Embodied Conversational Agents with Situation Awareness for Training in Virtual Reality

Abstract

Embodied conversational agents have a great potential in virtual reality training applications. This paper investigates the impact of conversational agents on users in a first responder training scenario. We integrated methods for automatic speech recognition and speech synthesis with natural language processing into a VR training application in the Unity game engine. Additionally, we present a method for enabling situation awareness for agents in a virtual environment. Finally, we conducted a between-subject lab experiment with 24 participants which investigated differences between conversational agents and agents with pre-scripted audio. Several metrics were measured in the experiment including presence, subjective task performance, learning outcome, interaction quality, quality of information presentation, perceived realism, co-presence, and training task duration. Our results suggest that users trying our conversational agents condition experienced significantly higher level of copresence than users with pre-scripted audio. Additionally, significant differences in subjective task performance and training duration were discovered between genders. Based on the results of our qualitative analysis, we provide guidelines that can facilitate future design of VR training applications and research studies with embodied conversational agents.