드론의 경사면 착륙을 위한 PPO 기반 도메인 랜덤화 강화 학습 환경 및 보상함수 설계

초록

Landing a drone on a slope is a critical task in unpredictable environments such as disaster relief or military operations. Traditional control methods have limitations handling nonlinear slope dynamics. This study utilized PPO to train slope landing, designing a combined sparse-dense reward function for position, velocity, and orientation alignment, and applying domain randomization with random slope angles ranging from 0 to 30 degrees. Results showed 99.67% landing success rate across the range of 0 to 30 degrees, outperforming curriculum learning methods that exhibited overfitting, with PPO superior to SAC and A2C algorithms. This research presents a generalized drone landing approach without curriculum learning, with future work on 3D environment extension and real-world validation.

키워드