Prediction of soft manipulator's dynamic position using long-short-term memory neural networks

Official URL: https://dx.doi.org/10.1109/SIU66497.2025.11111784

Soft robotics has recently emerged as a rapidly growing and promising paradigm within the broader robotics domain. A few features that distinguish soft robotics from classical rigid robotics are the former's utilization of low-stiffness materials, bio-mimetic designs, compliant nature, suitability for unstructured environments, and higher degrees of freedom. Despite these unique characteristics, modeling soft robotics using physics-informed methods can pose several challenges due to the innate attributes of the robots' deformable materials. This can result in complex behavior due to nonlinearity in the robots' dynamics. One particularly appealing approach to overcome such obstacles is employing data-driven models based on machine learning techniques. In this paper, we use time-series models such as LSTM to predict the 2D position of the tip of a soft manipulator. We compare the results against a baseline polynomial regression and ARIMA models.