ITSC 2025 Paper Abstract

Close

Paper FR-LM-T38.4

Zhang, Yifeng (National University of Singapore), Gong, Ping (National University of Singapore), He, Weiyi (National University of Singapore), Liu, Yilin (Beijing university of posts and telecommunications), Mingfeng, Fan (National University of Singapore), Sartoretti, Guillaume (National University of Singapore)

V2XFormer: A Multi-Stage Transformer for Multi-Agent Reinforcement Learning in V2X-Enabled Traffic Signal Control

Scheduled for presentation during the Regular Session "S38a-Towards Scalable and Trustworthy AI in Connected Mobility" (FR-LM-T38), Friday, November 21, 2025, 11:30−11:50, Coolangata 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Keywords AI, Machine Learning for Dynamic Traffic Signal Control and Optimization

Abstract

Connected vehicles (CVs), enabled by vehicle-to-everything (V2X) communication, provide more fine-grained and real-time traffic information, offering new opportunities to enhance network-wide adaptive traffic signal control (ATSC). However, effectively integrating heterogeneous CV and infrastructure data and leveraging such information for cooperative ATSC across multiple intersections remain critical challenges in CV-enabled traffic environments. To address this, we propose V2XFormer, a multi-stage Transformer framework designed to fuse features from CVs and intersections via vehicle-level temporal encoding, lane-level interaction modeling, and intersection-level coordination modeling, and jointly optimize traffic prediction and MARL-based signal control for cooperative control in CV-enabled environments. Specifically, at the lane level, we design a dual-encoder Transformer that utilizes temporal vehicle information to extract features from cooperative and competitive lanes, and enhance feature aggregation via an adaptive gated fusion mechanism guided by intersection-level context. At the intersection level, we introduce a decoder-only Transformer that adaptively integrates local and neighboring intersection features to enable broader cross-intersection coordination. This hierarchical design allows V2XFormer to capture both fine-grained lane interactions and high-level intersection dependencies, leading to more consistent and stable control policies over time. Experimental results show that our method consistently outperforms various baselines in network-wide traffic optimization, with notable improvements under high-demand and complex shared-lane scenarios, highlighting its effectiveness for large-scale ATSC in CV-enabled environments.

 

 

All Content © PaperCept, Inc.


This site is protected by copyright and trademark laws under US and International law.
All rights reserved. © 2002-2025 PaperCept, Inc.
Page generated 2025-10-18  21:25:24 PST  Terms of use