Inertia-1: Open Motion Models for Wearable Devices

Researchers published an Inertia-1 preprint on arXiv in July 2026 — a fully open foundation model of motion for wearable devices, trained on data from more than 18.2 million hours of accelerometer recordings from global sources.

Why such a model is needed

Wearable motion sensors provide a continuous and scalable window into human behavior and health, making them a natural candidate for foundation models. The authors of the preprint note that the principles of pretraining and scaling for such sensors have been poorly studied to date. According to them, previous work examined isolated solutions — for example, sensor placement or sampling frequency — typically with fixed settings and on narrow applied tasks that do not reflect the real diversity of sensor conditions.

What exactly the authors studied

Inertia-1 is a controlled framework for studying the complete lifecycle of motion foundation models: from data selection to architecture and training strategy. The work covers three groups of decisions and relies on a large-scale corpus of data collected from global sources.

• Training corpus — more than 18.2 million hours of accelerometer data from global sources
• Data parameters to study — sensor modality, device placement on body, sampling frequency, temporal window length
• Model parameters to study — architecture and model size
• Training parameters to study — pretraining objective and scale of training data
• Testing — 15 datasets from different application domains

This scope, by the authors' design, allows them to isolate the effects of each individual decision from overall model quality — that is, to understand which data, architectures, and training strategies truly determine the quality of a motion foundation model, and which only seem important in narrow experiments. Previously, according to the authors, such decisions were studied in isolation and almost always with fixed settings, which did not allow direct comparison of their contribution to the final model quality.

What tasks the researchers evaluated

The authors evaluated Inertia-1 on 15 datasets covering three classes of applied tasks: human activity recognition, freezing-of-gait detection — a symptom of Parkinson's disease, and disease prediction from movement data. According to them, extensive experiments on these 15 datasets revealed unexpected patterns important for building motion models that generalize across different tasks and sensor conditions, rather than working only in the narrow scenario for which they were initially trained.

"Inertia-1 not only offers advanced recipes for various applied tasks,

but also serves as a comprehensive, practical, and open cookbook for training motion representations based on wearable sensors," states the annotation of the preprint on arXiv.

What this means

The open publication of the complete cycle of experiments — data, architectures, and training objectives — gives wearable device and medical application developers a proven set of recipes for building their own motion models instead of fragmented and poorly reproducible solutions that were typical for this field previously. This is particularly important for health-related tasks — such as identifying symptoms of Parkinson's disease from accelerometer data, where model quality directly affects diagnostic accuracy.