Robust autonomous navigation for UAVS in urban environments using machine learning

Abstract

Urban canyons, characterized by their high-rise buildings and dense infrastructure, pose significant challenges to Unmanned Ariel Vehicle (UAV) navigation primarily due to the obstruction and interference of Global Navigation Satellite Systems (GNSS) signals. These challenges include multipath effects, where signals bounce off multiple surfaces before reaching the receiver, leading to inaccurate positioning data, and direct signal blockage, which causes intermittent loss of satellite visibility. The reliance on GNSS for UAV navigation in such environments is further compromised by potential denial of service, whether intentional or unintentional, through interference. These limitations highlight the need to explore alternative or supplementary navigation technologies to ensure the safe and efficient operation of UAVs in urban settings. To address these challenges, this research begins by focusing on reducing sensor errors in both GNSS and Inertial Navigation Systems (INS). Specifically, data-driven approaches have been proposed to mitigate errors from individual sensors. For INS, a Gated Recurrent Units (GRU)-based error prediction algorithm, trained on labelled sensor data, is utilised to reduce the effects of random walk caused by the sensor’s stochastic noise, which accumulates over time. For GNSS, a GRU classification algorithm is employed to detect and exclude Non-Line-Of-Sight (NLOS) signals from the list of available GNSS satellites. Once NLOS signals are removed, the remaining signals are ranked based on their contribution to Geometric Dilution of Precision (GDOP). The top 10 ranked signals are then used to compute the receiver's position, resulting in improved positioning accuracy and greater reliability over time. Building on these initial advancements, this thesis proposes a robust federated multi- sensor fusion architecture to further enhance positioning accuracy. The proposed architecture integrates GNSS and INS positioning, with additional enhancements provided by GRU-based corrections. Furthermore, Monocular Visual Odometry (VO) and a barometer are incorporated to refine positioning in GNSS-degraded environments. A novel Bayesian Long Short-Term Memory (LSTM) model with Monte Carlo dropouts is introduced to generate stochastic outputs, enabling the system to estimate navigation integrity by predicting position distributions and calculating Protection Levels (PL). By quantifying the reliability of navigation data, this approach ensures the safety and accuracy required for autonomous UAV operations in complex urban environments. The proposed architecture was validated using a comprehensive simulation setup. This includes GNSS Hardware-in-the-Loop (HIL) simulations with the Spirent GSS7000 and OKTAL-SE Sim3D for realistic modelling of multipath effects, as well as high-fidelity VO sensor data generated through AirSim and Unreal Engine. These simulations replicate urban canyon scenarios, accounting for variations in GNSS signal quality and adverse weather conditions, providing a robust environment to test navigation performance. The results indicate significant advancements in UAV navigation, including a 22.1% reduction in the 95th percentile horizontal error compared to state-of-the-art federated fusion approaches and a 98% reduction in misleading integrity information relative to traditional Extended Kalman Filters (EKF). These improvements demonstrate the architecture’s ability to deliver accurate, reliable navigation with a quantifiable measure of integrity, even in the most challenging urban environments. This research has profound implications for the future of Urban Air Mobility (UAM). UAM applications, such as passenger air taxis, last-mile deliveries, and emergency response, rely on precise and reliable navigation systems to ensure safe operation in urban settings. The ability to provide integrity information—quantifying the reliability of positioning data—is crucial for regulatory compliance, operational safety, and public trust in autonomous aerial systems. By addressing the challenges of GNSS-degraded environments and offering a robust framework for navigation with measurable integrity, this study lays a strong foundation for UAM. It enables scalable, reliable, and safe integration of UAVs into urban airspaces, bringing the vision of autonomous urban transportation closer to reality.