Incremental Learning of EMG-Based Control Commands by Introducing Closed Loop Training and Validation

Abstract

EDAN (EMG controlled daily Assistant), is a fully integrated wheelchair-based manipulation aid. It can be controlled by a joystick, or via electromyographic (EMG) signals and is designed to perform activities of daily living, supported by shared control capabilities in combination with whole-body impedance control. EMG helps to create an interface to generate muscular activity to command assistive devices in health care applications. At present, surface EMG decoding is impelmented using Gaussian Process Regression. The current state of the art provides Open Loop training and testing where the system records EMG data as the user tries to follow the motion based on Open Loop imagery of training cursor moving along the cardinal axes, followed by training and calibratng the model based on this data. The user provides simple muscular input with no feedback to verify user movement intention. Open Loop training also faces the problem of degradation in signal data over time and leads to noises and non-stationaries. The process labels are generated only based on EMG data along the mimicked direction of movement und user intention is not discerned. To overcome this problem, the thesis introduces a supervised Closed Loop training procedure intending to update and recalibrate the predictive EMG model when necessary, based on estimated user intention. The work improves upon the control and accuray the user has while using the EMG-based interface, and subsequent results and comparisons are thus provided.