Abstract: To solve the problem of multi-AGV route planning and conflicts in unmanned warehouses, with the objective of minimizing the total travel time, a multi-AGV route planning model is established, and an improved DQN algorithm based on dynamic decision-making is proposed. An empirical knowledge model based on static route planning of a single AGV is designed to guide the learning and exploration direction of AGVs. It avoids conflicts and obstacles for AGVs in advance, and accelerates the convergence of the proposed algorithm. Also, a conflict resolution strategy based on the shortest total travel time is proposed to fundamentally solve the problem of multi-AGV route conflicts and deadlocks. Finally, a grid map of an unmanned warehouse is established for simulation experiments. Results show that, compared with other DQN algorithms, the convergence speed of the proposed model and algorithm is increased by 13.3%, and the average loss value is reduced by 26.3%. This result indicates that the model and algorithm are conducive to avoiding and resolving the conflicts of multi-AGV route planning in unmanned warehouses, reducing the total travel time of multiple AGVs and having important guiding significance to improve the efficiency of unmanned warehouse operations.