Abstract: For the dimensionality curse problem of belief state space scale of partially observable Markov decision process (POMDP), a factored belief states space compression (FBSSC) algorithm based on dynamic Bayesian network (DBN) is proposed according to the decomposable features and dependent relationship of the belief state variables. Based on the building of the graph of dependent relationship among variables, the algorithm removes the redundant edges by detecting the dependent relationships, and decomposes the joint probability of transition function into the product of several conditional probabilities, which realizes the lossless compression of belief states space. Comparison experiments and RoboCupRescue simulation results show that the algorithm has the characteristics of lower error rate, higher convergence, and general applicability.