Robotic Arm Visual Grasping Enhanced by an Improved U-Net Network

This paper proposes a robotic arm vision-based grasping method utilizing an enhanced U-Net architecture and an Efficient Channel Attention (ECA) mechanism, aimed at addressing the localization challenges of soft belts in industrial automation grasping. Traditional approaches suffer from grasping point localization errors when handling non-rigid objects due to deformation and boundary ambiguity issues. This study enhances U-Net's feature extraction capabilities by embedding an ECA module, combined with image moment theory and edge analysis, achieving precise extraction of target geometric features and optimized grasping point selection. Experimental results demonstrate that the improved U-Net_ECA model achieves mIoU scores of 93.59% and 88.47% on the training and validation sets, respectively, significantly outperforming conventional approaches. Furthermore, the proposed grasp point localization algorithm effectively resolves positioning errors caused by object deformation, validating its practicality and robustness in industrial settings.