Memories of places often include landmark cues, i.e., information provided by the spatial arrangement of distinct objects with respect to the target location. To study how humans combine landmark information for navigation, we conducted two experiments: To this end, participants were either provided with auditory landmarks while walking in a large sports hall or with visual landmarks while walking on a virtual-reality treadmill setup. We found that participants cannot reliably locate their home position due to ambiguities in the spatial arrangement when only one or two uniform landmarks provide cues with respect to the target. With three visual landmarks that look alike, the task is solved without ambiguity, while audio landmarks need to play three unique sounds for a similar performance. This reduction in ambiguity through integration of landmark information from 1, 2, and 3 landmarks is well modeled using a probabilistic approach based on maximum likelihood estimation. Unlike any deterministic model of human navigation (based e.g., on distance or angle information), this probabilistic model predicted both the precision and accuracy of the human homing performance. To further examine how landmark cues are integrated we introduced systematic conflicts in the visual landmark configuration between training of the home position and tests of the homing performance. The participants integrated the spatial information from each landmark near-optimally to reduce spatial variability. When the conflict becomes big, this integration breaks down and precision is sacrificed for accuracy. That is, participants return again closer to the home position, because they start ignoring the deviant third landmark. Relying on two instead of three landmarks, however, goes along with responses that are scattered over a larger area, thus leading to higher variability. To model the breakdown of integration with increasing conflict, the probabilistic model based on a simple Gaussian distribution used for Experiment 1 needed a slide extension in from of a mixture of Gaussians. All parameters for the Mixture Model were fixed based on the homing performance in the baseline condition which contained a single landmark. from the 1-Landmark Condition. This way we found that the Mixture Model could predict the integration performance and its breakdown with no additional free parameters. Overall these data suggest that humans use similar optimal probabilistic strategies in visual and auditory navigation, integrating landmark information to improve homing precision and balance homing precision with homing accuracy.