Context-based state estimation for hybrid systems with intermittent dynamics

Robust state estimation is a key enabling technology for reactive control of hybrid systems, such as high performance mobile robots. Advanced mobile robots often exhibit intermittent interactions with their environment, resulting in high-order, high-dimensional dynamics that undergo high-frequency discrete changes. Estimating the state of such systems is complicated by the difficulty of modeling these rapidly evolving hybrid dynamics. This thesis makes state estimation more effective by recognizing that the hybrid dynamics can be approximated by a collection of simple models and identifying situations when specific models are applicable. Conventionally, state estimation for such systems is performed with multiple-model filters, but these filtering systems do not scale well as the number of models grows, and perform poorly in the face of high-frequency discrete dynamics. To overcome these limitations, this thesis develops an estimation framework that leverages available sensor information to improve the accuracy and scalability of multiple-model systems. The framework represents the dynamics with a hierarchy of contexts, and uses discrete-state estimation techniques to robustly identify the current context. The contextual information is then incorporated into multiple model filters to improve their accuracy and scalability. This thesis shows both experimentally and through simulation that integrating discrete and continuous estimation techniques enables accurate and scalable estimation for hybrid systems such as mobile robots with rapidly switching intermittent dynamics.