Unconstrained and constrained estimation of a linear EMG-to-force mapping during isometric force generation

Abstract

EMG-driven robotic devices require the estimation of the forces exerted by the human operator from muscle activity. Approximating the relation between EMG and force with a linear mapping may be accurate enough for numerous real-time applications, such as controlling exoskeletons or prostheses. However, while a linear mapping from the EMG activity to endpoint force may be identified by minimizing the error without any constraint, introducing some constraints may be helpful to determine a mapping which is more anatomically accurate. The presence of noise and the muscle redundancy may introduce errors in the estimation achieved by the unconstrained optimization. Contrarily, anatomical constraints, estimated from an accurate musculoskeletal model, would limit the effect of noise, but they would increase the algorithm complexity and its computational costs. This study compares the two algorithms (unconstrained and constrained) for the estimation of the forces exerted by a human participant from the EMG activity of several upper limb muscles. The two algorithms were tested on data collected during an isometric force generation task performed during multiple sessions spanning two days. Accuracy and consistency across sessions of the reconstructed forces were assessed. Data showed that the unconstrained algorithm allowed for a better reconstruction of the exerted forces, but the constrained mapping is more robust across sessions. Further studies will investigate which of the two algorithms reconstruct a mapping perceived by the participants as more natural during EMG-driven control.