Inertia-1 represents a significant stride toward a unified motion foundation model, addressing the fragmentation within motion-based AI. The field has historically suffered from diverse datasets and task-specific models, hindering interoperability and generalizability.

Inertia-1 tackles this by exploring the entire motion model lifecycle—data, sensing, objectives, and scale—within a single, controlled framework. The core innovation is a general representation that adapts across different body placements, devices, and tasks, functioning effectively even in settings beyond its initial training.

The study reveals that a model pretrained on wrist data can be applied to other body parts and sensor types like gyroscopes and magnetometers without retraining. Furthermore, incorporating additional data streams, such as from multiple body placements or different sensors, enhances the representation's accuracy and clarity. Key findings also highlight the importance of sensing design, with practical guidelines provided for sampling rates (1 Hz for activity recognition, higher for health signals), window lengths (30–60 seconds), input format (triaxial vs. vector magnitude), and modeling domain (time-domain for gait and health). The Inertia-1 pipeline involves pretraining at scale using over 18 million hours of self-supervised accelerometry data, followed by light tuning for transfer across diverse settings and deployment across various applications like fitness tracking and clinical screening.

This initiative is an open invitation for collaboration, aiming to build a comprehensive motion understanding model.