Abstract
Multimodal wind disturbances complicate quadrotor trajectory tracking through partial observability and shifts in disturbance distributions. To address these challenges, this study develops hierarchically structured distributionally robust proximal policy optimization (HSDR-PPO), integrating observation history with separate Kullback–Leibler constraints on wind-class probabilities and within-class disturbance distributions. The resulting adversarial probabilities weight policy updates while rollout collection maintains fixed candidate coverage, enabling history-conditioned robust control without wind measurements or scenario labels. Simulations using a Crazyflie 2X model with prescribed external forces cover five training seeds and eight test groups. HSDR-PPO achieves the lowest mean episode cost and mean physical failure rate, reducing cost by 13.0–19.2% relative to the best baseline in each group. Under intensity-matched colored-disturbance shifts, worst-decile cost decreases by 32.9% relative to Joint-KL PPO. Ablations support contributions from observation history and hierarchical reweighting, while evaluations under unseen chirps, a figure-eight reference, and an extended task duration demonstrate transfer without retraining. A policy trained in simulation configured for the experimental platform is subsequently deployed on a physical quadrotor, achieving a three-dimensional position root mean square error of 0.071 m in a representative 48 s outdoor flight. These results demonstrate improved robustness under the evaluated simulation conditions and successful deployment for outdoor trajectory tracking.