Multi-Agent Advantage Actor-Critic Learning For Message Content Selection in Cooperative Perception Networks

Abstract

Recent advancements in autonomous vehicle perception have exposed limitations of onboard sensors such as radar, lidar, and cameras, which road obstacles and adverse weather conditions can impede. Connected and Autonomous Vehicles (CAVs) are leveraging wireless communications to share perception information through a process called Cooperative Perception (CP), aiming to provide a more comprehensive understanding of their environment. However, this can result in excessive redundant and useless information in the network, as the same road objects may be detected and exchanged simultaneously by multiple CAVs. This not only consumes more network resources but also may overload the communication channel, reducing the delivery of perception information to CAVs and ultimately decreasing the overall CP awareness in the network. This paper introduces MCORM, a multi-agent learning method based on the advantage actor-critic algorithm to maximize object usefulness and reduce redundancy in the network. Our evaluations demonstrate that through this method, CAVs learn optimal CP message content selection policies that maximize usefulness. Furthermore, our proposal proves to be more effective in mitigating object redundancy and improving network reliability in comparison to existing approaches.