Multiple universities propose the robot safety benchmark RoboRisk and the guardrail LEASH; harm rates across six models range from 9.06% to 76.92%
GPT控制机械臂太危险!把菜刀悬在人手上,手机丢进水里
Xi'an Jiaotong-Liverpool University, Duke Kunshan University, National University of Singapore and other universities jointly propose the robot safety benchmark RoboRisk and the model-agnostic safety guardrail LEASH; six frontier large models have harm rates of 9.06% to 76.92% in robotic arm tasks, and after adding LEASH, four models have zero harm across 13 real-robot tasks, with safety success rates all at 84.62%. RoboRisk contains 92 simulated tasks and 13 real-robot tasks, covering five categories of risk, with metrics of harm rate HR and safety success rate SSR; each task is executed in a closed loop in simulation or on a real robot, and the model sees camera images and issues real actions. LEASH does not modify the model or retrain it; it consists of three safety gates—task gate, action gate and submission gate—and a shared safety policy, each review is based on the current camera image, and it uses the faster-responding decision model Jev, which does not generate free text, for scoring.
The paper reports the harm rates of six frontier models in robotic arm tasks, and shows how a guardrail that does not modify the model or retrain it reduces real-machine harm to zero for four models.