Event

StepTrigger: Foot-ground contact triggers robot backdoor

1 reports · 1 sources

Jiageng Zhang, Doniyorkhon Obidov, and Kaichen Yang propose a backdoor attack named StepTrigger, targeting legged robots driven by vision-language models (VLMs). Its trigger signal is not a prompt or a visible marker, but rather the foot-ground contact pressure patterns produced when a Unitree Go1 quadruped robot walks over dense terrain. In a hierarchical offline evaluation, the authors report that the trained planner maintained 98.75% normal behavior, a false-trigger rejection rate of 92.50%, a true-trigger activation rate of 76.25%, and an overall parsed behavior accuracy of 89.17%. These numbers come from offline testing, not from real-robot validation results.