Inference of 6-DOF Robot Grasps using Point Cloud Data

Abstract

Recent decade, grasping pose perception for robot grasping field takes an important role in developing intelligent robotic applications. In this paper, we present a novel pipeline to infer grasp configuration for 3-finger gripper from point cloud data of a working environment. Most of the previous works are on 2D or 2.5D inputs to perform the 2D grasps. However, the 2D grasp configuration is not really compatible with the physical robot structure, the grasp successful rate on a robot system is only around 73% to 80%. To improve the robustness and computation, our goal is to propose an approach that can generate reliable grasping configurations in the cluster space. To accomplish this, we adopt a point cloud deep network to learn the shapes of objects and recognize their surfaces from the partial view. Afterwards, the grasp candidates are inferred based on the object models and reference grasping configuration sets. In our experiment, the proposed point cloud network had a better successful grasping rate compared to previous approaches in the context of varied and imbalanced data. And the grasping inference was highly viable in the complicated working environment.