A relatively simple architectural space was modeled and used to compare the effects of spatial training in simulations versus training in the real world. Thirty-five subjects were trained in one of the following conditions: real world (RW), virtual environment (VE), nonimmersive virtual environment (NVE), and model (Mod). The VE condition made use of a head-mounted display to view the simulated environment, while the NVE condition used a desktop monitor. In the Mod condition, the subject viewed and could manipulate a 3-D model of the space, viewed from a desktop display. The training-transfer tasks, performed after brief unstructured exposure to the actual space or to one of the simulations, consisted of estimating the bearing and range to various targets in the real space from various spatially distributed stations, each such pair of estimates constituting a subtask of the overall transfer task. Results obtained from each of the four training conditions proved to be roughly the same. Training in any one of the simulations was comparable to training in the real world. Independent of training condition, there was a strong tendency among subjects to underestimate range. Variability in range errors was dominated by differences among subjects, whereas variability in bearing errors was dominated by differences among subtasks. These results are discussed in the context of plans for future work.