Open Loop Dynamic Trajectory Tracking Control of a Soft Robot using Learned Inverse Kinematics combined with a Dynamic Model

Abstract

As new soft robotic applications emerge, control requirements increase. Therefore, precise control methods for soft robots are needed. The main challenge in controlling soft robots is that soft robots are often underactuated and redundantly actuated at the same time. In addition, modeling is usually difficult due to large elastic deformations, unknown material parameters, and manufacturing inaccuracies. In soft robotics, so-called kinematic controllers, which neglect the dynamics of the system, are mainly used. In particular, data-driven controllers are very popular. However, more advanced applications of soft robots require increasingly faster and more accurate movements. Here, kinematic controllers are not sufficient anymore. A direct extension of existing data-driven kinematic controllers to dynamic control is usually not practical due to the huge amount of training data required. This paper presents a new open-loop dynamic trajectory tracking control of a redundantly actuated soft robot. A combination of a kinematic data-driven controller based on neural networks and a dynamic model-based control approach based on model inversion with the servo-constraints approach is used. This combined approach preserves the advantages of learning-based kinematic controllers for the dynamic control of soft robots while keeping the amount of training data required low. Experimental results show the strength of this approach.