Based on the analysis of the relevant intelligent prediction algorithm, it is found that existing algorithms don’t consider the potential relevance of the relationship among the virtual scene node, and they reduce the prediction accuracy of the algorithm. To tackle this problem, based on a new cube virtual scene model，an intelligent roaming algorithm is proposed. The algorithm establishes complete BiTree structure for each node, uses the potential association between the nodes to predict the object which may be operated next, protracts the predicted object in advance and then transfers it into the memory. This algorithm effectively compensates for the shortcomings of the related algorithm, thereby imptoves the efficiency and real time of roaming virtual scene. The algorithm is highy accurate.