부산대학교 도서관

Navigation of mobile robots within crowded environments is an essential task in various use cases, such as delivery, health care, or logistics. Common navigation approaches have weaknesses when deployed as a standalone system. For instance, global planners excel in planning collision-free paths in static environments when the map is perfectly known but can not consider dynamic or unknown obstacles. Learning-based local planners have shown superior performance in dynamic obstacle avoidance but can not handle long planning horizons due to their myopic nature. To address these issues, we adopt a hierarchical motion planning framework to handle complex long-range navigation problems. Three modules are designed for different planning horizons leveraging different observations. First, an extended hybrid A-Star approach is proposed to efficiently search for an optimal solution in the time-state space and produce reasonable landmarks for the subsequent modules. Second, an intermediate planner is proposed, which utilizes Delaunay Triangulation to encode obstacles and provides safer and more robust subgoals for the third module, the learning-based local planner trained using Deep Reinforcement Learning. The proposed approach is compared to two baseline navigation systems and outperforms them in terms of safety, efficiency, and robustness.