Many training applications of virtual environments (VEs) require people to be able to transfer spatial knowledge acquired in a VE to a real-world situation. Using the concept of fidelity, we examine the variables that mediate the transfer of spatial knowledge and discuss the form and development of spatial representations in VE training. We report the results of an experiment in which groups were trained in six different environments (no training, real world, map, VE desktop, VE immersive, and VE long immersive) and then were asked to apply route and configurational knowledge in a real-world maze environment. Short periods of VE training were no more effective than map training; however with sufficient exposure to the virtual training environment, VE training eventually surpassed real-world training. Robust gender differences in training effectiveness of VEs were also found.