Hierarchical Reinforcement Learning for Navigation among Movable and Immovable Obstacles

Abstract

Path planning is a key technique for vehicle navigation, and significant research has focused on how to manage obstacles. Previous methods have addressed either removing movable obstacles or avoiding immovable ones. However, real-world environments contain both movable and immovable obstacles. To address this, we propose a path planning system for Navigation Among Movable and IMmovable Obstacles (NAMIMO) based on hierarchical reinforcement learning. In our system, lower-level agents generate paths for the vehicle to either avoid or remove obstacles, while the higher-level agent dynamically selects which lower-level agent to utilize based on the movability of the obstacle, incorporating visual and linguistic knowledge.