In this paper, a novel localization algorithm robust to the unmodeled systematic odometry errors is proposed for low-cost non-holonomic mobile robots. It is well known that the most pose estimators using odometry measurements cannot avoid the performance degradation due to the dead-reckoning of systematic odometry errors. As a remedy for this problem, we tty to reflect the wheelbase error in the robot motion model as a parametric uncertainty. Applying the Krein space estimation theory for the discrete-time uncertain nonlinear motion model results in the extended robust Kalman filter. This idea comes from the fact that systematic odometry errors might be regarded as the parametric uncertainties satisfying the sum quadratic constrains (SQCs). The advantage of the proposed methodology is that it has the same recursive structure as the conventional extended Kalman filter, which makes our scheme suitable for real-time applications. Moreover, it guarantees the satisfactoty localization performance even in the presence of wheelbase uncertainty which is hard to model or estimate but often arises from real driving environments. The computer simulations will be given to demonstrate the robustness of the suggested localization algorithm.