Advancing mixed autonomy traffic networks: motion planning and coordination for autonomous vehicles in real-world environments
File(s)
Author(s)
Adan, Fahmy
Type
Thesis
Abstract
The imminent introduction of autonomous vehicles (AVs) to existing road networks presents unique challenges and opportunities for traffic management, particularly at unsignalised intersections. These intersections, while comprising a small portion of road networks, contribute disproportionately to accidents and suffer from underutilized capacity due to current right-of-way allocation mechanisms. This thesis addresses these challenges by developing novel approaches to Autonomous Intersection Management (AIM) in mixed-autonomy scenarios.
The research presents three key methodological contributions. First, this thesis develops an enhanced model-free multi-agent reinforcement learning method that achieves traffic-compliant coordination in mixed environments through policy space constraints rather than reward shaping. This approach significantly outperforms independent methods with individual objectives, validating the effectiveness of V2V communications and shared rewards. Second, this thesis proposes a novel distributed trajectory-based optimization system that enhances state-of-the-art sampling-based algorithms with decentralization capabilities and an inherent traffic prioritization mechanism. This method achieves exceptional performance in fully autonomous scenarios but shows limitations in mixed scenarios due to simplified trajectory prediction. Finally, this thesis demonstrates how combining machine learning models with model-based paradigms can overcome these limitations, particularly in challenging mixed traffic scenarios. The ML-enhanced dynamics approximation significantly improves the sampling-based algorithm's performance. The thesis also considers practical implementation aspects, including state estimation and autonomy stack details for field testing evaluation.
The research presents three key methodological contributions. First, this thesis develops an enhanced model-free multi-agent reinforcement learning method that achieves traffic-compliant coordination in mixed environments through policy space constraints rather than reward shaping. This approach significantly outperforms independent methods with individual objectives, validating the effectiveness of V2V communications and shared rewards. Second, this thesis proposes a novel distributed trajectory-based optimization system that enhances state-of-the-art sampling-based algorithms with decentralization capabilities and an inherent traffic prioritization mechanism. This method achieves exceptional performance in fully autonomous scenarios but shows limitations in mixed scenarios due to simplified trajectory prediction. Finally, this thesis demonstrates how combining machine learning models with model-based paradigms can overcome these limitations, particularly in challenging mixed traffic scenarios. The ML-enhanced dynamics approximation significantly improves the sampling-based algorithm's performance. The thesis also considers practical implementation aspects, including state estimation and autonomy stack details for field testing evaluation.
Version
Open Access
Date Issued
2025-02-11
Date Awarded
2025-12-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Ochieng, Washington
Publisher Department
Department of Civil and Environmental Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)