RL-Godot pedestrian simulation: curriculum-based reinforcement learning for pedestrian simulation

Abstract

The paper describes a research effort aimed at developing RL-Godot, a reinforcement learning (RL) based open source software system supporting the study of pedestrian dynamics. We first introduce the curriculum based RL approach to pedestrian simulation adopted for RL-Godot, then describe its system architecture and report a preliminary experimental application that evaluates the framework’s ability to qualitatively reproduce low-to-medium density pedestrian behaviors reported in the literature.