Work

A Multi-Agent Reinforcement Learning Framework for Autonomous Traffic-Light-Free Intersection Management

Year: 2025
Type: article
Abstract: Traffic signal control remains one of the most persistent sources of delay, fuel consumption, and inefficiency in urban road networks. Even adaptive signal systems rely on predefined phases that fail to exploit real-time vehicle-level intelligence. With the rapid emergence of connected and autonomous vehicles (CAVs) and vehicle-to-everything (V2X) communication, traffic-light-free intersections have become a viable alternative. This paper presents a novel Multi-Agent Reinforcement Learning (MARL) framework that enables vehicles to autonomously negotiate intersection passage without traffic lights. Each vehicle operates as an intelligent agent, coordinating with others through V2X communication while a lightweight Intersection Coordination Server (ICS) enforces safety constraints. A Graph Attention Network (GAT) captures dynamic spatial interactions among conflicting vehicles, and Multi-Agent Proximal Policy Optimization (MAPPO) ensures stable cooperative learning under partial observability. Extensive simulations conducted in Simulation of Urban Mobility (SUMO) and Car Learning to Act (CARLA) demonstrate substantial performance improvements, including up to 76% reduction in average delay, 48% increase in throughput, and up to 41% reduction in fuel consumption compared to adaptive signalized intersections. Results indicate that MARL-based, signal-free intersection control offers a scalable and safe pathway toward next-generation smart mobility.
Language: English
Open Access status: diamond
OpenAlex ID: W7217457847
FWCI:
Cites: 15
Cited by:
Related to: 10