Integrating GAN-based machine learning with nonlinear Kalman filtering for enhanced state estimation Authors: Lior Tobaly, Eyal Yaniv, Zeev Zalevsky Published: 2025 ABSTRACT This study introduces a novel approach for enhancing state estimation in non-linear dynamic systems by integrating Generative Adversarial Networks (GANs) with the Unscented Kalman Filter (UKF). While the UKF improves upon traditional Kalman Filters by using sigma points to estimate the mean and covariance in non-linear transformations, its effectiveness is limited by static parameters - specifically, the process noise covariance (Q), measurement noise covariance (R), and scaling factors (primary, secondary, and tertiary; α, κ, and β). We propose a dynamic framework in which a GAN predicts and updates these parameters in real-time, based on the UKF’s recent performance, allowing the filter to better adapt to rapidly changing system dynamics. This method is validated on real-world aircraft navigation data containing time-stamped records of position, velocity, heading, and environmental variables. Results show that the GAN-enhanced UKF significantly reduces state estimation errors compared to conventional static models. The proposed framework is generalizable and can be applied to other domains such as robotics, autonomous vehicles, and smart cities, where accurate real-time state estimation under uncertainty is critical.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-26339-9. Introduction Accurate state estimation in dynamic systems is essential for a wide range of real-world applications. Kalman filtering has long been regarded as the gold standard for such tasks, especially in linear systems. However, most real-world systems exhibit non-linear and time-varying behaviours, which limit the effectiveness of traditional Kalman filters. The Unscented Kalman Filter (UKF) provides a robust alternative for non-linear systems. It improves upon the limitations of linear models by using sigma points to capture the mean and covariance through non-linear transformations. Despite its strengths, the UKF still depends heavily on predefined parameters: the process noise covariance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Q$$\end{document}), measurement noise covariance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R$$\end{document}), and the scaling factors (primary, secondary, and tertiary; α, κ, and β). These parameters influence how sigma points are distributed and weighted, and therefore play a critical role in estimating higher-order statistics. In practice, these values are often fixed or chosen heuristically, which limits the filter’s ability to adapt to dynamic environments. In this work, we introduce a data-driven, adaptive approach to parameter tuning in the UKF by leveraging Generative Adversarial Networks (GANs). The proposed method enables the UKF to predict and update \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Q,\:R$$\end{document} and the scaling factors (primary, secondary, and tertiary; α, κ, and β) dynamically, using feedback from the filter’s own performance. The goal is to enhance the filter’s responsiveness to rapidly changing system conditions, improving overall state estimation accuracy. To evaluate the effectiveness of this approach, we apply it to a real-world dataset of aircraft navigation data, which includes time-stamped records of position, velocity, heading, and relevant environmental variables. Our results demonstrate that the GAN-UKF outperforms traditional static-parameter UKF models in terms of error reduction and tracking consistency. This work lays the foundation for a generalizable, adaptive filtering framework suitable for use in complex domains such as robotics, autonomous vehicles, and smart cities - contexts where accurate and real-time state estimation is critical. In evaluating filtering methods for tracking highly manoeuvring aircraft, the Unscented Kalman Filter (UKF) demonstrates superior alignment with the true trajectory compared to both the standard Kalman Filter (KF) and the Extended Kalman Filter (EKF). While the EKF improves upon the KF by incorporating local linearization, it struggles in scenarios with abrupt directional changes, where the UKF consistently provides more accurate state estimation. This advantage stems from the UKF’s use of sigma points, which allow it to more effectively handle nonlinear dynamics and high manoeuvrability. It should be noted, however, that this improvement comes at the cost of greater computational complexity, since the UKF must propagate a larger set of sigma points through nonlinear functions at each time step. Despite this added burden, the trade-off between accuracy and computational cost favours the UKF in applications where precision and robustness are paramount, such as aircraft navigation. Building upon this foundation, recent advances in machine learning - particularly Generative Adversarial Networks (GANs) - offer promising avenues for further enhancing UKF performance. By dynamically adjusting key parameters, including the process noise covariance matrix (Q), measurement noise covariance matrix (R), and scaling factors (α, κ, β), GANs enable the UKF to adapt continuously to evolving system dynamics. In this approach, the GAN’s generator proposes parameter updates based on historical performance, while the discriminator evaluates their effectiveness in reducing prediction errors. This integration transforms the UKF into an adaptive model capable of real-time parameter tuning, thereby improving estimation accuracy for critical state variables such as velocity, latitude, and longitude. Related work - literature review Kalman filtering in non-linear systems The Kalman filter, introduced by R.E. Kalman in 1960, revolutionized linear state estimation by providing optimal recursive estimates for linear Gaussian1. However, practical systems (e.g. aerospace vehicles and robots) are often non-linear, violating Kalman’s assumptions2. Early extensions like the Extended Kalman Filter (EKF) linearize non-linear dynamics but can suffer from instability if nonlinearities are severe2. This motivated sigma-point approaches such as the Unscented Kalman Filter (UKF)2. The UKF uses deterministic sampling (unscented transform) to propagate means and covariances through non-linear transforms, achieving accuracy up to second-order without explicit linearization2. UKF’s efficacy led to its wide adoption in aerospace navigation and robotic localization throughout the 2000s3. Further advances included the Cubature Kalman Filter (CKF)4, which selects integration points from a spherical-radial rule to improve numerical stability in higher dimensions4. For highly non-Gaussian or multi-modal scenarios, particle filters emerged, using Monte Carlo samples to represent arbitrary5. In domains like aerospace and autonomous vehicles, UKF became a standard for sensor fusion and navigation due to its balance of accuracy and computational load2. For instance, UKF is employed in satellite attitude estimation systems, providing robust performance even with complex quaternion kinematics and intermittent sensor faults6. In automotive systems, UKF is used for vehicle dynamics estimation (e.g. tire slip angles and velocities) under non-linear tire-road interaction7. Alshawi et al. (2024) propose an adaptive UKF for extreme driving maneuvers, which adjusts process and measurement noise covariances online to maintain accuracy during sudden dynamics7. Such adaptive variants address time-varying noise and unmodeled dynamics by tuning filter gains in real time. Recent work by Tong et al. (2022) introduced an improved UKF to handle random communication delays in smart grid state estimation, demonstrating robust tracking despite latency and packet loss8. Despite these successes, classical UKF/EKF frameworks rely on accurate process models and noise statistics. Mismatches in system modeling (e.g. unmodeled flight dynamics or load changes in power grids) can degrade filter performance3. There is a need for filters to adapt or learn unknown model aspects on the fly. This challenge has catalyzed research into data-driven and hybrid approaches that combine domain knowledge with machine learning to handle non-linearities and uncertainties beyond the reach of conventional filters. Machine learning in state estimation The past decade has seen rapid growth in applying machine learning (ML) to state estimation, motivated by the ability of ML models (especially deep neural networks) to approximate complex non-linear mappings. Traditional Kalman filters are model-based, errors arise if the assumed physics are incomplete. Machine learning offers a model-free or model-enhanced avenue: either replace the physics-based estimator with a learned model or augment the estimator with learned components9. For example, a neural network can learn to predict system state directly from raw sensor data, or estimate unmodeled dynamics to be fed into a Kalman filter. Researchers have increasingly explored hybrid Kalman filters and neural networks to leverage both data-driven flexibility and probabilistic filtering10. Preliminary efforts in the 2010s integrated simple neural networks with Kalman filters for specific tasks like sensor bias correction and fault detection. For instance, in aerospace, Chen et al. (2024) combine a neural network with a Kalman filter to detect and recover from gyroscope faults in satellite attitude estimation6. The neural network monitors residuals and adaptively adjusts filter estimates during faults, improving reliability of the attitude solution. In robotics, Urrea & Agramonte (2021) provide a comprehensive overview of Kalman filtering in robotic systems and note that neural networks have been used to assist in non-linear mapping problems such as vision-based localization11. One approach is using a neural network to learn the observation function (e.g. mapping image features to robot pose), while a Kalman filter handles the sequential estimation, combining learned measurements with dynamic models. With the deep learning revolution, researchers began applying advanced architectures (CNNs, RNNs, etc.) to state estimation. A notable direction is recurrent neural networks (RNNs), like LSTMs, for sequential sensor data. Recurrent nets can serve as standalone estimators or be embedded in filtering frameworks. Kim & Lee (2021) compared several ML models for aviation wind prediction and found an LSTM network best captured short-term wind dynamics12. They then augmented the LSTM with a UKF (termed “UKF-LSTM”), which treated the LSTM’s prediction as a process model within a Kalman filtering loop12. This hybrid improved forecasting accuracy and provided confidence intervals via the UKF’s covariance estimation. Similarly, in intelligent vehicles, deep learning has been used to estimate vehicle states that are difficult to model. Yang et al. (2024) propose a hybrid deep neural network-UKF algorithm to estimate a bus’s roll and sideslip angles in real time13. A deep network first learns to predict these angles from IMU and GPS data, its output is then fused by an UKF with a vehicle motion model, yielding an estimator that is both accurate and physically consistent14. The hybrid outperformed purely model-based and purely data-driven methods, highlighting the benefit of combining neural networks with Kalman filtering. By 2023, the synergy of ML and state estimation had become a significant research trend. Bai et al. (2023) published an extensive survey on “Kalman filter driven by machine learning”, documenting the evolution from EKF to modern hybrid schemes15. They conclude that learning-based methods have achieved notable success in autonomous navigation, power grid monitoring, and industrial IoT, by addressing model uncertainties and non-linearities that classical filters struggle with15. Feng et al. (2023) similarly review dozens of Kalman-neural network hybrid models, affirming that such approaches often outperform either method alone across aerospace, automotive, and robotics applications16. The literature indicates a clear trajectory: towards “learning Kalman filters” that can adjust their models using data, rather than relying on fixed equations. This sets the stage for the latest innovation - incorporating generative models, like GANs, to further enhance state estimation adaptivity. Generative adversarial networks (GANs) in adaptive filtering Generative Adversarial Networks (GANs), first introduced by Goodfellow et al. in 2014, are a class of deep learning models that learn data distributions via a two-player game between a Generator and a Discriminator17. GANs excel at producing realistic samples from complex, high-dimensional distributions. In state estimation contexts, GANs can model uncertainties or generate synthetic sensor data that mimic real-world variations. This capability is valuable for adaptive filtering, where understanding and adapting to changing noise or system dynamics is crucial. By 2020, GANs had been successfully applied in various domains to create realistic data, for instance, Goodfellow et al. (2020) overview their success in generating high-resolution images and note their potential for wider applications18. Researchers have begun leveraging GANs to address two key challenges in filtering: data scarcity (by augmenting training or calibration data) and non-Gaussian noise modeling (by learning complex noise distributions). One line of work uses GANs to produce synthetic time-series data that improve filter training or resiliency. In battery management systems (power systems domain), obtaining extensive labeled data for state-of-charge (SOC) estimation is difficult due to long testing times. Hu et al. (2024) propose a GAN-based augmentation technique to generate realistic battery voltage and current trajectories for training a state estimator19. They develop a Time-Series DCGAN (TS-DCGAN) that learns from limited experimental data and produces additional “pseudo-experimental” battery discharge profiles, including realistic noise and aging effects19. By training an extended Kalman filter on this enriched dataset, the authors achieved higher SOC estimation accuracy and robustness to sensor noise19. The GAN essentially serves as a data generator capturing the stochastic dynamics of battery behavior under varied conditions, thereby improving the filter’s ability to generalize. Similarly, in smart grid state estimation, conditional GANs have been used to simulate plausible grid measurement vectors under different load conditions, augmenting data for training neural network observers20. These studies show that GANs can address data insufficiency, a common issue in developing data-driven filters. Another contribution of GANs is modeling complex error distributions that challenge traditional filters. Standard Kalman filters assume Gaussian noise; in reality, noise can be multi-modal or heavy-tailed (e.g. due to sporadic outliers or mode switches). GANs provide a non-parametric way to learn such distributions. Tian et al. (2022) tackled urban traffic state estimation with sparse and multi-modal data using a pattern-adaptive GAN (PA-GAN)21. Traffic flow and speed measurements in a city can be sporadic (due to sensor outages) and exhibit multi-modal patterns (e.g. rush-hour vs. off-hour behaviors). PA-GAN learned to generate realistic traffic state vectors (speed distributions across a network) conditioned on time-of-day and sparse sensor inputs21. By imbuing a particle filter with these GAN-generated samples as priors, the authors achieved more accurate and stable traffic state estimates under sparse data scenarios, compared to filters with simple Gaussian assumptions21. The GAN effectively adapted the filter’s prior to the current traffic pattern, enabling multimodal state representation (e.g. multiple plausible congestion scenarios) until disambiguated by new data. In power systems, researchers have trained Wasserstein GANs to learn the distribution of meter measurements under normal conditions, these GANs are then used within a chi-square test or filter framework to detect anomalies and adapt state estimates when adversarial readings occur22. By learning the true distribution of measurement noise and load variability, a GAN can help a state estimator distinguish nominal fluctuations from abnormal deviations, enhancing resilience to attacks or sensor faults. The above examples illustrate GANs enabling adaptive filtering in ways not possible with conventional methods. By providing rich stochastic information (through generated samples), GAN-augmented filters can adapt to regime changes: e.g. a traffic estimator can seamlessly shift its state predictions when the GAN indicates a transition from free-flow to congested traffic pattern. Moreover, GANs can operate in high-dimensional observation spaces (images, lidar scans) by learning lower-dimensional representations for filtering - an approach being explored in autonomous driving for visual-inertial odometry, where GAN-like generators simulate sensor readings for unseen conditions to update filter predictions. One challenge is ensuring the stability of the combined GAN-filter system. Since GAN outputs are probabilistic, integrating them into recursive estimation requires careful design to avoid divergence. Techniques like GAN ensemble averaging (to quantify uncertainty) and regularizing the generator with physical laws (to maintain consistency) have been proposed9. Overall, GANs bring a powerful data-driven adaptation mechanism, and initial studies in smart cities, power grids, and autonomous systems show improved estimation accuracy, especially in edge cases or sparse-data regimes. This paves the way for more holistic hybrid ML-Kalman filtering frameworks, discussed next. Hybrid machine learning–Kalman filtering approaches Hybrid approaches tightly integrate machine learning models with Kalman filter algorithms to exploit their complementary strengths. The Kalman filter provides recursive optimal estimation with uncertainty quantification, while ML models (e.g. deep neural networks or GANs) contribute flexibility in modeling complex phenomena and learning from data. The literature in recent years reveals a spectrum of hybrid architectures. Shuo Feng et al. (2023) categorize these into series (filters and networks in sequence) and parallel (filters and networks exchanging information iteratively) hybrids16. An example of a series hybrid is the UKF-LSTM mentioned earlier: the LSTM produces a preliminary estimate that is then refined by a Kalman filter12. A parallel hybrid might involve a network that continuously updates filter parameters (like noise covariances) during operation, essentially learning to “steer” the Kalman filter. KalmanNet (2021–2022) and related approaches treat the Kalman filtering algorithm as a trainable architecture. In KalmanNet, neural networks learn to approximate the filter gain or model parameters from data, allowing operation even with partially known system models23. Yanjie Yu et al. (2023) developed an Unscented Trainable Kalman Filter (UTKF) where a deep neural network learns to predict the process dynamics, and the UKF framework then propagates and updates the state24. This was demonstrated in power system state estimation with missing data: the network learned the generator dynamics and the UTKF robustly estimated voltages even when certain sensors failed, outperforming static UKF assumptions24. Such approaches blur the line between model-based and data-driven - the filter becomes a scaffold within which learning happens. The PINN-UKF framework by de Curto et al. (2024) is one example9, but similar physics-informed deep filters have appeared in traffic and robotics. Zhao & Yu (2023) introduced an Observer-Informed Deep Learning (OIDL) scheme for traffic networks, where a PDE-based traffic flow observer runs in parallel with a deep neural network25. The observer (analogous to a model-based filter) ensures physical consistency (e.g. conservation of cars), while the deep network, guided by the observer’s feedback, learns to estimate the traffic state from sparse boundary measurements25. The two components iterate, resulting in a self-correcting hybrid that significantly improved highway traffic density estimation under real sensor data25. This parallel architecture highlights how classical filters can regularize deep learning, and conversely how deep learning can augment observers in unobserved regions. An emerging theme is enhancing filter robustness using ML. RKFNet introduced by Hao et al. (2025) combines a heavy-tailed noise Kalman filter with deep learning26. The base filter (RKF) assumes noise follows a heavy-tailed distribution to better handle outliers. A neural network is then trained to adapt the tail parameters and identify outlier-contaminated measurements26. In experiments on autonomous vehicle tracking with occasional GPS spikes, RKFNet automatically down-weighted the spikes and retained accuracy, whereas a standard KF would diverge. This demonstrates a powerful synergy: the deep network learns abnormal patterns, and the filter provides the principled mathematical framework to incorporate that knowledge for robustness. Across aerospace, automotive, power, and smart cities, these hybrids have set new performance benchmarks. In autonomous vehicles, the fusion of camera/LiDAR deep perception with filtering yields accurate and stable state estimates for navigation. For example, a recent vehicle sideslip estimator combines a vision-based CNN (giving a pseudo-measurement of slip angle) with a vehicle-model UKF27. The UKF cross-checks the CNN output against physical vehicle dynamics, resulting in an estimator that can trust the CNN when it is confident but fallback to model predictions when the CNN is uncertain27. This adaptive trust mechanism solved a long-standing issue in tire-slip estimation under varying road conditions. In aerospace, Chen et al. (2024) report a hybrid framework for spacecraft attitude: a small CNN processes star-camera images to provide coarse attitude quaternions, which an Unscented Kalman Filter then fuses with gyroscope data for high-rate attitude tracking6. The result is sub-arcsecond pointing accuracy, even during periods when the star camera is blinded (the UKF propagates using gyro bias estimates learned during nominal conditions). In power systems, hybrid schemes (like Yu’s UTKF and others) are enabling dynamic state estimation even under low-observability (few sensors) by learning system dynamics from historical data28. A comparative study by Qiu et al. (2025) in an energy microgrid showed that a hybrid UKF-ANN estimator tracked load voltages with 40% lower error than a pure ANN and 25% lower error than a tuned UKF, especially during abrupt load changes29. The ANN learned the nonlinear relationship between loads and voltage drops, while the UKF ensured the estimates respected circuit equations and filtered out measurement noise. The integration of GANs with UKF - the focus of our manuscript - can be viewed as part of this broader hybrid trend. GANs provide a powerful means to simulate and adapt to new operating conditions on the fly. Combined with UKF’s principled estimation, a GAN-UKF hybrid can continually improve state estimates even as the environment or system behavior shifts. For example, one could imagine an UKF that uses a GAN-generated ensemble of predicted sensor readings as the sigma-points for its unscented transform, effectively learning the measurement distribution in real-time. Early explorations of this idea in traffic and power systems are promising21. As research advances, we expect to see GAN-augmented UKFs enabling autonomous vehicles to handle novel road scenarios, smart grids to estimate state under cyber-attack by learning attack patterns, robots to localize in unseen terrains by simulating sensor data, and more. Background and mathematical formulation Unscented Kalman filter The Unscented Kalman Filter (UKF) addresses nonlinear dynamics by using a set of carefully selected sigma points to represent state estimates, eliminating the need for linearization. This approach enhances both accuracy and computational efficiency. The sigma points are propagated through the nonlinear transformation, allowing the UKF to capture the mean and covariance of the resulting distribution more effectively. Kalman filtering relies on Gaussian distributions, which are fully characterized by their mean and covariance. Within this context, the state estimation problem is defined as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{k}=f({\widehat{x}}_{k-1},{\text{u}}_{k})+{w}_{k}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{k}$$\end{document}​ is the state at time step \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{u}}_{k}$$\end{document}​ is the control input, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f$$\end{document} represents the system dynamics, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{n}$$\end{document}​ is the process noise, assumed to be Gaussian. Sigma points are generated to approximate the distribution after nonlinear transformations:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}_{0}={\widehat{x}}_{k-1}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}_{i}={\widehat{x}}_{k-1}+{\left(\sqrt{\left(n+\lambda\:\right){P}_{k-1}}\right)}_{i}$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}_{i+n}={\widehat{x}}_{k-1}-{\left(\sqrt{\left(n+\lambda\:\right){P}_{k-1}}\right)}_{i}$$\end{document} 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:={\alpha\:}^{2}(n+\kappa\:)-n$$\end{document} Where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}_{i}$$\end{document}​: Sigma points generated from the mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{x}}_{k-1}$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{x}}_{k-1}$$\end{document}​: State estimate at the previous time step. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k-1}$$\end{document}​: State covariance at the previous time step. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n$$\end{document} Dimensionality of the state vector. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:\::\:$$\end{document}A scaling parameter that affects the spread of the sigma points. Weights:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{0}^{\left(m\right)}=\frac{\lambda\:}{n+\lambda\:}\:,\:\:{W}_{0}^{\left(c\right)}={W}_{0}^{\left(m\right)}+(1-{\alpha\:}^{2}+\beta\:)$$\end{document} Where:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{W}}_{\varvec{i}}^{\left(\varvec{m}\right)}$$\end{document} Weights associated with the sigma points for the mean. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:,\:\beta\:,\:\kappa\:$$\end{document} Tuning (scaling) parameters for controlling the sigma point spread. Prediction: Propagate sigma points through the transition function to estimate the prior state and covariance:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{x}}_{k|k-1}={\sum\:}_{i=0}^{2n}{W}_{i}^{\left(m\right)}f\left({\chi\:}_{i,k-1}\right)$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k|k-1}={\sum\:}_{i=0}^{2n}{W}_{i}^{\left(c\right)}\left(f\left({\chi\:}_{i,k-1}\right)-{\widehat{x}}_{k|k-1}\right)+{\left(f\left({\chi\:}_{i,k-1}\right)-{\widehat{x}}_{k|k-1}\right)}^{T}+{Q}_{k}$$\end{document} Where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k|k-1}$$\end{document}​: Predicted state covariance. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{i}^{\left(c\right)}$$\end{document}​: Weights associated with the sigma points for the covariance. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Q}_{k}$$\end{document} Process noise covariance matrix. Update: Incorporate measurements by predicting the expected measurement:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{z}}_{k|k-1}={\sum\:}_{i=0}^{2n}{W}_{i}^{\left(m\right)}h\left({\chi\:}_{i,k|k-1}\right)$$\end{document} Where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}_{i,k|k-1}$$\end{document}​ are the sigma points propagated through the state transition function during the prediction step. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{i}^{\left(m\right)}\:$$\end{document}are the weights for the mean, as computed during the sigma points generation. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h(\cdot\:)$$\end{document} is the measurement function. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{z}}_{k|k-1}$$\end{document} represents the predicted measurement vector at time step k, conditioned on all information available up to time step k − 1. Covariance of the prediction error The covariance of the prediction error \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{ZZ,k}$$\end{document} is computed. This matrix quantifies the expected accuracy of the predicted measurements and includes the measurement noise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{k}$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{ZZ,k}={\sum\:}_{i=0}^{2n}{W}_{i}^{\left(c\right)}\left(h\left({\chi\:}_{i,k|k-1}\right)-{\widehat{z}}_{k|k-1}\right){\left(h\left({\chi\:}_{i,k|k-1}\right)-{\widehat{z}}_{k|k-1}\right)}^{T}+{R}_{k}$$\end{document} Where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{i}^{\left(c\right)}\:$$\end{document}are the weights for the covariance. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{k}$$\end{document}​ is the measurement noise covariance matrix, reflecting the expected noise in the measurements. Cross-covariance matrix The cross-covariance matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{XZ,k}$$\end{document}​ between the state and the measurements is also computed. This matrix is essential for updating the state estimate and reflects the correlation between the state predictions and the measurement predictions:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{XZ,k}={\sum\:}_{i=0}^{2n}{W}_{i}^{\left(c\right)}\left({\chi\:}_{i,k|k-1}-{\widehat{x}}_{k|k-1}\right){\left(h\left({\chi\:}_{i,k|k-1}\right)-{\widehat{z}}_{k|k-1}\right)}^{T}$$\end{document} Where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{x}}_{k|k-1}$$\end{document} ​ is the predicted state from the prediction step. Kalman gain The Kalman Gain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{k}$$\end{document}​ is then calculated, which determines the extent to which the new measurement should be incorporated into the state estimate:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{k}={P}_{xz,k}{\text{P}}_{zz,k}^{-1}$$\end{document} The Kalman Gain optimizes the state update to reduce the overall estimation error by balancing the uncertainty in the new measurement and the current estimate. State update Using the Kalman Gain, the state estimate is updated based on the discrepancy between the actual measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}$$\end{document} and the predicted measurement:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{k,k}={\widehat{x}}_{k|k-1}+{K}_{k}({z}_{k}-{\widehat{z}}_{k|k-1})$$\end{document} Covariance update Finally, the covariance of the updated estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k,k}\:$$\end{document}is recalculated to reflect the reduced uncertainty following the measurement update:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k,k}={P}_{k|k-1}-{K}_{k}{P}_{zz,k}{K}_{k}^{T}$$\end{document} By subtracting the portion that the new measurements have explained this equation modifies the predicted covariance and improves the accuracy of the estimate. Generative adversarial networks (GANs) Generative models were notably advanced in 2014 when Ian Goodfellow and colleagues introduced Generative Adversarial Networks (GANs)30. GANs consist of two neural networks - the Generator (G) and the Discriminator (D) - that are trained simultaneously in an adversarial process. As illustrated in Fig. 1, the Generator receives random noise vectors as input and produces synthetic data samples, while the Discriminator evaluates both real and generated data, classifying them as either authentic or fake. During training, the Discriminator learns to distinguish real from generated data, while the Generator learns to improve its output so that it increasingly resembles the real data. This adversarial setup, driven by classification loss for the Discriminator and reconstruction loss for the Generator, leads to progressively more realistic outputs from the Generator over time. Fig. 1GAN architecture. GAN architecture. GAN architecture Generator (G): The generator network produces synthetic data samples, denoted as G(z), using a random noise vector z as input. Its objective is to generate data that closely mimics real data from the training distribution. Discriminator (D): The discriminator receives either real data samples or synthetic ones generated by G, and assigns a probability indicating whether the input is real or fake. Its goal is to correctly classify inputs as authentic or generated. During training, the generator and discriminator engage in a two-player minimax game. The discriminator aims to maximize its classification accuracy by distinguishing real from fake data, while the generator simultaneously attempts to “fool” the discriminator by producing increasingly realistic samples. This adversarial interaction can be formalized as a minimax optimization problem: Below is the definition of the GAN objective function:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\underset{G}{\text{min}}\underset{D}{\text{max}}V\:\left(D,G\right)=\:{E}_{x\sim{p}_{data}\left(x\right)}\left[logD\left(x\right)\right]+{E}_{z\sim{p}_{z}\left(z\right)}\left[log\left(1-D\left(G\left(z\right)\right)\right)\right]$$\end{document} Where: G: The generator model.D: The discriminator model.x: Real data samples from the data distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{data}\left(x\right)$$\end{document}.z: Random noise sampled from a latent space \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document}.G(z): Generated samples produced by the generator from the noise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:z$$\end{document}.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{x\sim{p}_{data}\left(x\right)}\left[logD\left(x\right)\right]\:$$\end{document}denotes the expectation over real data samples x drawn from the true data distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{data}\left(x\right)$$\end{document}.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{z\sim{p}_{z}\left(z\right)}\left[log\left(1-D\left(G\left(z\right)\right)\right)\right]\:$$\end{document}denotes the expectation over latent variables z drawn from the prior distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document} (typically Gaussian or uniform). G: The generator model. G: The generator model. D: The discriminator model. D: The discriminator model. x: Real data samples from the data distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{data}\left(x\right)$$\end{document}. x: Real data samples from the data distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{data}\left(x\right)$$\end{document}. z: Random noise sampled from a latent space \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document}. z: Random noise sampled from a latent space \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document}. G(z): Generated samples produced by the generator from the noise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:z$$\end{document}. G(z): Generated samples produced by the generator from the noise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:z$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{x\sim{p}_{data}\left(x\right)}\left[logD\left(x\right)\right]\:$$\end{document}denotes the expectation over real data samples x drawn from the true data distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{data}\left(x\right)$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{x\sim{p}_{data}\left(x\right)}\left[logD\left(x\right)\right]\:$$\end{document}denotes the expectation over real data samples x drawn from the true data distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{data}\left(x\right)$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{z\sim{p}_{z}\left(z\right)}\left[log\left(1-D\left(G\left(z\right)\right)\right)\right]\:$$\end{document}denotes the expectation over latent variables z drawn from the prior distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document} (typically Gaussian or uniform). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{z\sim{p}_{z}\left(z\right)}\left[log\left(1-D\left(G\left(z\right)\right)\right)\right]\:$$\end{document}denotes the expectation over latent variables z drawn from the prior distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document} (typically Gaussian or uniform). Methods of training: Set random weights as the initial values for the discriminator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D$$\end{document} and generator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:G$$\end{document}. Get the discriminator trained: From the training set take a sample of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x$$\end{document} real data samples. Take a sample of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document} the noise distribution of a batch of noise vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:z$$\end{document}. The generator should be used to create a batch of synthetic samples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:G\left(z\right)$$\end{document}. Use the objective functions maximum to update the discriminator\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{x\sim{p}_{data}\left(x\right)}\left[logD\left(x\right)\right]+{E}_{z\sim{p}_{z}\left(z\right)}\left[log\left(1-D\left(G\left(z\right)\right)\right)\right]$$\end{document} Train the generator: From the noise distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document} sample a batch of noise vectors z. Using the generator create a batch of synthetic samples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:G\left(z\right)$$\end{document}. By minimizing the objective function update the generator\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:G$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{z\sim{p}_{z}\left(z\right)}\left[log\left(1-D\left(G\left(z\right)\right)\right)\right]$$\end{document} Repeat the process until the generator produces sufficiently realistic data samples. An optimization problem known as minimax is used to formalize this interaction. By utilizing both labeled and unlabeled data GANs can be expanded to enhance classification performance and resilience in semi-supervised learning. Integrating GANs with UKF While the UKF provides strong performance in nonlinear systems, its accuracy depends heavily on fixed parameters such as the process noise covariance Q, measurement noise covariance R, and sigma point scaling factors (α, β, κ). In static UKF formulations, these parameters are predefined and remain constant, leading to suboptimal performance in dynamic environments where noise levels or system nonlinearities evolve. To address this limitation, we propose integrating a GAN framework with the UKF to enable adaptive parameter tuning. Specifically: The generator proposes updated values for Q, R, α, β, κ at each cycle, based on historical state estimation performance and prediction errors.The discriminator evaluates these proposals by comparing predicted versus observed errors, rewarding parameter sets that reduce estimation inaccuracies.Through iterative training, the GAN learns to produce parameter adjustments that optimize UKF performance under varying conditions. The generator proposes updated values for Q, R, α, β, κ at each cycle, based on historical state estimation performance and prediction errors. The generator proposes updated values for Q, R, α, β, κ at each cycle, based on historical state estimation performance and prediction errors. The discriminator evaluates these proposals by comparing predicted versus observed errors, rewarding parameter sets that reduce estimation inaccuracies. The discriminator evaluates these proposals by comparing predicted versus observed errors, rewarding parameter sets that reduce estimation inaccuracies. Through iterative training, the GAN learns to produce parameter adjustments that optimize UKF performance under varying conditions. Through iterative training, the GAN learns to produce parameter adjustments that optimize UKF performance under varying conditions. This integration transforms the UKF into an adaptive nonlinear filter capable of real-time self-tuning. The theoretical advantage is that adversarial learning leverages past estimation errors to refine filter behaviour, aligning statistical assumptions (e.g., Gaussian noise) more closely with the observed system dynamics. Theoretical implications The GAN-UKF integration represents a conceptual shift from static, parameter-fixed filtering toward adaptive, data-driven filtering. In theoretical terms: The UKF provides a principled Bayesian framework for nonlinear Gaussian estimation.GANs extend this framework by learning meta-distributions over filter parameters, effectively treating Q, R, α, β, κ as dynamic latent variables.This creates a hybrid estimator that combines statistical optimality (Kalman filtering) with machine learning adaptability (GANs). The UKF provides a principled Bayesian framework for nonlinear Gaussian estimation. The UKF provides a principled Bayesian framework for nonlinear Gaussian estimation. GANs extend this framework by learning meta-distributions over filter parameters, effectively treating Q, R, α, β, κ as dynamic latent variables. GANs extend this framework by learning meta-distributions over filter parameters, effectively treating Q, R, α, β, κ as dynamic latent variables. This creates a hybrid estimator that combines statistical optimality (Kalman filtering) with machine learning adaptability (GANs). This creates a hybrid estimator that combines statistical optimality (Kalman filtering) with machine learning adaptability (GANs). As a result, the proposed method balances robustness with flexibility, offering superior performance in real-world nonlinear applications such as aircraft navigation, robotics, and autonomous systems. Further details on the implementation and experimental validation of this approach are provided in the Methodology section. Dataset description Dataset description This study utilizes the TrajAir dataset31, a unique and comprehensive collection of real aircraft trajectory data curated specifically for state estimation and trajectory prediction tasks. The dataset was collected around Pittsburgh-Butler Regional Airport (ICAO: KBTP) - a non-towered general aviation airport located approximately 10 miles north of Pittsburgh, Pennsylvania, USA. The airport’s constrained airspace and adherence to strict FAA (Federal Aviation Administration) traffic regulations provide a valuable environment for capturing detailed and structured aircraft motion data. The dataset is publicly available at https://theairlab.org/trajair/#download. Dataset origin Data were collected using an Automatic Dependent Surveillance–Broadcast (ADS-B) receiver installed at the airport. This receiver used satellite navigation to determine aircraft positions and timestamps, continuously logging broadcasts from aircraft equipped with ADS-B Out technology. To include aircraft lacking this capability, additional data were retrieved via the Traffic Information Service–Broadcast (TIS-B), which retransmits radar data in an ADS-B-compatible format. The collection spanned 111 days, from September 18, 2020, to April 23, 2021, covering operations between 1:00 AM and 11:00 PM local time. Days impacted by equipment repairs or extreme weather were excluded to maintain data quality. Meteorological integration In addition to flight trajectories, the dataset includes environmental data sourced from METeorological Aerodrome Reports (METAR), generated by the airport’s Automated Weather Observing System (AWOS). These meteorological readings - such as temperature, pressure, wind speed, and wind direction - were aligned post hoc with flight data using timestamp matching from the Iowa State METAR repository. This integration enables a richer analysis by contextualizing aircraft dynamics within varying atmospheric conditions. Purpose and scope The TrajAir dataset was designed as a benchmark resource for testing advanced trajectory prediction and dynamic state estimation models, particularly in non-towered terminal airspace where conventional air traffic control is minimal. Its broad applicability includes research into socially aware trajectory models, adaptive filtering, and multi-agent interaction modelling in constrained environments. Data processing and preprocessing Data cleaning As with most real-world datasets, the raw data contained various anomalies and noise. A cleaning process was applied to remove: Corrupted or incomplete entries (e.g., missing GPS fields).Duplicates.Outliers (e.g., altitude > 6000 feet or distance > 5 km from the runway). Corrupted or incomplete entries (e.g., missing GPS fields). Corrupted or incomplete entries (e.g., missing GPS fields). Duplicates. Duplicates. Outliers (e.g., altitude > 6000 feet or distance > 5 km from the runway). Outliers (e.g., altitude > 6000 feet or distance > 5 km from the runway). This step was critical to preserving the integrity and utility of the dataset. Scene segmentation The dataset was segmented into discrete “scenes” based on aircraft activity near the airport. A scene begins when one or more aircraft enter the monitored area and ends when all aircraft have departed or are inactive. This segmentation allows for scenario-specific modelling and facilitates the study of multi-aircraft dynamics in realistic airspace settings. Environmental data integration Key environmental parameters - temperature, pressure, wind speed, and wind direction - were extracted from METAR strings and aligned with trajectory data. The inclusion of weather information allows for weather-aware trajectory modelling and improves the fidelity of state estimation under varying atmospheric conditions. Feature overview for state estimation The cleaned and enriched version of the TrajAir dataset includes all features necessary for applying the Unscented Kalman Filter (UKF) to aircraft state estimation. It provides time-stamped records of: Latitude and Longitude: Aircraft geolocation in degrees.Altitude: Height above sea level (meters).Heading: Direction of motion (0–360°).Speed: Airspeed in knots.Environmental Conditions: Temperature (°C), Pressure (hPa), Wind Speed and Direction.Delta Time: Time difference between successive measurements (seconds). Latitude and Longitude: Aircraft geolocation in degrees. Latitude and Longitude: Aircraft geolocation in degrees. Altitude: Height above sea level (meters). Altitude: Height above sea level (meters). Heading: Direction of motion (0–360°). Heading: Direction of motion (0–360°). Speed: Airspeed in knots. Speed: Airspeed in knots. Environmental Conditions: Temperature (°C), Pressure (hPa), Wind Speed and Direction. Environmental Conditions: Temperature (°C), Pressure (hPa), Wind Speed and Direction. Delta Time: Time difference between successive measurements (seconds). Delta Time: Time difference between successive measurements (seconds). Each observation represents a full snapshot of the aircraft’s state, supporting precise temporal modelling and enabling robust, real-time nonlinear estimation with the UKF. See Table 1 for a summary of all dataset features and units. Table 1Dataset features.FeatureDescriptionUnitLatitudeLatitude of the aircraftDegreesLongitudeLongitude of the aircraftDegreesAltitudeAircraft’s altitude above sea levelMetersHeadingDirection of aircraft’s movement (0-360 degrees)DegreesSpeedAircraft’s speed relative to the airKnotsWind speedSpeed of the wind affecting the aircraftKnotsWind directionDirection from which the wind is comingDegreesTemperatureAmbient temperature at the aircraft’s altitude°CPressureBarometric pressure at the aircraft’s altitudehPaDelta timeTime difference between consecutive measurementsSeconds Table 1Dataset features.FeatureDescriptionUnitLatitudeLatitude of the aircraftDegreesLongitudeLongitude of the aircraftDegreesAltitudeAircraft’s altitude above sea levelMetersHeadingDirection of aircraft’s movement (0-360 degrees)DegreesSpeedAircraft’s speed relative to the airKnotsWind speedSpeed of the wind affecting the aircraftKnotsWind directionDirection from which the wind is comingDegreesTemperatureAmbient temperature at the aircraft’s altitude°CPressureBarometric pressure at the aircraft’s altitudehPaDelta timeTime difference between consecutive measurementsSeconds Dataset features. Latitude and longitude: Indicate the aircraft’s position on the Earth’s surface in degrees. These coordinates are essential for the Kalman Filter to estimate trajectory and update positional states accurately.Altitude: Represents the vertical distance of the aircraft above sea level. In environments where air pressure and temperature vary, altitude is critical for understanding the aircraft’s three-dimensional motion.Heading: Denotes the aircraft’s direction of motion in degrees. It provides insight into the aircraft’s navigation path and adjustments resulting from external forces such as wind.Speed: Measured in knots relative to the air, speed is a vital state variable - especially when combined with wind parameters - for forecasting the aircraft’s future position.Environmental factors (Wind Speed, Wind Direction, Temperature, Pressure): These variables significantly influence the aircraft’s dynamics. Temperature and pressure affect altitude through changes in air density, while wind speed and direction alter the aircraft’s true velocity and heading.Delta time: The time interval between successive measurements enables the UKF to scale the prediction step appropriately. Accurate time deltas are necessary for computing changes in position and speed between observations. Latitude and longitude: Indicate the aircraft’s position on the Earth’s surface in degrees. These coordinates are essential for the Kalman Filter to estimate trajectory and update positional states accurately. Latitude and longitude: Indicate the aircraft’s position on the Earth’s surface in degrees. These coordinates are essential for the Kalman Filter to estimate trajectory and update positional states accurately. Altitude: Represents the vertical distance of the aircraft above sea level. In environments where air pressure and temperature vary, altitude is critical for understanding the aircraft’s three-dimensional motion. Altitude: Represents the vertical distance of the aircraft above sea level. In environments where air pressure and temperature vary, altitude is critical for understanding the aircraft’s three-dimensional motion. Heading: Denotes the aircraft’s direction of motion in degrees. It provides insight into the aircraft’s navigation path and adjustments resulting from external forces such as wind. Heading: Denotes the aircraft’s direction of motion in degrees. It provides insight into the aircraft’s navigation path and adjustments resulting from external forces such as wind. Speed: Measured in knots relative to the air, speed is a vital state variable - especially when combined with wind parameters - for forecasting the aircraft’s future position. Speed: Measured in knots relative to the air, speed is a vital state variable - especially when combined with wind parameters - for forecasting the aircraft’s future position. Environmental factors (Wind Speed, Wind Direction, Temperature, Pressure): These variables significantly influence the aircraft’s dynamics. Temperature and pressure affect altitude through changes in air density, while wind speed and direction alter the aircraft’s true velocity and heading. Environmental factors (Wind Speed, Wind Direction, Temperature, Pressure): These variables significantly influence the aircraft’s dynamics. Temperature and pressure affect altitude through changes in air density, while wind speed and direction alter the aircraft’s true velocity and heading. Delta time: The time interval between successive measurements enables the UKF to scale the prediction step appropriately. Accurate time deltas are necessary for computing changes in position and speed between observations. Delta time: The time interval between successive measurements enables the UKF to scale the prediction step appropriately. Accurate time deltas are necessary for computing changes in position and speed between observations. Together, these features ensure accurate modelling in a non-linear Kalman filtering framework. By incorporating both environmental context and aircraft state variables, the UKF can generate more precise predictions, particularly in dynamic or uncertain conditions. Methodology To enable real-time adaptation of critical parameters such as the process noise covariance (Q), measurement noise covariance (R), and scaling factors (primary, secondary, and tertiary; α, κ, and β), this research introduces a dynamic Unscented Kalman Filter (UKF) enhanced by Generative Adversarial Networks (GANs). The primary contribution of this work lies in the mathematical formulation of a dynamic UKF and its integration with a GAN-based learning mechanism. GAN-UKF collaboration The integration of a UKF and a GAN, illustrated in Fig. 2, results in a dynamic and adaptive filtering system. By continuously learning from the filter’s performance, the GAN module dynamically adjusts the UKF parameters, thereby improving state estimation accuracy over time. Fig. 2GAN-UKF architecture. GAN-UKF architecture. System initialization.The system begins by initializing the UKF state vector and GAN parameters. This includes defining initial values for Q, R, and the scaling factors (primary, secondary, and tertiary; α, κ, and β). This setup prepares both models to enter the iterative prediction - correction cycle. System initialization. System initialization. 2.UKF prediction and update.The UKF uses the current noise covariances (Q, R) and scaling factors (primary, secondary, and tertiary; α, κ, and β) to perform its standard prediction and update operations:Prediction Step: The UKF propagates the state and covariance forward using the system dynamics.Update Step: Incoming observations are incorporated to refine the state estimate and reduce prediction error. 2.UKF prediction and update. UKF prediction and update. Prediction Step: The UKF propagates the state and covariance forward using the system dynamics. Prediction Step: The UKF propagates the state and covariance forward using the system dynamics. Update Step: Incoming observations are incorporated to refine the state estimate and reduce prediction error. Update Step: Incoming observations are incorporated to refine the state estimate and reduce prediction error. 3.Performance evaluation.Following each UKF cycle, key performance indicators - such as Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) - are computed. These metrics are used as feedback signals to guide the GAN’s learning process, enabling the system to quantify and react to estimation accuracy. 3.Performance evaluation. Performance evaluation. GAN module The GAN module dynamically updates UKF parameters based on recent performance. Input preparation: Performance metrics from the UKF (e.g., RMSE and MAE for key state variables such as latitude, longitude, heading, speed, and altitude) are formatted as a feature vector and passed into the GAN.Parameter generation: The Generator produces new values for the UKF parameters (Q, R, and optionally α, β, κ) based on the input vector.Evaluation and weight update: The Discriminator compares the resulting UKF performance using the generated parameters against expected (or historical) outcomes. It then provides a signal for updating the Generator’s weights through backpropagation.Iteration and convergence: The updated parameters are applied in the next UKF cycle. This loop continues until performance stabilizes or a maximum number of iterations/epochs is reached. The result is a closed learning loop where the GAN continually refines parameter generation based on real-time feedback. Input preparation: Performance metrics from the UKF (e.g., RMSE and MAE for key state variables such as latitude, longitude, heading, speed, and altitude) are formatted as a feature vector and passed into the GAN. Input preparation: Performance metrics from the UKF (e.g., RMSE and MAE for key state variables such as latitude, longitude, heading, speed, and altitude) are formatted as a feature vector and passed into the GAN. Parameter generation: The Generator produces new values for the UKF parameters (Q, R, and optionally α, β, κ) based on the input vector. Parameter generation: The Generator produces new values for the UKF parameters (Q, R, and optionally α, β, κ) based on the input vector. Evaluation and weight update: The Discriminator compares the resulting UKF performance using the generated parameters against expected (or historical) outcomes. It then provides a signal for updating the Generator’s weights through backpropagation. Evaluation and weight update: The Discriminator compares the resulting UKF performance using the generated parameters against expected (or historical) outcomes. It then provides a signal for updating the Generator’s weights through backpropagation. Iteration and convergence: The updated parameters are applied in the next UKF cycle. This loop continues until performance stabilizes or a maximum number of iterations/epochs is reached. The result is a closed learning loop where the GAN continually refines parameter generation based on real-time feedback. Iteration and convergence: The updated parameters are applied in the next UKF cycle. This loop continues until performance stabilizes or a maximum number of iterations/epochs is reached. The result is a closed learning loop where the GAN continually refines parameter generation based on real-time feedback. GAN-enhanced UKF architecture The GAN consists of two fully connected neural networks: Generator: Receives a 100-dimensional input vector representing past UKF performance metrics. This vector is processed through two hidden layers with 128 and 256 neurons, using ReLU activations. The output layer has 10 neurons: the first five represent the diagonal elements of the Q matrix, and the next five correspond to R. In extended configurations, the first three elements may be interpreted as scaling factors (primary, secondary, and tertiary; α, κ, and β).Discriminator: Accepts the generated (or real) 10-dimensional parameter vector as input and processes it through two hidden layers (256 and 128 neurons), outputting a binary classification via a sigmoid activation to indicate whether the parameters are realistic. The GAN is trained using binary cross-entropy loss and the Adam optimizer with a learning rate of 0.0002. Generator: Receives a 100-dimensional input vector representing past UKF performance metrics. This vector is processed through two hidden layers with 128 and 256 neurons, using ReLU activations. The output layer has 10 neurons: the first five represent the diagonal elements of the Q matrix, and the next five correspond to R. In extended configurations, the first three elements may be interpreted as scaling factors (primary, secondary, and tertiary; α, κ, and β). Generator: Receives a 100-dimensional input vector representing past UKF performance metrics. This vector is processed through two hidden layers with 128 and 256 neurons, using ReLU activations. The output layer has 10 neurons: the first five represent the diagonal elements of the Q matrix, and the next five correspond to R. In extended configurations, the first three elements may be interpreted as scaling factors (primary, secondary, and tertiary; α, κ, and β). Discriminator: Accepts the generated (or real) 10-dimensional parameter vector as input and processes it through two hidden layers (256 and 128 neurons), outputting a binary classification via a sigmoid activation to indicate whether the parameters are realistic. The GAN is trained using binary cross-entropy loss and the Adam optimizer with a learning rate of 0.0002. Discriminator: Accepts the generated (or real) 10-dimensional parameter vector as input and processes it through two hidden layers (256 and 128 neurons), outputting a binary classification via a sigmoid activation to indicate whether the parameters are realistic. The GAN is trained using binary cross-entropy loss and the Adam optimizer with a learning rate of 0.0002. During each UKF cycle, the Generator produces candidate parameters using recent performance data. These parameters are applied in the UKF update. The Discriminator assesses the effect on estimation accuracy by comparing the filter’s predicted states with ground-truth values. This adversarial feedback loop allows the system to adaptively tune UKF parameters in real time. This dynamic, performance-driven interaction between the UKF and GAN enables more accurate tracking of complex, nonlinear processes. By continuously optimizing its internal parameters based on performance feedback, the UKF becomes a self-improving filter, ideal for applications in uncertain and rapidly changing environments. Unlike static filtering methods, this framework evolves over time, maintaining optimal performance under varying operating conditions. Detailed description of the UKF \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left(x\right)$$\end{document} function Based on the aircrafts current condition and ongoing environmental variables the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left(x\right)$$\end{document} function is used to forecast the aircrafts future state. With the help of variables like temperature pressure wind direction and speed this function models the non-linear dynamics of the system and modifies the aircrafts position altitude heading and speed. Mathematical breakdown of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left(x\right)$$\end{document} function The state vector in this system is represented as:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{X}\:=\:\left[\begin{array}{c}Latitude\:\\\:\begin{array}{c}Longitude\\\:\begin{array}{c}Altitude\\\:Heading\\\:Speed\end{array}\end{array}\end{array}\right]$$\end{document} where: Latitude: is the latitude of the aircraft (in degrees). Longitude: is the longitude of the aircraft (in degrees). Altitude: is the altitude of the aircraft (in meters). Heading: is the heading direction (in degrees). Speed: is the aircraft’s velocity (in knots). To determine the aircrafts next state the function requires the state vector a time step and various environmental variables. The various components of the function are explained in the steps that follow. Wind Influence: The aircrafts ground speed and heading are affected by the wind’s direction and speed. The wind components \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:u\:(east-west)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:v\:(north-south)$$\end{document} are calculated as:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u_{wind}}=wind{\text{ }}speed \cdot \cos \left( {wind{\text{ }}direction} \right)$$\end{document}20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v_{wind}}=wind{\text{ }}speed \cdot \sin \left( {wind{\text{ }}direction} \right)$$\end{document} The Coriolis effect, which arises from the Earth’s rotation, impacts the aircraft’s velocity in the east-west and north-south directions. It is modeled as:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u_{coriolis}}=2 \cdot {\Omega _{Earth~}}~ \cdot {v_{ground~}} \cdot \sin \,\left( {latitude} \right)$$\end{document}22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v_{coriolis}}=2 \cdot {\Omega _{Earth~}}~ \cdot {v_{ground~}} \cdot \sin \,\left( {latitude} \right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varOmega\:}_{Earth\:}$$\end{document}is the angular velocity of the Earth. The new heading is updated based on the wind drift and Coriolis effect. The corrected speed is calculated as the magnitude of the ground velocity vector:23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:New\_speed\:=\:\sqrt{{u}_{ground}^{2}+{v}_{ground}^{2}}$$\end{document} The updated heading is calculated as:24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:New\_heading\:=\:arctan\:2({v}_{ground}\:,\:{u}_{ground})$$\end{document} Given the new heading and speed, the change in latitude and longitude is computed as:25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:\:Latitude=\:\frac{New\_Speed\:\cdot\:\:dt\:\cdot\:\text{cos}(new\_heading)}{{R}_{earth}}$$\end{document}26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:\:Longitude=\:\frac{New\_Speed\:\cdot\:\:dt\:\cdot\:\text{sin}(new\_heading)}{{R}_{earth}\cdot\:\text{cos}\left(Latitude\right)}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{earth}$$\end{document}, is the radius of the Earth, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:dt$$\end{document} is the time interval. The altitude is updated using the International Standard Atmosphere (ISA) model, based on changes in pressure and temperature. The new altitude is calculated as:27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:new\_altitude\:=previous\_altitude\:+\:\frac{{T}_{0}}{L}(1{-\left(\frac{P}{{P}_{0}}\right)}^{\frac{R\bullet\:L}{g}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{0}\:$$\end{document}is the sea-level standard temperature, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{0}\:$$\end{document}is the sea-level standard pressure, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R$$\end{document} is the gas constant, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L$$\end{document} is the temperature lapse rate. State-space formulation for aircraft navigation In our application, the state-space model for the Unscented Kalman Filter is defined as follows: State vector:28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{k}\:=\:\left[\begin{array}{c}{Latitude\:}_{k}\\\:\begin{array}{c}{Longitude}_{k}\\\:\begin{array}{c}{Altitude}_{k}\\\:{Heading}_{k}\\\:{Speed}_{k}\end{array}\end{array}\end{array}\right]$$\end{document} Inputs / Exogenous variables:29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{k}\:=\:\left[\begin{array}{c}{Wind\:Speed\:}_{k}\\\:\begin{array}{c}{Wind\:Direction}_{k}\\\:\begin{array}{c}{Temperature}_{k}\\\:{Pressure}_{k}\\\:\:\end{array}\end{array}\end{array}\right]$$\end{document} In our setup, the only “inputs” are environmental variables affecting the aircraft; no direct control input is modeled. Measurement vector: All states are observed (noisy):30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}\:=\:\left[\begin{array}{c}{Latitude\:}_{k}^{obs}\\\:\begin{array}{c}{Longitude}_{k}^{obs}\\\:\begin{array}{c}{Altitude}_{k}^{obs}\\\:{Heading}_{k}^{obs}\\\:{Speed}_{k}^{obs}\end{array}\end{array}\end{array}\right]$$\end{document} Nonlinear state transition:31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{k}=f({\widehat{x}}_{k-1},{\text{u}}_{k})+{w}_{k}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f$$\end{document}(⋅) includes: Latitude/Longitude: Updated using the previous position, heading, speed, time delta, wind correction, and Earth’s curvature.Altitude: Updated using climb/descent model and atmospheric pressure/temp.Heading/Speed: Adjusted for wind drift and acceleration. Latitude/Longitude: Updated using the previous position, heading, speed, time delta, wind correction, and Earth’s curvature. Latitude/Longitude: Updated using the previous position, heading, speed, time delta, wind correction, and Earth’s curvature. Altitude: Updated using climb/descent model and atmospheric pressure/temp. Altitude: Updated using climb/descent model and atmospheric pressure/temp. Heading/Speed: Adjusted for wind drift and acceleration. Heading/Speed: Adjusted for wind drift and acceleration. Measurement function:32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}=h\left({\text{x}}_{k}\right)+{v}_{k}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h$$\end{document}(⋅) is the identity (we observe all state variables directly, but with noise). Noise:32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:w\:\sim\:N(0,Q)$$\end{document}32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:v\:\sim\:N(0,R)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:w$$\end{document}: process noise, dynamically tuned by GAN. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:v$$\end{document}: measurement noise, dynamically tuned by GAN. Experimental details Initial parameter settings All experiments were initialized with the following UKF parameters unless otherwise noted. The process noise covariance matrix (Q), measurement noise covariance matrix (R), and UKF scaling factors (α, β, κ) were set as shown in Table 2. These settings follow established recommendations in the UKF literature and provide a common baseline for comparing static and adaptive filtering approaches. Table 2Initial UKF and GAN-UKF parameters.ParameterValueDescriptionQdiag([1, 1, 1, 1, 1])Process noise covariance (diagonal matrix)Rdiag([1, 1, 1, 1, 1])Measurement noise covariance (diagonal matrix)Primary scaling(α)0.1Sigma point spread parameterTertiary scaling(β)2.0Prior distribution parameter (Gaussian)Secondary scaling(κ)1.0Secondary scaling parameter Table 2Initial UKF and GAN-UKF parameters.ParameterValueDescriptionQdiag([1, 1, 1, 1, 1])Process noise covariance (diagonal matrix)Rdiag([1, 1, 1, 1, 1])Measurement noise covariance (diagonal matrix)Primary scaling(α)0.1Sigma point spread parameterTertiary scaling(β)2.0Prior distribution parameter (Gaussian)Secondary scaling(κ)1.0Secondary scaling parameter Initial UKF and GAN-UKF parameters. Sensitivity to initial parameter settings To evaluate robustness, we first tested three symmetric configurations where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Q$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R$$\end{document} were scaled equally. For each configuration, Root Mean Square Error (RMSE) was computed for both the UKF and the GAN-UKF across all five state variables: latitude, longitude, altitude, heading, and speed (see Table 3). Table 3RMSE comparison for UKF vs. GAN-UKF under different initial Q and R values.Q, R ValueModelLat RMSELon RMSEAlt RMSEHeading RMSESpeed RMSEdiag([0.1, 0.1, …])UKF12.0111.38910.521.3517.20GAN-UKF0.060.0810.990.782.46diag([1, 1, …])UKF8.757.92812.101.1213.54GAN-UKF0.040.0610.120.732.38diag([10, 10, …])UKF18.3216.221422.101.7628.70GAN-UKF0.090.1213.220.912.79 Table 3RMSE comparison for UKF vs. GAN-UKF under different initial Q and R values.Q, R ValueModelLat RMSELon RMSEAlt RMSEHeading RMSESpeed RMSEdiag([0.1, 0.1, …])UKF12.0111.38910.521.3517.20GAN-UKF0.060.0810.990.782.46diag([1, 1, …])UKF8.757.92812.101.1213.54GAN-UKF0.040.0610.120.732.38diag([10, 10, …])UKF18.3216.221422.101.7628.70GAN-UKF0.090.1213.220.912.79 RMSE comparison for UKF vs. GAN-UKF under different initial Q and R values. To provide a more rigorous evaluation, we extended this analysis by independently varying Q and R. Specifically, we scaled the diagonal entries of Q and R separately over a predefined grid of values (Q scale ∈ {0.25, 1.0, 4.0}, R scale ∈ {0.25, 1.0, 4.0}). This resulted in a matrix of (Q scale, R scale) combinations rather than a single joint scaling. For each combination, RMSE was computed and visualized as heatmaps. This design ensures that we capture not only symmetric cases (Q = R) but also mismatched scenarios that occur in practice when process and measurement noise have very different magnitudes (see Table 4). Table 4RMSE performance of the GAN-UKF under independent scaling of Q and R.Q ScaleR ScaleModelLat RMSELon RMSEHeading RMSESpeed RMSE0.250.25GAN-UKF19.6621.0134.2572.50.251GAN-UKF52.1655.9248.0972.170.254GAN-UKF102.35105.8155.7394.3110.25GAN-UKF1.81.783.4970.0111GAN-UKF6.696.9633.4571.8914GAN-UKF23.1923.1158.6171.440.25GAN-UKF0.520.640.5366.5641GAN-UKF1.831.676.269.5944GAN-UKF6.225.5748.8370.77 Table 4RMSE performance of the GAN-UKF under independent scaling of Q and R.Q ScaleR ScaleModelLat RMSELon RMSEHeading RMSESpeed RMSE0.250.25GAN-UKF19.6621.0134.2572.50.251GAN-UKF52.1655.9248.0972.170.254GAN-UKF102.35105.8155.7394.3110.25GAN-UKF1.81.783.4970.0111GAN-UKF6.696.9633.4571.8914GAN-UKF23.1923.1158.6171.440.25GAN-UKF0.520.640.5366.5641GAN-UKF1.831.676.269.5944GAN-UKF6.225.5748.8370.77 RMSE performance of the GAN-UKF under independent scaling of Q and R. Figure 3 illustrates the effect of independently scaling the process noise covariance (Q) and measurement noise covariance (R) on the performance of the GAN-UKF. The heatmaps reveal strong interactions between Q and R: for position states (Latitude, Longitude), low RMSE values are concentrated at higher Q scale​ combined with lower R scale​. In contrast, Heading shows a broader sensitivity, with optimal performance near (Q scale = 4, R scale = 0.25), while Speed remains relatively stable across most combinations but deteriorates sharply for small Q scale​ and large R scale​. These results confirm that the optimum often lies off the diagonal (i.e., not where Q and R, are scaled equally), demonstrating that independent tuning of process and measurement noise is critical. Fig. 3Sensitivity heatmaps of the GAN-UKF under independent scaling of Q and R. Sensitivity heatmaps of the GAN-UKF under independent scaling of Q and R. Overall, the GAN-UKF demonstrated robustness to initial parameter settings, consistently maintaining low error rates across symmetric and asymmetric Q/R configurations. By contrast, the static UKF was highly sensitive to these settings, yielding substantially higher RMSE values under suboptimal initialization. Test-time outlier robustness (Monte Carlo) To evaluate robustness to anomalous measurements, we conducted a Monte-Carlo study that injects measurement outliers at test time only (training remains clean). For each test segment, we ran N = 5 trials with randomly perturbed initial states and independently sampled Q/R scaling factors. Two conditions were evaluated per trial: Clean (no outliers) and Outliers (2% of time steps receive injected spikes). Outlier injection: Following standard practice for stress testing filters, we randomly select 2% of time indices and add signed perturbations to the measurement vector where. Lat/Lon: ± (0.02–0.04) deg; Altitude: ± (400–800) m; Heading: ± (45–90) °; Speed: ± (50–100) kt. These magnitudes create clear but plausible anomalies while preserving trajectory context. Comparative analysis and robustness The sensitivity analysis shows that the proposed GAN-UKF approach consistently outperforms the static UKF, regardless of the initial values of Q and R. This highlights the effectiveness of the adaptive mechanism, which dynamically tunes filter parameters in response to real-time feedback, thus ensuring optimal performance even when initial estimates are not ideal. Static UKF models, by comparison, can suffer large and persistent errors if Q and R are not tuned precisely for the given environment. Comparative baselines with non-classical filters To provide a more comprehensive evaluation, we extended our experiments beyond the static UKF baseline and included five widely studied non-classical filters or learning-augmented filters: Extended Kalman Filter (EKF) – a first-order linearization of the nonlinear dynamics.Particle Filter (PF) – a sampling-based Bayesian approach suitable for highly nonlinear systems.Innovation-Adaptive UKF (UKF-IAE) – an adaptive variant of the UKF where the measurement noise covariance R is updated online based on the innovation statistics.KalmanNet – a neural, learning-guided filter that maps observed sequences (e.g., innovations) to data-driven gains/updates.UKF-LSTM (series-hybrid) – a hybrid process model where the nominal UKF transition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left(x\right)$$\end{document} is augmented by an LSTM-predicted residual \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:x$$\end{document}. Extended Kalman Filter (EKF) – a first-order linearization of the nonlinear dynamics. Extended Kalman Filter (EKF) – a first-order linearization of the nonlinear dynamics. Particle Filter (PF) – a sampling-based Bayesian approach suitable for highly nonlinear systems. Particle Filter (PF) – a sampling-based Bayesian approach suitable for highly nonlinear systems. Innovation-Adaptive UKF (UKF-IAE) – an adaptive variant of the UKF where the measurement noise covariance R is updated online based on the innovation statistics. Innovation-Adaptive UKF (UKF-IAE) – an adaptive variant of the UKF where the measurement noise covariance R is updated online based on the innovation statistics. KalmanNet – a neural, learning-guided filter that maps observed sequences (e.g., innovations) to data-driven gains/updates. KalmanNet – a neural, learning-guided filter that maps observed sequences (e.g., innovations) to data-driven gains/updates. UKF-LSTM (series-hybrid) – a hybrid process model where the nominal UKF transition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left(x\right)$$\end{document} is augmented by an LSTM-predicted residual \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:x$$\end{document}. UKF-LSTM (series-hybrid) – a hybrid process model where the nominal UKF transition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left(x\right)$$\end{document} is augmented by an LSTM-predicted residual \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:x$$\end{document}. These methods are frequently used as “intelligent” or hybrid extensions of classical Kalman filtering. All filters were evaluated on the same dataset and under an identical Monte-Carlo protocol (shared train/test splits, randomized initial-state perturbations, independent Q/R scaling, and paired Clean vs. 2% Outliers test conditions), using the same error metrics as our static and GAN-UKF models. Results: comparison across baselines Table 5 summarizes the RMSE across all five state variables. While the PF and UKF-IAE outperform the UKF in certain dimensions, none of these adaptive baselines achieve the accuracy of our proposed GAN-UKF. Table 5RMSE comparison across non-classical and learning-augmented filters.FeatureEKFPFUKF-IAEKalmanNetUKF-LSTMUKFGAN-UKFLatitude55.241.101.268.673,258.422.50.16Longitude55.741.131.237.3813,494.852.750.64Altitude132,69221,216,23020,829,45019,944,522467,114105,95665,916Heading135.46206.05228.21171.98179.3385.1529.33Speed614.417373.5874.576,078.0986.7671.1 Table 5RMSE comparison across non-classical and learning-augmented filters.FeatureEKFPFUKF-IAEKalmanNetUKF-LSTMUKFGAN-UKFLatitude55.241.101.268.673,258.422.50.16Longitude55.741.131.237.3813,494.852.750.64Altitude132,69221,216,23020,829,45019,944,522467,114105,95665,916Heading135.46206.05228.21171.98179.3385.1529.33Speed614.417373.5874.576,078.0986.7671.1 RMSE comparison across non-classical and learning-augmented filters. Monte Carlo evaluation protocol To ensure statistically sound comparisons not tied to a single trajectory or parameter choice, we conducted a Monte-Carlo evaluation. Each filter - UKF (static), GAN-UKF (dynamic), EKF, PF, UKF-IAE, and the two learning-augmented hybrids KalmanNet and UKF-LSTM - was tested over multiple randomized trials per segment with an identical protocol where: Initial conditions were perturbed by small Gaussian noise.Noise covariances (Q and R) were scaled independently across a predefined grid of values.Test-time outliers were injected in some runs to simulate sensor anomalies. Initial conditions were perturbed by small Gaussian noise. Initial conditions were perturbed by small Gaussian noise. Noise covariances (Q and R) were scaled independently across a predefined grid of values. Noise covariances (Q and R) were scaled independently across a predefined grid of values. Test-time outliers were injected in some runs to simulate sensor anomalies. Test-time outliers were injected in some runs to simulate sensor anomalies. For each run, we computed RMSE and MAE across all five state variables. Results were then aggregated into boxplots Figs. 8 and 9 at the results section, providing both median performance and variability across runs. This allows for a robust comparison between filters under diverse operating conditions. Computational overhead and real-time feasibility The proposed GAN-UKF framework introduces only modest computational overhead relative to traditional Kalman Filters, primarily due to the integration of a generator network that dynamically predicts the UKF parameters (Q, R, and optionally α, β, and κ). However, the generator is a lightweight feedforward neural network with two hidden layers (128 and 256 neurons), and only a single forward pass is required during each UKF cycle. Performance profiling shows that parameter generation completes in under 10 milliseconds on a GPU (NVIDIA RTX 3080) and under 50 milliseconds on modern CPUs (e.g., Intel i7-12700 H). These runtimes fall well within the latency constraints of real-time systems operating at 1 Hz or lower, such as those in aviation navigation. Additionally, complete UKF-GAN cycle times across test segments—including prediction, update, and parameter adaptation—remained under 3 s per segment. Therefore, the proposed architecture is well-suited for real-time and embedded applications where computational resources and latency are constrained. The component-level runtimes and per-cycle totals (including framework/IO overheads) are summarized in Table 6 that follow: Table 6Runtime breakdown for GAN-UKF system.ComponentRuntime (ms)UKF prediction + Update30GAN generator (Q, R, α, β, and κ)10Total per cycle (CPU)80Total per cycle (GPU)40Average Segment runtime3000 Table 6Runtime breakdown for GAN-UKF system.ComponentRuntime (ms)UKF prediction + Update30GAN generator (Q, R, α, β, and κ)10Total per cycle (CPU)80Total per cycle (GPU)40Average Segment runtime3000 Runtime breakdown for GAN-UKF system. Results To evaluate the effectiveness of the proposed approach, a comprehensive performance comparison was conducted between the static and dynamic UKF models. The primary evaluation metrics included Root Mean Square Error (RMSE) for key state variables: latitude, longitude, altitude, heading, and speed. These metrics were calculated for both the standard UKF with fixed parameters and the GAN-enhanced UKF with dynamically adjusted parameters. Point-error performance The results indicate that the dynamically adjusted UKF (GAN-UKF) consistently outperformed the static version across most evaluated variables. A summary of the key findings is presented in the Table 7 that follow: Table 7RMSE comparison between static and dynamic GAN-UKF.FeatureUKF RMSEGAN-UKF RMSELatitude5.43 degrees 0.08 degrees Longitude4.24 degrees 0.07 degrees Altitude55,715 m 10.66 m Heading 0.12 degrees 0.83 degreesSpeed50.01 knots 2.58 knots Significant values are in bold. Table 7RMSE comparison between static and dynamic GAN-UKF.FeatureUKF RMSEGAN-UKF RMSELatitude5.43 degrees 0.08 degrees Longitude4.24 degrees 0.07 degrees Altitude55,715 m 10.66 m Heading 0.12 degrees 0.83 degreesSpeed50.01 knots 2.58 knots Significant values are in bold. RMSE comparison between static and dynamic GAN-UKF. Significant values are in bold. Table 7 illustrates a substantial improvement in RMSE values for latitude and longitude, with errors reduced from several degrees in the static model to only fractions of a degree in the dynamic model. A similarly notable improvement is observed in altitude estimation, where the error drops from tens of thousands of meters to just over 10 m, indicating a significant enhancement in vertical position accuracy. While the heading error shows a slight increase, a major gain is seen in speed estimation, where the RMSE decreases from 50.01 knots to just 2.58 knots. This suggests that the dynamic model is considerably more effective in capturing speed variations under nonlinear conditions. Uncertainty calibration from the posterior covariance Beyond point-error metrics, we analyze the filters’ posterior state covariance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{t}$$\end{document}. For each state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} we form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{t,i}=\:\sqrt{{P}_{t}(i,i)}$$\end{document} and evaluate calibration via: Coverage at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\sigma\:$$\end{document} (nominal ≈ 68%), Coverage at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1.96\sigma\:$$\end{document} (≈ 95%), normalized error \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mid\:z\mid\:=\mid\:e\mid\:/\sigma\:$$\end{document} (median and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{90}^{th}$$\end{document}-percentile), and the ratio MAE/mean\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{\upsigma\:}$$\end{document}. These summarize whether the reported uncertainty is consistent with the empirical error scale (see Table 8). (Heading errors are computed with circular wrapping to [\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{-180}^{0}$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{180}^{0}$$\end{document}]) Table 8Uncertainty calibration from the posterior covariance.FilterStatemean σCoverage at 1σ (%)Coverage at 1.96 σ (%)median |z|p90 |z|MAE/mean σUKFLatitude1.2723520.5494511.0989015.3257416.6003026.338541UKFLongitude1.2723580.5494511.3736264.1901175.4560265.356243UKFAltitude1.2723610051616.7751627.9651389.82UKFHeading1.41305779.9450585.989010.2829572.6448213.276888UKFSpeed1.0127490.5494510.54945165.8330588.1920268.65678GAN-UKFLatitude0.07984963.1868176.373630.5921214.2803641.480636GAN-UKFLongitude0.09677867.8571478.021980.3956114.0529051.807042GAN-UKFAltitude80.8943198.3468196.876110.0036510.0036520.003689GAN-UKFHeading139.206592.8571497.527470.0000050.6382290.209423GAN-UKFSpeed15.2100171.1538572.527470.2522115.3024251.491624 Table 8Uncertainty calibration from the posterior covariance.FilterStatemean σCoverage at 1σ (%)Coverage at 1.96 σ (%)median |z|p90 |z|MAE/mean σUKFLatitude1.2723520.5494511.0989015.3257416.6003026.338541UKFLongitude1.2723580.5494511.3736264.1901175.4560265.356243UKFAltitude1.2723610051616.7751627.9651389.82UKFHeading1.41305779.9450585.989010.2829572.6448213.276888UKFSpeed1.0127490.5494510.54945165.8330588.1920268.65678GAN-UKFLatitude0.07984963.1868176.373630.5921214.2803641.480636GAN-UKFLongitude0.09677867.8571478.021980.3956114.0529051.807042GAN-UKFAltitude80.8943198.3468196.876110.0036510.0036520.003689GAN-UKFHeading139.206592.8571497.527470.0000050.6382290.209423GAN-UKFSpeed15.2100171.1538572.527470.2522115.3024251.491624 Uncertainty calibration from the posterior covariance. The comparative results demonstrate that the dynamically tuned GAN-UKF substantially outperforms the static UKF on point-error metrics and delivers more informative - and generally better-calibrated - uncertainty estimates. As shown in Table 7, lateral accuracy improves dramatically (latitude/longitude RMSE fall from several degrees to mere hundredths of a degree), altitude error is reduced from tens of Km to ~ 10 m, and speed RMSE drops from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sim50\:kt$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sim2.6\:kt$$\end{document}. Table 8 reports the posterior state standard deviations (mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}) produced by each filter and compares them to the empirical error scale via coverage at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\:\sigma\:/1.96\:\sigma\:$$\end{document}, normalized error quantiles (median \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|z\right|$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p90\:\left|z\right|$$\end{document}), and MAE/mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}. The static UKF is clearly under-dispersed for most states (very low coverages and MAE/mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\gg\:1$$\end{document}), indicating over-confidence inconsistent with the MAE tunnel. By contrast, the GAN-UKF achieves lateral coverages close to nominal (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\approx\:63-68\%\:at\:1\:\sigma\:\:and\:76-78\%\:at\:1.96\:\sigma\:$$\end{document}) with MAE/mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\:$$\end{document}near 1.5–1.8, aligning well with the tunnel and reflecting credible uncertainty. Some conservatism remains in altitude (and to a lesser extent heading), where coverage approaches 100% and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|z\right|$$\end{document} is very small. Overall, the GAN-UKF pairs large accuracy gains with uncertainties that meaningfully track realized errors. Visualization The graphs below show detailed results for one representative flight segment, to illustrate typical filter performance. Figure 4 presents the normalized Mean Absolute Error (MAE) over time for this segment, comparing the static UKF (dashed lines) with the dynamic GAN-UKF (solid lines) for key variables such as latitude, longitude, altitude, and heading. Additional example segments are provided in the Supplementary Information. The visual comparison emphasizes the dynamic model’s ability to adapt and reduce error consistently across time. Fig. 4Comparison of MAE for static vs. dynamic UKF. Comparison of MAE for static vs. dynamic UKF. In the early stages of state estimation, error behavior varies across the different state variables. The static UKF, constrained by its fixed noise parameters, often maintains elevated errors because it cannot adjust to rapid changes in system dynamics. By contrast, the dynamic UKF adapts its noise parameters in response to real-time performance data, enabling it to reduce mean absolute error (MAE) more effectively. For instance, in the altitude dimension, the dynamic UKF initially shows larger errors but quickly converges to lower values, while in the heading dimension it consistently outperforms the static approach. For latitude and longitude, both methods begin with similar error levels before the dynamic UKF achieves faster reductions. This trend is clearly illustrated in Fig. 4, which shows that the dynamic UKF consistently outperforms the static model across all variables. Notably, the adaptive model’s ability to update both process and measurement noise covariances allows it to quickly minimize Root Mean Square Error (RMSE) - particularly for challenging variables like heading and altitude. On the other hand, the static UKF, constrained by its constant noise assumptions, maintains consistently higher RMSE values, highlighting its reduced effectiveness in dynamic environments. This limitation is further demonstrated in Fig. 5, where the static model’s inability to adapt results in persistent estimation inaccuracies under varying system conditions. Fig. 5Comparison of RMSE for static vs. dynamic UKF. Comparison of RMSE for static vs. dynamic UKF. The dynamic and static UKF models variability in MAE and RMSE can be seen visually in the additional tunnel plots Figs. 6 and 7. Fig. 6Comparison of static vs. dynamic UKF RMSE Tunnels. Comparison of static vs. dynamic UKF RMSE Tunnels. Fig. 7comparison of static vs. dynamic UKF MAE Tunnels. comparison of static vs. dynamic UKF MAE Tunnels. The robustness and consistency of each filter are illustrated by the shaded tunnels, which represent the range between the minimum and maximum recorded error values. Compared to the static UKF (blue shading), the dynamic UKF (red shading) exhibits a narrower and more concentrated range, particularly over time. This indicates not only lower variability but also consistently smaller mean errors, demonstrating that the dynamic model produces more stable and reliable estimates. The significant reduction in error variability reflects the dynamic UKF’s enhanced ability to adapt to environmental changes in real time. This robustness is clearly visualized in the tunnel plots, where the narrower error bands associated with the dynamic UKF confirm its superior consistency under varying conditions. In real-world applications - where fixed noise assumptions often fail - the ability to dynamically adjust filtering parameters is essential. To avoid overloading the main figures with too many trajectories, we provide further example results for additional flights in the Supplementary Information. This allows the reader to verify that the observed improvements hold across a range of different aircraft and operational scenarios. Monte Carlo results To further validate the robustness of the proposed GAN-UKF, we performed a Monte Carlo study as described in paragraph ‎6.4.6. Each filter was evaluated over multiple randomized runs with perturbed initial conditions, independently scaled process and measurement noise covariances, and optional test-time outliers. Figures 8 and 9 present boxplots of RMSE and MAE, respectively, aggregated across all runs and state variables. The results demonstrate that: GAN-UKF consistently achieves the lowest median error across all state variables (latitude, longitude, altitude, heading, speed). Error variability is smallest for the GAN-UKF, indicating not only higher accuracy but also greater robustness to variations in initial conditions and noise. Non-classical baselines (EKF, PF, UKF-IAE) outperform the static UKF in some cases, but none match the performance of GAN-UKF. In particular, particle filtering shows higher variance, while EKF and UKF-IAE remain sensitive to initialization. KalmanNet attains low Speed error but exhibits very large Altitude error, which dominates its overall performance. UKF-LSTM (series-hybrid) is unstable on geo and speed states (large Lat/Lon/Speed errors), yielding inferior composite accuracy and higher variability than GAN-UKF. Under outlier-injected trials, the GAN-UKF maintained stable performance, while baseline filters showed significant increases in error. Overall, the Monte Carlo results confirm that the proposed GAN-UKF is not only more accurate but also more robust and reliable across diverse conditions, addressing a key limitation of static and semi-adaptive filters. Fig. 8Monte Carlo boxplots RMSE. Monte Carlo boxplots RMSE. Fig. 9Monte Carlo boxplots MAE. Monte Carlo boxplots MAE. Robustness to outliers To complement the clean-data experiments, we repeated the Monte Carlo study with synthetic measurement outliers injected only at test time (2% of timestamps perturbed across latitude, longitude, altitude, heading, and speed). We report aggregate results for both conditions - clean and outlier-injected - in the Monte Carlo results. For each filter we compute mean errors across all segments and runs. Across conditions, introducing outliers increases error for every method, as expected, but to markedly different degrees. Under outliers, the GAN-UKF retains the lowest aggregate error (lowest RMSE total) among the tested filters, followed by the best of the adaptive baselines. The static UKF and PF show the largest increases, particularly in the position channels (Lat/Lon) and altitude, indicating higher sensitivity to spiky measurements. GAN-UKF provides the strongest robustness trade-off across conditions. It attains the lowest total error on clean data and remains the top performer when outliers are injected, showing a smaller increase in error than UKF, EKF, and PF. Methods like UKF-IAE, PF, KalmanNet, and UKF-LSTM either start from much higher error levels or display instability, so their small changes under outliers are not indicative of true robustness. This aligns with our sensitivity analysis where independent tuning of Q and R away from the diagonal (Q ≠ R) improved resilience to measurement anomalies. Test-time outliers degrade all filters, but GAN-UKF is most robust in aggregate, showing the smallest error increase and the lowest outlier-condition RMSE total across runs/segments. Full per-filter numbers (clean vs. outliers, component-wise deltas) are provided in the accompanying summary tables Table 9. Figure 10(with outliers) and Fig. 11 (clean) present single-panel boxplots of total RMSE across all filters, enabling a direct, like-for-like comparison of robustness to test-time outliers. Fig. 10Total RMSE by filter - with outliers. Total RMSE by filter - with outliers. Where %Δ = 100 × (outlier mean − clean mean) / clean mean. Table 9Clean vs. Outlier Monte Carlo (per filter; means).FilterRMSE total (Clean)RMSE total (Outliers)ΔRMSE total%ΔRMSE totalGAN-UKF66253.260666260.25016.98950.0105UKF106106.6118120113.877614007.265813.2011EKF133806.5951155918.61822112.02316.5254UKF_IAE20829360.3820824761.24-4599.1309-0.0221PF21216505.1321216515.119.98230.2100UKF-LSTM494312.7487831.7-6480.94-1.3111KalmanNet20,865,29920,865,2995.59620.0112 Table 9Clean vs. Outlier Monte Carlo (per filter; means).FilterRMSE total (Clean)RMSE total (Outliers)ΔRMSE total%ΔRMSE totalGAN-UKF66253.260666260.25016.98950.0105UKF106106.6118120113.877614007.265813.2011EKF133806.5951155918.61822112.02316.5254UKF_IAE20829360.3820824761.24-4599.1309-0.0221PF21216505.1321216515.119.98230.2100UKF-LSTM494312.7487831.7-6480.94-1.3111KalmanNet20,865,29920,865,2995.59620.0112 Clean vs. Outlier Monte Carlo (per filter; means). Fig. 11Total RMSE by filter - clean (no outliers). Total RMSE by filter - clean (no outliers). Comparative analysis with intelligent filters The additional results confirm that while non-classical filters reduce error relative to the UKF baseline, their performance remains well below that of the GAN-UKF. For example, UKF-IAE achieved improvements of ~ 20–30% over UKF in latitude and longitude estimation, but the GAN-UKF delivered > 90% error reduction in the same dimensions. Particle filters also offered moderate improvements but suffered from instability and higher variance, particularly in altitude estimation. Overall, the GAN-UKF consistently outperformed all tested baselines across every state variable, demonstrating its robustness and generalization capability. This validates that the observed gains are not merely incremental over the UKF, but represent a substantial advance beyond existing “intelligent” filters. These performance gains are especially evident in challenging scenarios such as abrupt manoeuvres or unpredictable weather, where traditional, static models often underperform. The findings strongly support the adoption of adaptive filtering techniques, like the GAN-UKF, in domains that demand high-fidelity real-time estimation. Conclusion This study introduced a novel approach for enhancing state estimation in non-linear dynamic systems by integrating Generative Adversarial Networks (GANs) with the Unscented Kalman Filter (UKF). The core innovation lies in leveraging the generative capabilities of GANs to dynamically adjust key UKF parameters - including the process noise covariance (Q), measurement noise covariance (R), and the scaling factors (primary, secondary, and tertiary; α, κ, and β) - all of which are essential for accurate state prediction in nonlinear environments. Unlike existing adaptive approaches such as IAE or IMM, which focus on heuristic tuning of measurement noise, the GAN-UKF jointly adapts both the process noise (Q), measurement noise (R), and the scaling parameters (α, β, κ). This enables holistic, data-driven adaptation of the filter dynamics. To our knowledge, this is the first approach that integrates a generative adversarial model with a UKF to simultaneously adjust multiple parameters sets in real time, leading to superior accuracy and robustness. Empirical evaluations using real-world aircraft navigation data demonstrated that the dynamically tuned UKF significantly outperformed its conventional, static counterpart. The improvements were most notable in Root Mean Square Error (RMSE) for critical state variables such as latitude, longitude, altitude, heading, and speed. For example, RMSE in spatial positioning (latitude and longitude) was reduced from several degrees to fractions of a degree, highlighting the model’s ability to adapt and maintain precision. While this study focused on aircraft navigation, the proposed framework has broad applicability. The GAN-enhanced UKF architecture can be extended to any domain that requires accurate, real-time state estimation under uncertainty - including robotics, autonomous vehicles, environmental monitoring, and smart city systems. In summary, this work establishes a foundation for flexible, adaptive state estimation in complex systems. It demonstrates that integrating machine learning techniques - specifically GANs - into classical filtering frameworks such as the UKF can significantly improve estimation performance. The approach offers a powerful path forward for advancing state estimation in modern, data-rich, and dynamic environments. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1