The scrum is one of the most iconic and demanding phases of rugby union. Opposing eight players from each team, it represents both a major tactical challenge and a complex biomechanical context, characterized by high levels of physical effort and an increased risk of serious injury. This phase, therefore, carries a dual performance and health stake, requiring tools capable of finely characterizing the forces produced by players. Yet existing studies are often dated and generally provide global measurements obtained under conditions far from actual on-field scrummaging conditions. It is in this context that the present thesis was conducted, in collaboration with Stade Toulousain, with the primary objective of developing and validating an embedded instrumentation system enabling the simultaneous estimation of individual 3D pushing forces for each of the eight players in a pack, under real scrummaging conditions. Attention was focused on ground reaction forces (GRF) as the primary mechanical
quantity for quantifying individual pushing efforts. Among the available solutions for measuring GRF in field conditions, instrumented insoles were identified as the most suitable given the constraints imposed by the context of this thesis. However, these sensors only measure forces perpendicular to their surface and exhibit an overestimation bias when highly flexed inside the shoe, as is the case during scrummaging. A Machine Learning (ML) approach was therefore adopted to correct this bias and estimate all three components of the GRF. A proof of concept was first established on an initial group of subjects, demonstrating the feasibility of estimating three-dimensional (3D) pushing forces during scrummaging with errors comparable to those reported in the literature for walking and running. A comparative study of several ML architectures (Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), Random Forest (RF)) and training strategies was then conducted, including elite development players from Stade Toulousain. A personalized MLP model trained on data from these players achieved the best performance, with body-weight-normalized errors below 8.5% across all axes, including during critical phases such as foot repositioning. By extending the training dataset to both subject groups, model generalization was improved with no significant loss of performance. At the end of this study, two operational models, one generic and one personalized, were provided and are directly applicable to Stade Toulousain players. These models were then deployed during three experimental field sessions under real scrummaging conditions, enabling an analysis of pushing forces at three levels collective, individual, and intra-individual. The results illustrate the potential of this tool for performance monitoring and the preservation of players’ physical integrity in a professional context. In summary, this thesis proposes a comprehensive and validated methodological framework for the synchronized, individualized measurement of 3D pushing forces during rugby scrummaging under real conditions, combining embedded instrumentation and ML. It thereby contributes new elements to the biomechanical analysis of the scrum and opens the way for future developments, in particular the integration of individual kinematic information for an even more complete characterization of the mechanisms underlying collective and individual scrum performance.