Modelling of a segmented pneumatic soft actuator with integrated sensors using neural networks

Abstract

Abstract The recent decade has seen a tenfold increase in publications on soft actuators for medical applications. One of the devices which has sprung from this research is the soft rehabilitation glove. However, the gloves developed as of writing has been limited by the lack of accurate control, in turn stemming from a lack of feedback or closed loop controllers. This project aims to evaluate the feasibility of utilising sensor information & neural networks to develop a model for actuator angle prediction which could be used to close the control loop. During this project a prototype pneumatic soft robotic system was developed, featuring a finger like segmented PneuNet actuator with two integrated flex sensors. Data was then gathered from this actuator which was used for training several neural networks to predict the three joint angles of the actuator. The results of this project can not definitively state the feasibility of using neural networks for joint angle prediction on a PneuNet actuator with integrated flex sensors. However, the results do suggest that with further development and a broader dataset such a system could possibly be realized. The results also suggest that such a model requires several sources of sensor input, including not only current but also past information.