新智元 (微信公众号)T2·Agents
74 pts
A multi-institution team from XJTLU, Duke Kunshan and NUS introduces RoboRisk, a robotic safety benchmark, plus LEASH, a safety guardrail. Six frontier models including GPT, Claude, Gemini and Grok were tested driving robot arms directly and all failed, producing hazards such as holding a knife over a human hand or dropping a phone into water. With LEASH, four models achieved zero harm in real-robot tasks, showing that conventional collision-avoidance safety cannot cover semantic-level danger.