Racing line optimization algorithms for high performance and/or automated vehicles

The lap time optimisation is the process to calculate the best trajectory to achieve the minimum lap time, while considering the boundaries due to the vehicle dynamics. Traditional approaches have leaned on physics-based control models, which, although effective in offline settings, fall short when applied to online racing scenarios. These models not only burden computational resources but also lack the adaptability required to respond promptly to dynamic variations in vehicle behaviour and track conditions. The work of this thesis consists of a reinforcement learning algorithm capable to act as a path re-planner with access to previously stored information coming from previous manoeuvres. This introduces the way for adaptive re-planning that can dynamically respond to changes in the vehicle's behaviour and evolving track conditions during the course of a racing session. Such dynamic changes include variations in friction coefficients, alterations in tire temperature, and fluctuations in the dynamic condition of the vehicle—all of which are common occurrences during a racing session. The primary objective of this study is to optimize lap times by considering real-time vehicle information, gathered through data collection by sensors(e.g: accelerometers, yaw rate and wheel speed snesors) from previous manoeuvres. To achieve this, two critical components are developed and integrated into the system: a highly representative 13-degree-of-freedom (DOF) vehicle model and a data buffer for storing historical performance data from prior laps. The 13-DOF vehicle model serves as a comprehensive representation of vehicle dynamics, forming the foundation upon which the re-planning process operates. The trajectory derived from this process guides the vehicle's path in the current state through a previously tuned Feedforward-Feedback (FF-FB) path tracking controller. Data extrapolation and retrieval rely on the use of buffers—databases storing dynamic information of the vehicle, including steering actions, lateral and longitudinal accelerations, as well as vehicle states such as position and yaw. This stored data is then accessed by the RL algorithm within a spatial window determined by the vehicle's current position. By focusing on localized data retrieval around the vehicle, the overall procedure of reading the database is facilitated, allowing for more efficient agent learning. In summary, the novelty of this research lies in the creation of an architecture that enables online learning based on data recorded from previous manoeuvres. This approach promises to enhance the performance and safety of autonomous racing vehicles, ultimately contributing to the advancement of autonomous driving technology in high-speed and dynamic environments.