Google DeepMind releases Gemini Robotics 2.5 and open ASIMOV-Jailbreak safety benchmark
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Language-model-driven robots completed real factory trials with 96% success, and alongside growing capabilities comes a common way to measure how hard they are to steer toward dangerous actions.
DeepMind has released a new version of its robot control system, Gemini Robotics 2.5. According to the company, task success rates rose 30–100% versus the previous version and speed by 20–30%, while the model is noticeably better at small, precise movements: the Apptronik Apollo 2 humanoid ties a bow, and the model handles serving food with tongs and working with a pipette. In trials outside the lab, Apollo 2 robots at Mercedes-Benz plants in Europe assembled auto parts with 96% success — which is still not enough for industrial use, DeepMind notes. The simulation environment for training was generated with Gemini 4 Argon, but the model itself runs on cloud servers, meaning there is latency. For safety testing, an open set called ASIMOV-Jailbreak was released: 1,500 prompts that provoke a robot into unsafe actions; with Argon in the loop, the system blocks 98.8% of attacks. The model is available to partners — Agile Robots, Apptronik and Boston Dynamics.
Sources 1