Data-Driven Multi-Modal Learning Model Predictive Control

We present a Learning Model Predictive Controller (LMPC) for systems with multi-modal dynamics performing iterative control tasks. Our goal is to use historical data from previous task iterations to design a data-driven control policy for the multi-modal system when the current mode is unknown. We first propose a novel method for using data to learn local affine time-varying (ATV) models of the multi-modal system dynamics. Then we present how to construct data-driven LMPC terminal components from multi-modal historical data, and how to design the resulting LMPC policy. The effectiveness of our method is demonstrated in a simulation example of an autonomous vehicle driving on a friction-varying track.