Léa Dubreil

Thesis: Processing of degraded GNSS measurements by means of Machine Learning as input of a hybrid positioning algorithm Supervising team: Gaël Pages (researcher, PhD, HdR, ISAE-SUPAERO, Toulouse), Samy Labsir, and Etienne Rouanet-Labé (navigation engineer, Thales Alenia Space, Toulouse)

Data-driven techniques like neural networks open the doors of improvements in system operating in harsh environments, where model-based techniques falter due to degraded hypotheses. This thesis is inscribed at the crossroads of Bayesian estimation and deep learning. There, we research how to tailor machine learning for the specificity of urban navigation. Indeed, off-the-shelves machine learning models are not sufficient to sustain the robustness, frugality and adaptability required in Global Navigation Satellite Systems (GNSS) receivers. Léa has been working on a novel method of retaining the physical meaning of the navigation solution based on geometrical and algebraic mathematical modelling compatible with a deep neural network. This produces interpretable results, paving the way to a certifiable algorithm. Her main research questions are now oriented towards benchmarking of these methods with regard to other hybrid architectures and classical Bayesian estimators such as Kalman Filters.