robot would reach into supply bins, pick up metal brackets and place them in the correct sequencing containers, while accounting for any gaps. The robot’s hands and arms were controlled by the Gemini Robotics 2.5 vision-language-action model while DeepMind’s embodied reasoning model tracked the workflow of the robot. Apollo 2 uses the five-finger, 22-degree-of-freedom SharpaWave hand. 

After just a few weeks of post-training on the new models, Apptronik’s CEO Jeff Cardenas says that they have improved significantly and are being used on Mercedes-Benz’s production lines with very little drop-off. He says that the models could be used to do repetitive tasks such as preparation work for build, freeing up technicians to do more complex work. 

After testing the model on more than 200 tasks in total, the model’s performance was compared to that of previous baselines. On average, the success rate increased by 30 percent to 100 percent depending on the task type, with the tasks also completed on average 20 percent to 30 percent faster than before. 

In addition to improving the model’s performance on tasks, DeepMind also looked at the safety of the model’s actions. A new reasoning-guidance technique was developed to increase the model’s ability to complete tasks safely around obstacles and people. This technique increased the model’s ability to complete tasks safely around obstacles by 33 percent and around people by 43 percent. Additionally, DeepMind released a new benchmark, ASIMOV-Jailbreak, to measure the model’s ability to detect and resist jailbreak attempts. The Gemini 4 Argon model achieved a jailbreak resistance of 99.5 percent on the ASIMOV-Jailbreak test, a significant improvement over the previous baseline of 82 percent.