강화학습 보상 설계를 통한 드론 함상 자율 착륙 강인성 개선
Reinforcement Learning Reward Design for Robust Autonomous Shipboard Drone Landing
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
The autonomous landing of multirotor UAVs on ship decks is challenging due to wave-induced deck motion, degraded visibility, Global Positioning System (GPS) interference, and communication uncertainties. To address this, a reward design framework based on reinforcement learning for vertical drone landing on a heaving shipborne platform using the Deep Deterministic Policy Gradient (DDPG) algorithm is developed in this study. The training environment combines a simplified vertical UAV dynamics model with wave profiles generated from the JONSWAP spectrum to enable randomized and realistic heave motion in each episode. To enhance training stability and policy robustness, we introduce a distance-based reward, a strict terminal penalty for failed landings, and hyperparameter scaling consistent with vehicle dynamics. MATLAB simulation results show that the proposed reward design achieves stable policy convergence and landing performance across diverse wave conditions. These results demonstrate the proposed reward model effectively improves the learning efficiency and robustness of autonomous shipboard landing systems.
Abstract
The autonomous landing of multirotor UAVs on ship decks is challenging due to wave-induced deck motion, degraded visibility, Global Positioning System (GPS) interference, and communication uncertainties. To address this, a reward design framework based on reinforcement learning for vertical drone landing on a heaving shipborne platform using the Deep Deterministic Policy Gradient (DDPG) algorithm is developed in this study. The training environment combines a simplified vertical UAV dynamics model with wave profiles generated from the JONSWAP spectrum to enable randomized and realistic heave motion in each episode. To enhance training stability and policy robustness, we introduce a distance-based reward, a strict terminal penalty for failed landings, and hyperparameter scaling consistent with vehicle dynamics. MATLAB simulation results show that the proposed reward design achieves stable policy convergence and landing performance across diverse wave conditions. These results demonstrate the proposed reward model effectively improves the learning efficiency and robustness of autonomous shipboard landing systems.