Labeled synthetic data generated by physics engines are commonly combined with real data for grasp point prediction in robotic clothes manipulation tasks. However, domain adaptation from synthetic to real data remains challenging due to two major sources of uncertainty: (1) visual discrepancies between synthetic and real images, and (2) inconsistencies in label scales or units. To address these issues, we propose an Uncertainty Calibration Network (UCN), which employs fine-grained feature learning and difference calibration to reduce the domain gap in image appearance. To handle label uncertainty, UCN is trained in an unsupervised manner without using synthetic labels, enabling synthetic images to enrich feature diversity and improve domain-shared representations. This approach significantly enhances the model’s generalization ability across domains, achieving robotic grasping and unfolding success rates of 84% and 75% in the real world, respectively.