Hardware components | ||||||
 |
| × | 1 | |||
 |
| × | 5 | |||
Presentation for in-depth explanation
https://docs.google.com/presentation/d/1zE8nWY0bpSRBJ2TIWo7ft2jMeYyQw4ZU0FRp0BMwnxY/edit?usp=sharinghttps://docs.google.com/document/d/1OIRwq_jamqlTe98g70cl0Gf0FeHmRAnJ3TDWw1jMYuo/edit?usp=sharing It began when I got a call from my uncle, who lives halfway around the world in India. He told me how his daughter—my young cousin—struggled with daily tasks because her prosthetic arm could only open and close in a single rigid motion. Though he hadn’t yet tried my design, he trusted my work, so I dove into creating a solution that would bring intuitive, multi-gesture control to his family.
2. Defining Success: Goals Across ContinentsI needed clear targets, both for myself and to show my uncle that this system would work:
Gesture Recognition: At least six distinct motions (open hand, fist, pinch, thumb-to-middle, thumb-to-ring, wrist rotation) with over 80% accuracy.
- Gesture Recognition: At least six distinct motions (open hand, fist, pinch, thumb-to-middle, thumb-to-ring, wrist rotation) with over 80% accuracy.
Instant Response: Under 400 ms end-to-end latency so it feels like a natural extension of the body.
- Instant Response: Under 400 ms end-to-end latency so it feels like a natural extension of the body.
Comfort & Portability: Lightweight dry electrodes and a smartwatch-style form factor.
- Comfort & Portability: Lightweight dry electrodes and a smartwatch-style form factor.
Affordability & Accessibility: A total parts cost under $120 and fully open-source designs for anyone to build.
- Affordability & Accessibility: A total parts cost under $120 and fully open-source designs for anyone to build.
I revisited core electromyography (EMG) principles:
Signal Noise: Skin impedance, movement artifacts, and muscle crosstalk.
- Signal Noise: Skin impedance, movement artifacts, and muscle crosstalk.
Electrode Trade-Off: Wet gel pads give clean signals but irritate skin; dry electrodes are more user-friendly.
- Electrode Trade-Off: Wet gel pads give clean signals but irritate skin; dry electrodes are more user-friendly.
User Variability: Every person’s EMG signature is unique, so a calibration routine is crucial.
- User Variability: Every person’s EMG signature is unique, so a calibration routine is crucial.
I sketched a board with two EMG channels, each using:
An INA333 amplifier for high common-mode rejection.
- An INA333 amplifier for high common-mode rejection.
A 4th-order band-pass filter (20–450 Hz).
- A 4th-order band-pass filter (20–450 Hz).
A 12-bit ADC interface to the microcontroller.
- A 12-bit ADC interface to the microcontroller.
After three PCB revisions—optimizing trace layouts and switching to gold-plated dry pads—I had a slim, 12 g sensor board costing $5.20 in parts.
A custom 3D-printed wrist cuff houses the sensor board and detachable battery pack. I tested multiple mounting approaches before settling on a comfortable silicone liner and adjustable Velcro strap.
5. Collecting Data: Proof of ConceptSince my uncle couldn’t test in India yet, I ran sessions on myself and four volunteers:
Place the sensor over flexor digitorum and extensor digitorum muscles.
- Place the sensor over flexor digitorum and extensor digitorum muscles.
Perform six gestures: open, fist, pinch, thumb-middle, thumb-ring, wrist rotate.
- Perform six gestures: open, fist, pinch, thumb-middle, thumb-ring, wrist rotate.
Record 60 repetitions each (2 s hold, 2 s rest).
- Record 60 repetitions each (2 s hold, 2 s rest).
This yielded 2, 880 labeled EMG segments—enough to train a robust model.
Each 200 ms window produced 10 features:
RMS amplitude per channel
- RMS amplitude per channel
Zero-crossing count
- Zero-crossing count
Spectral centroid from a 64-point FFT
- Spectral centroid from a 64-point FFT
Wavelet energy in the 100–200 Hz band
- Wavelet energy in the 100–200 Hz band
A lightweight TensorFlow network:
Input: 10 features
- Input: 10 features
Hidden Layers: 64 ReLU units, then 32 ReLU units with 30% dropout
- Hidden Layers: 64 ReLU units, then 32 ReLU units with 30% dropout
Output: 6-way softmax
- Output: 6-way softmax
On validation, accuracy reached 90.5%. With session-based cross-validation (simulating a new user or day), it averaged 83%, exceeding the 80% goal.
I ported the model to a PSoC™ 6 AI microcontroller using TensorFlow Lite for Microcontrollers:
Quantized weights (8-bit) fit in 3 KB flash.
- Quantized weights (8-bit) fit in 3 KB flash.
Inference took 4.2 ms on the on-chip AI engine.
- Inference took 4.2 ms on the on-chip AI engine.
End-to-end latency—from EMG sampling to servo command—averaged 280 ms.
- End-to-end latency—from EMG sampling to servo command—averaged 280 ms.
In Fusion 360, I modeled a tendon-driven, 3D-printed hand:
Snap-fit servos for easy maintenance.
- Snap-fit servos for easy maintenance.
Internal channels guiding braided fishing-line tendons.
- Internal channels guiding braided fishing-line tendons.
Lightweight PLA shell for each phalanx.
- Lightweight PLA shell for each phalanx.
After 40 hours of print time, I assembled five micro-servos, the PSoC board, and a 1, 000 mAh Li-ion battery into the wrist cuff.
I tested the fully assembled arm on myself:
Coffee Mug Grip: 95% success on the first try.
- Coffee Mug Grip: 95% success on the first try.
Piano Key Strike: Individual keys struck accurately.
- Piano Key Strike: Individual keys struck accurately.
Shoelace Tie: Completed unassisted after two tries.
- Shoelace Tie: Completed unassisted after two tries.
Knowing it works on me, I’m ready to ship the kit to my uncle in India for his family’s first hands-on trial.
10. Reflections & Next StepsCalibration Routines: Streamlining setup for new users.
- Calibration Routines: Streamlining setup for new users.
Force Sensing: Adding fingertip FSRs for variable grip strength.
- Force Sensing: Adding fingertip FSRs for variable grip strength.
Adaptive Learning: Online updates to tailor to each user’s EMG signature.
- Adaptive Learning: Online updates to tailor to each user’s EMG signature.
Community Workshops: Enabling families worldwide to build and customize their own prosthetics.
- Community Workshops: Enabling families worldwide to build and customize their own prosthetics.
Comments