LEASH and RoboRisk target unsafe foundation-model robot control
The paper introduces RoboRisk, a closed-loop safety benchmark with 92 simulated and 13 physical-robot tasks spanning task refusal, plan repair, execution hazards, injection and sequential memory. Tests of six foundation models found harm rates from 9.06% to 76.92%, while LEASH added task, action and commit gates that reduced harm across all six models while largely preserving task completion.
The work provides a concrete benchmark and model-independent safety layer for robots controlled by foundation models, with the benchmark, simulation results and implementation reported as open source.
Computer Science > Robotics
arXiv:2611.14115 (cs)
[Submitted on 10 Nov 2026]
Title:LEASH: Stay Safe in Agentic Robotics
Authors:Qinfeng Zhu, Weiguang Zhao, Ke Sun, Xihao Guo, Yunxi Jiang, Rui Zhang, Lei Fan, Kaizhu Huang
View a PDF of the paper titled LEASH: Stay Safe in Agentic Robotics, by Qinfeng Zhu and 7 other authors
#LEASH#RoboRisk#GPT-6 Astra#GPT-6.1 Sol#GPT-6 Luna#Claude Opus 5.5#Gemini 4 Argon#Grok 4.7#Dream API#PokeBot RoboAPI#Jev#OpenAI Decisions API