Improving task reliability by fusion of redundant homogeneous modules using voting schemes

In this paper we present a technique that can help us design system components (behaviors) with reliability in mind. Our approach is similar to N-version programming. A team of functionally redundant behaviors vote for a set of possible actions. A voter then fuses the votes and selects the action which has received the maximum number of votes. We draw some design guidelines and show that with careful design, the selected action will have higher probability of success (reliability) than the reliability of each of the behaviors. A comprehensive set of experiments together with results are presented in which a team of tracking behaviors control a stereo camera head in the active vision task of smooth pursuit. The results show that the approach dramatically improves tracking performance