Inertia-1 and wearable motion foundation models
Everything the vault has read from the Inertia-1 paper (arXiv:2607.06617) — the paper that treats continuous wrist/body accelerometer streams the way LLM pretraining treats text, and the vault's current anchor for wearable-sensing foundation models.
The model
- claim-inertia-1-pretrained-18-million-hours-accelerometer — the pretraining scale (18.2M hours)
- claim-inertia-1-design-sweep-15-downstream-datasets — the sweep's scope: three families of design choices, evaluated across 15 datasets and three task families
What the sweep found (2026-07-18 promotion)
- claim-inertia-1-data-scale-beats-model-size — data scale beats model capacity as a transfer lever
- claim-inertia-1-triaxial-beats-vector-magnitude-input — raw triaxial input beats vector-magnitude compression
- claim-inertia-1-no-pretraining-objective-dominates — objective choice barely matters for easy tasks, more for hard ones
- claim-inertia-1-self-supervised-beats-supervised-gap-widens — SSL vs. supervised is a bigger lever than which SSL objective
- claim-inertia-1-downsampling-robustness-is-task-dependent — pretraining buys sampling-rate robustness for activity recognition, not disease prediction
- claim-inertia-1-no-single-window-length-is-optimal — window length has no universal optimum either
Adjacent wearable-sensing work
- claim-sc2-tournament-wrist-tremor-declined-over-the-day — same sensing modality (wrist accelerometry), a much smaller-scale study
- claim-bed-rest-under-runs-spaceflight-bone-loss-depends-on-readout — adjacent wearable/physiological-monitoring thread: a different signal, but the same "the design choice determines the readout" shape