How EchoMotion Tracks Your Whole Body with Ultrasound from a Smartwatch

Cornell researchers built EchoMotion, a system that tracks full-body pose from a smartwatch by bouncing inaudible ultrasound off the wearer's limbs. On an unmodified Samsung Galaxy Watch 6 it predicted 12 joints with average error under 4 inches, falling to about 2.4 inches after a 12-minute calibration. This guide explains how active acoustic sensing works, why ultrasound beats cameras and IMUs for this job, and how students build an honest lab-scale version that classifies a defined set of exercises instead of reconstructing arbitrary pose.

Written by Projectech17 min readPublished
For Electronics, IoT and embedded-systems students who want to understand how a smartwatch tracks full-body pose with ultrasound and how to build a student-scale classifier. Topics: Ultrasound sensing, active acoustic sensing, TinyML, ESP32, MEMS microphones, ultrasonic transducers, time-of-flight, exercise classification
Technical illustration of a smartwatch on a wrist emitting concentric ultrasonic waves toward a person in a squat, with echo arcs returning from the arms and legs.
Illustration generated for this guide.
In this guide

How EchoMotion Tracks Your Whole Body with Ultrasound from a Smartwatch

A smartwatch already knows your heart rate and your step count. What it has never known is what your body is actually doing. Cornell University's EchoMotion, presented at UbiComp/ISWC in Shanghai on 13 October 2026, changes that: it turns a smartwatch into a full-body motion tracker by bouncing inaudible ultrasound off the wearer's arms and legs and reading the echoes. On an unmodified Samsung Galaxy Watch 6, tested with 15 new volunteers, it predicted the positions of 12 joints with an average error under 4 inches (10 cm) — and a one-time 12-minute calibration cut that error by more than a third, to about 2.4 inches (6.1 cm). The system was built by the Cornell SciFi Lab, directed by Cheng Zhang, with Saif Mahmud as lead author and François Guimbretière as co-author, and published in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (DOI 10.1145/3832001). This guide explains the engineering underneath: how active acoustic sensing works, why ultrasound instead of a camera or motion sensors, what the power and privacy trade-offs really are, and why a student build classifies a defined set of exercises rather than reconstructing arbitrary pose. The student-scale version of this idea is the EchoMotion-Inspired Ultrasonic Smartwatch for Exercise Tracking and Form Feedback using TinyML.

Why a watch should know your pose

Fitness tracking today is mostly counting. A watch counts steps, counts heartbeats, and guesses calories from those counts. What it cannot do is tell you that your knee is collapsing inward during a Bulgarian split squat, or that your back is rounding on a deadlift, or that your left arm is doing more work than your right during a shoulder press. Those are form errors, and form errors are what cause injuries and stall progress — yet catching them has always needed either a camera watching you or a set of motion sensors strapped to your body.

Cameras work, but they are the wrong tool for a wrist. A camera on a watch points at whatever the wrist points at, which is rarely the whole body, and it records video of everything around you. Motion sensors — the accelerometers and gyroscopes already in every watch — can track the wrist itself very well, but the wrist is only one point on a body with many joints. Inferring where your elbow, knee and ankle are from wrist motion alone is guesswork with weak evidence.

EchoMotion's insight is that the wrist is not just a place to put a sensor; it is a good vantage point for sound. Sound travels out from the wrist in all directions, bounces off the arms, torso and legs, and comes back carrying information about where those body parts are. The watch does not need to see your body. It needs to listen to it.

How active acoustic sensing works

EchoMotion is an active acoustic sensing system. "Active" means it sends out its own signal and listens for the return, the way a bat echolocates or a sonar maps the seabed, rather than passively listening to sounds the world makes. The signal is ultrasound — sound above the range of human hearing — so the wearer hears nothing while the watch is pinging away.

The hardware in the Cornell prototype is deliberately simple. The researchers built a 3D-printed watch replica with a microphone-and-speaker pair on each side: two speakers emit inaudible ultrasound, and two microphones pick up the echoes from the wearer's limbs. The production version ran on an unmodified Samsung Galaxy Watch 6, using the watch's own speaker and microphone, which is the detail that makes the idea credible as a real product rather than a lab curiosity. The same trained model also worked on a Galaxy Watch 7 and a Pixel Watch 3 without retraining, which suggests the echo patterns the system relies on are properties of human bodies and movement, not quirks of one watch model.

The sensing loop has four stages: emit, reflect, measure, infer.

Emit. The speakers send out short bursts of ultrasound. Because the bursts are inaudible, they can run continuously during a workout without annoying the wearer or anyone nearby.

Reflect. The sound waves travel outward and bounce off the nearest surfaces — the forearm, the upper arm, the torso, the thigh, the shin. Each body part is at a different distance from the wrist, so each one returns its echo at a different time.

Measure. The microphones record the returning echoes. The key measurement is time of flight: how long each echo took to travel out and back. Sound moves through air at roughly 343 metres per second, so the distance to a reflecting surface is the speed of sound multiplied by the round-trip time, divided by two. An echo that returns after 3 milliseconds came from something about half a metre away. The system also measures how strong each echo is and how its pattern shifts over time, because a moving limb changes the echo continuously.

Infer. The raw echo pattern is not a skeleton. It is a messy, overlapping set of reflections from several moving body parts at once, and turning it into joint positions is a pattern-recognition problem. EchoMotion solves it with a small machine-learning model trained on real data: 10 volunteers performing 18 exercises, with a camera system recording the true joint positions as ground truth. The model learns the mapping from echo pattern to the positions of 12 joints. Once trained, it takes live echoes as input and outputs a pose estimate in real time.

That training setup is worth understanding, because it defines what the system can and cannot do. The model was trained on 18 specific exercises, and it predicts 12 specific joints. It is not a general-purpose body scanner. It is a learned mapping from echo patterns to poses, and it works best on movements similar to the ones it was trained on.

Why ultrasound instead of a camera

The comparison that matters most is against a wearable camera, because a camera is the obvious way to see a body. EchoMotion's sensing uses about 12 milliwatts of power — roughly 40 times less than a wearable camera, according to the Cornell team. That is not a small difference; it is the difference between a feature that runs for a whole workout and a feature that kills the battery. In the team's tests, a 30-minute session used 5–6% of the Galaxy Watch 6 battery, which is in the same range as an ordinary workout tracking session. A camera-based tracker at 40 times the sensing power would not survive the same workout.

Power is only half the argument. The other half is privacy. A camera records images — of the wearer, of the room, of anyone else who walks into frame. An acoustic system records echoes, which are numbers describing distances, not pictures of people. Echoes can still leak information in principle, but they are far less revealing than video, and they are much easier to process entirely on the device. For a fitness feature that runs in a gym full of other people, that difference matters. The Cornell team also found a real-world limit on this front: hallway echoes bouncing off other people nearby degraded performance slightly, a reminder that sound does not respect personal space the way a camera's field of view does.

Ultrasound also has a practical advantage over the other obvious alternative, which is strapping inertial sensors to every limb. A full IMU suit tracks pose well, but nobody wears one to the gym. EchoMotion needs no extra hardware on the body at all — one watch, worn normally, is the whole sensor suite. That is the design lesson students should take from it: the best sensing system is often the one that asks the user to change nothing.

Why the student build classifies exercises instead of reconstructing pose

Here is the honest part, and it is the part that makes the difference between a good student project and a doomed one. EchoMotion reconstructs continuous pose — the 3D positions of 12 joints, frame by frame — because its researchers had a camera ground-truth system, 10 volunteers, and the compute to train a regression model on thousands of labelled examples. A student team does not have that. What a student team does have is a microcontroller, a small ultrasonic transducer, a MEMS microphone, and a few weeks.

So the student build reframes the problem. Instead of asking "where exactly is every joint?", it asks "which exercise is being performed, and is the form right?" That is a classification problem, not a regression problem, and classification is dramatically easier to do well at student scale. The build is the EchoMotion-Inspired Ultrasonic Smartwatch for Exercise Tracking and Form Feedback using TinyML: a wrist-worn device that emits ultrasound, records the echoes, and runs a tiny machine-learning model on the microcontroller to recognise a defined set of exercises and flag form errors.

Classification works because different exercises produce different echo signatures. A squat moves the thigh and shin in a large, slow arc; a bicep curl moves the forearm in a short, fast one; a jumping jack moves both arms and legs symmetrically. The echo pattern for each is distinct enough that a small model can tell them apart, and the model only needs to know the handful of exercises it was trained on — the same 18-exercise logic as the Cornell work, scaled down to a set a student team can actually collect data for. Form feedback follows the same logic: a Bulgarian split squat performed with the knee collapsing inward produces a different echo pattern from one performed cleanly, because the leg is in a different place at the bottom of the movement. The model does not need to know the knee's exact coordinates; it needs to recognise the difference between the correct pattern and the wrong one.

The second reason classification is the right student target is data. Training a pose-regression model needs synchronised camera ground truth for every training sample — a motion-capture setup a college lab rarely has. Training a classifier needs labelled examples: perform the exercise correctly, record the echoes, label the file "squat-correct"; perform it with the common error, record, label "squat-knee-in". A phone camera and a careful teammate are enough ground truth for that, because the label is the exercise name, not a set of joint coordinates. A student team can collect a usable dataset in an afternoon. That is what makes the project buildable rather than aspirational.

The third reason is compute. A pose-regression model that outputs 12 joint positions in real time needs more memory and processing than a small microcontroller comfortably provides. A classifier that outputs one of a handful of exercise labels, running on a TinyML model, fits on an ESP32-class board with room to spare. The student build's honesty is in matching the ambition to the hardware, not in shrinking the Cornell paper until it fits.

What the numbers actually mean

The headline figure — average joint error under 4 inches — deserves a careful read, because students quoting it in a report should know what it does and does not promise. An average error under 10 cm across 12 joints and 15 new volunteers means the system usually knows roughly where each joint is, not that it knows exactly. For rep counting and form correction, roughly is enough: the system does not need millimetre precision to tell a good squat from a bad one. For anything that needed exact joint angles — clinical gait analysis, for example — it would not be enough, and the Cornell team does not claim otherwise.

The calibration result is the more interesting number. A one-time 12-minute calibration session cut the average error by more than a third, to about 2.4 inches (6.1 cm). That tells you the system has a large person-specific component: bodies differ in size and shape, and the echo patterns differ with them. A short personal calibration lets the model adjust to the individual wearer, and the improvement is large enough that any student build should include a calibration step rather than treating it as optional. It also tells you something about the failure mode: without calibration, the model is averaging across body types, and the error is the cost of that averaging.

The cross-device result — the same model working on the Galaxy Watch 7 and Pixel Watch 3 without retraining — is the number that matters for anyone thinking about a product. It says the learned mapping is not overfitted to one watch's speaker and microphone characteristics. For a student build, the equivalent lesson is to test on more than one wrist: collect training data from several teammates, not just the one who built the device, or the demo will work for its builder and fail for the examiner.

The power and privacy trade-offs, honestly

EchoMotion's power story is strong but specific. Sensing at about 12 mW, roughly 40 times less than a wearable camera, is a genuine advantage, and 5–6% battery for a 30-minute session is a number a student can quote with confidence because it comes from the paper's own tests on a real watch. But two caveats belong in any honest write-up. First, the 12 mW figure is the sensing cost, not the whole system cost — the processor running the model, the display, and the radio all draw power too, and the 5–6% session figure is the total that includes them. Second, the student build will not match the Cornell numbers exactly. A microcontroller board with a discrete ultrasonic transducer and a MEMS microphone has a different power profile from a Galaxy Watch, and the honest approach is to measure the student build's own current draw and report it, not to copy the paper's figures onto different hardware.

The privacy story is strong in the same qualified way. Echoes are less revealing than video, and processing on the device means no raw data has to leave the watch. But "less revealing" is not "revealing nothing": an echo stream is a record of body movement, and a system that logs it is logging biometric data. A student build should keep its echo data on the device, delete training recordings after the model is trained, and say so in the report. The hallway-echo finding — that echoes off other people nearby degraded performance slightly — is also a privacy-adjacent warning: the sensor hears the room, not just the wearer, so a demo in a crowded hall will perform worse than a demo in an empty room, and the report should record which one it was.

Building the student version: what to get right

The student build follows the same four stages as EchoMotion — emit, reflect, measure, infer — with student-scale parts at each stage. The wrist unit is an ESP32-class microcontroller with an ultrasonic transducer and a MEMS microphone, worn on a strap. The transducer emits short ultrasonic bursts; the microphone records the returning echoes; the microcontroller extracts time-of-flight and echo-strength features; and a TinyML classifier, trained on the team's own labelled recordings, identifies the exercise and flags the common form error for that exercise.

Three build decisions matter more than the rest. First, collect training data from more than one person, because the Cornell calibration result proves the model is person-sensitive and a single-person dataset will not generalise. Second, choose a small, well-defined exercise set — five or six movements with clearly different echo signatures — rather than trying to cover everything, because each additional exercise makes the classification harder and the dataset thinner. Third, design the demo around what the system actually does: show the exercise being recognised, show the rep count incrementing, and show the form warning firing on the deliberate error. A demo that shows those three things working is a better project than one that claims full-body pose reconstruction it cannot deliver. The full parts list, firmware behaviour, setup guide and review kit are in the EchoMotion-Inspired Ultrasonic Smartwatch for Exercise Tracking and Form Feedback using TinyML.

Frequently asked questions

How does EchoMotion track your body without a camera?
It uses active acoustic sensing. Two speakers in the watch emit inaudible ultrasound, the sound bounces off the wearer's arms, torso and legs, and two microphones record the returning echoes. Because sound travels at a known speed, the time each echo takes to return reveals how far away the reflecting body part is. A small machine-learning model, trained on volunteers performing exercises with a camera recording the true joint positions, turns the pattern of echoes into an estimate of 12 joint positions.

How accurate is it?
On an unmodified Samsung Galaxy Watch 6 with 15 new volunteers, the average joint error was under 4 inches (10 cm). A one-time 12-minute calibration cut that by more than a third, to about 2.4 inches (6.1 cm). That is accurate enough for rep counting and form correction, which is what the system is designed for — not for clinical measurement, where exact joint angles would be needed.

Why ultrasound instead of a camera or motion sensors?
Power and privacy. EchoMotion's sensing uses about 12 mW, roughly 40 times less than a wearable camera, and a 30-minute session used 5–6% of the Galaxy Watch 6 battery. Unlike a camera, it records echo distances rather than images of the wearer and the room, and it needs no extra sensors strapped to the body — one watch, worn normally, is the whole sensor suite. Motion sensors in the watch can track the wrist well but cannot see where the elbows, knees and ankles are.

Can a student build this?
Yes, at lab scale, by reframing the problem. The Cornell system reconstructs continuous pose for 12 joints, which needs camera ground truth and serious training data. The student build classifies a defined set of exercises and flags form errors, which needs only labelled echo recordings — a far easier dataset to collect — and a TinyML classifier that fits on an ESP32-class board. The full build is the EchoMotion-Inspired Ultrasonic Smartwatch for Exercise Tracking and Form Feedback using TinyML.

Does it work on other watches?
The Cornell team's model worked on a Galaxy Watch 7 and a Pixel Watch 3 without retraining, after being developed on the Galaxy Watch 6. That suggests the echo patterns the system relies on come from human bodies and movement rather than from one watch's hardware. For a student build, the equivalent test is to train on several teammates' data and check the model still works on a wrist it has not seen.

What are its limits?
It is trained on a defined set of exercises, so it works best on movements similar to its training data. Echoes from other people nearby degrade performance slightly, so a crowded room is harder than an empty one. Average error under 4 inches is good for form feedback but not for clinical measurement. And the student build inherits all of these limits plus its own: a small exercise set, a discrete transducer rather than a watch-grade speaker, and a classifier that recognises patterns rather than reconstructing exact joint positions.

Limitations

  • Every EchoMotion figure in this guide comes from the Cornell paper and its TechXplore coverage of 11 November 2026, not from independent measurement: average error under 4 inches, about 2.4 inches after calibration, about 12 mW sensing power, and 5–6% battery for a 30-minute session are reported figures from the research team.
  • The system is trained on a defined set of 18 exercises and predicts 12 joints; it is not a general-purpose body scanner, and performance on movements outside its training set is unproven.
  • Average error under 4 inches is suitable for rep counting and form correction, not for clinical gait analysis or any application needing exact joint angles.
  • Echoes from other people nearby degraded performance slightly in the Cornell tests, so the system performs best in a clear space and worse in a crowded room — a limit the student demo should show, not hide.
  • The student build classifies a defined set of exercises rather than reconstructing continuous pose, and its accuracy, power draw and exercise set are its own measured values, not the Cornell figures. Do not quote the paper's numbers as the student build's results.
  • Ultrasound sensing records body-movement data. Keep echo recordings on the device, delete training data after the model is trained, and run demos only with the consent of everyone in the room.

More project guides

More in IoT & Embedded

How Electronic Noses Work: From Scentronic's Scent Messages to a Student E-Nose

Ohio State's Scentronic sends smells over the internet by sending only a word. This guide explains how: the tin-dioxide physics inside an MQ gas sensor, why six sensors together form a fingerprint one sensor cannot, how a small k-NN classifier turns that pattern into a scent name, why only a label travels over MQTT, and how a wristband turns the label back into mist — plus the practical protocol for recording your own fingerprints and the honest limits of the whole idea.

Read guide

Best 10 Predictive Maintenance & Equipment Health Monitoring Projects Students Actually Build

Machines that warn you before they break — predictive maintenance is the most employable theme in industrial IoT. This guide lists 10 real published project builds: a NASA C-MAPSS turbofan RUL model, a live 3D digital twin, ISO-10816 vibration monitoring, TinyML anomaly detection, transformer and capacitor-bank health monitors, LoRaWAN structural monitoring, acoustic leak detection, RO purifier health tracking, and underground cable-fault location. Every entry links to its live project page.

Read guide

Best 10 Waste Management & Recycling Projects Students Actually Build

Ten real, published projectech.in builds across the waste chain — IoT segregation bins, fill-level dustbin nodes, GPS garbage-truck tracking, CNN waste classifiers, an edge-AI compost monitor, a kitchen-waste biogas digester and a plastic shredder — each with the problem it solves, how it works and why students pick it.

Read guide