AGV Path Planning Algorithm Based on Improved DQN

In the context of rapid growth in the logistics and e-commerce industries, improving the efficiency of automated warehousing systems, particularly in sorting and throughput, has become crucial. Automated guided vehicles (AGVs) have increasingly replaced human workers in handling and sorting tasks, leading to significant improvements in efficiency and accuracy. Path planning, a core AGV technology, plays a vital role in optimizing logistics operations and reducing transportation costs. Common path planning algorithms, such as Dijkstra and A*, each have strengths and limitations. While Dijkstra’s algorithm guarantees finding the shortest path by exploring all possible nodes, it suffers from high time and space complexity. A*, though more efficient, sometimes compromises optimality by neglecting certain nodes. Current algorithms overlook the impact of excessive turns, which increases travel time. Deep reinforcement learning (DRL), by combining deep learning and reinforcement learning, offers a promising solution to path planning in dynamic, complex environments. DRL can make intelligent decisions without pre-constructed maps, adapting to the environment through training. In this paper, an improved DQN-based algorithm for AGV path planning is proposed, introducing the concepts of a turning factor and a smoothness factor to address the time cost of turns. By enhancing DQN’s state and reward design, the algorithm minimizes the number of turns while maintaining the shortest possible path. Experimental results demonstrate that this approach outperforms traditional A*, modified A*, and standard DQN algorithms, reducing transport time by approximately 21.3% compared to the best-performing baseline, the improved A* algorithm.