A Unified Approach for Decision Making and Trajectory Planning for Automated Driving using Time-Varying Potential Fields in Roundabouts

In the last decades, the Automated Driving (AD) has become a cornerstone in modern research on transportation, since it can significantly improve urban mobility, reducing traffic congestion, while lowering pollution and energy consumption, and enhancing road safety by avoiding human failures that most of the time are crashes source. Specifically, the present work focuses on SAE Level 4 (i.e. High Driving Automation) autonomous vehicles (AVs) that travel within urban roundabouts, proposing new strategies to handle the Motion Planning (MP) in such complex scenarios. The state-of-the-art underlying this thesis introduced a three-layers structure for the MP architecture, given by a Global Path Planner (GPP), a Local Trajectory Planner (LTP) and a Nonlinear Model Predictive Control (NMPC) controller, merging together in a complete AD system. The GPP works offline and provides the global reference path from the starting point to the target one. The LTP plans locally the optimal trajectory by solving a NLP, whose cost function involves static Artificial Potential Fields (APFs) to model the road path and the Adaptive APFs (AAPFs) to manage the other vehicles in the driving scenario taking into account their time-varying position. At last, the NMPC controller computes the optimal controls for tracking the local reference trajectory, ensuring safe and comfortable maneuvers. Thus, the main purpose of this thesis is to extend the work discussed above in roundabout scenarios. Firstly, Mirrorable APFs are introduced for the dynamic implementation of road APFs, in order to comply with the traffic regulation within the roundabouts and perform proper lane-changing, while ensuring the recovery of lane-keeping behavior. Then, the prediction step for the implementation of the AAPFs, which model the surrounding actors, has been adjusted according to the actors' directions of motion, which are not known a priori. This re-adaptation is due to the complexity of the scenario, as the roundabout infrastructure is a central intersection connecting multiple roads. Therefore, the actors' positions relative to the reference path must be defined correctly in order to compute the actors' directions of motion; consequently, different $Prediction$ $Modes$ for actors' trajectories are identified. As a result, AAPFs will model the actors' motion along a suitable predicted trajectory. Furthermore, since many actors may circulate within the roundabout, every motion maneuver must be planned in safety conditions while respecting the right-of-way rule: thus, barrier functions can be embedded in the NLP so that the vehicle is prompt to perform the planned yielding maneuver properly. Therefore, all the strategies proposed in this work have been implemented in the complete AD system and tested on a road network given by four main roads intersecting in a central roundabout, in order to validate the overall operational performances. Moreover, the driving scenario has been designed through the Driving Scenario Designer tool by the MATLAB AD Toolbox. Future work may improve the tuning of LTP parameters in order to make quasi real-time computations and/or enhance Vehicle-To-Infrastructure (V2I) communication for achieving more precise and reliable data exchange about the surrounding environment.