Hybrid model-based and data-driven disturbance prediction for precise quadrotor trajectory tracking
Abstract: Precise quadrotor trajectory tracking is required to safely perform agile and maneuverable tasks, even in confined and complex places, against external influences, such as aerodynamic effects. In this study, we propose a hybrid model-based and data-driven disturbance prediction scheme to compensate for the mutual effects of quadrotor dynamics and aerodynamics intricately intertwined, and hence effectively achieve precise quadrotor trajectory tracking. Specifically, to accurately predict aerodynamic effects, we employ spectrally-normalized neural networks with rectified linear unit (ReLU) activation functions, trained on a dataset derived from actual flight data. The employed spectral-normalization ensures stable training while imposing a Lipschitz continuity constraint on the network. From the predicted disturbances, feasible state and control trajectories were generated using the differential flatness property of a mathematical quadrotor model. The disturbance prediction and trajectory generation were repeated until convergence. The convergence is also proved mathematically based on the Lipschitz continuity constraint of the network. The acquired state trajectories are then used as a feed-forward term for a cascaded nonlinear feedback controller to track a given position reference trajectory. To validate the effectiveness of the proposed approach, trajectory tracking experiments were conducted with a lab-built quadrotor, and the results were compared with those from control-, model-, and learning-based disturbance rejection approaches. The experiments have confirmed that the proposed scheme provides highly predictive and well-trained control efforts for precise quadrotor maneuvering. Consequently, it achieves a maximum improvement of 30.24% compared to other methods.
Loading