Abstract: When a global positioning system (GPS) aided strapdown inertial navigation system (SINS) aligns in-motion, the filtering system will be nonlinear because of the large initial attitude error. In addition, when the GPS signals are disturbed, the outliers appearing in the observation will reduce the filtering accuracy and even cause the filtering divergence. Aiming at the problem of filtering for nonlinear systems with outliers, an improved particle filter algorithm is designed in this paper, which can extract particles from the posterior distribution of states by using Gaussian sum approximation algorithm (GSA), Bayesian formula, Markov chain Monte Carlo algorithm (MCMC) and ensemble Kalman filter algorithm (EnKF) synthetically. Furthermore, according to the fuzzy theory, an outlier constraint function is added into the improved algorithm to construct the robust ensemble particle filter (REnPF) proposed in this paper. The simulation results of the GPS aided SINS in-motion alignment show that the REnPF can effectively avoid false alarm and missing detection problems, and provide good filtering accuracy.