Learning model predictive control for autonomous racing

Abstract: In this work, an improved Learning Model Predictive Control (LMPC) architecture for autonomous racing is presented. The controller is reference-free and is able to improve lap time by learning from history data of previous laps. A terminal cost and a sampled safe set are learned from history data to guarantee recursive feasibility and non decreasing performance at each lap. Improvements have been proposed to implement LMPC on autonomous racing in a more efficient and reliable way. Improvements have been done on three aspects. Firstly, system identification has been improved to be run in a more efficient way by collecting feature data in subspace, so that the size of feature data set is reduced and time needed to run sorting algorithm can be reduced. Secondly, different strategies have been proposed to improve model accuracy, such as least mean square with/without lifting and Gaussian process regression. Thirdly, for reducing algorithm complexity, method like combining different model construction strategies was proposed. Also, running controller in a multi-rate way has also been proposed to reduced algorithm complexity when increment of controller frequency is necessary. Besides, the performance of different system identification strategies have been compared, which include strategy from Newton’s law, strategy from classical system identification and strategy from machine learning. Factors that can possibly influence converged result of LMPC were also investigated, such as prediction horizon, controller frequency. Experiment results on a 1:10 scaled RC car illustrates the effectiveness of proposed improvements and the difference of different system identification strategies.

This work has been carried out at the Model Predictive Control lab at the Department of Mechanical Engineering, University of Berkeley, CA.

Examiner: Mikael Johansson, KTH