Google Releases Gemini Robotics 2.5, a Visual-Language-Action Model with Enhanced Capabilities for Industrial Applications

The fusion of artificial intelligence and physical robot technology has once again reached a milestone. Recently, Google officially released the most capable strongest Vision-Language-Action (VLA) robot model to date—Gemini Robotics 2.5. This model has achieved a comprehensive leap in three key areas: the breadth of robot skills, learning speed, and core dexterity, reliability, and safety. It is currently being fully opened to Google's robot partners.

Core performance has significantly leaped: task success rates have soared, and it natively adapts to multiple types of robots.

As a breakthrough achievement of the new generation, Gemini Robotics2.5 has completed rigorous testing of over 200 different complex tasks.

Test data shows that, compared to the previous generation baseline model, the success rate of this model in various tasks increased significantly by 30% to 100%, and the completion speed of executing tasks also accelerated by 20% to 30%. More importantly, this model possesses extremely strong versatility, capable of natively adapting to multiple different types of robot hardware carriers, demonstrating excellent cross-platform adaptability.

Mercedes production line field verification: Humanoid robots efficiently complete component picking and packing

To test its practical value in real industrial environments, Gemini Robotics2.5 has currently conducted a practical Proof of Concept (PoC) on the production line of the well-known German automaker Mercedes.

In a complex automobile manufacturing workshop site, the humanoid robot equipped with this model successfully undertook the tasks of picking and packing automobile parts, with an actual operation success rate as high as 96%. This striking data indicates that the model does not only remain at the theoretical stage in the laboratory but also possesses practical potential to handle industrial-grade high-intensity, high-precision operations.

Safety mechanisms fully upgraded: equipped with physical hazard avoidance, jailbreak resistance rate reaches 99.5%

In terms of safety and robustness, Gemini Robotics 2.5 also delivered an outstanding performance.

The model is equipped with an advanced real-time physical hazard avoidance mechanism at its base, which significantly improved the completion rate when the robot handles complex tasks involving safety. At the same time, to ensure the safety of the model in open environments, Google also open-sourced the jailbreak robustness testing benchmark ASIMOV-Jailbreak, whose measured jailbreak resistance rate reaches as high as 99.5%.

According to information, the rapid development and performance breakthroughs of Gemini Robotics 2.5 are largely attributed to the powerful acceleration in underlying computing power and reasoning capabilities of the new flagship large model Gemini 4Argon, which also laid a solid foundation for the subsequent development of embodied intelligence.