SCDA: A Semi-supervised Adaptation Framework with Multi-modal Normalization for Cross-Domain Grasp Detection

In cross-domain grasp detection, distributional variations across datasets often lead to severe performance degradation when models trained on fully labeled data are applied to new domains. While existing domain adaptation methods partially alleviate this problem, they still suffer from adaptation instability. To address this, we propose a Semi-supervised Cross-domain grasp Detection Adaptation (SCDA) framework that incorporates a small amount of labeled target data into a teacher-student model to enhance cross-domain adaptability. The student model is jointly supervised on source and target domains, while the teacher generates pseudo-labels from unlabeled target data to guide learning. Furthermore, we propose the Multi-modal Rectified Batch Normalization with Channel Attention (CA-MRBN) module, integrated into SCDA, which leverages depth information and channel attention to enhance feature representation, stabilize cross-domain distributions, and mitigate domain shift. Extensive experiments demonstrate that SCDA consistently outperforms existing methods, achieving superior adaptability in cross-domain grasp detection tasks.