Design of a Layered System for Drone Inspection Path Optimization and Autonomous Obstacle Avoidance on Construction Sites: Global A* and Local Dueling Double DQN Methods

Abstract

Conducting periodic aerial inspections of construction sites using unmanned aerial vehicles (UAVs) can significantly improve the monitoring of construction progress and safety oversight. However, the chaotic and partially observable construction site environment, combined with moving workers, vehicles, and swinging tower crane loads, makes purely geometric path planning methods highly vulnerable. This paper proposes a hierarchical planning and control framework that decouples the problem into global and local levels. The global layer runs an octile-based A* search on a static occupancy grid constructed from the site model to generate an energy-efficient reference path that traverses all inspection viewpoints; The local layer models close-range obstacle avoidance as a Markov decision process and trains a Dueling Double deep Q-network (D3QN) with experience replay, whose composite reward simultaneously rewards progress toward the goal, penalizes collisions, and accounts for nearby hazards and sudden changes in heading. Next, the Chaikin corner-cutting phase is used to refine the discrete strategy output into a flyable trajectory. The framework was implemented and evaluated in a simulated construction site measuring 80m×80m and containing six dynamic obstacles. In 100 random trials, the proposed method reached the target viewpoint with a success rate of 86.0% and a collision rate of only 14.0%, whereas static A* achieved 69.0%/31.0%, the Artificial Potential Field (APF) achieved 59.0%/27.0%, and standard Deep Q-Network(DQN) achieved 74.0%/18.0%;Furthermore, its path length (208.7 m) and trajectory roughness (16.4 degrees/step) are close to the geometric optimum. The results show that combining global optimality with a locally learned reactive obstacle-avoidance strategy yields safer and smoother patrol flights than either component alone. This study combines global path planning, deep reinforcement learning, and autonomous UAV control to provide a feasible solution for safe decision-making and trajectory optimization in intelligent inspection robots operating in complex dynamic environments. It is of great significance for advancing Artificial Intelligence(AI)-driven intelligent manufacturing, mechatronic equipment, and autonomous robotics technology.