Robots are still limited to controlled conditions, that the robot designer\nknows with enough details to endow the robot with the appropriate models or\nbehaviors. Learning algorithms add some flexibility with the ability to\ndiscover the appropriate behavior given either some demonstrations or a reward\nto guide its exploration with a reinforcement learning algorithm. Reinforcement\nlearning algorithms rely on the definition of state and action spaces that\ndefine reachable behaviors. Their adaptation capability critically depends on\nthe representations of these spaces: small and discrete spaces result in fast\nlearning while large and continuous spaces are challenging and either require a\nlong training period or prevent the robot from converging to an appropriate\nbehavior. Beside the operational cycle of policy execution and the learning\ncycle, which works at a slower time scale to acquire new policies, we introduce\nthe redescription cycle, a third cycle working at an even slower time scale to\ngenerate or adapt the required representations to the robot, its environment\nand the task. We introduce the challenges raised by this cycle and we present\nDREAM (Deferred Restructuring of Experience in Autonomous Machines), a\ndevelopmental cognitive architecture to bootstrap this redescription process\nstage by stage, build new state representations with appropriate motivations,\nand transfer the acquired knowledge across domains or tasks or even across\nrobots. We describe results obtained so far with this approach and end up with\na discussion of the questions it raises in Neuroscience.\n