BROADMIND AI R&D
Continuous 3D hand decoding from eight dry channels — and the ten-minute path from a clean install to a first stream.
Most of what a research kit does is meant to be invisible. A stream arrives, the numbers are boringly correct, the file opens in the tool you already use. The work that actually costs time sits one layer above that — and it rarely survives contact with a specification table.
This is the first of a short, recurring note from the R&D side of the bench: what we are measuring, what the numbers are a number for, and what we have not solved yet.
What you are looking at
Eight dry contacts sit over sensorimotor cortex — CP1, CP2, C3, C4, CP5, CP6, FT7 and FT8 — sampling at 250 Hz, 24-bit. The model reads that stream and regresses continuous 3D hand position: not a gesture label, not a discrete command, a position in space, frame by frame.
Nothing is averaged on the way in. The montage is chosen for motor-imagery work rather than for a full-head survey, and the raw stream is what the model sees, which is the only reason the number below can be checked by anyone else.
The number, and what it is a number for
Our decoder reaches up to 85% on this task. Published results from other groups land in the same band, which is the useful part of the comparison: it puts our number where the literature is, rather than somewhere the literature does not go.
It is a peak, and we label it as one. Accuracy in BCI is a property of a model on a task, not a property of a headset: a better decoder moves that figure, and it does not move the noise floor, the sample rate or the montage. Subjects, sessions and the train/test split ship with the v1.0 artefact — until that exists this is a peak on our own capture, not a benchmark. If you point your own model at the same stream, expect your own number.
That distinction is also the argument for an unrestricted raw stream. Acquisition is the part we can specify on a datasheet. Decoding is the part you should be able to re-run, replace, or disagree with.
From install to first signal
The other half of the work is making the first ten minutes boring. What follows is the shape of that path: open the port, take the stream, watch eight channels come up.
import serial
PORT = "/dev/tty.usbmodem14201" # the kit enumerates as a USB serial device
kit = serial.Serial(PORT, 921600)
# 8 dry channels, 250 Hz, 24-bit — nothing filtered for you
SECOND = 8 * 3 * 250 # 6 000 bytes of raw samples
for i in range(3):
frame = kit.read(SECOND)
print(len(frame), "bytes", round(len(frame) / 1024, 1), "kB")
Two things worth saying about that snippet: it is schematic, not the SDK call — the interesting part is how little sits between a port and a number — and the 6 000 bytes per second is the whole stream. There is no lossy step hiding in the middle.
Where this is going
Two things we want to publish properly: a fixed evaluation set for the hand decoder, so that the 70–85% band can be re-measured by anyone holding a kit, and a calibration routine that survives a different head. Until those exist, the honest position is the one above — acquisition is specified, decoding is demonstrated.
If you want to point a model at the stream yourself, the specifications are here and the kit is the whole kit.