2025/09/13 by Yang, Meiyi, Jichao Zhang, Shijie Zang
Engineering · #Computer Engineering #Engineering #Robotics
paper · doi:10.17605/osf.io/yw3t2
This project investigates a data mining–driven path planning algorithm for intelligent mobile robots operating in dynamic and complex environments. Traditional approaches such as A* and RRT often face limitations in predicting obstacle behaviors, maintaining real-time adaptability, and balancing multiple objectives including energy efficiency, safety, and trajectory smoothness. To address these challenges, we propose a hierarchical fusion architecture consisting of three layers. Cognitive layer: is multimodal perception through LiDAR–vision fusion and enhanced fuzzy C-means clustering, enabling semantic map construction and dynamic obstacle behavior prediction. Decision layer is a two-level Hierarchical Deep Deterministic Policy Gradient (HDDPG) framework for adaptive global task planning and local resource allocation. And Execution layer is a hybrid strategy combining an improved Artificial Potential Field (APF) with Q-learning–based parameter adaptation in the Dynamic Window Approach (DWA), ensuring both global path optimality and local obstacle avoidance. Extensive evaluations in ROS/Gazebo simulations and TurtleBot3 physical experiments confirm significant improvements over baseline algorithms. Results demonstrate higher success rates under high obstacle density, reduced energy consumption (7.1–20.8 Wh vs. 25–48 Wh), lower path curvature, and consistently safer navigation distances. This work contributes a new paradigm for data-driven, adaptive path planning in robotics, bridging multimodal perception, reinforcement learning, and dynamic optimization. It provides technical evidence for robust autonomous navigation in real-world scenarios such as logistics, urban mobility, and disaster response.