3D Grasping Pose Detection Method Based on Improved PointNet Network

Jiahui Chen, Yunhan Lin, Haotian Zhou, Huasong Min

Abstract: In this paper, a 3D grasping pose detection method based on improved PointNet network was proposed. We introduced a skip connection into the backbone layer of PointNet network and compared the collocation of the latest different activation functions and loss functions to solve the problem that the lightweight PointNet network did not consider the local geometric information of the point cloud. This method effectively ensured the geometric information of the original point cloud, realized feature reuse, and made the transfer of features and gradients more effective. We applied the improved PointNet to robotics grasping, and our proposed method could directly deal with the original point clouds obtained by sensors. Compared with similar algorithms GPD and PointNetGPD, experiment results showed that the proposed grasping pose detection method had improved the test accuracy on the data set and the actual success rate of grasping.