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

Cybersickness, a motion sickness like discomfort, is a major barrier to the usability of virtual reality (VR) systems. While prior work has focused mainly on predicting cybersickness severity, practical mitigation requires not only detecting how sick a user feels but also deciding whether a countermeasure is beneficial and determining its appropriate intensity. In this paper, we propose a two phase multitask learning framework that jointly models cybersickness severity, blur effectiveness, and blur intensity. In Phase 1, we pretrain temporal deep learning backbones on two single label datasets with only severity annotations. In Phase 2, we pro-gressively finetune the models on a multi-label dataset containing severity, blur effectiveness, and blur level labels. We evaluate three backbone architectures a Time-Series Transformer, Deep Temporal Convolutional Network, and TS-Mamba under a 10-fold block aware cross validation scheme. Results show that two phase training significantly outperforms single phase baselines, with the Time Series Transformer achieving best performance (FMS MAE = 0.57, R2 = 0.87; Blur Level MAE = 0.49, 2 = 0.95; Blur Preference ACC = 99.5%). Unlike prior rule based reduction frameworks that rely on static heuristics, our approach provides a data-driven "detect-decide-dose" pipeline that adapts blur mitigation dynamically to individual users. This demonstrates that single label pre-training is an effective strategy for developing multitask VR safety models under limited labeled data. To our knowledge, this is the first framework that unifies cybersickness prediction and adaptive reduction in a single model.