Refugio: a voice-driven generative virtual reality “safe place” for personalized emotion regulation – feasibility and usability study

Abstract

Introduction: Virtual Reality (VR) has emerged as a pivotal tool for mental health interventions, paving the way for innovative approaches to emotional regulation. While the customization of therapeutic “safe places” is believed to enhance outcomes, existing applications often depend on inflexible, pre-configured content. This paper introduces Refugio, a groundbreaking VR application designed to bridge this gap through an integrated, real-time personalization pipeline powered by a self-hosted Large Language Model (LLM) and an efficient voice-to-3D generative pipeline. Methods: We conducted a feasibility and usability study with 30 non-clinical participants to evaluate the feasibility and user acceptance of this natural language-driven approach, while preliminarily exploring its potential to support emotional outcomes and immersion while minimizing cognitive workload. We assessed usability (using the System Usability Scale - SUS), cognitive load (NASA-TLX), immersion (RJPQ), and emotional state before and after (Self-Assessment Manikin - SAM). Results: The study provided compelling preliminary evidence of the system’s feasibility and high user acceptance. Refugio achieved an “Excellent” mean SUS score of 86.33, alongside low reported mental demand and frustration. Participants showed strong engagement, creating 147 objects, and reported positive emotional shifts, with the majority experiencing increased valence (21 out of 30) and decreased arousal (19 out of 30). Discussion: This study demonstrates that Refugio’s architecture is a viable and well-received method for implementing deep generative personalization. Our findings suggest that this natural language driven approach adeptly balances creative freedom with high usability and low cognitive load, thereby establishing a robust foundation for future clinical validation.