Unscented KalmanNet Enhances Nonlinear State Estimation

State estimation in nonlinear dynamic systems often relies on the Unscented Kalman Filter (UKF), which provides posterior covariance but can suffer from degraded accuracy and calibration due to unknown noise statistics or model mismatches. A new hybrid recursive estimator, Unscented KalmanNet (UKN), has been developed to address these limitations. UKN enhances the UKF by integrating two distinct learned components: NoiseNet, which predicts time-varying noise covariances, and GainNet, which applies a residual correction to the analytical gain. Crucially, UKN preserves the UKF's explicit sigma-point covariance recursion, ensuring positive definiteness of noise predictions. Its training objective is calibration-aware, combining state error with covariance and innovation consistency terms through adaptive weights, optimizing both accuracy and calibration. Benchmarked against UKF, KalmanNet, and Bayesian KalmanNet, UKN consistently achieved the lowest aggregate state-estimation error across synthetic systems and real-flight data, demonstrating substantial RMSE reductions and superior fold-to-fold variability.