Conditional Generative Adversarial Network-based Finger Position Estimation for Controlling Multi-Degrees-of-Freedom Myoelectric Prosthetic Hands

As most existing single-degree-of-freedom myoelectric prosthetic hands cannot perform complex movements, multi-degrees-of-freedom prosthetic hands have been made commercially available owing to the development of robotics technology. However, their control strategies are either weak in intuitiveness or require considerable training data. To perform complex movements with minimal training data, the movements should be represented as complex motions with primitive movement bases. This warrants the estimation of finger joint angles rather than the movement patterns. In this paper, we propose a control method for estimating the finger joint angles based on a conditional generative adversarial network. Surface electromyography signals and finger angles were simultaneously acquired from four subjects using three electrodes and a leap motion. The results indicate that the finger position can be accurately estimated using the corresponding electromyography. In the future, the proposed control strategy can be used to estimate unlearned finger positions.