Nowadays virtual humans such as non-player characters in computer games need to have a real autonomy in order to live their own life in persistent virtual worlds. When designing autonomous virtual humans, the action selection problem needs to be considered, as it is responsible for decision making at each moment in time. Action selection architectures for autonomous virtual humans should be individual, motivational, reactive and proactive to obtain a high degree of autonomy. This paper describes in detail our motivational model of action selection for autonomous virtual humans in which overlapping hierarchical classifier systems, working in parallel to generate coherent behavioral plans, are associated with the functionalities of a free flow hierarchy to give reactivity to the hierarchical system. Finally, results of our model in a complex simulated environment, with conflicting motivations, demonstrate that the model is sufficiently robust and flexible for designing motivational autonomous virtual humans in real-time