Google DeepMind Releases Gemini Robotics 2.5 for Industrial Collaboration

Compile/VR Gyroscope

On November 10, Google DeepMind officially released the next-generation vision-language-action model Gemini Robotics 2.5, improving robots' dexterous manipulation, task execution, and safety control capabilities in complex environments.

According to Google, Gemini Robotics 2.5 has been tested in over 200 tasks, covering delicate operations such as using grippers to pick up food, tying gift bows, and liquid transferring. Compared to the previous generation model, its success rate in different tasks increased by 30% to 100%, and task completion speed increased by 20% to 30%, especially strengthening the ability to handle flexible objects, multi-step operations, and complex tool use.

In terms of industrial applications, Google collaborated with Apptronik and Mercedes to conduct concept validation on automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The Apollo 2 humanoid robot equipped with Gemini Robotics 2.5 can autonomously grasp metal parts and complete sorting, ordering, and boxing according to production requirements.

Google disclosed that the success rate of robots in the sorting tasks of the aforementioned components reached 96%. Among them, Gemini Robotics ER is responsible for understanding tasks and multi-step planning, while Gemini Robotics 2.5 controls the robotic arms and dexterous hands to execute actions and make adjustments based on on-site changes.

At the same time, Google also strengthened the physical safety capabilities of the models. Official tests show that the new model's performance in completing safety tasks near obstacles and personnel has improved, and the open-source evaluation benchmark ASIMOV-Jailbreak for robot safety instruction attacks was launched. Additionally, Gemini 4 Argon is also used to generate simulated training scenarios to assist robots in learning new skills.