Google Releases SensorFM: Wearable Health Foundation Model Pretrained on Trillions of Sensor Minutes

SensorFM uses a ViT-1D masked autoencoder architecture pretrained on over 1 trillion minutes of unlabeled sensor signals from 5 million consenting participants. The team tested scaling across four model sizes and data amounts, finding special behaviors when model capacity exceeds data volume. On 35 health tasks, frozen embeddings plus PCA-50 linear probing beat handcrafted feature baselines in 34 tasks. They also designed an agent classroom to auto-search 30,516 prediction heads and validated personal health agent utility via clinician evaluation. For teams building wearable health apps, SensorFM provides a ready-to-use foundation model reducing reliance on complex feature engineering. Its linear probing achieves SOTA, enabling resource-limited teams to build high-accuracy health monitoring systems. The pretrained model adapts easily to different sensor types and populations.