Empirical modeling of uncertainty in vision systems for industrial robotic applications

Many industrial robotic applications concern a manipulator that pick up objects from a working area, being their position in space estimated through a vision system. Although many methods have been focused on the uncertainty related with them. Example of factors that can influence the uncertainty are calibration errors, image quality, particular setting of the algorithm used, particular physical features of the objects and environmental conditions. With such factors it is clear that try to find out a precise and deterministic model for the uncertainty can be a difficult task to perform. In this paper, after analyzing the factors involved, we present an experimental methodology that allows to build an empirical model of the uncertainty and to identify the factors that actually influence it. The methodology is based on two statistical tools: Design of Experiment and Process Modeling. Due to the many constraints on the experimental space, the high number of factors involved and their different nature, we direct our attention to the so called "optimal" designs. By comparing different models and criteria, we obtained a model which allowed us to reduce the error on the estimate of objects position.