A glove-type wearable Hangul keyboard — type Korean anywhere, in any posture, without looking.
Team 손보드 (HandBoard) · Department of Mechanical Engineering, Seoul National University 🏆 Grand Prize, Mechatronics Competition (SNU Dept. of Mechanical Engineering / HD Hyundai, Dec 2025) Entry for the 15th SNU College of Engineering Creative Design Festival (창의설계축전, 2026)
16 tactile switches sit on the finger phalanxes of a pair of gloves and are pressed by the same hand's thumb. The device enumerates as a standard USB HID keyboard sending Dubeolsik (두벌식) keycodes — no driver, no companion app; the OS IME composes Hangul exactly as it would for a desktop keyboard. Total bill of materials: ₩14,184.
A within-subject study with 20 participants compared KIH against a TV-style cursor on-screen keyboard (the input method used on smart TVs and in XR). Each participant trained 10 minutes per condition, then completed 15 one-minute trials, typing words drawn at random from a 300-word pool without looking at the device. Condition order was counterbalanced.
| KIH (glove) | TV on-screen keyboard | |
|---|---|---|
| Speed, trial 1 → 15 | 2.3 → 10.6 WPM (4.6×) | 5.6 → 7.5 WPM |
| Power-law fit | WPM = 1.55·n0.602 (R² = 0.636) | WPM = 5.80·n0.079 (R² = 0.142) |
| Error rate, trial 1 → 15 | 17.2% → 4.1% (−76%) | 4.3% → 5.6% (correction rate, no trend) |
Three things the data shows:
- KIH keeps learning; the cursor keyboard does not. A power-law exponent of 0.602 against 0.079 means the baseline was already saturated at first contact — participants had used TV cursor keyboards before, and that is as fast as the method gets.
- Speed and accuracy improved together. No speed–accuracy trade-off appeared: by trial 15 users were 4.6× faster and made 76% fewer uncorrected errors.
- The crossover is early. KIH passes the baseline at around trial 9 — roughly nine minutes of practice — and the curve is still rising at trial 15. Extrapolating the fit gives ~12 WPM at 30 trials and ~16 WPM at 50.
The final 10.6 WPM is on par with recent hands-/eyes-free text entry (ankle gestures 11–13 WPM, CHI 2026) and above the ~8 WPM reported for commercial cursor-selection keyboards — while leaving both the user's gaze and posture free.
WPM here follows the standard convention, computed as jamo-per-minute ÷ 5.
Mapping efficiency (KSPC 1.27, 74.2% single-tap) is in analysis/RESULTS.md.
Keyboards have kept the "board on a desk" form factor for over a century. That form binds the input coordinate frame to a surface outside the body, which constrains posture and gaze at the same time — and that constraint, not any individual device's performance, is the actual problem.
- XR — with an HMD on, expert typists lose ~36% of their speed on a physical keyboard, and a review of 176 XR text-entry techniques names text input as the area where XR still falls short of desktop and mobile.
- Accessibility — Korean braille notetakers cost about ₩6,000,000, yet only ~13% of Korean visually-impaired people read braille. There is no low-cost, braille-free tactile input device for the rest.
- Posture — 45.5% twelve-month prevalence of neck pain among office workers, with prolonged sitting and forward neck flexion as significant risk factors.
So the project reframes the goal as posture-free, eyes-free text entry and puts the keyboard on the hand instead of on the desk.
-
Phalanxes are built-in keycaps. People can distinguish and operate multiple buttons on their own fingers using touch and proprioception alone, without looking (DigitSpace, CHI '16). 16 buttons (8 per hand) expand on that capacity without exceeding it. Positions were chosen from three eyes-free layout candidates using responses from 30 participants plus the HCI literature. Because the switches ride on the fingers, no desk or input surface is needed, and because a switch closes an electrical contact, there is no gesture to estimate — no camera, no recognition, no probabilistic error.
-
Multi-tap = stroke addition (가획). Base jamo are one tap; derived jamo come from tapping the same button again — two taps for aspirates and y-vowels (ㄱ→ㅋ, ㅏ→ㅑ), three for tense consonants (ㄱ→ㄲ). This is the 가획·병서 principle of Hunminjeongeum mapped onto repetition count, so the only rule to memorize is "same family, same button, press again."
Hangul's base letters largely coincide with its highest-frequency jamo, so following the derivation principle also buys frequency efficiency for free: on a 199,806-jamo corpus the layout scores KSPC 1.27 (−26% vs. a frequency-blind layout) with 74.2% of input completed in a single tap.
-
Sequential, never chorded. Chorded keyboards never reached general adoption because of the memorization barrier. Left-consonant / right-vowel follows Dubeolsik, so existing Korean typists carry their layout knowledge over; compound vowels and final clusters are typed sequentially and composed by the OS IME, exactly as on a desktop keyboard.
-
Per-user calibration. The multi-tap window (default 300 ms) is measured per user by
firmware/keyboard_glove/tap_calibration.py: the firmware is flipped into a raw-tap mode, the user types two sentences, and every inter-tap interval is labelled intentional multi-tap vs separate keystroke that reuses the button by aligning observed keydowns against the target sequence frommapping.json. The tool picks the threshold that minimises misclassification and rewritesTAP_WINDOW_DEFAULTin the .ino directly — bumping aCAL_STAMPso the board's stored EEPROM value cannot shadow the newly calibrated one. No serial port, no extra packages.
firmware/
keyboard_glove/
keyboard_glove.ino Arduino Leonardo firmware (USB HID, multi-tap engine, raw-tap
measurement mode, calibration block, CAL_STAMP + EEPROM)
tap_calibration.py per-user multi-tap window calibration — lives next to the sketch
because it edits it
mapping.json copy of experiments/mapping.json so the calibration tool works
standalone; keep both in sync (the tool warns if they diverge)
test/ host-side test of the multi-tap engine — stubs the Arduino API,
#includes the sketch and drives it on a virtual clock, so the
timing logic is verified without a board (`make both`)
legacy/ original course-project sketch (development history)
calibrate_window.py optional serial route: tune the window live without re-flashing
experiments/
PROTOCOL.md experiment protocol (Korean): button reach-time & mapping cost,
learning curve, usability/posture, cursor-OSK comparison,
eyes-free — standard text-entry methodology (MacKenzie et al.)
speed_test.py 1-minute random-word speed test GUI (the main study tool)
tv_remote_sprint.py TV-remote cursor keyboard, 60-second word sprint — the matched
baseline (same word pool, same WPM definition)
tv_osk_test.py same input method, sentence-transcription variant
logger.py transcription/tapping logger GUI
analyze.py metrics & figures: CPM/WPM, MSD error rate, learning curve,
mapping cost vs. 10,000 random layouts
mapping.json 16-button ↔ jamo mapping (single source of truth)
logs/
speed_vs_tv_all.csv raw per-trial records for the study — 20 participants × 2
conditions × 15 trials (600 rows), participants pseudonymous
README.md data dictionary; why the two conditions carry different columns
analysis/
mapping_analysis.py corpus → jamo frequency, KSPC, same-button bigram rate
verify_results.py recomputes every figure in RESULTS.md from the raw data and
exits non-zero on any mismatch
STUDY_DESIGN.md study design, metric definitions and limitations
COST_MARKET.md bill of materials, scale economics and recommended price
RESULTS.md quantitative results
docs/
BUILD.md build guide: parts, button placement, assembly, bring-up
images/ photos, diagrams and the study figure
Every Python tool is standard-library only and ships with a --selftest that runs without a GUI;
the firmware has a host-side test suite that needs no hardware. The study's raw data is in the
repository, so the numbers below can be recomputed with python analysis/verify_results.py.
| Part | Qty | Unit | Note |
|---|---|---|---|
| Tactile switch (6×6 mm) | 16 | ₩134 | 2 per finger (index–ring), 1 pinky + 1 function key, per hand |
| Arduino Leonardo (ATmega32u4) | 1 | ₩8,500 | native USB HID |
| Leather gloves | 1 pair | ₩1,000 | |
| Wiring, heat-shrink | 1 set | ₩1,040 | |
| Adhesive, insulating tape | 1 set | ₩500 | |
| Total | ₩14,184 | ~₩12,230 at 15+ units (−13.8%) |
Flash firmware/keyboard_glove/keyboard_glove.ino (Arduino IDE, board: Leonardo).
Serial at 115200 baud: W<ms> set tap window · S save to EEPROM · R1/R0 raw-tap mode ·
C1/C0 tap stream · ? status.
# Python 3.8+; GUIs use tkinter (stdlib), no pip packages
python experiments/speed_test.py # 1-min speed test
python experiments/tv_remote_sprint.py # TV-remote baseline, 60s word sprint
python experiments/tv_osk_test.py # baseline, sentence-transcription variant
python experiments/logger.py --mode transcribe --participant P01 --session S1
python firmware/keyboard_glove/tap_calibration.py # tap-window calibration → patches the .ino
python experiments/analyze.py transcribe # figures (needs matplotlib)
# verification, no hardware required
python experiments/speed_test.py --selftest
make -C firmware/keyboard_glove/test both손보드 (HandBoard) — Department of Mechanical Engineering, Seoul National University
| Kijun Kim (김기준) | Hardware | glove integration, switch mounting, prototype build |
| Dongwon Lee (이동원) | User studies & market | study design, data analysis, cost/market analysis |
| Hongjun Ha (하홍준) | Software | firmware, multi-tap engine, calibration tooling |
MIT — hardware design, firmware and experiment tools.