Active Semi-supervised Grasp Pose Detection with Geometric Consistency

摘要

Learning-based pose detection in robotic grasping has been widely studied because of its generalization ability to deal with unknown objects. However, collecting a large labeled dataset is of great difficulty. There are countless objects in our lives, and there may be a considerable number of grasp poses for each object. Thus, it is impossible to label all the grasp poses. In this paper, we propose an active semi-supervised grasp pose detection strategy, in which we use the feature of grasp geometric consistency for data selection and training. Our method can select the most valuable samples to annotate based on the geometric consistency. As far as we know, this is the first work that leverages active learning and semi-supervised learning to solve the problem of grasp data. We experimentally verify that our method, which uses 66% selected data, outperforms random selection, which uses all labeled data, and achieves the best performance compared with the baseline and other well-known methods.