Development of a pure pursuit lane keeping controller for a 1:10 scale autonomous vehicle

Research on Autonomous Vehicles (AVs) has experienced an increasingly significant growth of interest in the last few years because of its potential to enhance safety, efficiency and convenience of automotive transportation. Anyway, due to its complexity, there are still a lot of technical and social challenges to be tackled in this field. The driving task of an AV can be grouped mainly into five parts: Sensing & Input, Perception & Scene Understanding, Behavior Planning & Selection, Motion & Control and Actuation. Path Tracking Controllers (PTCs) lay in the Motion layer and play a key role in controlling the lateral and longitudinal dynamics of the vehicle by generating suitable control signals for the low-level actuation part, in order to follow the planned path. This thesis stems from the participation of the DEI-Unipd Team in the Bosch Future Mobility Challenge 2022 and focuses on the development of a lane keeping PTC for a 1:10 scale vehicle: lateral control is based on the traditional Pure Pursuit strategy and longitudinal control is obtained by means of a PID controller with a variable velocity reference. In this scenario, the Lookahead Heading Error (LHE) — input to the PP-based controller — is computed by means of a Convolutional Neural Network (CNN) which is trained by means of a Gazebo simulator: virtual augmented images taken from a perturbed path are used in a supervised strategy to characterize the LHE. Throughout this thesis the LHE will be assumed to be available at a fixed distance. A pure camera based controller is then proposed, specifically addressing the effects of the delay in the steering actuation mechanism and the "corner-cutting" effect, characteristic of the Pure Pursuit strategy. Thanks to a low-noise input provided by the CNN, a derivative action and a velocity reference generation technique, uncommon in PP-type approaches, were introduced in order to cope with the presence of limited steering dynamics and constraints imposed by the LHE estimation.