Abstract: A novel extended particle algorithm is proposed aiming at solving the unmodeled motion problem, which is inextricable merely using conventional MCL. KLD sampling is utilized to measure the approximation error by Kullback-Leibler distance between the true distribution and its sampled representation by generating a dynamically sized set of samples according to the exact posterior being estimated. At the same time, the over-convergence and uniformity validations are introduced to verify correspondence between sample distribution and sensor information for timely re-sampling which highly saves computational resource and enhances localization efficiency. A further experiment, obtained with a real robot in an indoor environment, illustrates the favorable performance of this approach.