Improving Robustness of Vision-Language-Action Models by Restoring Corrupted Visual Inputs