Neural network-based pose estimation for fixtureless assembly

Abstract

A prototype fixtureless robotic assembly workcell will require a machine vision system to locate randomly fed parts without the use of models or camera calibration. The Feature CMAC artificial neural network has been shown to solve the 3-DOF pose estimation problem for simple target parts. In this paper, the network is extended to handle an unmodified industrial target part. A tradeoff between neural network accuracy and generalization results from the number and quality of features extracted from the image. As a result, the accuracy of Feature CMAC pose estimation is dependent on the choice of feature detection algorithm. Three such algorithms were evaluated to minimize pose estimation error. RMS errors were found to be less than 0.13 of the training interval (1.0 mm in position, and 1.2/spl deg/ in orientation), with an average worst-case grasp point error of 2.8 mm. A discussion of optical-axis bias and orientation loss is included.