Google DeepMind launches Gemini Robotics 2.5 with 30–100% higher success

Gemini Robotics 2.5 builds on Google DeepMind's line of vision-language-action models, which translate visual input and language instructions into robot actions. The company reports that the new version was evaluated across more than 200 tasks, including handling delicate objects, using tools, manipulating flexible materials, and performing precision assembly operations. Across these tasks, success rates were 30 to 100 percent higher than previous versions, and operations were completed 20 to 30 percent faster, according to Google DeepMind.

Beyond raw task performance, Google DeepMind introduced reasoning-based safety controls that increased safe task completion by 33 percent around obstacles and 43 percent when people were nearby. It also released ASIMOV-Jailbreak, an open-source benchmark for evaluating whether robotics AI systems resist instructions that could lead to unsafe actions. In those evaluations, the Gemini 4 Argon model achieved 99.5 percent resistance, compared with 82 percent for the previous baseline.

Gemini 4 Argon also supports robotics research by generating simulated training scenarios and digital twins, and by helping researchers investigate task failures. Google DeepMind says these capabilities accelerated development of Gemini Robotics 2.5 by reducing the amount of physical training data required. The model is now being made available to the company's robotics partners as it continues developing AI systems intended to operate across different robot platforms and real-world applications.