<p>Traditional adaptive and robust control methods sometimes struggle to address the complex and dynamic uncertainties inherent in quadrotor flight, frequently resulting in overly conservative or unstable behavior due to reliance on presumed uncertainty bounds. To overcome these limitations, we propose a unified online-learning-based control framework that integrates data-driven uncertainty estimation with a nominal flight controller. Furthermore, we systematically evaluate four representative learning methods, namely Extreme Learning Machines (ELMs), Radial Basis Function Neural Networks (RBFNNs), Echo State Networks (ESNs), and Gaussian Processes (GPs), spanning both parametric and non-parametric techniques. Through extensive simulation and hardware experiments, we benchmark these methods in terms of (i) uncertainty estimation accuracy, (ii) trajectory tracking performance, and (iii) real-time computational efficiency. Our results reveal key trade-offs: GPs yield highest accuracy but require approximations for real-time use. ELMs and RBFNNs enable fast inference but suffer initialization sensitivity. ESNs provide the best balance, offering both stability and real-time performance. This work underscores the critical role of model selection in learning-based flight control, providing insights for deploying these algorithms in resource-constrained aerial systems operating under uncertainty.</p>