Grasp Pose Detection Based on Point Cloud Shape Simplification

Abstract

In the robot grasping environment, grasping unknown objects that have neither model data nor RGB data is very important, we present a new approach for grasping unknown objects. Firstly, depth camera is used to obtain the partial 3D data of target object. Then, we perform 3D segmentation and the segmented parts are simplified to a cylinder, a sphere, an ellipsoid or a parallelepiped based on the geometric and semantic shape characteristic. Grasping constraints are used to obtain grasp candidates based on the simplified shape. The grasp descriptions are used to describe each candidates and CNN is used for grasping training. We evaluate our grasp networks both in simulation and robotic experiment, and the experiment result shows that the grasp quality score using simplified constraints is more robust than without simplified constraint when the gripper posture is uncertain.