DeMaVLA: A Vision-Language-Action Foundation Model for Generalizable Deformable Manipulation