Learning from a trillion minutes of sensor data
To build the pre-training set, we sampled de-identified data from five million people who agreed to use their data for health and wellness research, captured between September 2024 and September 2025. The data set spans more than 100 countries, all 50 US states, and more than 20 Fitbit and Pixel Watch models. We extracted several weeks’ worth of data from each person, yielding more than two billion hours – more than a trillion minutes – of fine-grained signals.
The SensorFM accommodates 34 combined 1-minute features derived from five sensing modalities: photoplethysmography (PPG), accelerometry, electrodermal activity (EDA), skin temperature, and altimetry. Together, these capture heart rate and heart rate variability, blood oxygen saturation, sleep stages, movement and steps, skin conductivity, and temperature over a full 24-hour period.
Instead of relying on labels, SensorFM learns through self-supervised reconstruction, based on LSM-2 Approach and Adaptive and Inherited Invisibility (AIM) framework. This is a critical design choice, because missing and fragmented data (e.g., periods of time when data is not available) are the norm with wearable devices, due to a variety of factors such as sensor power cycling, devices coming off the wrist, operating modes to save power, and turning sensors on and off. Traditional self-supervised methods assume complete and uninterrupted inputs and are therefore forced to either account for gaps (which can introduce bias) or ignore incomplete windows (which throw away valuable data). AIM takes no path: it treats real-world invisibility as a natural artifact and learns directly from incomplete recordings, combining tokens I inherited Of the real gaps with those Artificially The goal of reconstruction is hidden and the two are treated equally. The result is a perceptive representation of lack through construction. SensorFM not only tolerates fragmented data, but uses it productively, as the generative results below show.
