Swin-MaskGNet: a robust robotic grasp pose detection model for densely occluded environments

Current robotic grasp poses detection methods in densely occluded environments primarily rely on holistic object information, often neglecting the importance of local feature analysis, which can lead to suboptimal detection accuracy. In response, we propose Swin-MaskGNet, an end-to-end grasp detection model specifically designed for such challenging conditions. This model utilizes the sliding window attention mechanism and the multi-scale feature extraction capabilities of the Swin Transformer, effectively integrating both local and global information. This integration enhances the identification of local grasp points while simultaneously maintaining a comprehensive understanding of global dense occlusions. Furthermore, we introduce a multi-task learning module that concurrently performs instance segmentation and grasp detection, significantly improving the model’s comprehension of object boundaries and occlusion relationships, thereby elevating overall grasp detection accuracy. In addition, we have integrated an attention mechanism into the model, aimed at further optimizing the representation of channel and spatial features, enabling the model to focus more effectively on critical grasping areas, thereby producing more precise grasping information. Experimental results demonstrate that the model achieves detection accuracies of 93.37% on the OCID-Grasp dataset, 97.93% on the Cornell Grasp dataset, and a validation accuracy of 95.31% on the Jacquard dataset. In real-world grasping experiments, Swin-MaskGNet attains a practical grasp success rate of 85.5%, underscoring its robust adaptability to densely occluded environments.