To improve the picking efficiency of AGVs in intelligent warehouse systems, aiming at two sub-problems of AGV-shelf task allocation and multi-AGV conflict-free path planning, this paper introduced order splitting strategy according to the order characteristics and built a mathematical model with the objective of minimizing the total time for AGVs to complete all orders. Firstly, it designed the ASTA to solve the matching problem by determining the shelf priority. Secondly, it proposed an improved Q-Learning algorithm with a greedy parameter and embedded conflict elimination strategy to obtain the optimal conflict-free picking path scheme under the splitting strategy. Finally, through experimental comparative analysis of order set values in the 40 m×40 m warehouse layout, the proposed algorithm was compared with the two existing algorithms. The results indicate that the proposed algorithm reduced the total time for AGVs to complete the all orders by an average of 11.63% and 26.74% respectively, which verified the effectiveness of the splitting strategy. And it was verified that the splitting strategy and the proposed algorithm can effectively alleviate the congestion, reduce the length of traveling paths and improve the picking efficiency, through the comparison of three indexes of the number of AGVs used, the time of complete orders and the percentage of waiting time for path conflict. Furthermore, for the sensitivity on the number of AGVs, it tested the influence of different numbers of AGVs on the travel time and path conflict waiting time. It was found that the number of 19 AGVs is the optimal configuration. These results verify the feasibility of the model and the effectiveness of the proposed algorithm.