From Simulation to Dialogue: A Quasi-Experimental Mixed-Methods Study of Conversational-AI Supported VR Fire Escape Training

Virtual reality fire escape drills can support procedural learning. However, many implementations rely on scripted interaction that limits learner-initiated inquiry during evacuation. In this study, a Conversational AI Supported VR Fire Escape System ( CAVR) was developed as a Unity based virtual reality drill integrated with the Live Character API through Web Real-Time Communication (WebRTC) for low-latency speech interaction and with retrieval augmented generation grounded in a curated fire safety knowledge base. A quasi-experimental mixed methods design was applied with 163 first year undergraduates from two universities. The experimental group used the CAVR System and the control group used a virtual reality fire escape system without conversational support. Results indicated higher post test knowledge achievement in the experimental group after adjustment for pre-test scores with a significant group effect. Higher behavioral intentions were observed across dimensions with small to medium effect sizes. Trust levels were substantially higher in the experimental group with a large effect size. Interview themes indicated that learner-initiated inquiry during embodied action continuous dialog verification through persistent subtitles and perceived grounding of guidance through retrieval augmented generation contributed to the observed outcome pattern. The implications of using a CAVR System for fire safety education are discussed.