Artificial Intelligence (AI) has been revolutionising large parts of society. Its subset, Machine Learning (ML), has recently been leveraged to begin improving Global Navigation Satellite System (GNSS) position, navigation and timing (PNT) performance. Early attempts at replacing standard optimal estimation methods directly with ML proved limited and adoption by industry tepid. As ML algorithms rapidly evolve, ML use in GNSS research has grown substantially and initial effective approaches can be seen in areas such as measurement classification and positioning.

This research focusses on position level information and GNSS feature engineering required for advanced ML algorithm use. The objectives are: Can the resulting model 1) generate stable and useful corrections under specific circumstances, and 2) reveal patterns between feature property and model efficiency. Precise Point Positioning (PPP) GNSS process is performed on some of the popular Google Smartphone Decimeter Challenge (GSDC) public datasets, as smartphone GNSS data are the most difficult to model with conventional means. A York University PPP engine is used for this solution, and multiple advanced ML algorithms are selected in ensembled learning: a) convolutional neural network (CNN) + long short-term memory (LSTM) and b) CNN + transformer. a) is good at modelling time series and is quick, but can overfit data and has reduced effectiveness over long sequences. b) is more efficient and good at capturing global dependencies, but is less efficient on small datasets and involves more complex tuning.

Initial results with a subset of GSDC training datasets shows few-decimetre-level improved ML-based positioning from PPP-only solutions. As expected, individual dataset results varied based on local collection environments, e.g., open sky, urban canyons, etc. CNN+LSTM produces an overall horizonal rmse of 1.21 +- 0.82 m and CNN+transformer 1.17 +- 0.49 m. For some datasets, both models produced impressive ~50 cm horizontal rmse. Therefore, while both ensemble learning approaches produce statistically similar positioning results, the latter does so with less individual solution variation. These are preliminary experiments. Numerous questions remain as to the effectiveness of individual input GNSS features and their use. Future work will involve further feature engineering, determination and understanding of the correlation between the features and resulting spatial information, and attempting to understand how the ML models are interpreting systematic and gross errors in the data.

Autonomous navigation systems, such as Unmanned Ground Vehicles (UGVs), operating in GNSS-denied environments must rely on accurate and consistent onboard sensing to support robust state estimation. The prediction stage using Extended Kalman Filter (EKF) employs motion model equations based on inertial measurement unit (IMU) data, with high-frequency updates from the IMU to propagate the system state; sampling frequencies are typically between 100 Hz and 1 kHz. However, long correction intervals can introduce errors during inertial prediction, especially when state updates are limited by the low update frequencies of LiDAR navigation measurements, which typically operate between 10 Hz and 20 Hz, thus limiting the accuracy of the state estimation. The research explores a strategy center on the generation of data using LiDAR sensor measurements to densify observation updates and improve state estimation. This work designs a synthetic observation framework to generate oversampled LiDAR measurements, closely synchronized with inertial prediction states. Oversampled data are integrated into State Estimation pipelines using the Extended Kalman Filter (EKF) to obtain more frequent correction updates and consistent temporally. This paper presents a systematic investigation of different machine learning techniques for generating synthetic LiDAR observations to evaluate each technique in terms of predictive fidelity, uncertainty representation, and computational requirements, and conducts a comparative evaluation to determine their effectiveness in enhancing state estimation. A simulation developed in a Python environment is employed to analyze different GNSS-denied configurations for autonomous vehicles. The study further examines key balancing effects related to data smoothness, computational load, and predictive uncertainty, providing insights into the practical integration of generated synthetic data within state estimation pipelines. Results show that synthetic observations generated based on ML methods significantly enhance estimation robustness and stabilize the state during extended periods of sparse sensing prior to correction, thereby limiting error growth before subsequent update events. Overall, the findings present synthetic data generation as a scalable approach to improve the navigation performance of UGVs in GNSS-denied environments.

This work investigates the application of Machine Learning techniques to augment low-cost Microelectromechanical Systems (MEMS) based Inertial Measurement Units (IMU). In recent years, significant advancements in ML, particularly deep learning, have been realised enabling promising solutions for mitigating IMU errors such as bias, cross-coupling, scale factor, and noise.

A literature review identified three candidate architectures for evaluation: Convolutional Neural Networks (CNN), Temporal Convolutional Networks (TCN), and a hybrid CNN–Long Short-Term Memory–Attention model. Additionally, a Multilayer Perceptron (MLP) was implemented as a baseline.
An experimental dataset was collected using composite IMU which contains an ensemble of 16 MEMS IMUs. In its usual mode of operation, the data from 16 IMUs are averaged to form a composite output whose performance is significantly improved compared to a single IMU (a factor of 4 reduction in stochastic error terms).

The input to the ML models was a contiguous window of inertial measurements from a single IMU from the ensemble, the target output of the ML models was the corresponding window of inertial measurements averaged across the remaining 15 IMUs in the ensemble. The models were trained and evaluated using the Root-Mean-Square-Error (RMSE) between the model output and target output as the primary metric.

Results show that the TCN architecture achieved the best overall performance, reducing acceleration and angular rate errors, relative to the ensemble, by factors of 1.8 and 1.35 on the dataset respectively. The CNN and hybrid models failed to reduce errors, while the MLP demonstrated modest improvements in acceleration error. Bayesian optimisation of the TCN model’s hyper-parameters further improved performance, though at the expense of a significant increase (approximately 300x) in model size.

The study concludes that TCNs offer the most promising balance of accuracy and computational efficiency for MEMS based IMU augmentation – in our case enabling a single IMU to more closely approximate the performance of the full ensemble. Future work will focus on expanding the training dataset with synthetic data, refining evaluation metrics, and implementing real-time solutions on low size, weight and power embedded platforms.

In urban environments, GNSS positioning performance is severely degraded by signal obstructions and multipath effects, leading to significant 3D positioning errors. This research, developed as part of a doctoral thesis at the Università di Genova, addresses these limitations by introducing an adaptive machine learning (ML) framework designed to improve navigation robustness through real-time error prediction and mitigation.
The core of this work is a Random Forest Regressor trained to predict 3D positioning errors by analyzing real-time signal features. The model utilizes a specific set of input variables, including Signal-to-Noise Ratio (SNR), the number of visible satellites, the Code minus Carrier (CMC) etc.
By processing these features, the regressor provides a reliable estimate of the expected error under varying environmental conditions.
A key innovation of this approach is the implementation of a dynamic parameter selection mechanism. The system leverages the model’s predictions to evaluate multiple processing configurations in real-time, automatically selecting the parameters that minimize the predicted 3D error. This adaptive strategy allows the navigation system to maintain high robustness even when signal quality degrades.
Beyond error correction, the framework is designed to calculate real-time risk levels, providing a quantitative assessment of positioning reliability.
Validation through field tests in multipath-rich urban environments shows that this ML-based approach significantly improves positioning accuracy, achieving an improvement of up to 50% in the most challenging scenarios.
Ultimately, this work provides a scalable solution for high-precision navigation in urban scenarios by establishing a robust, ML-driven framework for active error mitigation.

Today, society has become largely dependent on Position, Navigation and Time (PNT) information derived from Global Navigation Satellite Systems (GNSS). These systems provide accurate estimates of position and time, but are vulnerable to interference (jamming) due to their weak signal strengths. In military contexts, this poses a significant threat to operational effectiveness, as falling back to conventional navigation methods such as maps and compasses is time-consuming and error-prone. Finding satisfactory alternative navigation solutions is particularly challenging for military personnel on foot, as they require tight size, weight, and power constraints, while covertness necessitates passive sensors. Smaller inertial measurement units (IMUs) satisfy these demands, but the pure integration of their measurements accumulates errors that render them unusable within minutes. Thus, external aiding is necessary to limit these errors.

Recent research for improved IMU-based pedestrian navigation is based on deep learning approaches. Research in this field is mainly focussed on civilian use-cases and often on indoor navigation. A promising approach by Liu et al. named TLIO (tight learned inertial odometry), tightly couples an extended kalman filter (EKF) with a convolutional neural network (CNN). This set-up allows for EKF state predictions using an IMU to be corrected by outputs from the CNN. These outputs consist of displacement estimates and estimated uncertainties, based on an input buffer of IMU measurements. This current work combines this with a similar approach inspired by Bajwa et al., called DIVE (deep inertial-only velocity aided estimation). In contrast to TLIO, the CNNs output 3D linear velocities based on body-aligned frame IMU measurements, allowing better integration into our existing KF framework and requiring less processing.

This research aims to show operational value of the NN-aided INS in a military context. As such, real-world outside data were used exclusively. These data were gathered on multiple test days, with multiple test `subjects', movement speeds, and terrain types. By utilising RTK GPS measurements as references, the performance of the system's position estimation can be more reliably and quantitively analysed, compared to earlier indoor navigation research. A custom hardware setup was used, which consists of a sensor box mounted to a tactical vest.

The test set consists of a person running through a mix of loose sand and walking on crushed seashell paths. The total distance covered is about 1.5 km. The total accumulated position estimation error at the end of the test trajectory is circa thirty meters. These results indicate that the deep learning-aided inertial navigation solution is feasible for outdoor applications. The longer distances and different terrain types, including rough terrain like loose sand, give confidence that this is a promising solution for a military context.