Adaptive and interpretable reinforcement learning-enhanced unscented Kalman filtering for autonomous driving state estimation
Applied Soft Computing, vol.202, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 202
- Publication Date: 2026
- Doi Number: 10.1016/j.asoc.2026.115924
- Journal Name: Applied Soft Computing
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Compendex, INSPEC
- Keywords: Adaptive noise covariance estimation, Autonomous driving, Interpretable reinforcement learning, Reinforcement learning, State estimation, Unscented Kalman filter
- Yıldız Technical University Affiliated: Yes
Abstract
Accurate state estimation under noisy sensor measurements is a fundamental requirement for autonomous driving systems. This paper proposes an interpretable reinforcement learning (RL)-enhanced Unscented Kalman Filter (UKF) framework that improves estimation accuracy while preserving the structure of classical model-based filtering. By maintaining physically meaningful covariance adaptation, the proposed approach ensures transparency and interpretability in the estimation process. The framework introduces a structured interaction mechanism in which an RL agent adaptively adjusts the process and measurement noise covariance matrices in real time based on measured velocity information and its temporal variation. This enables dynamic and explainable uncertainty adaptation under varying motion and noise conditions without compromising the stability of the underlying filter. The proposed method is evaluated in a Python-based simulation environment using a kinematic vehicle model across multiple motion scenarios. Experimental results demonstrate that the RL-UKF consistently outperforms classical and fixed-parameter filtering approaches, achieving approximately 13.2% improvement in RMSE and 4.6% improvement in mean absolute error. Statistical significance is validated using the Wilcoxon signed-rank test, yielding p-values significantly below conventional thresholds (p≪0.001), confirming that the observed improvements are robust and not due to random variation. Overall, the results demonstrate that reinforcement learning can effectively complement model-based estimation by enabling adaptive, interpretable, and statistically reliable performance improvements for autonomous driving applications.