Estimation of Non-Convex Pareto Front in Collision Avoidance at Urban Intersections

Abstract

Inexperienced drivers, such as the elderly and novice drivers, are prone to cause traffic accidents due to human error. Autonomous driving is expected to reduce traffic accidents by assisting their recognition, judgment, and operation. However, it is only effective in situations where drivers can easily make decisions, such as driving on highways, and is still challenging in urban areas. In this paper, we focus on safe and efficient autonomous driving in situations where multiple moving obstacles simultaneously exist, assuming an urban intersection. Since it is difficult to construct a driving model for such a situation, we introduced a reinforcement learning method that does not require a driving model. This paper proposes a collision-avoidance problem as a multi-objective sequential decision-making problem. We propose a method for learning a non-convex Pareto front concerning safety and speed using the multi-objective reinforcement learning algorithm, Pareto-DQN. The proposed method's performance through computer experiments is verified in a T-intersection environment. We confirmed the acquisition of multiple Pareto-optimal driving policies that could not be achieved using conventional methods with linear scalarization. The proposed method is helpful for system designers because it provides a more detailed representation of the driver's non-convex preferences.