Inertia-1

A wearable motion foundation model — arXiv:2607.06617, "Inertia-1: An Open Exploration of Wearable Motion Foundation Models" (Xu, Anand, Jiang, Zhuang, Shuai, Sankararaman, Yang; submitted 7 July 2026) — pretrained on over 18.2M hours of accelerometer data and evaluated across 15 downstream datasets spanning human activity recognition, freezing-of-gait detection, and disease prediction. It anchors the vault's wearable/health-foundation-model cluster: the paper's own systematic design sweep (data choices, model choices, training choices) is now the vault's most-documented single source on what actually moves the needle for wearable-sensor pretraining.

References

• claim-inertia-1-pretrained-18-million-hours-accelerometer · claim-inertia-1-design-sweep-15-downstream-datasets · claim-inertia-1-data-scale-beats-model-size · claim-inertia-1-triaxial-beats-vector-magnitude-input · claim-inertia-1-no-pretraining-objective-dominates · claim-inertia-1-self-supervised-beats-supervised-gap-widens · claim-inertia-1-downsampling-robustness-is-task-dependent · claim-inertia-1-no-single-window-length-is-optimal
• Related hubs: entity-yuzhe-yang · entity-freezing-of-gait-detection