Abstract: Kalman filter has been widely used in GNSS/INS integrated navigation system，and manyadaptive Kalman algorithms have been developed to avoid divergence caused by various exceptions，such as abnormal observations and dynamic model disturbance etc. In the GNSS/INS loosely integrated navigation system，the adaptive filter is generally constructed based on the predicted residual because the observations are deficient compared to the state parameter number. Thus，the GNSS/INSloosely integrated navigation system is highly dependent on observations. The system performanceand stability will degrade when the impact of abnormal observations cannot be decreased effectively.To improve the filter robustness in the GNSS/INS loosely integrated navigation system，the neuralnetwork algorithm is employed to monitor the abnormal observations，and combined with the traditional adaptive algorithm to form enhanced adaptive filter. Based on the proposed method，four adaptively robust Kalman filtering algorithms are developed，which is tested on the raw GNSS/INS data.Results show that，the new algorithms can not only improve the filter estimation accuracy，but alsoreduce the impacts of abnormal observations and dynamical model disturbance.