I see emotions as analogous to the value function in RL. It is essentially a prediction of future rewards based on current state and action plan. Artificial RL agents learn emotion as it is related to their tasks and environments.