Movement Onset Detection and Target Estimation for Robot-Aided Arm Training

Abstract

Abstract This paper presents a motion intention estimation algorithm that is based on the recordings of joint torques, joint positions, electromyography, eye tracking and contextual information. It is intended to be used to support a virtual-reality-based robotic arm rehabilitation training. The algorithm first detects the onset of a reaching motion using joint torques and electromyography. It then predicts the motion target using a combination of eye tracking and context, and activates robotic assistance toward the target. The algorithm was first validated offline with 12 healthy subjects, then in a real-time robot control setting with 3 healthy subjects. In offline crossvalidation, onset was detected using torques and electromyography 116 ms prior to detectable changes in joint positions. Furthermore, it was possible to successfully predict a majority of motion targets, with the accuracy increasing over the course of the motion. Results were slightly worse in online validation, but nonetheless show great potential for real-time use with stroke patients.