A Self-Supervised Learning Manipulator Grasping Approach Based on Instance Segmentation

Abstract

Automatic grasping is playing an important role in robotics, and traditional grasping approaches cannot deal with both occlusions and texture-less objects well. To improve the stability and accuracy of grasping in a novel and occlusion environment, a grasping approach based on instance segmentation and self-supervised learning pose estimation network is proposed to grasp objects with the manipulator in this paper. The approach can be divided into three phases: instance segmentation, pose estimation, and pose transformation. Instance segmentation predicts classification masks for each pixel. Masks which stand for the contour of objects provide a heuristic knowledge for the self-supervised learning pose estimation network. Pose estimation network has two fully connected layers and regards estimation as a self-supervised classification problem. Then, the pose can be transformed from value in pixel coordinates to actual value relative to the base coordinate of the manipulator. As a result, the manipulator can be operated to grasp by the given actual value of pose. With the help of the approach proposed in this paper, we improve the grasping accuracy by 35%, compared to the former grasping approach based on the pose estimation network on the grasping dataset of CMU. Besides, grasping with this approach on hardware also shows a high success rate. Therefore, the proposed approach is a more robust and more accurate way of grasping.