One of the major challenges of AI is the design of a suitable representation to organize knowledge and exploit it. 
In the context of the work presented, I focus on the autonomous design of this representation, by an embodied agent 
that exploits its interactions with its environment in order to build a representation that it can use to achieve 
its own objectives. An ideal study framework for this development relies on an open and unknown environment prior 
to the system's design. Some challenges are discussed and illustrated throughout this research thesis, from the 
desired qualities for this representation support, to its dynamics which require defining rewards and priming mechanisms. 
Through different projects and different centralized or decentralized learning methods, this thesis addresses targeted 
issues for the development of such an agent, from the perception of patterns from raw continuous low-level signals, 
to the learning of algorithms in an endogenous representation, but also to the online learning of semantic concepts.