Sixth-generation (6G) networks are expected to deliver positioning and sensing as native services, supported by large antenna arrays, wide bandwidths, and carrier frequencies that make the propagation channel highly informative about geometry. Classical localization and sensing estimators exploit this structure explicitly, but degrade under array imperfections, coherent multipath, and non-line-of-sight propagation. Deep learning architectures are more tolerant to such effects, at the cost of large labelled datasets, opaque behaviour, poor transfer across environments, and long training times. The goal of this project is to develop model-based machine learning methods for positioning and sensing. Under this paradigm, physical models, such as propagation channel or array response models, are incorporated into the design, training, or optimization of the learning method, so that learning is restricted to the components that these models fail to capture. The expected benefit is a substantial reduction in the amount of labelled data required, improved robustness to model mismatches, and estimators whose behaviour remains interpretable. GPU resources are required at two stages. First, in order to train and evaluate the performance of the developed neural architectures over multiple array geometries, carrier frequencies, and signal-to-noise ratios configurations. Second, realistic data are generated through ray tracing over dense receiver grids, a time-consuming process that is considerably shortened on GPUs.