Generating Social Contexts with Virtual Agents to Foster Interruptibility Research for Socially Interactive Agents

The social context of a situation is crucial for people to decide how to behave and interact. For artificial agents to be truly socially interactive and unobtrusive it is equally important to understand the social context of a situation. Training these agents to reliably recognize social context is a difficult task, as big data sets of training data are required and a wide variety of different social context cues exist. In this contribution, we describe the creation of virtually simulated situations using virtual agents that demonstrate different social contexts. With this approach, we provide a fast and cost-effective way to create diverse and adaptable training data, by being able to generate video material of the virtual scenes. A suitable context detection system can then be used to support the classification of social context and decide upon, for example, whether a user is available for an interaction, or not. Further, these virtual social context situations can be also used for virtual reality studies to investigate how users react to interruptions in social contexts.