A reinforcement learning approach to automated driving of shuttle vehicles

The work aims at investigating a possible alternative solution to the control systems currently utilized for Driving Automation, through the implementation of a system based on Reinforcement learning for the control of a level 4 oriented shuttle solution for dedicated roads. The Driving problem will be, therefore, structured as a Markov Decision Process, in order to comply with the RL problem formulation. A model-based simulation will be implemented in MATLAB and Simulink, in order to build a feasible and reliable training environment. The study will introduce the Driving Automation paradigm, as well as the Reinforcement Learning approach. The employed DDPG RL Actor will be, then, analyzed, along with the state of the art of RL application to AD. The central part of the exposition will present the main aspects of the developed simulation and control solution, from the structure of the Deep Neural Networks, to the definition of the dynamic simulation environment and the elaborated solutions to data elaboration challenges like the virtual lane definition for lane change in curving roads, to data interpolation for missing data lectures and intersection management. The training sequences and the learning results will, then, be analyzed, leading to the extraction of the conclusions and of the prospective for the future researches.