Google has launched Gemini Robotics 2.5, its most capable vision-language-action robot model to date, achieving breakthroughs in three key areas: dexterity, reliability, and safety—significantly expanding the breadth of robot skills and accelerating learning speed. The model has been tested across more than 200 different tasks, achieving a task success rate 30% to 100% higher than the previous-generation baseline and completion speeds 20% to 30% faster, while natively adapting to multiple robot types. In a real-world proof of concept on the Mercedes production line, the model helped a humanoid robot complete the task of picking and packing automotive parts, achieving a 96% success rate. The model features a real-time physical hazard avoidance mechanism that significantly improves the completion rate of related safety tasks, and is accompanied by the open-source jailbreak robustness test benchmark ASIMOV-Jailbreak, with a jailbreak resistance rate of 99.5%. This R&D progress was accelerated by Gemini 4 Argon, and Gemini Robotics 2.5 is currently being made available to Google's robotics partners.