In unmanned warehouse systems, it is important to solve the collision, deadlock and path planning problems between automated guided vehicles (AGVs). This paper presents a method of modeling AGV system in warehouse environment with Petri net, which effectively solves the problem of conflict when AGV transports goods. On this basis, a multi-agent deep reinforcement learning AGV path planning framework is proposed. The AGV path planning problem is regarded as a partially observable Markov decision process, and the deep deterministic policy gradient algorithm is extended to multi-agent systems. Observation space, state space, action space and reward function of AGV are designed to realize AGV conflict-free path planning in Petri Net. Due to the addition of feedback after Petri Net trigger condition when setting the reward function, the congestion generated when AGV transported goods is greatly reduced, and the total amount of delivery in the warehouse is increased within the specified time. In addition, because the proposed framework sets the path branch points as agents, it can effectively cope with the random generation of multiple task starts and the change of the number of AGVs in the environment, and improve the generalization ability of neural networks. Simulation experiments were carried out on AnyLogic software platform. The feasibility and effectiveness of the path planning method were verified by comparing the cargo transportation situation under different AGV scales and the control experiments with or without Petri Net condition positive and negative feedback in the reward function.