A smartwatch already has a surface that stays in contact with the body for hours: its back. It is used for pulse, temperature and conductivity measurements, but rarely as an interaction surface. EITWatch asks a practical question: can hand gestures be read through the small patch of skin pressed by an ordinary watch, without a camera looking into the room and without an electrode band extending beyond the case?
Xuanyou Liu, Novel Alam and Karan Ahuja answer with a prototype of eight planar electrodes arranged in a 31 mm ring and driven by an ESP32-S3.1 The UIST paper describes 35 impedance measurements at 48 Hz, a multi-depth acquisition protocol and a study with twelve participants.2 The numbers are interesting, but they do not describe a shipping smartwatch. They describe a shift in interface design: the first component is not the screen but the geometry of an electrical field.
Seeing without a camera, but not without assumptions
Electrical impedance tomography injects a small alternating current between electrodes and measures resulting voltages. The current does not travel in a straight line; it bends through tissue. Muscle conducts differently from fat and bone, and when muscles and tendons change shape during a gesture, the measured distribution changes too.2
This is not a tiny electrical camera producing a clean picture of a hand. EITWatch receives a group of signals whose variations can be associated with gesture classes. That distinction matters for design: the sensor does not need to understand everything the hand is doing. It needs to produce a signature stable enough for a chosen interaction to be recognised.
The approach avoids several camera problems. An onboard camera can capture a room, another person or a face when the gesture does not require it. Optical systems also deal with occlusion, light and wrist orientation. EITWatch sees neither the room nor an image of the hand; it reads a local disturbance under the skin. The privacy promise is real but bounded: physiological data remains sensitive data even when it is not a photograph.
The authors do not present EIT as magic. They start with a form-factor problem: earlier systems generally use a band around the wrist or separate analogue electronics, giving their electrodes more geometry. EITWatch removes that convenience and keeps the patch a watch can actually occupy.2
What a bracelet loses when it becomes a watch
In a circumferential band, electrodes surround the wrist. They can inject and measure around a cross-section, with paths crossing the volume from multiple angles. A watch back is planar. Eight electrodes look from one side, and the electronics must fit a shape that was not designed for tomography.
The acquisition protocol is therefore the centre of the work. With neighbouring electrodes, the field is mostly sensitive to shallow tissue. EITWatch fixes one electrode as a source and moves the return across the other seven. Separations range from about 11.9 to 31 mm; in a homogeneous model, characteristic depth grows with distance, roughly from 6 to 15.5 mm. This is not a penetration boundary, and the authors say so: the field decays continuously and a real wrist is heterogeneous.2
The board produces 35 measurements per frame rather than simply rotating adjacent measurements. Different paths sacrifice some channels while reaching deeper tissue and different directions. In a saline tank, anomalies at 10 and 20 mm retain stronger relative responses under multi-depth scanning than under adjacent injection. The authors avoid assigning the whole difference to depth: separation, direction and channel count change together.2
That caution matters. A sensor demo can easily turn correlation into a physical explanation. Here, the paper proposes a useful hypothesis — varying paths partly compensates for the missing ring — and supplies an experiment that supports it without pretending to isolate every variable.
Building a signal in a case that is still too large
The prototype uses a 40 × 60 mm board. The electrode ring fits the 40 mm watch-back area while analogue electronics extend another 20 mm. The paper specifies a 50 kHz sine generator, current source, amplifiers, digital potentiometers, multiplexers and a 12-bit ADC; a XIAO ESP32-S3 sequences the measurements.2
The phrase “smartwatch-integrated” needs precision. The electrodes follow a watch-back geometry and the microcontroller is onboard. The prototype is not yet the complete miniature electronics of a commercial watch. That does not weaken the research. It identifies its actual object: show that a planar patch can carry discriminative signal before solving the entire product.
The authors’ hardware and firmware repository makes the claim inspectable.3 That matters for a maker audience. The form is not only a research image; it is schematics, layouts, a bill of materials and firmware. Reproducible does not mean anyone can build it in a weekend. It means the conversation can return to electrodes, gains, frequencies and tradeoffs.
Calibration is the hidden interface
A gesture is not a fixed property of a body. The watch is remounted slightly differently, pressure changes, skin becomes dry or wet, and anatomy varies. EITPose, an earlier system in the same family, makes the problem measurable. Its eight-electrode band consumes about 0.3 W, runs around 10 Hz and was evaluated on 22 participants; mean per-joint error was 11.06 mm within a session, 17.81 mm across sessions and 18.91 mm across users.45
EITPose initialises the device by adjusting injection current and gain for each electrode pair. A waveform checker rejects noisy measurements when contact degrades. Calibration becomes part of the product rather than a user failure.5
EITWatch exposes the same issue differently. Participants perform gestures, return two days later and are evaluated without personal training data. Six static gestures reach 92.5% in-session, then fall to 73.2% after remounting two days later and 63.1% for a new person with no training. For five micro-gestures plus rest, the reported figures are 91.5%, 70.4% and 55.3%.12
The conclusion is not “92% accuracy.” It is that donning and person are first-order variables. The interface must choose between a command that works immediately but recognises little, and a richer command that asks for learning, recalibration or an adaptive model.
Macro gesture and micro gesture are different products
EITWatch’s six macro gestures — fist, thumbs-up and pointing among them — are easy to show in a demo. Its five micro gestures — thumb slides, pinches, rapid hand opening and a wrist flip — are closer to commands that could be issued without interrupting an activity.1
A micro gesture is harder to specify. It must be distinct for the model, comfortable for the body, rare enough not to trigger accidentally and memorable enough for a person. Sensor physics shapes the interface grammar. An electrode does not recognise “next”; it reads a particular movement with particular muscular constraints.
The authors use filtered signals, sliding windows, 175 features and an ExtraTrees classifier.1 That choice is as revealing as the sensor. The prototype does not depend on a giant opaque model; it extracts features and performs lightweight classification. That is promising for embedded devices, but it also means the learner must absorb wrist noise rather than only the ideal gesture in a lab protocol.
Comparing systems without mixing measurements
The wrist-interface field now includes approaches that do not see the same thing. WatchHand uses a smartwatch’s speaker and microphone to emit inaudible FMCW signals and reconstruct a 3D hand pose; its 2026 CHI paper reports a 7.87 mm mean per-joint position error in a cross-session test with device remounting.6 It follows another philosophy: obtain rich geometry through acoustics, with active emission and a different signal-processing burden.
OpenWatch is not a competing sensor so much as a multimodal benchmark. It contains more than ten hours of IMU and PPG data from 50 participants, 59 labelled gesture sequences and subject-independent evaluations.7 Its most useful lesson for EITWatch is methodological: performance and memory efficiency do not automatically rise together, and a specialised architecture can outperform a much larger general model on a wearable task.
Cameras, EMG, ultrasound, radio, inertial sensing and EIT all move the tradeoff. A camera gives an intuitive relation between image and gesture but captures an environment. An IMU is cheap and already present in a watch but mostly sees wrist motion. EMG reads muscle activity more directly but has its own electrodes and calibration. EIT reads a changing tissue volume without imaging it, then has to work with more indirect physics.
Earlier high-resolution EIT work had already shown that electrode count, sampling speed and reconstruction method change recognition and image quality.8 EITWatch adds a product constraint: the device must resemble a watch before it can become a watch interface.
Privacy and power are not bonuses
EITPose reports a 0.3 W profile and frames the absence of a camera as a privacy advantage.5 EITWatch’s project page reports 8.6 hours of continuous sensing on a small 300 mAh battery.1 These are prototype measurements, not commercial battery promises. They point to a direction: continuous interaction makes energy part of the definition of a gesture.
A camera-free sensor is not automatically neutral. Signals may reveal movement, habits or physiological properties. Not filming a kitchen does not mean the data needs no protection. A sensible architecture could keep inference on the watch, transmit only an event and let a user inspect or delete raw samples.
That is compatible with a lightweight model. Local classification can avoid transmitting 35 raw channels continuously. It also complicates debugging: understanding a false trigger requires keeping a useful sample or summary.
The interface starts before the pixel
EITWatch is interesting because it does not add a new animation to an existing watch. It changes where the watch can sense. The 31 mm constraint shapes the field; the field shapes the protocol; the protocol produces signatures; signatures limit the vocabulary of gestures; that vocabulary eventually determines the visible interface.
Digital products often reverse this order. A team draws an interaction and then looks for a sensor that can reproduce it. Here, the material comes first. A good interface does not hide physics; it translates physical constraints into gestures that are legible, comfortable and reliable.
The project has not solved universal wrist input. Its own results show a large gap between repetition within a session and transfer across days or people. That is exactly why it deserves attention. EITWatch is not merely a new way to scroll cards. It is a board, an electrode ring, a calibration problem and a user study showing that every interface begins with a meeting between a body and a material.