This paper presents a rules-based model in which virtual agents have di.erent individualities and can acquire social abilities as a function of rewards. Moreover, the agents can group with others they had interacted in past time, and these formed groups are characterized based on groups cohesion (generated as a function of its members). One of the goals of this model aims to provide different groups animation (density, formation and motion) based on groups characteristics, consequently, the realism of the rules used in the model can impact the visual result. We propose an integration with a physics model in order to generate the required information to be used in these rules. Finally, some results and discussions about applying this methodology in a combat game are presented.