Speed-Adaptive Gait Phase Estimation Using Dynamic Movement Primitives

Abstract

This study presents a method for adaptive human gait phase estimation using adaptive dynamic movement primitive(aDMP) based on thigh angle and its integral. The phase variable represents the user's gait phase progress from 0 to 100, continuously and monotonically. As robots cannot directly recognize the user's intuition and condition, this phase variable delivers important information to them. Previous research has focused on estimating the phase variable using user kinematic data, but this paper proposes an adaptive approach that takes gait patterns from five gait speed groups. The aDMP model is generated through locally weighted regression(LWR) based on the user's kinematic data, which allows for a more precise and user-friendly estimation of the phase variable. Using the error between the estimated thigh angle and measured thigh angle, check whether the measured phase was predicted slower or faster than the true value and adjust it. The model was trained and validated with a multimodal dataset of human gait at different walking speeds and showed improved R square value and root mean square error between an estimated phase variable trajectory and a linear line of true value phase variable over one gait cycle (ground truth of phase variable). This method has the potential to enhance the control of lower limb assistant robots such as exoskeletons and prostheses.