Wind disturbance-resilient flight control for small and medium-sized UAVs in plateau canyon environments using deep reinforcement learning

基于深度强化学习的中小型无人机高原峡谷抗风飞行控制

School of Aeronautics, Northwestern Polytechnical University, Xi'an 710072, China

Received: 28 April 2025

Abstract

Fixed-wing unmanned aerial vehicles (UAVs) featured in long endurance and extended range, demonstrate notable advantages in performing wide-area surveillance missions over complex terrains such as plateau canyons. However, the presence of strong and highly variable wind fields in such environments poses serious challenges to flight safety and trajectory stability. The present study focuses on lateral trajectory control in typical plateau canyon wind environments and proposes a compensation-based deep reinforcement learning (DRL) control strategy grounded in the L1 guidance law framework. To achieve both model fidelity and efficient training, the control policy is trained in an environment composed of a simplified dynamics model and a wind field model retaining key canyon characteristics, guided by a reward function tailored to lateral trajectory control. Then the trained policy is successfully transferred to a six-degree-of-freedom high-fidelity model and a hardware-in-the-loop (HIL) simulation platform for validation. The results show that the present control strategy effectively suppresses wind-induced disturbances in plateau canyon environments. Under extreme lateral wind conditions with a maximum crosswind speed of 16 m/s, the trajectory deviation is reduced to only 28.6% comparing with that by using the traditional L1 method. The results further highlight the method's strong transferability, robustness, and practical feasibility for engineering applications.

摘要

固定翼无人机具有续航时间长、距离远等特点, 在高原峡谷等复杂地形区域的大范围巡察任务中更具优势。然而该类地形常伴随强烈多变的风场, 严重威胁无人机的飞行安全与轨迹稳定性。针对以上问题, 开展了面向典型高原峡谷风场环境的深度强化学习(DRL)横航向轨迹控制研究, 提出一种以L1制导律为基准的补偿式DRL控制策略。以控制侧向轨迹为目标设计评价函数, 基于简化动力学模型与保留峡谷风场特征的简化风场模型开展策略训练, 在保证建模精度的同时实现快速训练, 所训练策略成功迁移至六自由度高保真模型及半物理仿真试验平台进行验证。试验结果表明, 所提轨迹控制策略对高原峡谷风场扰动具有良好的抑制效果: 在最大侧向风速达16 m/s的扰动环境中, 其轨迹偏差仅为传统L1方法的28.6%, 同时展现出良好的迁移特性、鲁棒性与工程可实现性。