SensorFM: A Foundation Model for Wearables Trained on a Trillion Minutes of Data

Google Research has presented SensorFM, a Large Sensor Foundation Model pre-trained on over one trillion minutes of multimodal sensor signals from five million consented participants. The model learns via self-supervised reconstruction using the LSM-2 approach and Adaptive and Inherited Masking (AIM) framework, treating real-world missing data as a natural artifact. Scaling experiments show that increasing both data and model size yields near-linear improvements in pre-training loss and downstream performance, with the largest model (SensorFM-B) winning on 33 of 35 tasks. Evaluated across 35 health tasks from three independent studies, SensorFM's frozen embeddings outperform supervised baselines, and an agentic "classroom" of LLM agents automatically designs prediction heads that beat linear probes on most tasks. When integrated into a Personal Health Agent, SensorFM's predictions significantly improve health summary quality over baseline, with no statistical difference from using ground-truth measurements.