WAGRL: A Wind Adaptive Guidance Framework for Uncertainty-Aware UAV Trajectory Control

Abstract

To address the challenges of wind-induced uncertainty, decision inconsistency, and safety risks in quadrotor trajectory tracking, this study presents a Wind Adaptive Guidance Reinforcement Learning (WAGRL) framework. Unlike conventional controllers relying on fixed aerodynamic models, WAGRL integrates ensemble-based policy learning with uncertainty-guided reinforcement learning to adaptively manage diverse wind conditions. A set of digital expert policies is trained to capture heterogeneous wind-response strategies, and an ensemble-based estimation mechanism quantifies epistemic uncertainty for robust action selection. An entropy-driven adaptive intervention module monitors real-time decision confidence and activates expert guidance when uncertainty exceeds adaptive thresholds. To validate the proposed framework, extensive simulations are conducted under both nominal and wind-disturbed settings, covering square-loop tracking, geometric path following, and gust rejection scenarios. The results show that WAGRL achieves faster convergence, enhanced trajectory accuracy, and smoother attitude responses. Real-world flight tests further confirm that WAGRL maintains stable and reliable performance under outdoor wind disturbances, demonstrating its potential for safe and adaptive UAV control in complex aerial environments.