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

Abstract

Soft bioinspired manipulators have a theoretically infinite number of degrees of freedom, providing considerable advantages. However, their control is very complex, making it challenging to model the elastic elements that define their structure. Finite elements (FEA) can provide a model with sufficient accuracy but are inadequate for real-time use. In this context, Machine Learning (ML) is postulated as an option, both for robot modeling and for its control, but it requires a very high number of experiments to train the model. A linked combination of both options (FEA and ML) can be an approach to the solution. This work presents the implementation of a real robot made up of three flexible modules and actuated with SMA (shape memory alloy) springs, the development of its model through finite elements, its use to adjust a neural network, and the results obtained.