中国科学院自动化研究所机构知识库

Aquatic environment has been damaged by human activities for years. Trash on the water surface is the prominent problem, e.g., dumped plastic bottles are ubiquitous. To clean water surface by aquatic robots, grasp detection is crucial task in the process of collection. Owing to lack of labeled data for training and the limited computational capacity of aquatic robots, supervised grasp detection method is intractable. In this context, an unsupervised grasp detection strategy is proposed in this paper for bottle grasping. Firstly, the bottle is approximately abstracted as an elongated cylinder, then a grasping model is constructed, in which a graspable position is described by the central point, long axis direction angle, and the maximum width. Secondly, to extract location points of bottles on the water surface, a high speed background elimination method based on RGB histogram is utilized in the gridding bounding box. Next, principal component analysis (PCA) with an outliers removing algorithm is employed to estimate parameters of the graspable position. Finally, experimental results of grasp detection show that estimated position is significant to grasp control. Especially, the accuracy of long axis direction angle is up to 3◦, and computation speed of the background elimination method is 1.47 times the speed of Canny