REAL-TIME HIERARCHICAL CLASSIFICATION OF MOTION INTENT FOR WEARABLE ROBOTICS

Abstract

Robotic exoskeletons and prostheses could address many global health challenges associated with the aging population. Accurate sensing of the pilot’s motion intent is a major challenge in the control of these robotic devices; it allows the robots to deliver assistive forces at the pilot’s joints with the right timing and magnitude.

This thesis presents a novel strategy for classifying locomotion modes for motion intent detection. Based on the idea that human motion can be described at multiple levels of specificity, a deep neural network based hierarchical classifier is developed that performs classification of locomotion mode labels at multiple levels of specificity simultaneously.

The main contributions of this thesis are (1) a dataset of lower limb motion for developing algorithms for lower limb motion intent detection, (2) a novel deep learning based classification strategy for motion intent detection that is evaluated in real-time on an ankle exoskeleton robot and (3) high performance real-time hardware for sensing and real-time inference.