Model predictive control for intelligent lower limb robotic assistance

Abstract

Loss of mobility or balance resulting from neural trauma is a critical consideration in public health. Robotic exoskeletons hold great potential for rehabilitation and assisted movement, yet optimal assist-as-needed control remains unresolved given pathological variation among patients. This thesis develops, simulates and experimentally evaluates a Model Predictive Control (MPC) architecture for lower limb exoskeletons. It uses a Fuzzy Logic Algorithm (FLA) to bridge the gap in human-robot synergy by identifying modes of assistance (passive, active-assist, and active-resist) based on human involvement.

Muscle activity, mapped through Electromyography (EMG) or Mechanomyography (MMG), is known to precede the onset of human joint torque for modelling and prediction of human-exoskeleton movement. This thesis investigates the complementary and competing benefits of MMG and EMG as a means of human joint torque prediction. A qualitative and quantitative comparison is presented using three biomechanics agnostic machine learning approaches: linear regression, polynomial regression, and neural networks. At the expense of training and implementation complexity, the neural network models performed best, achieving a normalised mean absolute error of 0.063 with MMG and 0.048 with EMG.

The controller is evaluated in hardware with three subjects on a seated 1-Degree of Freedom (DOF) knee exoskeleton tracking a sinusoidal trajectory with human relaxed, assistive, and resistive. The controller is also demonstrated with one subject assisting the swing phase of walking. Human joint torque is predicted using a linear regression model with EMG signals to inform the MPC and assistance mode selection by the FLA. Experimental results show quick and appropriate transfers among the assistance modes and satisfactory assistive performance in each mode. Results illustrate an objective approach to lower limb robotic assistance through on-the-fly transition between modes of movement. This provides a new level of human-robot synergy for robotic rehabilitation, mobility assistance, and human-robot interaction more broadly.