Inertia-1 is a wearable-motion foundation model pretrained on over 18.2 million hours of accelerometer data

Inertia-1 is a foundation model for wearable motion sensing, introduced in "Inertia-1: An Open Exploration of Wearable Motion Foundation Models" (arXiv:2607.06617, submitted 7 July 2026, by Zongzhe Xu, Aakarsh Anand, Sarah Jiang, Chuntung Zhuang, Zitao Shuai, Sriram Sankararaman, and Yuzhe Yang). Its defining characteristic is the pretraining scale: the pretraining draws on "massive corpora of accelerometer data from global sources spanning more than 18.2M hours" (abstract). (The capture's further gloss — that this corpus is far larger than prior wearable-sensing pretraining efforts — is plausible but is not a comparison the abstract itself makes.)

The paper positions passive body-motion sensing as an under-exploited substrate for the same self-supervised-pretraining logic that produced large language models — "Wearable motion sensing provides a continuous and scalable window into human behavior and health, making it a natural fit for foundation models, yet its pretraining and scaling principles remain poorly understood." The move is to treat continuous accelerometer streams the way LLM pretraining treats text: a large, cheap, unlabeled signal from which a general-purpose representation can be learned and then transferred to specific downstream tasks (see claim-inertia-1-design-sweep-15-downstream-datasets).

This makes the wearable-pretraining scaling question a training-time counterpart to the inference-time scaling explored for LLMs in claim-test-time-compute-can-substitute-parameters: both ask where marginal compute or data is best spent, but Inertia-1 addresses the pretraining side — how much accelerometer data, at what sampling rate and window length, feeding what model size, yields a transferable representation.