Gaussian process (GP) models are widely used for regression tasks. Recently, several authors have developed GP models that fuse information from sources of different fidelity levels, with successful application in building a surrogate model for the dynamics of a rover wheel traversing soft soil. In its current version, noise in the inputs, which are given as signals, is removed by a low-pass filter. In this thesis, we explore an alternative approach by treating the inputs as randomly distributed around the smoothed signals, which might recover valuable information in the high-frequency components of the input signals. Since standard GP models assume deterministic inputs, they are not suited for this task. Nonetheless, multiple authors have developed GP regression models that specifically address uncertain inputs. We review a set of these models and evaluate their predictive performance against standard GP regression models trained on smoothed input signals in a synthetic example. Furthermore, we discuss two multi-fidelity GP regression models and develop extensions to make them account for uncertain inputs. Using these models, we conduct another synthetic experiment to assess the predictive performance of our extensions in comparison to the baseline models trained on smoothed input signals. The results of both synthetic examples suggest that employing input-uncertainty-aware models is the preferable choice when the input data consists of noisy signals. Furthermore, we apply our best-performing multi-fidelity model to rover wheel-soil interaction data. Preliminary results suggest that our model may serve as a surrogate model with enhanced predictive performance. However, further experiments should be conducted to substantiate this claim. Moreover, the insights gained from our synthetic experiments may also assist researchers from other fields in making informed decisions about model selection for more general applications involving uncertain inputs.