Coordinated Urban Signal Control via Edge Native Federated Multi Agent RL: VEINS–Simu5G Results

Abstract

Introduction: Urban traffic congestion induces persistent delays, elevated emissions, and safety risks across arterial networks. Conventional fixed‑time and actuated control assume quasi‑stationary demand and limited interdependence, which mismatches modern conditions with sharp peaks, incidents, multimodal flows, and spillback between adjacent junctions.

Objectives: Design and evaluate an edge‑native intersection control framework that keeps perception and decision making on smart poles, enables privacy‑preserving cooperation via federated multi‑agent reinforcement learning, explicitly models communication effects, and targets simultaneous improvements in efficiency and emissions while maintaining safety and governance.

Methods: The architecture separates a millisecond‑level operational plane using IEEE 802.11p from a scalable orchestration plane using MQTT 5.0 for telemetry and model coordination. Lightweight agents learn locally from on‑pole sensors; periodic aggregation constructs a network‑level prior without exporting raw data. Evaluation employs a digital‑twin‑in‑the‑loop setup coupling VEINS (OMNeT++–SUMO) with Simu5G to represent 802.11p and NR‑V2X latency/loss; emissions are estimated with SUMO’s HBEFA‑based model. Baselines include fixed‑time, actuated, and non‑federated MARL.

Results: Across single‑junction and 3×3 grid scenarios, the proposed approach shows potential reductions in average delay and queue length with concurrent decreases in CO₂ and NOx relative to baselines, while maintaining safety via supervisory constraints and demonstrating robustness under realistic wireless impairment.

Conclusions: An edge‑native, federated design is operationally feasible for smart‑pole deployments, supports governance by keeping data local, and yields reproducible performance gains when communication dynamics are modeled alongside traffic flow.