강화학습을 사용한 도달 가능 집합 기반 자동 수직 주차 경로 계획

초록

Automated parking systems (APSs) require precise path planning, particularly in vertical parking scenarios where backward maneuvering is essential. Traditional geometric path planning methods often face challenges related to discontinuity and computational complexity. To overcome these limitations, this study proposes a reinforcement learning-based approach that leverages the concept of a reachable set to generate flexible and adaptive backward paths. Utilizing the deep deterministic policy gradient (DDPG) algorithm, the agent learns to navigate from diverse initial poses to predefined intermediate poses, enabling smooth and efficient trajectory generation. By interpreting the reachable set as a continuous region rather than discrete points, the method supports robust path planning even in constrained environments. This approach shows significant potential for real-time application and future extension to dynamic obstacle avoidance in complex parking scenarios.

키워드