New possibilities for research and training in sport could be offered thanks to virtual reality simulations. The literature is examined, analysing the advantages and properties of current simulators. This paper aims to present an original experiment in virtual reality simulation design based on autonomy, interaction and evolving concepts. Our main suggestion is that believability of simulation for the user needs both regularity and surprises (in terms of virtual agents' behaviour). A naturalistic approach was used to study experts' decision-making processes in soccer settings. The decision-making model represented a guideline for choosing and implementing computational models. The advantages of virtual reality simulation are discussed in relation to training and research in sports science.