We present a cost-effective pretraining paradigm for VLA models using only synthetic data, achieving direct sim-to-real transfer and strong zero-shot generalizability for robotic grasping. Key ...