Predicting Cybersickness Trend and Extent Based on FMS Labeled Dataset.

Cybersickness has been a hindrance to the widespread adoption of virtual reality. As cybersickness is dynamic, with its manifestation changing constantly, predicting and mitigating it in real time is key. However, previous research has utilized predictive models trained mostly with datasets whose sickness levels were measured and labeled only after experiencing the content. In addition, many datasets rely on physiological signals as input, which makes them difficult to apply in actual VR usage. This makes such timely predictions unreliable or practically infeasible. We have created a publicly available dataset where the ground truth sickness levels were densely marked every 0.5 seconds and adjusted/updated on-demand using the FMS, and purposely excluded the difficult-to-collect physiological data. We demonstrate that predictive models trained with such a dataset, comprising just the content motion profile and FMS data, can still produce comparably reliable sickness prediction, and more so, when user-specific parameters (e.g., age, gender) are added. The created dataset is available at https://github.com/pvtryan1024/2025_Cybersickness_dataset.