Optimising QoS for HD Map Delivery in Vehicular Networks Using Reinforcement Learning

Abstract

Autonomous vehicles (AVs) increasingly rely on high-definition (HD) maps to support safe manoeuvres and timely decision-making. Unlike continuous data streams such as voice or video, HD map updates are transmitted as large, event-driven bursts requiring both high reliability and low latency to prevent safety-critical failures caused by outdated map information in autonomous vehicles. However, the current IEEE 802.11p standard treats HD map traffic as best-effort, making it vulnerable to packet collisions, increasing latency in high-density, high-mobility vehicular ad hoc networks (VANETs). Existing queue management and reinforcement learning (RL)-based approaches address some aspects of resource allocation, but they exhibit constraints in three critical areas: mobility-awareness, scalability to large-scale networks, and multi-parameter optimisation of IEEE 802.11p.
This thesis develops an RL framework to improve the dissemination of HD maps in VANETs. First, it introduces a lightweight single-agent RL solution that incorporates sojourn time, the expected duration a vehicle remains within the coverage of a roadside unit, to enhance mobility-aware scheduling and reduce HD map transmission latency under heterogeneous traffic conditions. Second, recognising that single-agent approaches struggle with the exponential growth of the state–action space in dense environments, the thesis proposes distributed multi-agent RL strategies, showing significant improvements in HD map transmission latency and throughput as network density increases. Third, the work extends to a multi-task design where agents specialise in jointly optimising contention window (CW) and inter-frame space (IFS), both key IEEE 802.11p parameters that govern channel access timing and prioritisation. The RL model implicitly learns and exploits the interdependencies between these parameters, enabling improved traffic differentiation across HD map, video, voice, and best-effort flows while maintaining reliable delivery under mixed network loads. Finally, the thesis develops one of the first hardware-in-the-loop (HIL) testbeds for HD map in VANETs, integrating LiDAR and video data streams with OMNeT++ simulator, and Robot Operating System (ROS), to validate RL policies under realistic system setup, facilitating their transition toward production deployment.
The proposed strategies demonstrated lower latency for the delivery of an HD map across heterogeneous data traffic types. These contributions collectively advance the state of knowledge on intelligent wireless resource allocation for safety-critical vehicular applications.

Date of Award	19 Feb 2026
Original language	English
Awarding Institution	Northumbria University
Supervisor	Nauman Aslam (Supervisor), Juan Zhang (Supervisor) & Zhenhui Yuan (Supervisor)