A stereo disparity algorithm is implemented on an industrial robot vision system. One camera manipulated by an industrial robot is used to obtain stereo images of an object. After a pair of stereo images are processed, the points of interest are selected from each image. The similarities of three types of feature characters are used to assign initial weights of the, matching probabilities to a set of point candidates. Then a relaxation method, which is based on probabilities of local connectivity, similarity of feature characters, and smoothness of matching, is used iteratively to improve the initial probabilities of image matching. The positions of object points are determined by image disparities based on a triangulation method. Experiments performed on an AdeptOne Robot with based vision system show that the algorithm works well.