After limb loss, the motor cortex still issues commands for the missing hand, and the sensory cortex still expects to hear back from it. This project decodes that intention with AI, drives a prosthetic hand, and sends real feedback the other way — closing a loop most prosthetics leave open.
4-channel forearm sEMGSynthetic signal · simulated
CH1CH2CH3CH4
01 Signature demo
The closed loop
Select an intended gesture and watch it move through the complete control pipeline — sensing, decoding, actuation, and feedback — as one integrated sequence.
Intended gesture
Loop closed
t = 0ms · UI simulation timing, not measured system latency
Confidence derived deterministically from profile similarity to the five gestures' known activation signatures — a UI illustration, not a live model inference.
Sensor / feedback responseSimulated
Pressure0%
Slip0%
No contact
02 Architecture
Five stages, one continuous loop
Each stage is a real, ordered step in the control pipeline. Select one to see what it does and what's actually implemented.
Sense
Surface electrodes on the residual limb pick up electrical activity from forearm muscles still associated with the missing hand's movements.
Implemented as: synthetic sEMG generator
03 Adaptation experiment
Signal drift & recalibration
Electrode shift and muscle fatigue change a signal's statistical properties after calibration. Drag the slider to introduce simulated drift, then recalibrate — using the real accuracy curves measured in research/emg_simulation.py.
Values interpolated from measured results: without recalibration, accuracy fell from 96.3% to 77.7% across drift 0–0.5. With a brief recalibration step, accuracy at drift 0.5 recovered to 89.4%.
04 Research
Validating the computational pipeline
These results come from a synthetic sEMG signal generator designed to reproduce relevant signal characteristics such as noise, crosstalk, and drift. They validate the computational signal-processing and classification pipeline. They are not human-subject or clinical results.
What this is, honestly: no real EMG hardware was available for this stage, so these results come from a synthetic sEMG signal generator built to match known real-signal characteristics — not a human subject. It validates that the signal-processing and classification pipeline works; it is a computational proof of concept, not a clinical result.
0%
Baseline test accuracy
0% ±1.2
5-fold cross-validation
0
Gesture classes (Rest · Open · Close · Pinch · Point)
Gesture classification — confusion matrix
Synthetic sEMG · 150 held-out test windows · rows = actual, columns = predicted
Accuracy under simulated signal drift
Electrode shift / fatigue simulation · with and without recalibration
No recalibrationWith recalibration
Signal
Synthetic sEMG
4 channels, 1000 Hz, 250 ms windows
Processing
Time-domain features
MAV · WL · ZC · SSC · RMS
Model
Random Forest
200 trees · 75/25 split · 5-fold CV
Control
Gesture → action
Classified intent maps to a simulated prosthetic pose
Feedback
Pressure / slip
Simulated sensor response per gesture, not measured hardware
Adaptation
Drift + recalibration
Parameterized electrode-shift experiment, 0.0–0.5
05 Limitations
What this establishes — and what it doesn't
Stated plainly, because scientific honesty matters more than looking finished.