Quantifying the performance of Deep Neural Networks in predicting curvature and force output response of a Pneumatic Soft Actuator

Soft robotics promises new opportunities for solving problems that were limited by rigid robots due to their compliant physical structure. Pneumatic Soft Actuators (PSA) are a class of soft robots that have gained popularity due to their cost-effective manufacturing and high force-generating capability. Due
to their highly nonlinear dynamics, accurate prediction of their response to input pressure (curvature) as well as force output is challenging. Artificial Neural Networks were employed to this effect. This work shows a comparative analysis of three variants of the recurrent neural network used in this paper for predicting the behavior of soft robotic actuators. Each network was trained
using the same data set to predict the position, curvature, and applied force by the soft robotics on an external force. The result shows the different performance of the three models, evaluating them using Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The Long Shortterm Memory (LSTM) model outperformed both the Gated Recurrent Unit (GRU) and Recurrent Neural Network (RNN) models, producing the lowest error in the three metrics.