Optimization of neural network hyperparameters for modeling of soft pneumatic actuators

Abstract

Especially the field of medical robotics is faced with the challenge that contact between a robotic structure is not only tolerated but at the very core of the application, for example in minimally invasive surgery. Therefore, high demands regarding safety need to be met when operating a robotic structure in such a delicate environment. Soft material robotic systems offer the potential to surpass their rigid counterparts in this area due to their material inherent compliance and adaptability. Nonetheless, soft structures can also exhibit challenging characteristics like highly nonlinear deformation. This effect is often neglected when setting up deformation models. Recent approaches tackle this challenge by applying machine learning methods to learn the nonlinear behaviour. In previous research, we gained promising results from learning the kinematics of a soft pneumatic actuator (SPA) via a feedforward artificial neural network (ANN). To overcome a trial and error approach when designing an ANN, in this article we introduce Bayesian optimization to find suitable hyperparameters. This way an ANN architecture is found that exceeds the accuracy of our previous studies by more than a factor of 20 [9].