Hybrid Reinforcement Learning-Based Eco-Driving Strategy for Connected and Automated Vehicles at Signalized Intersections

Bai, Zhengwei; Hao, Peng; ShangGuan, Wei; Cai, Baigen; Barth, Matthew J. · 2022 · OpenAlex-citations

DOI: 10.1109/tits.2022.3145798

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 optimizing energy efficiency and traffic throughput for Connected and Automated Vehicles (CAVs) at signalized intersections within mixed traffic environments. Existing eco-driving strategies often rely on rule-based or model-based algorithms that depend on idealized assumptions, such as traffic-free conditions or 100% vehicle connectivity, which limits their applicability in real-world scenarios. Furthermore, many conventional approaches focus solely on longitudinal maneuvers, ignoring the potential benefits of lateral control. To overcome these limitations, the authors propose a Hybrid Reinforcement Learning (HRL) framework that integrates rule-based policies with deep reinforcement learning (DRL) to handle the complexity of long-term intersection-based eco-driving. The proposed HRL framework consists of three main components: a rule-based driving manager that coordinates rule-based policies and the RL policy, a multi-stream neural network that extracts features from vision and Vehicle-to-Infrastructure (V2I) data, and a DRL-based policy network that generates longitudinal and lateral eco-driving actions. The system utilizes a Long-Short Term Reward (LSTR) model to balance conflicting objectives, such as speeding up and saving energy. The ego-vehicle perceives the environment through on-board sensors (camera, radar, OBD) and V2I communications providing Signal Phase and Timing (SPaT) information. The observation space includes a 12-dimensional vector of logical data and processed image data formatted into a multi-frame spatiotemporal format. The action space defines discrete longitudinal acceleration levels and target lane changes. The study was evaluated using a Unity-based simulator designed to replicate mixed traffic scenarios with human-driven vehicles exhibiting diverse dynamic characteristics. Numerical experiments compared the HRL method against state-of-the-art model-based eco-driving approaches. The results demonstrate that the HRL framework significantly outperforms the baseline models, reducing energy consumption by 12.70% and saving 11.75% in travel time. The framework successfully handles multi-logical tasks, including collision avoidance, lane changing, stop-in-red maneuvers, and start-in-green mechanisms, without relying on simplified environmental assumptions. The significance of this work lies in its ability to provide a robust, adaptive eco-driving strategy for CAVs in realistic mixed-traffic conditions. By combining the interpretability and stability of rule-based systems with the generalization power of deep learning, the HRL framework offers a practical solution for improving urban traffic efficiency and reducing energy consumption. This approach advances the field by addressing the limitations of traditional model-based methods and demonstrating the viability of hybrid AI techniques for complex, multi-objective driving tasks.

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.

• eco driving