Cerebellum-Inspired Model Predictive Control for Redundant Manipulators With Unknown Structure Information

When the structure information of a redundant manipulator is unknown, motion control methods that do not rely on its model are attractive. Due to the numerous advantages of model predictive control (MPC), such as the direct handling of constraints, this article proposes a model-free MPC algorithm for redundant manipulators with unknown structure information. In this article, a cerebellum-inspired model based on the echo state network (ESN) is employed to replace the kinematic model of the redundant manipulator, and an MPC algorithm based on the cerebellum model and neural dynamics (ND) approach is developed. Unlike existing studies, this work considers both performance optimization and system constraints of the redundant manipulator, and can achieve high-precision prediction and tracking by designing an online training algorithm for the cerebellum model. Furthermore, this article proposes an ND-based correction algorithm to modify the prediction model and an ND solver to solve the MPC scheme. Theoretical analyses confirm the convergence of both the ND-based correction algorithm and ND solver. Simulation and experimental results consistently demonstrate that the proposed cerebellum-inspired MPC (CIMPC) algorithm is effective and outperforms comparison algorithms in terms of tracking performance.

I. Introduction

In recent years, robotics technologies have flourished, stimulating various related research in industry and academia. As a typical type of robot, the manipulator plays an essential role in a variety of fields, such as minimally invasive surgery [1], motion assistant devices [2], and so on [3]. Redundant manipulators, i.e., manipulators with more degrees of freedom (DOFs) than required, are popular for their ability to accomplish complex tasks [4]. One class of practical solutions to trajectory tracking problems for redundant manipulators is to deal with their inverse kinematics at the velocity or acceleration layers [5], [6], [7]. Liu et al. design a motion control scheme in the form of quadratic programming (QP) at the velocity layer for the redundant manipulator, where the trajectory tracking task is embedded in the scheme as an equality constraint [5]. A kinematic control scheme at the acceleration layer for the redundant manipulator is presented in [6], which eliminates the joint drift and considers the joint acceleration constraint. However, this type of method essentially realizes the tracking control at the velocity or acceleration layers. Even though some of these methods add feedback compensation at the position layer, they still cannot deal with some sudden problems at the position layer in time [8]. In addition, another defect of this type of method is that it cannot directly deal with joint constraints, and it is necessary to transform joint constraints at different layers to the variable layer (velocity layer or acceleration layer). After undergoing these transformations, the original feasible region of the joint angle may be reduced accordingly [9].