Tuning path tracking controllers for autonomous cars using reinforcement learning

This article proposes an adaptable path tracking control system, based on reinforcement learning (RL), for autonomous cars. A four-parameter controller shapes the behaviour of the vehicle to navigate lane changes and roundabouts. The tuning of the tracker uses an 'educated' Q-Learning algorithm to minimize the lateral and steering trajectory errors, this being a key contribution of this article. The CARLA (CAR Learning to Act) simulator was used both for training and testing. The results show the vehicle is able to adapt its behaviour to the different types of reference trajectories, navigating safely with low tracking errors. The use of a robot operating system (ROS) bridge between CARLA and the tracker (i) results in a realistic system, and (ii) simplifies the replacement of CARLA by a real vehicle, as in a hardware-in-the-loop system. Another contribution of this article is the framework for the dependability of the overall architecture based on stability results of non-smooth systems, presented at the end of this article.