Terrain Feature-Driven Real-time Waypoint Generation via Deep Reinforcement Learning

초록

With the increasing demand for unmanned vehicles, the development of autonomous guidance and control technology is actively progressing. Path planning in complex terrain faces challenges with classical methods, including high computational complexity and difficulty in utilizing terrain features globally. Additionally, achieving mission execution efficiency, such as immediate path generation under new conditions, remains challenging. This study proposes a 3D path planning approach using deep reinforcement learning to address these challenges. We define an optimal waypoint generation problem based on height map and design an attention-based CNN (Convolutional Neural Network). The process of generating waypoints is trained through the SAC(Soft Actor-Critic) algorithm by extracting terrain features from height map. Performance validation is conducted through comparison with the PSO (Particle Swarm Optimization) algorithm and A* algorithm. Monte Carlo simulation in both simple and complex terrain scenarios demonstrates that the proposed method generates near-optimal waypoints while significantly improving computational efficiency.

키워드