PointNet++GPD:6-DOF grasping pose detection method based on object point cloud

To improve the accuracy and reliability of the grasp detection model for generating grasp poses, this paper proposes an end-to-end grasp pose detection network based on PointNet++ to address the shortcoming that PointNetGPD is challenging to learn local features in point clouds. Firstly, a dataset with a grasping quality score label is generated based on the BigBIRD dataset for model training. Then a grasping pose evaluation network based on PointN et++ is designed to classify the candidate grasping poses, to obtain the optimal grasping pose. Experimental results show that the classification accuracy of the grasping pose of the proposed model is higher than that of PointNetGPD and GPD, and the success rate of grasping objects in the real environment is higher than that of the above two models.