This dissertation describes, implements and analyzes a comprehensive system for perceptually-driven virtual reality simulation, based on algorithms which dynamically adjust level of detail (LOD) for entity simulation in order to maximize simulation realism as perceived by the viewer. First we review related work in simulation LOD, and describe the weaknesses of the analogy that has traditionally been drawn between simulation LOD and graphical LOD. We describe the process of "perceptual criticality modeling" for quantitatively estimating the relative importance of different entities in maintaining perceived realism and predicting the consequences of LOD transitions on perceived realism. We present heuristic cognitive models of human perception, memory, and attention to perform this modeling. We then propose the "LOD Trader", a framework for perceptually driven LOD selection and an online approximation algorithm for efficiently identifying useful LOD transitions. We then describe "alibi generation", a method of retroactively elaborating a human agent's behavior to maintain its realism under prolonged scrutiny from the viewer, and discuss its integration into a heterogeneous perceptually driven simulation. We then present the "Marketplace" simulation system and describe how perceptually driven simulation techniques were used to maximize perceived realism, and evaluate their success in doing so. Finally, we summarize the dissertation work performed and its expected contributions to real-time modeling and simulation environments.