<p>This paper addresses robust state estimation for radar-based localization in complex environments. In practical radar target tracking, model errors are often inevitable, and these errors typically degrade the estimation accuracy of conventional Kalman filters. To address this issue, a sensitivity penalization mechanism is introduced within the standard Kalman filtering framework. By incorporating a penalty term for model error suppression in the cost function, a robust state estimation algorithm is derived. Theoretical analysis shows that, under certain assumptions, the proposed estimator converges to a stable system. When the system is exponentially stable, the estimator is asymptotically unbiased, and the estimation error covariance remains bounded. Finally, numerical simulations show that, under fixed and time-varying model-error scenarios, the proposed method achieves average RMSE reductions of 15.35% and 23.35%, respectively, relative to the standard Kalman filter implemented with nominal parameters, further validating the effectiveness of the proposed method.</p>