To facilitate efficient and safe navigation of legged robots on compliant terrains, a robust predictive framework for their dynamic behavior is crucial. However, the modeling of locomotion on granular media is challenging due to the substrate’s deformation upon contact with robotic appendages, resulting in complex interactions. While high-fidelity physics-based models can capture these interactions, the associated computational demands hinder real-time implementation. In this work, a data-driven methodology is employed to determine robot appendage trajectories on uneven terrains. In particular, data from simulations is integrated with noisy sensor measurements to predict leg contact forces with granular media, in a hybrid approach involving both offline and online phases. During the offline phase, singular value decomposition is used to reduce the dimensionality of physics-based simulation data. Gaussian Process regression is then leveraged to model the evolution of parameters in a low-dimensional space with respect to various operational conditions, creating a surrogate model for preliminary force estimates. Finally, in the online phase, a reduced-order particle filter is used to assimilate experimental data for real-time updates. The efficacy of this framework is demonstrated through test cases, in which the gait responses of robotic appendages interacting with granular media are examined.