Deep Koopman Data-Driven Optimal Control Framework for Autonomous Racing

EasyChair Preprint 5622

Abstract

A model-based, data-driven control framework is introduced within the context of autonomous driving in this study. We propose a data-driven control algorithm that combines autonomous system identification using model-free learning and robust control using a model-based controller design. We present a full solution framework that is capable to automatically generate optimal path while performing system identification of a vehicle with unknown dynamics. We then design model-based control which is actively learned from a data-driven approach. Based on our new system identification algorithm, we can approximate an accurate, explainable, and linearized system representation in a high-dimensional latent space, without any prior knowledge of the system. To validate the algorithm, we conduct the model predictive control of an autonomous vehicle based on the augmented system identification on a scaled racing vehicle. The result indicates that we are able to design control in the lifted space to achieve tasks in path control and obstacle avoidance. The automatic path generation combined with the data driven control requires no a-priori knowledge of the vehicle and also proved to be effective that only requires less than 5 laps to design an optimal trajectory while identified a system that is able to achieve minimum lap time without extra learning episodes.

Keyphrases: Data-driven control, Deep Neural Network, Koopman operator, Koopman operator theory, Linear operator approach, Model Predictive Control, autonomous vehicle, controller design, system identification

@booklet{EasyChair:5622,
  author    = {Rongyao Wang and Yiqiang Han and Umesh Vaidya},
  title     = {Deep Koopman Data-Driven Optimal Control Framework for Autonomous Racing},
  howpublished = {EasyChair Preprint 5622},
  year      = {EasyChair, 2021}}