Autonomous exploration of mobile robot in unknown environment is an important problem in robot research. The exploration strategy of unknown environment has some problems, large localization uncertainty, long decision-making time, slow exploration rate and poor robustness. These problems can be effectively avoided by using graph neural network and deep reinforcement learning. Based on this architecture, a new robot autonomous exploration method is proposed in this paper. In this method, the concept of virtual landmarks is introduced to describe the environment map. Firstly, the existence probability of landmarks represented by virtual landmarks and its estimation of uncertainty are updated based on EM (expectation maximum) exploration strategy, and then the exploration map is constructed according to the information about the robot and environment, which is used as the data structure to represent the environmental information and improve the robustness of the exploration strategy to the map size. Then the exploration graph is input into the gated graph neural network to mine the hidden information between the data and help update the graph node information. Finally, a double deep Q-learning network (ddqn) is constructed based on the gated graph neural network, so as to reduce the impact of noise on action selection and improve the performance of the exploration strategy. We carried out experiments in the simulation environment, and compared with other autonomous exploration methods. Experiments show that the exploration strategy has short decision time and good robustness to map size changes, and it can improve the exploration speed and achieve higher map accuracy.