Multi-agent footstep steering with deep reinforcement learning
Date
2024
Authors
Peng, Kun
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Crowd simulation plays a crucial role in a wide range of fields, from digital media to urban planning. However, traditional particle-based algorithms often lack essential information to present realistic human bipedal locomotion. This research aims to propose a more realistic and efficient steering model for crowd simulation by combining Multi-Agent Reinforcement Learning (MARL) with bipedal locomotion modelling. This study explores the advantages of MARL and analyzes a mathematical approach to simplifying complex bipedal locomotion. The approach utilizes the Proximal Policy Optimization algorithm and trains the model in adjustable randomized maze-like environments. Assessment results of the model indicate that the model learns goal-reaching behaviours and learns to avoid static and dynamic obstacles. Furthermore, the agents can simulate complex steering behaviours such as side-stepping and turning-like behaviours with two feet. This research contributes to the advancement of the field of crowd simulation through a flexible and realistic approach to modelling human steering behaviours in complex and dynamic environments.
Description
Keywords
crowd simulation, multi-agent reinforcement learning, bipedal locomotion