Wearable devices collect vast amounts of personal health data, but understanding the context behind this data has been a challenge. This gap hinders the full potential of personalized health insights. Manually annotating sensor data with descriptive text is impractical due to cost and time. To address this, SensorLM, a family of sensor-language foundation models, has been developed. SensorLM is pre-trained on an unprecedented 59.7 million hours of multimodal sensor data from over 103,000 individuals. This allows it to interpret and generate human-readable descriptions from wearable sensor data. A novel hierarchical pipeline automatically generates descriptive captions, creating the largest sensor-language dataset to date. SensorLM offers capabilities like zero-shot sensor understanding, sensor-text alignment, and sensor caption generation. It demonstrates state-of-the-art performance in tasks like activity recognition and excels at generating coherent and factually correct captions. The model's performance consistently improves with more data, larger model sizes, and increased computation. SensorLM represents a significant advance in making personal health data understandable and actionable, paving the way for future digital health coaches and wellness applications.