ControlVLA: Few-shot Object-centric Adaptation for Pre-trained Vision-Language-Action Models