Social communication interactions are paramount in human daily-life. In virtual environments, virtual agents allow for more social presence to the users, and can trigger social behaviours on them. However, the complexity of real-life situations, notably with external stimuli perturbations (noise, visual occlusions etc.) is not handled by classical interaction models. We propose here a model that includes an acoustic and visual sensing system, to compute and evaluate agent-dependent perceptions from the agents and their surroundings, before triggering expressive actions and reactions.