Online Residual Learning Using Interpretable Reservoir Computing for Quadrotor Control

Quadrotors, valued for their mobility and cost-effectiveness, have found widespread use in applications such as aerial photography and infrastructure inspection. However, their complex and nonlinear dynamics make them sensitive to uncertainties. In dynamic scenarios with online data but little-to-no prior knowledge of these unknowns, data-driven approaches show promise in both system identification and controller design. Nonetheless, the black-box nature of deep learning poses challenges for trust and generalizability. In this paper, we introduce a novel tracking controller featuring an online learning module for quadrotor residual dynamics. This module, implemented using deep Echo State Network (ESN), enhances adaptability to unforeseen scenarios, thereby extending the applicability of the controller to a broader range of situations. Furthermore, we employ post-hoc interpretation techniques tailored for the ESN to improve trustworthiness. This is achieved through dynamic system analysis and visualization of network predictions. Simulation results demonstrate the effective tracking of a figure-8 trajectory with online learning compensating for various non-parametric uncertainties, which showcases the potential of the proposed approach and estab-lishes a foundation for the real-world testing in future.