Google Research has introduced SensorFM, a foundation model designed to read the body through the data that wearables already collect. It was pretrained on more than one trillion minutes of sensor signals gathered from about five million people, using de-identified data collected between September 2024 and September 2025, and the ambition behind it is straightforward to state. Do for physiological sensor streams what foundation models have done for language and images, pretrain a single large model on an enormous pile of unlabeled data, then adapt that one model to many specific tasks without needing a fresh labeled dataset for each.
What the model actually consumes is a compact but rich picture of the body over a day. SensorFM takes in 34 one minute aggregate features derived from five sensor modalities, photoplethysmography, accelerometry, electrodermal activity, skin temperature, and altimetry. Between them those signals capture heart rate and heart rate variability, blood oxygen saturation, sleep stages, motion and steps, skin conductance, and temperature across a full 24 hour window. From that continuous stream the model learns a general purpose representation of human physiology, a numerical summary of what a body is doing that can then be pointed at specific questions.
The scale is the point, on both the data and the results. The pretraining corpus spans more than 100 countries, all 50 US states, and over 20 Fitbit and Pixel Watch device models, adding up to more than two billion hours of minute resolution signal. Google reports that the representation transfers to 35 health prediction tasks and supports label efficient adaptation as well as data infilling, filling in gaps where a sensor dropped out. Its largest variant, trained on the full five million person corpus, cut reconstruction loss by 31 percent over the smallest version and improved downstream results by an average of 9 percent on classification tasks and 21 percent on regression tasks, the pattern you look for when co-scaling model size and data is working.
The application Google points to is a grounding tool for what it calls a Personal Health Agent, an assistant that could reason about your body's signals with a real model of physiology underneath it rather than guessing from raw numbers. That framing is worth taking seriously and also worth being careful about. This is a research release, not a product you can use, and it produces health predictions and representations, not clinical diagnoses, so it is a foundation to build on rather than a doctor in your watch. The reported gains are Google's own, and the whole thing rests on an enormous amount of de-identified personal health data, which carries the privacy weight that any project at this scale does.
What makes it matter is the direction more than any single number. The foundation model recipe, pretrain broadly then adapt narrowly, is escaping the text and image domains where it grew up and moving into the continuous physiological data that hundreds of millions of people already generate every day. If the approach holds up, a single pretrained model becomes a shared base layer that many health applications can sit on top of, and the Personal Health Agent framing shows where Google wants that to lead, an assistant grounded in the language of your own body. It is not a diagnosis engine today, but it is a clear signal that wearables are becoming a serious modality for large models, not just a source of step counts.