Advancing Cybersickness Prediction in Immersive Virtual Reality Using Pre-Trained Large Foundation Models

Cybersickness is one of the problems that users often encounter during virtual reality (VR) experiences. Existing machine learning (ML) and deep learning (DL) methods for predicting cybersickness require massive amounts of high-quality data for effective training, extended training times to achieve accurate predictions, and lack of transferability capability to new VR environments. To address this, we propose a novel approach using zero-shot and few-shot learning mechanisms to leverage the knowledge of pre-trained large foundation models, namely a time series generative pre-trained transformer (TimeGPT) and Chronos. Validated on two open-source VR cyber-sickness datasets, namely Simulations 2021 and APAL Head 2019 datasets, our fine-tuned TimeGPT model outperforms traditional DL models, achieving superior accuracy (RMSE of 0.26 and 2.69 for the Simulations 2021 and APAL Head 2019 datasets) and significantly reduced training times (up to 81% and 79% for the same datasets) compared to the trained-from-scratch Transformer models.