Automatic Intersection Management in Mixed Traffic Using Reinforcement Learning and Graph Neural Networks

Klimke, Marvin; Völz, Benjamin; Buchholz, Michael · 2023 · Crossref

DOI: 10.1109/iv55152.2023.10186800

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of automatic intersection management (AIM) in mixed traffic environments, where connected automated vehicles (CAVs) share the road with human-driven vehicles. While existing AIM approaches often assume fully automated traffic or require strict spatial separation, this work proposes a cooperative behavior planning scheme using reinforcement learning (RL) and graph neural networks (GNNs) that functions effectively with varying automation rates. The motivation stems from the reality that mixed traffic will be prevalent for the foreseeable future, necessitating planners that can account for the uncertainty and unpredictability of human drivers. The proposed method models cooperative planning as a multi-agent RL problem using a centralized training and execution paradigm. The traffic scene is represented as a directed graph where vehicles are nodes and interactions are edges. To handle mixed traffic, the graph representation is extended to include uncertainty regarding the maneuver intentions of non-connected vehicles; if an intention is unknown, edges are created for all potential conflict points to encourage cautious planning. The network architecture combines relational graph convolutional layers and graph attention layers to process vertex and edge features, including a new priority relation feature. The RL training incorporates a reward function that balances velocity, smoothness, and collision avoidance, with a specific "reluctance" penalty to prevent automated vehicles from unnecessarily stopping and disrupting traffic flow. Measurement uncertainties are modeled using noise processes tuned with real-world driving data. Experiments were conducted in a simulation environment using a kinematic bicycle model and human driving behaviors modeled by an extended intelligent driver model. The RL planner was evaluated against an enhanced first-in-first-out (eFIFO) baseline across scenarios with varying traffic densities and automation levels. Results demonstrate that the learned planner significantly increases vehicle throughput and reduces interaction delays compared to eFIFO as the share of automated vehicles increases. Notably, non-automated vehicles benefit from the cooperative maneuvers virtually as much as automated ones. The RL planner also showed robust performance, with throughput rising monotonically with automation levels, whereas the eFIFO baseline sometimes resulted in lower flow rates than simple precedence rules under high traffic demand. The study concludes that the proposed learning-based AIM scheme effectively improves traffic efficiency in mixed traffic without negatively impacting peak performance in fully automated scenarios.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness