Residual Disturbance Learning for Robust Quadrotor Trajectory Tracking Under Multi-Frequency Aerodynamic Perturbations

Abstract

Trajectory tracking remains difficult for outdoor quadrotors exposed to aerodynamic disturbances. Thermal updrafts, wind fluctuations and vibration degrade tracking accuracy and limit what conventional disturbance-rejection strategies can compensate. This paper studies how residual disturbance learning improves robustness within a hierarchical Backstepping–LPV-MPC control architecture. Rather than replacing the analytical disturbance observer, a BiLSTM–LSTM network learns the disturbance components that the observer leaves uncompensated, using the temporal structure of the flight data. A stability analysis shows that the resulting closed-loop system remains uniformly ultimately bounded. Simulations over diverse trajectories and disturbance conditions show improved disturbance estimation and more robust trajectory tracking, with consistent performance across operating scenarios. Learning the residual of an analytical observer is therefore a practical route to robust quadrotor operation under multi-scale aerodynamic disturbances.