Inertia-1: Open-source motion AI trained on 18.2M hours of wearable data

Inertia-1 is a fully open exploration of wearable motion foundation models, built by a team of researchers including Zongzhe Xu, Aakarsh Anand, Sarah Jiang, and others. Using massive corpora of accelerometer data from global sources spanning over 18.2 million hours, the project establishes a controlled framework covering the entire lifecycle: data choices (sensor modality, device placement, sampling rate, window length), model choices (architectures, model size), and training choices (pretraining objective, data scale). This systematic approach addresses the fragmentation in prior work, which often isolated design choices under fixed settings and narrow tasks.

Extensive evaluations across 15 datasets—including human activity recognition, freezing-of-gait detection, and disease prediction—uncover intriguing findings for building motion foundation models that generalize across tasks and sensing conditions. Inertia-1 not only presents state-of-the-art recipes for diverse downstream tasks but also serves as a comprehensive, practical, and open cookbook for wearable motion representation learning. The work is a significant step toward a unified foundation model for human behavior and health monitoring from wearable sensors.