Inverse Kinematics Modeling of Soft Manipulator Based on Optimized Particle Swarm Optimization-Back Propagation Neural Network

Abstract

In recent years, soft manipulators have attracted much attention in the field of robotics research due to their favorable flexibility, good environmental adaptability, and safety in human interaction. Despite this, the accurate and fast modeling of soft manipulators remains a difficult task subject to the uncertainties in their highly nonlinear. This article proposed an accurate, fast, and end-to-end modeling method based on data driven particle swarm optimization-back propagation neural network (PSO-BPNN). This network is targeted at the end-to-end tabular data and can easily realize accurate 3-D control of the soft manipulator. First, a simple, fast, and accurate inverse kinematic model of soft manipulator was generalized based on the BPNN, and then, a PSO algorithm was employed to optimize the BPNN, establishing an accurate and faster convergence inverse kinematic model mapping between 3-D end tip coordinates, and air pressures of three chambers applied in the soft manipulator. In addition, an external load was introduced to verify the data generalization of the model. Through trajectory tracking experiments, this method is proven to achieve good performance, which has the lowest average error of tip position (0.5590 and 1.2217 mm with load), best converge speed (converge at 221 epoch and converge at 256 with load), fastest training speed per epoch (0.0104 and 0.0242 s with load), the lowest standard deviation (0.1029) and the second lowest standard deviation (0.1199) with load, and the lowest average relative error (1.09% and 2.38% with load). Results demonstrated that the proposed model could effectively improve the training speed, end tip control accuracy, and data generalization of the soft manipulator.