Towards Intelligent Virtual Exposure Therapy: A Pilot Study on an Imitation-Learning-Based AI Therapist within the BRAVE Project

This thesis explores the development of an AI-driven therapeutic agent within the BRAVE project, an early-stage initiative aimed at advancing Virtual Exposure Therapy (VRET) for Specific Phobia. The project recreates phobic scenarios through Unity-based virtual reality environments while employing Multimotion, a machine learning system that processes biosignals collected during sessions to predict patient emotional states in real-time. This emotional monitoring supports therapists in their clinical decision-making, with the ultimate goal of training an AI agent capable of learning therapeutic strategies through observation of expert clinician behavior during VRET sessions. In my thesis I conducted a pilot feasibility study for developing an imitation learning-based AI agent capable of acting as a virtual therapist during exposure therapy sessions for patients affected by Aviophobia. Given the absence of real clinical data at this early project stage, I adopted a different approach: a synthetic therapy session dataset was generated using Large Language Models (LLMs). This generation process was conducted in close collaboration with an expert psychotherapist specialized in exposure therapy, who contributed both to prompt engineering and validation of the generated outputs to ensure clinical plausibility and therapeutic adherence. The synthetic dataset comprises detailed transcripts of patient-therapist interactions, corresponding emotional states represented as valence-arousal coordinates, and Unity simulation parameters that control environmental stressors in the virtual environment, such as turbulence intensity, rainfall, lighting conditions, and ambient audio cues. Rather than training the agent to directly predict individual Unity simulation parameters, I abstracted the complex parameter space into discrete severity levels representing the overall intensity of the exposure scenario. A neural network was trained to predict two key therapeutic decisions: when to modify the severity of the presented scenario (i.e., when to increase or decrease exposure intensity) and which severity level to transition to. This approach significantly reduces the dimensionality of the action space by mapping the multitude of possible Unity simulation states to a manageable set of hierarchical severity configurations. The model learns to mimic therapeutic decision-making by observing patterns in how graduated exposure should progress in response to patient emotional trajectories. The network architecture processes sequential data including valence-arousal trajectories and previous severity levels to generate contextually appropriate interventions that balance therapeutic challenge with patient tolerance, essentially capturing the clinical judgment of when a patient is ready to advance, maintain, or retreat in exposure intensity. This pilot study demonstrates the technical and methodological feasibility of developing AI therapeutic agents through imitation learning from synthetic data. The framework establishes a reproducible pipeline encompassing synthetic data generation with clinical oversight, feature engineering for emotional and environmental state representation, and supervised learning for therapeutic decision prediction. While current results are preliminary and limited by the synthetic nature of training data, they provide crucial validation of the conceptual approach and system architecture.