SSDC-Grasp: A Joint Semantic Segmentation and Depth Completion Approach for Robotic Grasping of Transparent Objects

The optical characteristics of transparent objects, specifically reflection and refraction, induce texture distortion and geometric information loss in RGB-D imagery, presenting critical challenges for existing RGB-D-based grasping approaches. To address this, we propose SSDC-Grasp, a synergistic perception framework integrating semantic segmentation with depth completion, whose core innovation lies in dualmodality feature fusion for transparent object reconstruction. Firstly, semantic features including color and texture in central regions are reconstructed via a semantic segmentation network (SSNet). Secondly, we develop DCNet, a novel lightweight depth completion network employing an encoder-decoder decoupling architecture to repair 3D geometric information from local to global scales. SSDC-Grasp synergizes SSNet and DCNet to generate complete transparent object point clouds with enriched semantics. This comprehensive feature reconstruction strategy collectively enhances representation capabilities in transparent regions, thereby enabling more reliable grasp pose detection. Real-world robotic experiments demonstrate an average grasping success rate of 81.81%. The project is available at: https://github.com/meiguiz/SSDC-Grasp.