We present a novel way to design a control system for a robot, using emotions as a way to produce richer internal states. We believe that using a single scalar as an evaluation of the quality of the policy and stating that the goal of the agent is to gather reward, as it is proposed by reinforcement learning, is not an appropriate granularity for creating an autonomous control system : even with a fine-tuned reward function, efficient on a specific task, it is often impractical to derive any useful knoweldge from it in order to build more flexible, neuromimetics control systems. A complete shift of paradigm is necessary for a bottom up approach\,: the robot is given pain and pleasure perception circuits and we examine how emotions arise and are the basis for respondant and operant conditioning. Inspired from the cerebral circuits of superior mammals responsible of this behavior, we propose an implementation in an autonomous robot using models of adaptive neural networks.