Hybrid Deep Reinforcement Learning Model for Safe and Efficient Autonomous Vehicles at Pedestrian Crossings

The autonomous vehicle is a particularly promising area of application for machine learning algorithms. Autonomous driving is challenging due to the complexity of the road network, the hardly predictable human behaviors, either pedestrians or drivers and the imperative need to ensure the safety of all road users. Deep Reinforcement Learning algorithms have garnered significant interest as a viable option for controlling autonomous vehicles. They allow defining a control policy by solely describing the observation space, the action space, and a reward function. Yet, the most common solutions face difficulties when the action space is hybrid (simultaneously continuous and discrete), and the reward is sparse, which is the case when a selection between choices must be made (like yielding or not), and the trajectory of the vehicle must be adapted according to this decision. This paper addresses these challenges by focusing on controlling an autonomous vehicle at pedestrian crossings. The vehicle adapts its speed (continuous decision) and the displayed signal (discrete decision) according to the pedestrian's state. The unique characteristic of the speed profile is to facilitate safe pedestrian crossings without requiring the vehicle to come to a complete stop. This raises the consistency problem between the speed profile and the signals issued by the vehicle. This paper introduces a new hybrid DRL model capable of making both decisions. It discusses the training method, the construction and necessary constraints on the reward, as well as the success of generalization to different speed limitations.