Design and intelligent control of a low-cost, sEMG-driven, 3D-printed robotic hand prosthesis: A three-tier benchmarking framework with novel robustness metrics

Upper-limb amputation and neuromuscular disorders impose functional limitations on millions of people worldwide. Commercially available myoelectric prostheses, controlled by surface electromyography (sEMG) signals from residual limb muscles, cost approximately USD 75,000–100,000 (about INR 62–83 lakh), limiting accessibility in resource-limited settings. We present a 3D-printed, tendon-driven research hand prototype with a component cost of USD 122–232 (about INR 10,100–19,300), built around a single MyoWare 1.0 sensor and an Arduino Nano, with gesture classification performed on a tethered host PC.

Five classifiers—Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), one-dimensional Convolutional Neural Network (1D CNN), and Long Short-Term Memory (LSTM)—were benchmarked across synthetic, Ninapro DB5, and live recordings of three gestures from ten able-bodied participants. Under subject-independent Leave-One-Subject-Out (LOSO) cross-validation, RF achieved the highest live accuracy (68.7%). Friedman and Nemenyi testing showed that RF was significantly better than SVM, KNN and LSTM (p < 0.05), while its advantage over the 1D CNN (60.7%) was not significant. Permutation importance identified waveform length, RMS and MAV as the most informative features.

We further propose two metrics: the Robustness Degradation Index (RDI), which quantifies the synthetic-to-real performance gap and was lowest for RF (43.7%), and a Cost–Performance Index (CPI), which was 13.77 for the prototype. The hand achieved an end-to-end latency of 285 ms and reliably held objects of up to 200 g. These results demonstrate well-above-chance three-class control with minimal hardware in able-bodied users and motivate evaluation with trans-radial amputees.