Context Identification for Efficient Multiple-Model State Estimation of Systems With Cyclical Intermittent Dynamics

Abstract

This paper presents an approach to accurate and scalable multiple-model (MM) state estimation for hybrid systems with intermittent, cyclical, multimodal dynamics. The approach consists of using discrete-state estimation to identify a system's dynamical and behavioral contexts and determine which motion models appropriately represent current dynamics and which individual and MM filters are appropriate for state estimation. Furthermore, the heirarchical structure of the dynamics is explicitly encoded, which enables detection not only of rapid transitions between motion models but of higher level behavioral transitions as well. This improves the accuracy and scalability of conventional MM state estimation, which is demonstrated experimentally on a mobile robot that exhibits fast-switching, multimodal dynamics.