Using deep learning models to accelerate the design of soft robots with genetic algorithms

Abstract

The motion of soft robots is intrinsically linked to their shape. Design of soft robots is then still a challenge, with a very large design space to explore in terms of possible shapes. Generative methods can be of interest, but they require intensive use of robots motion prediction. We assess the interest of using deep learning models to accelerate the synthesis. The case of pneumatically-actuated structures is considered. We show first that a Resnet model can accurately describe the structure motion after learning on a dataset based on finite element simulations. Second, we show that the model accuracy can be maintained during a synthesis, outside the initial dataset, using transfer learning.