An Onboard UAV Multi-task System for Trajectory Prediction and State Estimation Employing Transformer- and Reservoir-based Networks

With the increasing technological advancements and deployment of unmanned aerial vehicles (UAVs) in different applications (e.g., delivery services, surveillance, critical infrastructure monitoring, military operations, and search-and-rescue operations), a robust, lightweight, and accurate system for UAV state identification and trajectory prediction is becoming a mandate. This work introduces an onboard real-time multi-task learning framework that combines Transformer neural networks and reservoir computing (RC) architectures to enhance outdoor UAV operations. The proposed system employs the Transformer-based architecture to capture the temporal dependencies in sequential data for long-term horizons, while RC-based networks are utilized to ensure robust and real-time performance. Specifically, custom multi-task models are implemented and fine-tuned to collect multi-modal sensor measurements, aiming to enhance the two-fold objective of simultaneous UAV state identification and trajectory prediction through shared feature learning. A dataset comprised of measurements collected from real-world UAV operations (in 3D space) under various conditions is employed to train and evaluate the proposed system. Furthermore, a prototype onboard UAV system is implemented and tested in real-world field experiments. Extensive evaluations of the experimental results demonstrate that the onboard framework achieves accuracy close to the ground truth in both state identification and trajectory prediction, consequently showcasing its potential for practical applications.