SensorFM
Summary
SensorFM is Google’s large-scale wearable-health foundation model. It learns passive multivariate time-series representations from minute-resolution wearable sensor streams and transfers them to health prediction, missing-data infilling, downstream-head search, and Personal Health Agent grounding.
Role In The Wiki
SensorFM is the strongest current wearable-sensor evidence in this wiki for scaling self-supervised time-series representation learning on a very large real-world corpus. It is important for the foundation TSFM agenda because it combines data/model co-scaling, missingness-aware masked reconstruction, frozen-embedding transfer, and agent-facing prediction outputs.
The boundary is equally important: SensorFM is not an action-conditioned world model. It does not model treatments, behavioral recommendations, control inputs, or intervention consequences as first-class actions.
Official Artifacts
- Preprint: arXiv:2605.22759
- Official blog post: SensorFM: Towards a general intelligence and interface for wearable health data
- Public code: not verified at ingest time.
- Public weights: not verified at ingest time.
- Public pretraining data: not released; the source describes private consented Fitbit / Pixel Watch research data.
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in sensorfm-2026 rather than duplicating verdict rows here.
At the entity level, SensorFM anchors the wearable-sensor branch of passive time-series representation learning: large-scale unlabeled data, missingness-aware reconstruction, downstream health labels, and agent-grounding use. Its open gap is the same as most passive TSFMs: it lacks explicit actions, control inputs, interventions, and counterfactual rollout.