ONLINE LEARNING · DTW BAND ENGAGED

Custom gestures, learned in the wild.

CSI motion from paired ESP32-S3 seeds is enrolled per user, warped against the live template bank, and adapted continuously. The bank below shows the current cohort, their cost surface, and the past 60 seconds of recognition flux.

— templates — users enrolled — rec/min — avg conf ±—ms warp band

—

Templates in bank

+2 today · 6 cohorts

Users enrolled

3 active · 4 dormant

Recognitions / min

—

+0.0 vs 1m

Avg confidence

—%

band: ±—

DTW cost (median)

—

window 5s · warp ±25%

Seeds streaming

2/3

ESP32-S3 · CSI @ 100 Hz

Live Recognition — DTW band realtime · top match · time-warp ribbon

Recognition field acquiring

Confidence orbit · adaptive radius

LIVE

Confidence

Top match

—

DTW warp ribbon live ↔ template overlay

Cost matrix with optimal warping path · last 1.6s

cost trace

Path cost

—

Warp

—

Length

—

Band

±25%

Top candidates

Ranked by inverse DTW cost

Acceptance—

Thresholdτ ≤ 0.32

Adapt stepη 0.08

Template Constellation — gesture bank arranged by inter-template DTW distance

Template orbit force-directed · DTW metric

Lines: similarity > 0.6 · radius: usage frequency · halo: confidence

primary cohort adapted drifting user-specific

CSI motion flow 3 subcarrier groups · 100 Hz

Channel state information from seed-001 · windowed amplitude

100 Hz

Subcarrier α

—

Subcarrier β

—

Subcarrier γ

—

Enrollment & Adaptation — active learning sessions and DTW cost surface

Active enrollment sessions

Users currently demonstrating new gestures · samples required: 5

Adaptation surface cost gradient · 64×48

DTW distance field for "swipe-up" template · live ridge tracking

η 0.08

Template	swipe-up · v3 (adapted 14m ago)
Cost minima	—
Ridge stability	—
Updates queued	0 · last commit —

Event Pulse — radial timeline of gesture_learned · template_updated · recognition_event

Pulse map last 60s · radial decay

Each pulse fades outward · ring marks every 10 seconds

recognition template_updated gesture_learned

Event stream

Newest first · capped at 60

all recognition template learned

Seeds — paired ESP32-S3 motion sensors streaming CSI

Hardware

Seeds delivering CSI at 100 Hz · 802.11n band

Link health

Concentric ring · RSSI / packet loss / jitter

RSSI

—

Loss

—

Jitter

—

Build & Source

Cog: dtw-gesture-learn

Version	v1.2.0
Size	20 KB
Complexity	Medium
Target	ESP32-S3
Input	CSI motion
Algo	DTW + online adapt
Emits	gesture_learned · template_updated · recognition_event
Source	cognitum-one/cogs/src/cogs/dtw-gesture-learn

Settings — DTW band · acceptance · adaptation · enrollment · hardware · privacy

DTW recognition

Live values — changes propagate to running recognizer within 2s

applied

Sakoe-Chiba band warp window

±25%

Acceptance threshold τ lower = stricter match

0.32

Adaptation rate η online template update step

0.08

Drift sensitivity flag cohort drift

40%

Enrollment

User-side gesture demonstration policy

Samples required per template

Max templates per user

Enrollment timeout

60s

Auto-prune unused templates

remove templates with usage < 5% over 7d

Hardware · ESP32-S3 CSI

Channel state information capture

CSI sample rate Hz · per seed

100Hz

Active subcarriers 3 / 8 enabled

Amplitude smoothing

EWMA · α = 0.18

Phase compensation

remove CFO drift before DTW

Privacy & retention

Local-first · nothing leaves the appliance unless toggled

Raw CSI retention

Anonymize user labels

strip user ID from exported template bank

Share aggregate stats

cohort metrics only · no per-user data

Encrypted template bank

AES-256 at rest · key in TPM