Towards Calibration-Free Affective Modeling from Physiological and Behavioral Signals in Immersive VR Interaction

Understanding users' affective states is important for enabling adaptive Virtual/Extended Reality (VR/XR) interaction systems. However, reliable cross-subject affect prediction remains challenging due to substantial individual variability in physiological responses and behavioral patterns during interaction. We propose AMDA, an Asymmetric Modality Disentanglement Architecture for continuous Valence-Arousal-Dominance prediction from multimodal physiological and behavioral signals in VR teleoperation. AMDA models the two modalities asymmetrically: physiological signals are processed via personality-aware residual disentanglement to separate stable traits from affective dynamics, while behavioral signals are modeled through interaction-driven personality conditioning to capture task-dependent emotional expressions. Experiments on a multimodal VR teleoperation dataset collected under diverse network degradation conditions show that AMDA improves cross-subject continuous affect prediction, particularly for Valence and Dominance. The results further suggest a calibration-free direction for affect-aware VR/XR systems through structured multimodal representation learning.