This thesis aims to improve the Equimod model by integrating sensorimotor contingencies theory. The main objective is to explore how action and perceptual transformations can enrich machine learning representations. The work will focus on three main axes: analyzing the representation space structure, improving representations through sensorimotor contingencies, and developing active learning mechanisms. The study will seek to create a more flexible and performant model, capable of better generalization with few annotated data.