Control of a Soft Actuator using a Long Short-Term Memory Neural Network

Soft robots offer new opportunities because of their compliant physical structure and their wide range of applications. Currently the development of such robots is hampered by their low controllability. One of the main constituents of soft robots are soft actuators. The aim of this project is to improve the control of a non-linear system, the soft actuator, and its interaction with the environment, by training a long short-term memory (LSTM) neural network to accurately predict the actuator's position in space, its curvature, and the force applied by its end-effector on an external object. The increased performance of the trained network resulted in an error as low as $0.01\pm 0.005\ \mathrm{N}$ in estimating the force applied by the end effector on the external object. The results show significantly superior performance (on the order of 10 times) in the positional and curvature predictions of the LSTM network when using one marker per air-chamber.