Automated driving in low complexity scenarios through adaptive potential fields and model predictive control

In the field of control, automated driving (AD) makes up one of the most active and sought-after research areas due to its potential to transform transportation. Reducing carbon emissions, increasing safety, and improving logistics are only a small part of the major benefits autonomous vehicles (AV) can deliver to companies and citizens. This thesis introduces a complete control system for AD, focusing mainly on one of the core problems of AV: motion planning (MP). MP refers to the computational process that allows an AV to reach a destination from a starting point, following a series of valid configurations and avoiding obstacles. In literature, MP is usually divided into three distinct steps: deciding which driving maneuver to realize, finding an optimal trajectory to implement it, and finally using it as a reference for a suitably tuned control system. The goal of this work is to merge the first two steps, thus eliminating the need for a complex logic capable of adapting to all possible situations that may occur on a road. To achieve this aim, a Local Path Planner (LPP) employs Adaptive Artificial Potential Fields (AAPF). AAPF refers to a potential field that can adapt its shape to perfectly model the road and the surrounding vehicles, considering their evolution in time. AAPF formulate a suitable Nonlinear Optimization Problem (NLP) that computes the safest trajectory for the best maneuver that can be performed at any instance, considering all the possible trajectories of the surrounding AVs. A Nonlinear Model Predictive Control (NMPC) controller is then used to track that safe trajectory given by the LPP and, at the same time, control the lateral and longitudinal dynamics of the AV, as required by the higher levels of driving automation: SAE L4 and L5. The combination of an offline Global Path Planner (GPP), that computes a global reference path to reach the destination based on GPS data, the LPP, that can adapt it to the changing surroundings of the AV, and finally the NMPC, allows an AV to arrive at the destination in a safe, comfortable, and time-optimal way. The LPP definition and the use of AAPF are this thesis's main contributions, and extensive simulation tests are provided to verify their effectiveness in different maneuvers and conditions in both a highway and extra-urban road environment. In addition, further tests considering the whole control system for a shuttle bus in a restricted area are provided to prove one of the possible applications for this AD architecture and MP technique. Future work includes both the possibility of using Reinforcement Learning (RL) or other Machine Learning (ML) techniques to obtain a more effective tuning of the many AAPF parameters in a static or dynamic manner, as well as the possibility of generalizing this MP approach to control any dynamic system that interacts with evolving surroundings whose state uncertainty in time is bounded.