DeepMind Unveils Gemini Robotics 2.5 and Open-Sources Safety Benchmark

Google DeepMind has released Gemini Robotics 2.5, its most capable vision-language-action (VLA) model to date, delivering significant gains in dexterity, reliability, and safety. The model has been validated in a real-world proof-of-concept on Mercedes-Benz production lines in Hungary and Germany, where it drove Apptronik's Apollo 2 humanoid robot to pick and pack metal parts with a 96% success rate—a 30% to 100% improvement over the previous baseline and 20% to 30% faster task completion.

Technically, Gemini Robotics 2.5 covers over 200 tasks, including tying bows, pipetting, and manipulating flexible objects. DeepMind also used Gemini 4 Argon as a simulation engine for training, enabling inference of error causes from video and validation in both simulation and real-world settings. Additionally, the team open-sourced the ASIMOV-Jailbreak safety benchmark to standardize safety evaluation for human-robot collaboration. The model is now available to partners including Agile, Apptronik, and Mercedes-Benz.

Industry-wide, this marks the first time a humanoid robot has demonstrated near-practical success rates on a real automotive production line, a key step for VLA models moving from lab to unstructured environments. The open safety benchmark is expected to drive unified safety standards and accelerate the deployment of humanoid robots in manufacturing and logistics, with profound implications for the competitive landscape.