Robust State Estimation with Learning Observation Model-Based FGO