Onboard UAV State Estimation and Trajectory Prediction Using Multi-Task Reservoir Computing

The rapid advancements in unmanned aerial vehicle (UAV) technology have led to their use in different applications, ranging from critical infrastructure monitoring and search-and-rescue to remote sensing. However, UAV operations are easily affected by environmental conditions and sensor malfunctions that lead to the need for an efficient, accurate, and trustworthy state identification and trajectory prediction framework. This work proposes an innovative real-time UAV system with the two-fold objective of state identification and trajectory prediction, employing a lightweight multi-task learning framework based on reservoir computing (RC) network architecture to achieve reliable and robust UAV operations. Specifically, custom multi-task models are designed and finetuned to obtain multi-modal sequential data (related to drone movement) by exploiting the ability of shared feature learning in an RC-based network architecture to accurately achieve and enhance real-time and simultaneous drone state classification and trajectory prediction. A real-world dataset is also created to train and evaluate the proposed multi-task model, encompassing drone movements recorded during numerous outdoor experiments. Finally, a UAV prototype system is implemented and extensively tested in a real-world environment to demonstrate its enhanced performance in trajectory prediction and drone state identification compared to existing methods.