Adaptive agents exhibit tightly coupled interactions between nervous system, body and environment. Parisi recently suggested that the current focus on sensorimotor interaction between agent and environment needs to be complemented by an "internal robotics", i.e. modeling of the interaction between internal physiology and nervous system in, for example, emotional mechanisms. The dynamical systems notion of "collective variables" can help understanding such interactions. In emotions physiological states are key parameters that trace the global dynamic concern relevance of the situation. Such variables may be key, in adaptive systems, to monitoring and controlling the agent's interaction with the external environment. We show in a simple robotic simulation that the neural controller can self-organize to exploit the dynamical regularities traced by these variables. We conclude this can prove to be a useful technique in robots and animals, towards evolving emotion-based adaptive behaviors.