Cornell Builds 12-Joint Smartwatch Tracking With 6.1cm Error on Commercial Watches
Updated
Updated · T3 · Nov 14
Cornell Builds 12-Joint Smartwatch Tracking With 6.1cm Error on Commercial Watches
3 articles · Updated · T3 · Nov 14
Summary
EchoMotion let Cornell researchers track 12 body joints using only a smartwatch’s built-in speaker and microphone, with average error below 10cm across 15 first-time users.
A 12-minute calibration cut that error to about 6.1cm by using AI to interpret ultrasonic echoes reflected from the wearer’s body rather than camera footage or external sensors.
Samsung’s unmodified Galaxy Watch 6 ran the system, and the same model also worked on the Galaxy Watch 7 and Google Pixel Watch 3 without retraining.
Battery use stayed at 5% to 6% over a 30-minute Galaxy Watch 6 test, while privacy and lower data demands could make the approach useful for gyms, fitness coaching and rehabilitation.
The research project still struggles somewhat in narrow spaces where other people’s echoes interfere, and the team says training on data from thousands of users could improve performance in busier settings.