Inverse dynamics learned gait planning of an exoskeleton to negotiate uneven terrains using neural networks

Abstract

This paper deals with a methodology to create a mathematical model in order to analyze a novel design of a full-body powered pseudo-anthropomorphic exoskeleton (32 DoF). The expressions for torque used to generate a training data-set of kinematic and kinetic parameters of the system, are determined using Lagrangian and Denavit-Hartenberg joint parameters; inclusive of reaction force on the lower limbs by the upper limbs of the exoskeleton. This training data-set is used to train a multilayer feed-forward neural network for generation of the instantaneous torque values for joint actuation; the network is trained using Levenberg-Marquardt algorithm (LMA) to solve the mean squared deviation curve fitting. This method can serve as a replacement for the inverse dynamics model deployed to solve torque calculation problems within a fraction of second; and is tested by comparison of the output torque of lower torso with that of sample gait cycle data. This method is implemented for gait planning of the exoskeleton to traverse uneven terrains, i.e., staircases, sloping surfaces and ditches.