SensorLM: Learning the Language of Wearable Sensors

Summary

SensorLM introduces large-scale sensor-language pretraining that aligns wearable physiological sensor streams — including PPG, activity, skin temperature, heart rate, SpO2, and sleep — with natural language descriptions of health states. Trained on Google's large-scale wearable cohort (millions of Pixel Watch / Fitbit users), SensorLM enables zero-shot natural language queries over sensor data (e.g., "detect high stress periods" or "identify sleep disturbances"). The model learns cross-modal representations that support zero-shot, few-shot, and fine-tuned prediction across diverse health downstream tasks.