Chen Siang ; Tang Wei ; Xie Pengwei ; Yang Wenming ; Wang Guijin

Abstract

Fast and robust object grasping in clutter is a crucial component ofrobotics. Most current works resort to the whole observed point cloud for 6-Dofgrasp generation, ignoring the guidance information excavated from globalsemantics, thus limiting high-quality grasp generation and real-timeperformance. In this work, we show that the widely used heatmaps areunderestimated in the efficiency of 6-Dof grasp generation. Therefore, wepropose an effective local grasp generator combined with grasp heatmaps asguidance, which infers in a global-to-local semantic-to-point way.Specifically, Gaussian encoding and the grid-based strategy are applied topredict grasp heatmaps as guidance to aggregate local points into graspableregions and provide global semantic information. Further, a novel non-uniformanchor sampling mechanism is designed to improve grasp accuracy and diversity.Benefiting from the high-efficiency encoding in the image space and focusing onpoints in local graspable regions, our framework can perform high-quality graspdetection in real-time and achieve state-of-the-art results. In addition, realrobot experiments demonstrate the effectiveness of our method with a successrate of 94% and a clutter completion rate of 100%. Our code is available athttps://github.com/THU-VCLab/HGGD.