Online Reinforcement Learning-Based Real-Time Stabilization Framework for Quadrotors

Abstract

This study presents a framework based on adaptive learning in real-time for quadrotor stabilization, capable of optimizing continuous parameters in response to dynamic environmental conditions. Using high-fidelity sensor measurements, the proposed architecture employs an actor-critic paradigm to derive optimal stabilization policies without prior knowledge of system dynamics. Specifically, critic and actor networks facilitate the simultaneous evaluation and improvement of policies, allowing the system to mitigate aerodynamic perturbations - provided via high-velocity fan systems - without human intervention. To enhance computational efficiency and provide millimeter-level feedback, a high-precision localization system is integrated into the control loop, serving as the primary positioning reference. The Artificial Intelligence (AI) contribution lies in the model-independent Integral Reinforcement Learning (IRL) weight update laws, while the engineering application is demonstrated through real-time stabilization of a nano-quadrotor under varying aerodynamic loads. The efficacy of the framework was rigorously evaluated by varying the intensity of the wind fan within a controlled indoor flight facility. Experimental results demonstrate that the online Reinforcement Learning (RL) agent successfully identifies and adapts the control parameters to maintain stability, providing a robust foundation for model-independent Unmanned Aerial Vehicle (UAV) control in dynamic environments.