An efficient pose classification method for robotic grasping

Abstract

Background: The unstructured environment, the different geometric shapes of objects, and the uncertainty of sensor noise have brought many challenges to robotic grasping. PointNetGPD (Grasp Pose Detection) which was published in 2019 proposes a point cloud-based grasping pose detection method, which detects reliable grasping poses from the point cloud, and provides an effective process to generate and evaluate grasping poses. However, PointNetGPD uses the point cloud inside the parallel-gripper and the network only uses three channels of information when classifying grasping poses. Methods: In order to improve the accuracy of grasping pose classification, the concept of grasping confidence region was proposed in this paper, which shows the hotspot area of the object can be grasped successfully, and there will be higher success rate when performing grasping in this area. Based on the concept of grasping confidence regions, the grasping dataset in PointNetGPD is improved, which can provide richer information to the classification network. Using our dataset, we trained a scoring network that can score the point cloud collected by the depth camera. We added this scoring network to the classification network of PointNetGPD, and carried out the experiment of grasping poses classification. Results: The experimental results show that the classification accuracy increases by 4% after calculating the score channel on the original dataset; the classification accuracy increases by nearly 1% after using the trained scoring network to score the original dataset. Conclusions: The concept of positive grasp center area is proposed in this paper. Based on this concept, we improve the dataset in PointNetGPD, and use this dataset to train a scoring network to add the score information to the point cloud. The experiments show that our proposed method can effectively improve the accuracy of grasping poses classification network.