Two Phase Multi-Task Learning for Cybersickness Prediction and Adaptive Reduction.

Area of Science:

• Virtual Reality
• Human-Computer Interaction
• Computational Neuroscience

Background:

• Cybersickness, a significant usability barrier in virtual reality (VR), necessitates advanced mitigation strategies beyond simple severity prediction.
• Current VR systems lack dynamic, individualized approaches to reduce user discomfort.
• Effective mitigation requires predicting sickness severity, determining countermeasure benefit, and optimizing intensity.

Purpose of the Study:

• To develop a novel two-phase multitask learning framework for unified cybersickness prediction and adaptive mitigation.
• To jointly model cybersickness severity, blur effectiveness, and blur intensity.
• To create a data-driven pipeline for dynamic, personalized VR safety interventions.

Main Methods:

• A two-phase multitask learning approach was employed, pretraining on single-label severity data and finetuning on multi-label data (severity, blur effectiveness, intensity).
• Evaluated Time-Series Transformer, Deep Temporal Convolutional Network, and TS-Mamba architectures using 10-fold block-aware cross-validation.
• The framework integrates detection, decision-making, and dosage of blur countermeasures.

Main Results:

• The two-phase training significantly outperformed single-phase methods.
• The Time-Series Transformer achieved superior performance: FMS MAE=0.57, R²=0.87; Blur Level MAE=0.49, R²=0.95; Blur Preference ACC=99.5%.
• Demonstrated the effectiveness of single-label pre-training for multitask VR safety models with limited data.

Conclusions:

• The proposed framework offers a data-driven, adaptive "detect-decide-dose" pipeline for personalized cybersickness mitigation.
• This is the first model to unify cybersickness prediction and adaptive reduction.
• Single-label pre-training is a viable strategy for developing robust multitask VR safety systems.