State estimation refers to all algorithms calculating the internal state of a robot. Since the true state is always unknown, it has to be estimated based on a series of noisy and inaccurate measurements. This is a crucial task in mobile robotics and a more accurate state estimation almost always leads to an improved robot performance. At present, model-based state estimation algorithms are the state of art in legged robotics, while hybrid-driven approaches combining traditional model-based and novel data-based methods have not been comprehensively explored yet. Therefore, the aim of this thesis is the development, implementation and evaluation of a hybrid-driven state estimation algorithm for the RH5 humanoid robot. A data-based machine learning (ML) component should be integrated into the currently implemented model-based framework. For supervised training of the ML component, first synthetic data from simulation and after that real RH5 data will be used for transfer learning while the ground truth data will be gathered by the Qualisys motion capture system. Finally, some variants of the new hybrid-driven algorithm should be evaluated against the current model-based algorithm for certain motions such as walking and single leg balancing.