Comparison of Deep Reinforcement Learning Methods for Safe and Efficient Autonomous Vehicles at Pedestrian Crossings

These past years, the domain of Connected and Autonomous Vehicles (CAV) has been extremely flourishing, with fully autonomous self-driving cars being an active research area. The most challenging aspect of the driving task is the interaction between the driver agent and the environment, especially to handle conflicts with the other road users. The most vulnerable ones are the pedestrians. To ensure better safety at crosswalks, this paper focuses on planning a vehicle trajectory, to pass without risks, while cooperating with the pedestrian. According to recent works, Deep Reinforcement Learning can be used to efficiently control the vehicle in a complex environment. This paper assesses both widely held continuous action space approaches for controlling the vehicle: Deep Deterministic Policy Gradient and Proximal Policy Optimization. Both approaches are used to allow the vehicle to safely and efficiently interact with a pedestrian willing to cross the road at the crosswalk. The main contributions of the paper are the design and parameterization of these approaches for this single conflict use-case, the feasibility study of both approaches, and the detailed comparison showing their benefits and limitations.