As the supporting technology of advanced manufacturing industry, industrial robots have been widely applied in various manufacturing fields, such as assembly, grinding and other important processes. But all the reliable implementations of these processes depend on the accurate and stable grasping. For this reason, vision systems are often introduced to improve the autonomous and intelligence ability of industrial robots. But the precision, speed and stability of most vision-based robot grasping systems can not meet the requirements of industrial applications. Therefore the studies on vision-based robot grasping are of important academic significance and practical value. Parts detection and reaching are two basic procedures in vision-based robot grasping tasks. The main difficulties and problems are presented as follows: In the aspect of detection, grasping tasks require the detection stage can provide accurate pose information of the objects. In the aspect of grasping strategy, most methods divide the process in which the visual information of objects poses is transformed to the grasping poses into multiple calibration processes. These methods have accumulative errors and the results are easy to be disturbed by the environment. In view of the above difficulties and problems, this article conducts a detailed study. The main contributions are as follows: 1. In order to meet the precision and speed requirements of grasping tasks, we propose an algorithm which can detect the parts fast and estimate the pose accurately based on “coarse-to-fine” detection strategy. At first, candidate orientations are detected and a large number of wrong orientations are excluded so as to increasing the detection speed. Then, the Chamfer Matching algorithm based on distance transform is used twice for searching the exact position of grasping point. At last, the imaging model of camera is used for correct the grasping orientations. The experiments show that our methods can meet the precision and speed requirements of grasping tasks. 2. In order to improve the autonomous and intelligence abilities of robot grasping and reduce the accumulative errors of several calibration processes, we propose a learning-based object reaching approach by using visual information. At first, the feature points which show the state of parts are extracted exactly by hierarchical structure clustering. The feature points are robust and invariant under translation, rotation. Secondly, the relations...
修改评论