cheir (χείρ) "hand" + graph (γράφω) "writing" — a hand that writes itself.
A wearable gesture-tracking glove that reads the orientation of each finger, fuses it with a Madgwick filter, and classifies static hand-shapes / the fingerspelling alphabet into text or control signals — entirely on-device.
Coming soon — will be updated as the glove reaches a working state.
Cheirograph is a left-hand glove instrumented with six IMUs: five MPU-6050 modules (one per finger) multiplexed behind a PCA9548A, plus the XIAO nRF52840 Sense's onboard IMU as a rigid back-of-hand reference. Each sensor's raw accelerometer and gyroscope data is fused into an orientation quaternion with a Madgwick filter running at 100 Hz on the MCU. A finger's pose is expressed as its orientation relative to the hand — making the signal invariant to arm rotation. Those relative quaternions feed a small TinyML classifier (trained in Edge Impulse and deployed on-device) that outputs recognised letters over BLE HID.
This is a solo engineering project and portfolio flagship, built to understand every layer of the stack firsthand — from I²C bus multiplexing through quaternion math to on-device inference.
MPU-6050 × 5 ──┐
├── PCA9548A mux ──┐
XIAO onboard ──┘ ├── Madgwick fusion (100 Hz)
│ → q_hand, q_finger[0..4]
├── Relative orientation
│ q_rel = conj(q_hand) ⊗ q_finger
├── Feature vector (5 × quat)
├── TinyML classifier (Edge Impulse)
└── BLE HID → typed letter / control signal
Key design choices:
- MCU-side Madgwick (not the MPU-6050 onboard DMP) — uniform, transparent, works through a mux.
- Relative-to-hand quaternions — waving your arm changes nothing; curling a finger changes everything.
- Middle-phalanx placement for fingers (proximal for thumb) — captures combined MCP+PIP bend, richest single-sensor curl signal.
- Left hand — non-dominant, keeps the right hand free for soldering and laptop during live testing. Fixed for the dataset.
| Part | Role | Notes |
|---|---|---|
| Seeed XIAO nRF52840 Sense | MCU + hand-reference IMU + BLE | Onboard LSM6DS3 on internal I²C — not behind the mux |
| 5× MPU-6050 (GY-521) | Finger IMUs | All at I²C addr 0x68; sit behind the mux |
| PCA9548A | 8-channel I²C multiplexer | Addr 0x70; selects one finger IMU at a time |
| Half-finger glove (left) | Substrate | Finger IMUs on middle phalanx; thumb on proximal phalanx |
| Leukoplast tape / cable ties | Mounting + strain relief | Wires fatigue at knuckles — anchor every run |
Authoritative wiring + sensor-to-finger map: hardware/WIRING.md.
Cheirograph/
├── README.md — this file
├── CLAUDE.md — standing context for Claude Code
├── GENERAL_PLAN.md — roadmap, phases, Gantt
├── DOCUMENTATION.md — dated ledger of planned vs. achieved
├── DECISIONS.md — engineering forks and rationale
├── .gitignore
├── LICENSE
│
├── firmware/ — numbered milestone folders (never overwritten)
│ ├── 00_led_sanity_test/ — Phase 0: first flash, onboard RGB LED blink
│ ├── 01_xiao_imu_test/ — Phase 1: XIAO onboard IMU over serial
│ ├── 02_single_mpu6050_test/— Phase 2: single MPU-6050, direct I²C
│ ├── 03_mux_channel_test/ — Phase 3: PCA9548A mux bring-up
│ ├── 04_all_imus_raw/ — Phase 4: all 6 IMUs streaming @ 100 Hz
│ ├── 05_madgwick_fusion/ — Phase 5: Madgwick filter per IMU
│ ├── 06_relative_orientation/— Phase 6: relative quaternions + skeleton viz
│ ├── 07_full_glove/ — Phase 7: mounted on glove, strain-relieved
│ └── lib/ — shared library code (only after it stabilises)
│
├── tools/ — Python scripts: plotter, visualiser, data capture
├── data/ — labelled training samples
├── ml/ — Edge Impulse export, model artefacts
│
├── hardware/
│ ├── BOM.md — bill of materials
│ ├── WIRING.md — authoritative sensor ↔ mux-channel ↔ finger map
│ └── datasheets/ — board references, pinouts, spec sheets
│
└── docs/
├── REFERENCES.md — external sources, datasheets, libraries used
├── log/ — narrative devlog, one entry per phase
└── media/ — photos, GIFs, serial traces
Skills this project exercises, layer by layer (see GENERAL_PLAN.md for which phases are actually complete):
- Embedded C/C++ — Arduino framework on PlatformIO; real-time sensor read/fuse loop at 100 Hz.
- I²C bus multiplexing — PCA9548A channel switching; address collision resolution.
- Sensor fusion (Madgwick filter) — quaternion-based orientation from raw accel + gyro; gyro-bias calibration; coordinate frame discipline.
- Real-time data visualisation — Python / pyserial / matplotlib serial plotter; 3D hand-skeleton visualiser.
- TinyML / on-device inference — Edge Impulse pipeline; model quantisation; deploying to nRF52840.
- BLE HID — keystroke output over Bluetooth Low Energy.
The project is structured as eleven incremental, proven-before-advancing phases. Full roadmap, deliverables, and Gantt chart: GENERAL_PLAN.md.
This section will be expanded as phases land.
Toolchain (current — Arduino IDE):
- Arduino IDE with the Seeed board package — add
https://files.seeedstudio.com/arduino/package_seeeduino_boards_index.jsonunder File → Preferences → Additional Boards Manager URLs, then install Seeed nRF52 Boards via Boards Manager. - Tools → Board → Seeed XIAO nRF52840 Sense; select its COM port (double-tap RESET if the port doesn't appear).
- Open the
.inoin anyfirmware/NN_*/milestone folder and Upload. - Serial Monitor at 115200 baud.
Full board reference, pinout, and library notes:
hardware/datasheets/XIAO_nRF52840_Sense.md.
A PlatformIO migration (for
platformio.iniversion pinning) is planned once the milestone folders stabilise — seeDECISIONS.md. Until then, the board-package and library versions in use are recorded in each milestone README.
- Python 3 for
tools/— dependencies pinned intools/requirements.txt(lands with the first script).
Fingerspelling / static gestures only. Not full ASL translation — see GENERAL_PLAN.md for the scope boundary.
MIT — see LICENSE.