Soft Robot Positioning Using Artificial Neural Network

Abstract

The experiment investigated the performance of an artificial neural network in solving the inverse kinematic problem of a soft robot. For this purpose, a simple soft robot was designed of building blocks, stringed on three rubber hoses, and an actuating system, to provide the hydraulic pressure. An axial extending of a hose, while the others are in the relaxed state, results in bending of the robot. The network was employed, as a black box, to approximate the behavior of the system. In accordance with the purpose, the input consisted of the desired spatial coordinates and the output of the step motor angular displacements. The network was trained and tested using records collected at 200 randomly chosen robot positions. The relative testing error of positioning, about 5%, confirmed a predictable robot behavior. The solution proposed is competitive in terms of simplicity, safety and price of realization. The experiment provided basics for the future research of the design of modular soft robots.

Use of Finite Elements in the Training of a Neural Network for the Modeling of a Soft Robot

“…The number of hidden layers was decided considering the results presented in the work of Kovandžić [ 23 ], where it was shown that a multilayer network compared to a single hidden layer network did not present a significant improvement. For this reason, we decided to use six hidden layers in our network.…”