Google DeepMind began rolling out Gemini Robotics 2.5 to its robotics partners on November 10. The model turns camera input and instructions into movements for robots such as Apptronik’s Apollo 2 humanoid and the Franka Duo. DeepMind describes a partner rollout and has given no public access date or price.
A parts-sorting trial on active production lines
DeepMind, Apptronik and Mercedes-Benz ran a proof of concept at Mercedes plants in Kecskemét, Hungary, and Sindelfingen, Germany. The humanoid picked metal brackets from supply bins and placed them in trays in the order needed on the line. Gemini Robotics ER kept track of the sequence, while the 2.5 model controlled the robot’s hands and arms. DeepMind gave a single success figure for the sorting pilot and did not break it down by site.
What changed in version 2.5
The earlier Gemini Robotics 2 brought whole-body control to multiple robot designs. DeepMind tested 2.5 across roughly 200 tasks, including handling delicate objects and using tongs. In those evaluations, DeepMind reported relative gains in success rate of 30% to 100%, depending on the task, and said the model completed tasks 20% to 30% faster than earlier baselines.
