Multi-Agent Connected Autonomous Driving with Infrastructure Perception

Document Type

Conference Proceeding

Publication Date

1-1-2026

Abstract

We evaluated the driving performance of connected autonomous vehicles (CAVs) that utilize infrastructure-based perception and planning and compared them to unconnected autonomous vehicles (UAVs) that rely solely on onboard systems. Simulated environments were employed to model sensor noise and to train driving agents using multi-agent reinforcement learning (MARL). The modeled infrastructure-based automated driving system assigns a dedicated driving agent to each vehicle. These agents receive latent scene representations from a centralized infrastructure-supported situation understanding module and selected actions for each CAV in a distributed manner. In contrast, the baseline UAV driving agent operates independently, without any external input. We used identical reward functions and policy architectures to train each agent type. The results demonstrate that connectivity contributed to smoother and more efficient autonomous driving - even under a simulated infrastructure communication latency of 200 milliseconds. Specifically, CAVs achieved a 0.43% improvement in collision-free route completion (reaching 99%), reduced trip duration time by 11%, increased average velocity by 13%, and decreased average acceleration and jerk by 25% and 21%, respectively. These findings underscore the potential of infrastructure-connected autonomy to enhance traffic flow and improve overall driving performance.

Publication Title

IEEE Intelligent Vehicles Symposium Proceedings

ISBN

[9798331547936]

Link to Full Text

Share

COinS