On January 15, 2026, Google researchers detailed a large validation study of smartwatch-based estimation of advanced walking metrics. The work, led by Amir Farjadian and Ming-Zher Poh, compared consumer wearables with a laboratory-grade gait analysis system using data from volunteer participants at Google in Mountain View and at Kyoto University in Japan.
Key Details
The study validated Pixel Watch-based estimation of spatio-temporal gait metrics against pressure-sensing walkway measurements. Researchers used Pixel Watch 1 devices on both wrists and multiple Pixel 6 smartphones placed at different body locations.
- The cohort included 246 adults aged 18 or older without reported gait-affecting conditions or use of assistive devices.
- Researchers collected approximately 70,000 walking segments across all participants and test conditions.
- Ground-truth gait metrics came from a Zeno Gait Walkway pressure mat system in controlled laboratory settings.
- Walking tasks included a six-minute walk test, fast walking, and induced asymmetry using a hinged knee brace.
- Smartwatch data from both wrists contributed to model training, while smartphone evaluation focused on front and back pocket placements.
Google used a multi-output temporal convolutional network to estimate several gait metrics directly from smartwatch sensor signals. Inputs included raw 3-axis accelerometer and 3-axis gyroscope data at 50 Hz, plus user height.
The model produced bilateral gait speed and double support time, and unilateral step length, swing time, and stance time. Models were trained on five-second windows with one-second overlap and optimized mean absolute percentage error across all outputs.
Researchers applied a five-fold cross-validation strategy, assigning each participant's data to a single fold to avoid leakage between training and evaluation sets.
For most gait metrics, smartwatch estimates achieved Pearson correlation coefficients above 0.80 relative to walkway measurements. The intraclass correlation coefficient for these metrics also exceeded 0.80, while correlations for double support time ranged from 0.56 to 0.60.
Quantitative comparisons showed no statistically significant differences between smartwatch and smartphone performance for any metric, although the smartphone model used roughly twice as many walking segments. Both models significantly outperformed a naive baseline that predicted the training-set mean for each metric.
Background Context
Spatio-temporal gait metrics such as walking speed and step length are known to be vital biomarkers for multiple health conditions. Peer-reviewed research links gait characteristics with overall health status, fall risk, and progression of neurological or musculoskeletal diseases.
Historically, high-precision gait analysis has depended on laboratory systems such as pressure-sensing walkways and motion capture cameras. Smartphone sensors have offered a more portable option but rely on consistent placement, typically in a front or back pocket. Wrist-worn smartwatches provide fixed positioning and allow continuous data collection during walking without needing a nearby phone.
Previous temporal convolutional network approaches for gait estimation often focused on detecting gait events only, requiring separate numerical integration to derive spatial metrics. In contrast, the Google model outputs both spatial and temporal gait parameters directly, avoiding an additional integration step.
An ablation analysis showed that including user height as an input improved smartwatch estimates for gait speed and step length. Google states that future research will refine and expand smartwatch gait metrics based on these findings and explore broader clinical and wellness applications.
Source Citations
Official publications and resources for this work are listed below.
.svg)
.svg)
.svg)