The Pulse

Cornell researchers use one smartwatch to estimate full-body motion

Cornell researchers developed EchoMotion, a system that uses ultrasonic echoes and an AI model to estimate body positions from a smartwatch. In a test with new users, a short calibration reduced average joint-position error to about 2.4 inches.

AI.info Team · · fact-checked · against 2 sources ↓ · Published 12 Nov 2026

Cornell researchers have developed a system that estimates a person's full-body position using ultrasonic echoes from a single smartwatch and an artificial intelligence model. In tests with people new to the system, average joint-position error fell to about 2.4 inches after a one-time, 12-minute calibration, according to the Cornell Chronicle.

The system, called EchoMotion, uses a commercial watch's built-in speaker and microphone rather than a camera to sense movement. It remains a research project: the team is developing it for fitness tracking and exploring other uses.

Listening to movement

EchoMotion comes from Cornell's Smart Computer Interfaces for Future Interactions, or SciFi, Lab, directed by Cheng Zhang, an associate professor of information science. The system sends out inaudible ultrasonic signals and uses their reflections to estimate the wearer's body position.

Lead author Saif Mahmud, a doctoral student in information science, said conventional wearable sensors such as accelerometers and gyroscopes measure the motion of the device itself. Whole-body tracking has traditionally required more elaborate arrangements, potentially involving phones, earbuds or virtual reality headsets as well.

Some systems use cameras. Cornell said acoustic sensing avoids the need for a camera and produces less data to process, while preserving the privacy of the wearer and people nearby.

From prototype to commercial watches

The researchers first built a 3D-printed watch replica with a speaker and microphone pair on each side. Its speakers sent sound waves around the wearer, and its microphones picked up echoes reflected by moving limbs.

The researchers collected echo patterns from 10 volunteers wearing the prototype on their nondominant hand while performing 18 exercises. They used the data to develop a deep learning model that translates echo patterns into estimated movements.

To measure accuracy, the team compared its predictions for 12 joint positions, from the head to the ankles, with poses recorded by cameras. The cameras served as the comparison in the study; they were not part of the smartwatch sensing system.

Commercial smartwatches have a single speaker and microphone pair, so the researchers also tested EchoMotion on an unmodified Samsung Galaxy Watch 6 with 15 new volunteers. For people new to the system, the average error in each predicted joint location was under four inches; the one-time, 12-minute calibration reduced it by more than one third, to about 2.4 inches.

The same model also worked on a Samsung Galaxy Watch 7 and a Google Pixel Watch 3 without retraining, the researchers found. Sensing in the prototype drew about 12 milliwatts, roughly 40 times less than a wearable camera, while a 30-minute session on a Galaxy Watch 6 used 5% to 6% of its battery, according to Cornell.

Outside the lab

In a further test, 10 new participants performed the 18 exercises in a small room, a narrow hallway and beside a noisy road. Furniture and ambient noise did not interfere with the system, but performance suffered slightly in the hallway because of echoes from other people, the researchers found.

Mahmud expects a model trained on data from thousands of users, rather than dozens, to interpret movement better in crowded settings. That is a projection, not a result from the reported tests.

Fitness and other uses

The team is developing the technology for a personal fitness agent that could count repetitions and correct exercise form. The researchers are also exploring a combination of EchoMotion and first-person vision on smartglasses to track objects a person handles.

The research received support from the National Science Foundation. The paper was presented at the UbiComp/ISWC conference in Shanghai on October 13.