Design and Deployment of a Hybrid Neural Network for Real-Time GPS Correction in Slow Moving, Low SWaP, UAV Flight Systems

Abstract

The global surge in demand for unmanned aerial vehicles (UAVs), projected to reach $91 billion by 2033, spans diverse applications in urban environments. However, achieving accurate horizontal geo-localization in urban settings remains challenging due to poor GPS performance, creating a significant barrier to effective drone opera-tions, even in static or slow-moving deployments. Recreational-grade GPS systems, commonly used in low Size, Weight and Power (SWaP) UAVs, suffer from residual errors exceeding 100 meters in urban environments, severe-ly limiting real-time navigation accuracy and drone autonomy. Although capable of good accuracy in conven-tional aerospace deployments, existing traditional aerospace-ready GPS receivers are expensive, power-intensive, and heavy, making them unsuitable for low SWaP drone deployment. Existing performance-enhancing post-processing techniques for recreational GPS are limited by the real-time requirements and the noise, power, and computational constraints inherent to drone operations. This underscores the critical need for accurate, ro-bust, lightweight, and power-efficient urban localization solutions for UAVs. In this work, we present a hybrid Integrated Recurrent Neural Network–Deep Neural Network (IRNN-DNN) architecture, capable of running on em-bedded microcontrollers with recreational-grade L1C GPS receivers. Leveraging Long Short-Term Memory (LSTM) networks, the IRNN-DNN model captures temporal dependencies and spatial error patterns across multi-ple concurrent GPS receivers, allowing real-time error correction. Field tests with MTK3339 GPS receiver on UAVs in static test conditions showed, with a 2D residual error of 1.71 meters at 10 Hz, a 30% improvement over existing models—validate the IRNN-DNN approach as a promising embedded solution for enhancing urban lo-calization performance in constrained UAV applications.