The observable degree of navigation state has a significant effect on the state estimation of GNSS/INS. In order to improve the accuracy of heading of land vehicle, an improved robust cubature Kalman filter (RCKF) method is proposed. First, the resampling-free sigma-point update framework is employed to separate the cubature point update from the Gaussian information limitation, and thus improving the propagation efficiency of the information contained in instantiated points in the iteratively filtering period. Secondly, in order to improve the heading of land vehicle when it travels along a straight-line, the Gaussian process (GP) is introduced into the uncertainty calibration of moment approximation of system model based on state observability analysis. Simulation results indicate that GP-RCKF improves the heading angle obviously when the state observability is weak, and compared with RCKF the heading is improved by 28.9%.