Gasgoo Munich- On November 12, Google DeepMind unveiled its next-generation vision-language-action (VLA) model, Gemini Robotics 2.5, designed to enhance dexterity, reliability, and safety in complex environments. The model is now being deployed to robotics partners.
Image source: Google DeepMind
According to Google DeepMind, Gemini Robotics 2.5 has completed testing across more than 200 distinct tasks, spanning tool use, flexible object manipulation, and precision operations. Success rates for these tasks increased between 30% and 100% compared to previous models, while execution speeds improved by 20% to 30%.
Automotive manufacturing emerged as a key proving ground for the model validation.
Collaborating with Apptronik and Mercedes-Benz, Google DeepMind deployed Apollo 2 humanoid robots on Mercedes production lines in Kecskemét, Hungary, and Sindelfingen, Germany, to conduct parts sorting and kitting tests.
In these trials, the robot handled workflows using the Gemini Robotics embodied reasoning model, while Gemini Robotics 2.5 controlled its arms and dexterous hands to grasp, sort, and place loose automotive metal brackets into designated containers — achieving a 96% success rate. The project is currently in the proof-of-concept phase.
On the technical front, Gemini Robotics 2.5 is adaptable to various robot bodies, including Apollo 2, equipped with a five-finger, 22-degree-of-freedom dexterous hand, and the Franka Duo platform, which uses a two-finger gripper.
Furthermore, the new model improves safety during interactions with obstacles and personnel through guided reasoning. Google also introduced the Gemini 4 Argon model to support robotics development, generating simulation training data from real-world scenarios and analyzing task failures to accelerate iteration.
Notably, this upgrade extends beyond motion execution capabilities, further reinforcing the role of general large models in robot training, simulation, and task reasoning.
