Robot Grasping Detection Method Based on Keypoints

Abstract

This study introduces a novel keypoint-based grasp detection network, denoted as GKSCConv-Net, which operates on n-channel input images. The network architecture comprises three SCConv2D layers and three SCConvT2D layers. The SCConvT2D layers facilitate upsampling to maintain consistent dimensions between the output and input images. The resultant output consists of maps indicating left grasp points, right grasp points, and grasp center keypoints. The accuracy of predictions is enhanced through the incorporation of the keypoint refinement module and feature fusion module. To validate the model's generalization and applicability, comprehensive training, testing, and evaluation are conducted on diverse data sets, including the Cornell data set, Jacquard data set, and others representing real-world scenarios. Furthermore, ablation experiments are employed to substantiate the efficacy of the spatial reconstruction unit (SRU) and channel reconstruction unit (CRU) within the SCConv, exploring their impact on grasp keypoint detection outcomes. Real robotic grasping experiments ultimately affirm the model's outstanding performance in practical settings.