Goal-directed navigation requires integrating information from a variety of internal and external spatial cues, representing them internally, planning, and executing motor actions sequentially. However, a comprehensive computational account of how these processes interact in an ambiguous, uncertain, and noisy environment giving rise to biases and variability observed in navigation behavior is currently unavailable. In this paper, we introduce an optimal control under uncertainty model, which provides a computational-level explanation of how landmarks and path integration interact and are combined to reduce variability in navigation. We apply our model to trajectory and end-point data from three previously published studies that employ a variant of the triangle completion task with landmarks. Contrary to observer models, which attribute human endpoint variability to perceptual cue combination processes only, this dynamic Bayesian actor model provides a unifying account of a wide range of phenomena found in this task by considering variability in perception, action, and internal representations jointly. Taken together, these findings have wide-ranging implications for the analysis and interpretation of human navigation behavior, including resolution of seemingly contradictory results on cue integration in navigation.