The goal of the project is to explore the possibilities of multi-fidelity machine learning for wheel-soil interactions. The scarcity of high-fidelity data in the problem leads to attempts to enhance models based on lower-fidelity data, which is easier to obtain. With multiple overlapping features present in the Mars rover dataset, dimensionality reduction techniques are implemented to enhance performance and combat memory problems. The correlation between low and medium fidelity data has been explored to assess the propriety of multi-fidelity methods along with various Gaussian Processes kernels' fitness for the task. Along with Gaussian Processes, the Neural Network approach has been introduced for the multi-fidelity regression task of predicting the force of the Mars rover incorporating both tabular and image data. The models have been benchmarked and compared according to their accuracy, the number of required data points, and possible further improvements.