Modeling Interactions of Autonomous/Manual Vehicles and Pedestrians with a Multi-Agent Deep Deterministic Policy Gradient

Abstract

This article focuses on the development of a stable pedestrian crash avoidance mitigation system for autonomous vehicles (AVs). Previous works have only used simple AV–pedestrian models, which do not reflect the actual interaction and risk status of intelligent intersections with manual vehicles. The paper presents a model that simulates the interaction between automatic driving vehicles and pedestrians on unsignalized crosswalks using the multi-agent deep deterministic policy gradient (MADDPG) algorithm. The MADDPG algorithm optimizes the PCAM strategy through the continuous interaction of multiple independent agents and effectively captures the inherent uncertainty in continuous learning and human behavior. Experimental results show that the MADDPG model can fully mitigate collisions in different scenarios and outperforms the DDPG and DRL algorithms.