Hierarchical GNN-Based Multi-Agent Learning for Dynamic Queue-Jump Lane and Emergency Vehicle Corridor Formation

TL;DR

This work tackles the challenge of expediting emergency vehicle passage in congested traffic by proposing a hierarchical graph neural network architecture trained with multi-agent MAPPO. A global high-level planner determines strategic corridor formation, while local low-level controllers execute continuous trajectories on nearby subgraphs; a centralized GNN critic supports training. The approach leverages a dynamic vehicle interaction graph with attention-based message passing to handle variable vehicle counts and to capture spatial relationships, achieving substantial improvements in EMV travel time (up to 28.3% over MAAC and 44.6% over uncoordinated traffic), near-zero collision rates (0.3%), and maintaining 81% of normal traffic efficiency. Ablation and generalization studies confirm the value of the graph structure and hierarchical decomposition, showing robustness across densities, CV penetration rates, and road lengths, with strong sample efficiency and interpretability through attention visualizations. The results underscore the potential of integrating graph representations with hierarchical MARL to advance intelligent transportation systems and emergency response capabilities, while highlighting practical considerations for deployment, including communication reliability, multi-EMV scenarios, and real-world human driver modeling.

Abstract

Emergency vehicles require rapid passage through congested traffic, yet existing strategies fail to adapt to dynamic conditions. We propose a novel hierarchical graph neural network (GNN)-based multi-agent reinforcement learning framework to coordinate connected vehicles for emergency corridor formation. Our approach uses a high-level planner for global strategy and low-level controllers for trajectory execution, utilizing graph attention networks to scale with variable agent counts. Trained via Multi-Agent Proximal Policy Optimization (MAPPO), the system reduces emergency vehicle travel time by 28.3% compared to baselines and 44.6% compared to uncoordinated traffic in simulations. The design achieves near-zero collision rates (0.3%) while maintaining 81% of background traffic efficiency. Ablation and generalization studies confirm the framework's robustness across diverse scenarios. These results demonstrate the effectiveness of combining GNNs with hierarchical learning for intelligent transportation systems.

Figures (7)

• Figure 2: System architecture. Hierarchical processing: graph construction, high-level strategy planning (every 5 steps), and low-level control (every step).
• Figure 3: Graph construction pipeline. Vehicle states are transformed into a directed graph with three edge types.
• Figure 4: Detailed hierarchical architecture for emergency vehicle corridor formation. Left: High-level planner processes the global vehicle graph through 3 GAT layers to produce per-vehicle discrete strategies (e.g., yield left, yield right, maintain). Center: Low-level controller operates on local $k$-hop subgraphs, combining ego embeddings with strategy embeddings via concatenation to output continuous control actions. Right: Centralized critic (used only during training) aggregates global information for value estimation under the CTDE paradigm. Red dashed arrows indicate strategy conditioning flow from high-level to low-level modules.
• Figure 5: Curriculum learning stages. Training progresses from simple (few vehicles, high CV%) to complex (dense traffic, low CV%).
• Figure : (a) Before: Congested traffic blocks EMV