Enhancing Reinforcement Learning for Automated Driving through Virtual Lane Logic

This work investigates an alternative approach to current control systems for the Automated Driving (AD) of shuttle vehicles on dedicated roads. The proposed solution decouples the problem into two levels: a Deep Deterministic Policy Gradient (DDPG) Reinforcement Learning (RL) agent and a dedicated vehicle logic generating Virtual Lane (VL) data to eliminate redundancy and allow for smooth lane changes on curved roads. The training uses an environment defined through a model-based simulation, exploiting MatLab and Simulink tools, and has been conducted following a Curriculum Learning strategy. The performance of the introduced approach have been evaluated by testing the agent capabilities and exploring its behavior in the presence of external disturbances in the controlled states.