In this HDR I present some of my works from the last 10 years. My research spans several areas related to artificial intelligence such as computational neuroscience, developmental robotics, machine learning, multi-agent systems, psychophysical modelling, intelligent tutoring systems, etc. My main research question is how an autonomous agent can make sense of its environment through learning and perception influenced by the enactivist and constructivist approaches that emphase the role of action in the process. To this end, I propose some contributions with an approach that is bio-inspired, through the use of neural networks and some mechanisms coming from neuroscience and cognitive science, focusing mainly on the structuration of representations. I have therefore chosen to focus on three main research axis. First, which representations to learn and how to structure them with self-supervised learning. Second, how structured representations can shape the fusion of various modalities into a unified perception. Third, the interaction between the temporality of learning and representations in active or incremental learning.  Moreover, I defend in my research project the interconnection of these three different elements within the framework of the sensorimotor contingencies theory in order to obtain more robust, adaptive and generic agents.