Recurrent Neural Network for Modelling a Contractive Soft Actuator

Abstract:
Robotic stomach simulators are designed and employed to model stomach motility and gain insights into the digestive process. These simulators consist of multiple nonlinear soft actuators. Their modelling enhances the controllability and optimisation possibilities and shows their geometry variations and deformations. By leveraging a dynamic model, it becomes possible to design time-efficient model-based controllers that handle the nonlinearities of the soft robot. However, the modelling of soft actuators is challenging due to their infinite number of degrees of freedom. In this paper, we propose a novel approach that maps the dynamics of a contractive soft actuator to a neural network. The short calculation time of neural networks combined with their ability to model complex relationships make them a promising approach for controlling soft actuators. Since the contractive behaviour of the soft actuator is difficult to capture, training data can be generated using a physical simulation model. The results demonstrate that a simple recurrent neural network can model similar dynamics to a physical simulation model with significantly reduced calculation times. Since a fast and precise computer model is necessary to capture the complex dynamics of the soft actuators accurately for control purposes, the recurrent neural network is promising.