Zero-Shot Sim-to-Real VLA Learning for Mobile Manipulation