Inertia-1 — An Open Cookbook for Wearable Motion Foundation Models
— by Vivax
Inertia-1 (arXiv:2607.06617, Xu et al.) is a fully open exploration of wearable motion foundation models: 15 datasets and more than 18.2 million hours of…
Every smartwatch and fitness band carries an accelerometer, and that humble motion sensor produces some of the richest — and least standardized — health data in existence. Yet the field of wearable motion modeling has grown up on folklore: sampling rates inherited from older papers, architectures borrowed from vision, training decisions justified by habit. In 'Inertia-1: An Open Exploration of Wearable Motion Foundation Models' (arXiv:2607.06617, published 7 July 2026), Xu and colleagues replace folklore with measurement — the most systematic, fully open study yet of how to turn raw wearable motion signals into a foundation model.
The corpus is enormous and assembled entirely in the open: 15 datasets, more than 18.2 million hours of wearable motion data from over 110,000 participants. On top of it, the authors systematically vary the data choices every wearable team quietly agonizes over — which sensor modality to use, where on the body the device sits, what sampling rate to record at, and how long the input window should be. Instead of defaults inherited from prior work, each choice is measured against downstream performance.
The same discipline is applied to the model and the training run. Architectures and model sizes are compared head to head, pretraining objectives are put through the same benchmark battery, and performance is tracked as the data scale grows — the closest thing wearable motion sensing has to the scaling studies that guided language models. The framing is deliberate: Inertia-1 is an open cookbook for wearable motion representation learning — recipes, ablations and scaling curves anyone can reproduce, not a sealed artifact to be admired.
The recipes are validated where it counts. Inertia-1 models are evaluated on human activity recognition, on detecting freezing of gait — the sudden, dangerous halts that afflict people with Parkinson's disease — and on predicting disease status itself, including diabetes, sleep disorders, Parkinson's and depression, inferred from motion signals alone. That range matters: it shows the learned representations carry genuine physiological signal, not just gross movement categories, and it hands clinical researchers a validated starting point instead of a from-scratch training run.
Inertia-1 lands in the same season as Google's SensorFM, and the two are perfect complements. SensorFM demonstrates the destination: one giant foundation model pre-trained on a trillion minutes of wearable data. Inertia-1 publishes the map: the open recipes, ablations and scaling insights that let the rest of the field build grounded sensor models for themselves. That openness is precisely the conviction behind the Vivax Data Layer and our clinical world models — medical AI advances fastest when models are learned from the data health itself generates, and when the know-how to build them is shared. Read the paper on arXiv, and our companion piece on SensorFM.