An open, minimal sEMG wristband, like the Meta Neural Band. Eight bipolar surface-EMG channels around the wrist, a 24-bit biopotential front end, Bluetooth LE, and a module the size of a small watch face — designed from first principles to record data good enough to train gesture models, and cheap enough to just build.
Status: design complete, not yet built. Schematics, both PCBs (DRC-clean), mechanics, firmware (builds and links, untested on hardware), recorder and analysis code are done; everything below that says "simulated" or "estimated" has not been measured on a real board yet. See what's next.
| Channels | 8 bipolar (P–N 20 mm along the forearm) + a driven bias pad in every pod |
| Front end | TI ADS1298, 24 bit, gain 6, 4 kSPS acquisition, ~1.2 µVrms (20–850 Hz) |
| Electrodes | 24 dry 6.5 mm ENIG gold pads on a 2-layer polyimide flex, foam-backed, 1 mm proud |
| Radio | nRF52832, BLE 5 2M PHY, printed antenna; 2 kSPS × 8 ch, 16-bit stream |
| IMU | LSM6DSL accel + gyro, 208 Hz |
| Power | 100 mAh LiPo, ~10 h streaming; magnetic charger that only fits when the band is off the wrist |
| Module | 31.9 × 34.3 × 9.9 mm, ~15 g total, one size for 145–180 mm wrists |
| Cost | ≈ $195–240 for a first build incl. flex, prints and off-the-shelf parts (breakdown) |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
The boards as laid out (KiCad renders, main board top and bottom; 28.7 × 27 mm):
 |
 |
All four copper layers (python3 tools/copper_layers.py):
A 1080 × 1350 turntable video is in docs/media/turntable.mp4.
All images are rendered from the real CAD and KiCad files with python3 render/make_renders.py (Blender Cycles);
see docs/renders.md.
skin ─ 8 pods × (P, bias, N) gold pads ─ polyimide flex (GND-shielded) ─ 20-pin FPC
└─ main board 28.7 × 27 mm, 4 layers, one-sided assembly
ADS1298 ─ 16 × 4.7 kΩ inputs, bias amp (1 MΩ‖4.7 nF → all middle pads), filtered AVDD/2 reference
│ SPI
nRF52832 ─ printed 2.4 GHz antenna ─ BLE ─► nRF52840 USB dongle ─► host/neurband_ble.py ─► .npz
├ LSM6DSL IMU, button (wake/marker), LED
LiPo ─ XC6206 3.3 V ─ RC filter ─ AVDD charge pads (recessed) ─ TP4054
Like Meta's research band (A generic non-invasive neuromotor interface for human-computer interaction, Nature 2025), it has what matters: bipolar gold contacts 20 mm apart along the forearm, µV-level noise, 2 kHz data. It drops what only costs money: per-pod amplifiers, 16 channels, custom electrodes. The reasoning, decision by decision, is in docs/rev_c/README.md.
Four independent reviews and simulations went into this design:
- Signal chain (docs/science/signal_chain.md): noise budget, dry-contact impedance, mains pickup, sampling and BLE throughput, with pass/fail bench gates.
- Circuit simulation (docs/science/simulations.md, ngspice): bias-loop stability (57° worst-case phase margin), mains residual (≤ 1.5 µV on battery), RF filtering, radio-burst coupling — which is why the bias reference is filtered.
- Forearm volume-conductor model (docs/science/emg_forward_model.md): a 3-D finite-difference model of the distal forearm (bones, muscles, fat, skin, anisotropic muscle), lead fields of all 8 channels, travelling motor-unit action potentials. 8 channels separate single muscles at 98–100 %; pad size, spacing and band rotation are quantified.
- Data & models (docs/science/data_and_models.md): what 8-channel wrist EMG can decode, how to use Meta's open datasets, a first-week recording protocol, and a working pipeline.
 |
 |
Honest expectations (per user, with calibration): 6–8 hand/wrist postures at ~90–98 %, pinches and thumb swipes at ~75–90 %, 2-axis wrist control. Not expected: models that work across people without calibration, typing, fluent handwriting.
A first prototype can be built for around $200 (≈ $195–240 for two assembled boards, the electrode flex, prints and the off-the-shelf parts; breakdown).
- Order the main board (assembled) and the electrode flex through the fab files in
boards/neurband_c/andboards/neurband_c_flex/(*.fab.json). Printmech/rev_c/parts (MJF PA12 / TPU, or PETG/TPU on your own printer). Buy the off-the-shelf bits in docs/rev_c/cost.md (LiPo 401225, magnetic charge pins, nRF52840 dongle, Pico as SWD probe, 22 mm elastic webbing, foam discs). - Assemble: foam discs under each pad, pods onto the flex, flex tail into J1 (contacts toward the board), cell onto the battery pads, close the module. Details: docs/mech/mechanical_c.md.
- Flash
firmware_c/(Zephyr) over SWD with any CMSIS-DAP probe; follow the bring-up checklist in docs/firmware_c.md (ADC ID, test signal, shorted-input noise, then skin). - Record:
python3 host/neurband_ble.py -o rec.npz, and in a second terminalpython3 -m analysis collect basic --out data/s1. - Check and train:
python3 -m analysis report data/*.npz,python3 -m analysis train data/*.npz. Try the whole pipeline without hardware:python3 -m analysis synth /tmp/nb.
Safety. Wear it only while it runs on its own battery. Never connect a charger, USB cable, debugger or grounded instrument while the pads touch skin. This is a research prototype, not a medical device.
| Path | What |
|---|---|
boards/neurband_c/ |
main board: schematic generator (gen_c_sch.py), KiCad project, layout scripts |
boards/neurband_c_flex/ |
electrode flex: schematic + parametric PCB generator |
mech/rev_c/ |
FreeCAD scripts and STEP for module, pods, strap, charger, fit checks |
firmware_c/ |
nRF52832 Zephyr firmware |
host/ |
BLE recorder and packet tests |
analysis/ |
quality gates, features, baseline models, collection prompts, Meta dataset reader |
sim/ |
ngspice AFE simulations and the forearm EMG forward model |
render/ |
Blender pipeline for the images (python3 render/make_renders.py) |
docs/ |
design notes, reviews, science, layout and mechanical docs |
archive/ |
rev A (16 ch, ESP32, 54 × 34 mm) and rev B (compact 16 ch, unfinished) |
tools/ |
the small Python KiCad schematic generator used for every schematic |
- Rev C schematic, main board (DRC 0), flex (DRC 0), mechanics, fit checks on 145/160/180 mm wrists
- Firmware builds and links for the nRF52832 (Zephyr 4.4.2, 0 warnings, 30 % flash / 94 % RAM); host recorder and analysis pipeline tested on synthetic data
- Build two boards; bench gates: shorted-input noise, mains residual, radio-on noise, bias-loop stability
- Antenna tuning (π network on the board), BLE throughput on a real link
- First week of recordings with the protocol in
docs/science/data_and_models.md
Research summary of Meta's work in docs/research/meta_semg_research.md. Human model: Blender Studio "Human Base Meshes" (CC0). KiCad stock libraries. Meta's datasets referenced here are CC BY-NC 4.0.