1 paper · 1 filter
Xuan Dong, Zhe Han, Tianhao Niu +2
Vision-Language-Action (VLA) models have shown strong performance in language-conditioned robotic manipulation, yet their robustness to linguistic variation remains poorly understo…