1 paper
Heejae Suh, Jongwook Han, Zahra Gholami +1
Unreliable visual inputs can harm task performance and cause potential physical safety risks for vision-language-action (VLA) models. We analyze how π0.5 and GR00T models act und…