Proportional myoelectric control of multiple degrees-of-freedom (DOF) in active finger joints is important in replicating dexterous hand motion in robotic prostheses and orthoses. However, this is still difficult to achieve as current myoelectric control strategies often require the separate control of each joint and do not consider the strong correlations that exist between these joints. To address this problem, we propose using a shared low-dimensional encoding based on synergistic models to represent both the high-DOF finger joint kinematics and the coordination of muscle activities taken from electromyographic (EMG) signals in the forearm. A Bayesian Gaussian Process Latent Variable Model (GPLVM) is used to learn a shared latent structure model that not only allows the automatic selection of the dimensionality but also captures the information variance, both shared and specific to the observed EMG and hand kinematic data.
In the first part of this presentation, we show how using features obtained from an EMG-to-Muscle Activation is not only suitable for continuous and simultaneous estimation of finger kinematics, but is also shown to perform better than time-domain based features. In the next part, we demonstrate that the proposed shared model is able to reconstruct the full-joint continuous finger kinematics from muscle activation inputs, whose results are inferred from a shared latent manifold. We show that the proposed method outperforms commonly used simultaneous regression and linear dimensionality reduction methods.
Our proposed approach not only presents a viable solution for a myoelectric strategy for handling high-DOF finger control, but also aims to open new avenues in developing novel myoelectric interfaces for synergy development and long-term control and adaptation.