Inertia-1: A Public Exploration Toward a Unified Motion Foundation Model

Inertia-1 is a unified motion foundation model proposed to overcome the limitations of fragmented existing motion models. Unlike existing approaches where data sampling rates, window lengths, and sensor modalities vary widely, it studied data, sensing, objective functions, and scale in an integrated manner within a single controlled space.

Although this model was pretrained on wrist accelerometer data, it can transfer without retraining to other body parts such as the head, chest, and hips, and to other sensor types such as gyroscopes and magnetometers. It exhibits the characteristic that combining multiple sensor streams increases the accuracy of representation and produces clearer activity clusters.

The sensing design guidelines derived from the key research findings are as follows:

• Sampling rate: 1Hz is sufficient for activity recognition, but precise health signals require higher sampling rates
• Window length: 30-60 seconds is optimal, capturing both context and clarity
• Maintaining 3-axis data: Triaxial input consistently outperforms vector magnitude summaries
• Time-domain modeling: This is more advantageous than frequency-domain reconstruction for preserving gait and health indicators

Inertia-1 was built through self-supervision leveraging over 18 million hours of large-scale accelerometer data, and can be utilized across various healthcare applications ranging from fitness tracking to clinical screening.