ITSC 2025 Paper Abstract

Paper TH-EA-T25.1

Li, Peifeng (Beijing Institute of Technology), Tan, Huachun (Beijing Institute of Technology), Wen, Zoutao (Beijing Institute of Technology), Zhao, Yanan (Beijing Institute of Technology), Gao, Bolin (Tsinghua University)

Dynamic Grouping Enhanced Deep Reinforcement Learning for Cooperative On-Ramp Merging

Scheduled for presentation during the Regular Session "S25b-Cooperative and Connected Autonomous Systems" (TH-EA-T25), Thursday, November 20, 2025, 13:30−13:50, Cooleangata 4

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 Cooperative Driving Systems and Vehicle Coordination in Multi-vehicle Scenarios, Data Analytics and Real-time Decision Making for Autonomous Traffic Management, Cooperative Vehicle-to-Vehicle Data Sharing for Safe and Efficient Traffic Flow

Abstract

On-ramp merging continues to pose a significant challenge for connected and automated vehicles (CAVs), primarily due to suboptimal vehicle interactions that often lead to congestion and unstable merging sequences. To address this, we propose a Grouping Cooperative Merging Deep Reinforcement Learning (GCM-DRL) framework that integrates optimized control strategies with deep policy learning to enhance merging performance and coordination stability. The framework is equipped with an Adaptive Grouping Module, a dedicated component that dynamically partitions vehicles into behaviorally coherent groups based on real-time state features. Group-level guidance is derived through the Grouping Cooperative Controller, which assigns target passing times using a kinematically informed scheduling strategy. By integrating this control signal into a deep reinforcement learning (DRL) framework trained with proximal policy optimization (PPO), the approach improves both the stability and interpretability of the learned policies. Extensive simulations across various CAV flow rates demonstrate that GCM-DRL outperforms both rule-based and purely learning-driven approaches in terms of merging efficiency, safety, and energy consumption. These results underscore the potential of embedding interpretable control structures within DRL to enable robust real-time cooperation among fully CAVs.