NaptX

Overview

NaptX puts a vibrotactile actuator on the back of the fingernail instead of on the fingerpad. The pad stays free, so you can touch a screen, a key, or a real object while receiving haptic cues about it.

It exists as two devices built for two different jobs: NaptX, a self-contained wireless unit that lives on a single nail and is optimised for size, cost and wearability; and Napt10X, a wrist-worn research platform that drives ten actuators, reads ten accelerometers, and exists to run experiments quickly. Both came out of my PhD at Imperial College London.

Why the nail

Haptic gloves and fingertip devices deliver rich sensations by covering the one surface the brain relies on most. Bracelets and body suits leave the hands free but can't create the illusion of touching anything with the hand.

Rings and back-of-finger devices free the pad too, but they put the feedback a phalanx away from where the finger is acting.

The fingernail sits between those failures. It is rigid, coupled almost directly to bone, and one layer of tissue away from the pad - so it transmits vibration efficiently: about 70 % of a nail pulse's acceleration reaches the fingerpad. A small actuator there is clearly felt and precisely localised. The catch is that stimulation on the nail can feel like a foreign object rubbing against it, which is exactly what continuous buzzing from an ERM motor does.

Short impulses from an LRA behave differently. They are crisp, start within 6 ms, and in early demos many people described them as coming from the fingertip itself. That observation started the project. The measurements since then say something more careful: with the hand at rest, a nail pulse is mostly felt at the nail. What the nail reliably offers is a cue on the right finger with the pad left free. Whether that cue is attributed to the fingertip while the finger is touching something is the question now being tested.

NaptX - the single-nail unit

• 9 × 6 × 15 mm, 2.45 g - light enough that inertia doesn't tug on the nail during fast hand movement
• Origami-folded flex PCB wrapping around a coin LRA, bonded to a 3D-printed support that sits on the nail - the video above shows the KiCad flex viewer written to preview the fold in 3D before ordering
• ESP32-C3 for control and Bluetooth, with the antenna routed directly on the flex - reliable wireless at this size was one of the harder problems
• STSPIN240 current driver for precise amplitude and frequency control
• 3-axis accelerometer for finger motion, and an RGB LED plus push button so it can be used without another interface
• 150 mAh LiPo - several minutes of continuous vibration, aimed at short-session use
• Roughly £20 in parts

Napt10X - the research platform

Where NaptX is one channel done small, Napt10X is ten channels done flexibly.

• Teensy 4.0 for real-time waveform generation and 480 Mbit/s USB, with an ESP32-C3 for WiFi and Bluetooth
• 10 × Vybronics VG0640001D LRAs - 6 mm × 4 mm, 1 g, up to 1.3 G at their 230 Hz resonance, Q-factor around 5
• 5 × STSPIN240 drivers (dual bridge, 1.3 A), PWM at 100 kHz to stay out of both the audible and the tactile range, each with a low-pass output filter
• Two 8-bit shift registers collapse the 20 direction inputs down to 12 microcontroller pins
• 10 × LIS3DH accelerometers, multiplexed and read at 1 kHz, so the actual delivered vibration is measured rather than assumed
• SGTL5000 audio codec (16-bit, 44.1 kHz) and an EYESPI header for a small display - hooks for multimodal experiments that haven't been used yet
• Around £55 in parts, or about £90 with the audio and display options

Driving an LRA properly

An LRA is a magnet on a spring inside a coil. It is cheap, small, and a poor actuator for anything except a sine wave at its resonant frequency - the Q-factor above means that at 115 Hz or 345 Hz you get roughly half the amplitude for the same power. Getting controllable feedback out of it is the engineering problem.

Drive chain. PWM into an H-bridge driver, then a second-order LC filter (33 µH, 1 µF, corner around 28 kHz) to turn the switching waveform into something the actuator's tiny moving mass responds to linearly.

Model. Electrically the LRA is a 25 Ω, 1 mH coil in series with a back-EMF proportional to the velocity of the moving mass. Mechanically it is a mass-spring-damper driven by the coil force. Coupling the two gives a transfer function from filtered voltage to displacement that is a second-order resonance in series with an electrical first-order lag - which is the narrow passband you have to control around. The full derivation is in the repository under docs/lra_model.md.

Where is a nail pulse felt?

The first user study tested the assumption the project started from. Fifteen LRAs were taped to one hand - nail, fingertip, and medial phalanx on each of five fingers - and 18 participants localised and rated the intensity of single pulses over 90 trials.

A 2026 reanalysis of the recorded accelerations changed how this study reads. For the same command, the nail received about half the acceleration of the fingertip, so the original comparison was between unequal pulses. With that accounted for, the picture is asymmetric. A nail pulse is felt at the nail about three times in four and at the fingertip about one time in five. A fingertip pulse is reported at the nail far more often. Participants can tell the two sites apart at every intensity, and how often a given person attributes a nail pulse to the fingertip is a stable trait that ranges from 0 to 60 %. The knuckle, further away, is almost never confused with the fingertip.

So the nail is a well-localised site that shakes the pad, not a site that passes for the pad. An earlier version of this page concluded that people could not tell the two apart. That conclusion did not survive the reanalysis and has been withdrawn.

This study and the next were run with Qi Li, an MSc student I co-supervised with Yanpei Huang and Etienne Burdet.

Feeling a remote surface through your nail

The second study asked whether nail feedback is useful while the same finger is busy controlling something. A Kinova Gen3 arm carried a sensorised probe over silicone-covered boards with hidden relief underneath - the surgical analogy is a polyp under tissue. The friction-induced vibration picked up by the probe was replayed on the participant's nail, and the same finger's orientation drove the arm.

Three tasks of rising difficulty: rank the density of hidden bumps in four corners, locate four hidden squares on a 3 × 3 grid, and identify hidden shapes. Each was run twice - once with the participant teleoperating the arm, once with the arm scanning automatically - to test whether controlling with the finger you're also sensing through gets in the way.

It didn't. Accuracy in teleoperation was slightly higher in all three tasks (not significant), and the hardest task was completed faster under teleoperation. Most participants reported that actively controlling the exploration helped their perception. Using one finger to both act and feel, with nothing covering the fingerpad, works.

Demonstrations

Bump illusion. An Android app shows a grid of dots; sliding a finger over the screen fires a pulse on the nail at each dot. In informal demos most people describe the bumps as being at the fingertip, not the nail. That is an observation, not a measurement, and it happens while the pad is in contact with the screen - the case the localisation study did not test. The more telling observation: if someone else touches the screen while the wearer holds their finger in the air, the sensation is artificial and unpleasant - but if the wearer moves their raised finger in sync, some feel a bump anyway. Context and agency shape where a sensation lands.

Remote texture. Two NaptX units, one on each index finger. One records the natural friction vibration of sliding over a surface, band-pass filters it, and streams it to the other, which replays it on the opposite nail. Participants described it as realistic, particularly when the passive finger was resting on a surface.

Sensing force through the nail

The missing input is fingertip force - vibration propagation through the finger depends heavily on how hard it is pressing, so matching feedback to a touch needs to know that. Most force sensors go where the touch is, which defeats the purpose.

The approach being explored uses the nail itself as a dielectric: a thin copper plate on the device forms a capacitor with the fingertip through the nail, and pressing deforms the nail enough to change the capacitance measurably. Early prototypes give a clean real-time signal as finger pressure oscillates. Fringe effects and robustness across users are the open questions.

What's still open

With the hand at rest, about three in four nail pulses are felt at the nail. The demonstrations suggest that contact changes this: when the finger touches a surface at the moment of the pulse, the brain may attribute the vibration to the contact. That is the next experiment - calibrated stimuli, pulses timed to contact onset, and the delay varied.

The second open question is retention. Nail cues speed up finger selection while they are on. Whether the benefit stays once they are switched off has not been measured yet.

Student work on this platform

Two MSc projects at Imperial were completed on NaptX in 2026.

Piano training - ten nail actuators cue which finger to use. Against matched visual cues, twenty participants were 247 ms faster, and wrong-hand errors fell from 67 to 2. The same setup runs a teacher-to-student link for remote lessons.

Texture rendering - a piezo actuator on the nail with a custom five-channel high-voltage driver. Matching rendered textures to real materials rose from 24 % to 43 % once each material had its own pulse parameters.

Open hardware

Everything is in the repository: KiCad schematics and layouts for both boards, the bill of materials with supplier part numbers, a C++ library, a command-line demonstration app, and the project documentation under docs/.

Status

Active, as the haptics module of the CleverHand platform. Current work is a ten-channel wrist-worn driver kit for research labs, a piezo line for texture rendering, and the papers from the three studies above.