Methods for Real - Time

Methods for Real-Time Bin-Picking Using 2D Vision Computer vision has proved to be a valuable asset in robot automated applications. One such application is known as bin-picking, in which a robot is used to automatically pick up a number of items from a container. In some cases these “bins” contain overlapping items or even unsorted heaps. In this paper two methods are presented that use only 2D vision information from a single camera to establish the optimal pick order of such a heap. The work aims to investigate the possibilities to complete the functionality of the ABB PickMaster with bin-picking capabilities. Therefore Cognex’s CVL library has been used for object recognition, just like in the PickMaster. The goal is to find a method that can be trusted to sort out occasional occlusions. The result is a full description of two alternative procedures. They lack some functionality but can easily be extended to handle a more general group of objects. In the current edition, only scenes containing relatively flat objects without holes can be analyzed. Emphasis has been put on determining the order in which to pick the items, because this is considered to be the most difficult problem. Calculation of the z-coordinate and attitude angle of the pick points is considered secondary. In a more general context the bin-picking algorithms can be used together with other pattern recognition techniques based on point-to-point-, feature-to-featureor feature-to-pointmatching. It has not been of primary interest to optimize the computational work at this early stage; however some efforts have been made. The first algorithm processes most scenes in under 20 ms, whereas the second algorithm normally requires between 100 and 500 ms, in addition to the time it takes to perform the model matching.