The U.S. Army Research Institute is conducting a research program with the goal of using virtual environments (VE) to train dismounted soldiers. To accomplish this goal, the conditions necessary for transfer of training from VE to real world environments must be identified. This paper reports the results of two experiments investigating the use of VE for training spatial knowledge as it relates to learning routes through large buildings. This task is especially relevant to a hostage rescue situation or other missions performed by special operations forces. Both experiments used the same highly detailed computer model of a large office building. In the first experiment, 60 college students first studied directions and photographs of landmarks for a complex route, then rehearsed the route using either the VE model, the actual building, or verbal directions and photographs. Everyone was then tested in the actual building. Building-trained students made fewer wrong turns and travelled less distance than did VE-trained students, who in turn made fewer wrong turns and took less time to traverse the route than did verbally-trained students. In the second experiment, 64 students practiced a different route using either a landmark-oriented or a left/right direction-oriented instructional strategy, and with their field of view either linked solely to body orientation or controlled by both body orientation and head movements. These data indicate that the use of an instructional strategy that increases the amount of exploration of a VE tends to improve route learning. The use of head tracking, however, had no effect on learning. The results indicate that individuals can learn how to navigate through real world places by training in a VE. While the building model was not quite as effective in training subjects as the actual building, it was much better than verbally rehearsing route directions. The results also suggest that instructional strategy is an important determinant of learning in a VE.