Exploiting models of different granularity in robust predictive control

Abstract

The use of detailed models over long horizons in predictive control can be computationally challenging. Furthermore, always-present uncertainty renders the use of such sophisticated detailed models over long time horizons questionable due to the resulting variability of the trajectories. We propose a multi-stage scheme that combines the use of models of different granularity - using detailed models for short-term predictions, while performing long-term predictions with less detailed models. Using projection and invariance properties for the different model complexities and the transitions between them, we show that this scheme is recursively feasible. In a simulation study, we show how two models of different complexity can be combined for steering a mobile robot through a landscape with obstacles.