A Learning-based Approach for Adaptive Closed-loop Control of a Soft Robotic Arm

The characteristic compliance of soft/continuum robot manipulators entails them with the desirable features of intrinsic safety, low power to actuation ratio and adaptability to the environment. At the same time, it makes analytical models excessively slow for efficient use in control. We propose a recurrent neural network (RNN) approach for adaptive model based closed-loop control of a continuum robot. First, the forward dynamic model is trained offline on data obtained by continuous motor babbling, learning the relationship between the actuators’ inputs and the robot tip position. Then another network, named inverse model, is used as a closed loop controller and trained by minimizing the forward model tracking error. We show that using the trained controller, the continuum robot is able to track a circular task with a low RMS error, and to maintain its performance under an external load, after updating the networks’ weights