Novel Methods In Training Autonomous Vehicles For Urban Roads

This thesis presents the development and application of a novel platform to train autonomous vehicles (AV) for urban roads. Interactive and immersive virtual reality (VR) environments are developed for the collection of mobility preference, behaviour, movement, and orientation data. The resulting naturalistic data can be used directly to train AV control systems. This platform is exemplified in an end-to-end case study resulting in a multi-objective braking system which maximizes both pedestrian safety and passenger comfort. It begins with the development of an immersive VR pedestrian road-crossing environment and compilation of a unique, naturalistic dataset. A vehicle agent is then successfully trained against the dataset, learning a multi-objective brake control policy using deep reinforcement learning methods and reducing the negative influence on passenger comfort by half while maintaining safe braking operation. This platform offers the opportunity to simulate complex, human-in-the-loop scenarios AVs will inevitably face and train them for these scenarios.