Google Research has introduced SensorLM, a foundation model that pairs smartwatch signals with natural language to describe user activity and health. Announced on 28 July 2025 in Mountain View, the model was built from 59.7 million hours of Fitbit and Pixel Watch data.
What is SensorLM?
SensorLM belongs to a new family of sensor-language models trained to connect wearable data with text. Google used de-identified records from 103,643 volunteers in 127 countries collected between March and May 2024. The dataset spans heart-rate, motion and other physiological signals, totalling 2.5 million person-days.
The training approach combines contrastive learning to align sensor traces with text and a generative arm that produces captions automatically. The resulting model performs zero-shot and few-shot human activity recognition, cross-modal retrieval and caption generation.
- Model family: SensorLM (sizes to be disclosed)
- Training data: 59.7 million hours from Fitbit and Pixel Watch
- Key tasks: zero-shot classification of 20 activities, few-shot learning, cross-modal retrieval, captioning
Why it matters
Previous wearable studies relied on small, hand-labeled datasets. By generating captions automatically, Google scaled the corpus to tens of millions of hours, enabling the type of foundation models that have transformed vision and language research. Early tests show SensorLM outperforms earlier multimodal baselines and follows known scaling laws, improving accuracy as more data and parameters are added.
Source material
The company detailed the work in a Google Research blog post and in the accompanying paper released on arXiv.
SensorLM’s release also highlights the growing importance of consumer devices like Pixel Watch, which already log everyday fitness and sleep metrics for millions of users.
.svg)
.svg)
.svg)