The integration of artificial intelligence and physical robotics has reached another milestone. Recently, Google officially released its most capable visual-language-action (VLA) robot model to date—Gemini Robotics 2.5. The model achieves comprehensive upgrades across the breadth of robot skills, learning speed, and three core areas: dexterity, reliability, and safety. It is currently being made fully available to Google's robotics partners.
Core performance leaps forward: task success rates soar, native support for multiple robot types
As a breakthrough of the new generation, Gemini Robotics 2.5 has completed rigorous testing across more than 200 different complex tasks.
Test data shows that, compared with the previous baseline model, the model's success rate across tasks increased substantially by 30% to 100%, while task completion speed also increased by 20% to 30%. More importantly, the model is highly versatile and can natively adapt to many different types of robot hardware platforms, demonstrating excellent cross-platform adaptability.
On-site validation on Mercedes-Benz production line: humanoid robot efficiently completes parts picking and packing
To test its real-world value in industrial environments, Gemini Robotics 2.5 has now conducted an actual proof of concept (PoC) on the production line of the well-known German automaker Mercedes-Benz.
In the complex automotive manufacturing workshop, a humanoid robot equipped with the model successfully took on the task of picking and packing automotive parts, achieving an actual operational success rate of 96%. This striking figure indicates that the model is not merely at the theoretical laboratory stage; it also has the practical potential to handle high-intensity, high-precision industrial-grade operations.
Comprehensive safety upgrade: equipped with physical hazard avoidance, 99.5% jailbreak resistance rate
In terms of safety and robustness, Gemini Robotics 2.5 has also delivered an excellent result.
The model is built with an advanced real-time physical hazard avoidance mechanism at its foundation, significantly improving the robot's completion rate when handling complex safety-related tasks. At the same time, to ensure the model's safety in open environments, Google has also open-sourced the jailbreak robustness benchmark ASIMOV-Jailbreak, which measured a jailbreak resistance rate as high as 99.5%.
It is understood that the rapid development and performance breakthroughs of Gemini Robotics 2.5 are largely thanks to strong acceleration from the new flagship large model Gemini 4 Argon in underlying computing power and reasoning capabilities, laying a solid foundation for the subsequent development of embodied intelligence.
