A multi-AGV motion planning method, device and system are disclosed. The method of the present disclosure comprises: establishing an object model through reinforcement learning; building a neural network model based on the object model, performing environment settings including AGV group deployment, and using the object model of the AGV in a set environment to train the neural network model until a stable neural network model is obtained; setting an action constraint rule; and after the motion planning is started, inputting the state of current AGV, states of other AGVs and permitted actions in a current environment into the neural network model after trained, obtaining the evaluation indexes of a motion planning result output by the neural network model, obtaining an action to be executed of the current AGV according to the evaluation indexes, and performing validity judgment on the action to be executed using the action constraint rule.