Google DeepMind released an open-source benchmark for robotics jailbreak resistance Nov. 10, while reporting model results and a factory pilot involving its Gemini Robotics 2.5 system.
Called ASIMOV-Jailbreak, the benchmark focuses on whether robotics systems resist adversarial instructions that disguise unsafe requests as routine tasks. Google DeepMind described it as a test of jailbreak resistance in robotics, rather than a broad measure of robot safety.
Separately, the company reported 99.5% jailbreak resistance for Gemini 4 Argon and listed Gemini Robotics ER 2 as a previous comparison at 82%. Google DeepMind described Argon as a simulation engine that helped accelerate development of Gemini Robotics 2.5; it did not identify the 99.5% figure as a score on ASIMOV-Jailbreak.
The company also detailed a proof of concept with Mercedes-Benz using Apptronik’s Apollo 2 humanoid robot. Gemini Robotics ER tracked the workflow, while Gemini Robotics 2.5 controlled the robot’s hands and arms. Google DeepMind reported a 96% success rate for pick-and-pack operations.
Google cited unpredictable real-world conditions and the need for substantial physical data as challenges in developing robotics systems. It said efficient modeling and simulation can help address those demands. The company said automating repetitive kitting work could give technicians more time to focus on complex build tasks.
