Repetitive learning model predictive control: An autonomous racing example

Abstract

We propose an optimization based, data-driven framework to design controllers for repetitive tasks. The proposed framework builds on previous work of Learning Model Predictive Control and focuses on problems where the terminal condition of one iteration is the initial condition of the next iteration. A terminal cost and a sampled safe set are learned from data to guarantee recursive feasibility and non-decreasing performance cost at each iteration. The proposed control logic is tested on an autonomous racing example, where the vehicle dynamics are identified online. Experimental results on a 1:10 scale RC car illustrate the effectiveness of the proposed approach.