A Small Opening Workspace Control Strategy for Redundant Manipulator Based on RCM Method

This article presents a method for redundant manipulators working in the small opening workspace without collision. To achieve this aim, we began with an improved incremental radial basis function (RBF) neural network (RBFNN) method to estimate manipulator dynamic parameters, and then with the help of Lyapunov function, the control strategy could converge within a fixed time. To avoid the collision of workspace and constrain the posture of end-effector, we proposed a safety region convolutional neural network (CNN) method adapted with the Remote Center of Motion method inspired by the minimally invasive surgical manipulator. Torque observer is also implied to estimate the external force to resist external interference. Experiments on Baxter, a seven-degrees of freedom (DoF) redundant manipulator, demonstrate the feasibility of the proposed control strategy.

I. Introduction

In recent years, with the continuous research on neural networks, researchers have made remarkable progress in manipulator control with uncertain kinematics parameters or dynamics model [1]–[3]. Among all, the radial basis function (RBF) neural network (RBFNN) method has been proven that RBFNN is a reliable way of estimating unknown parameters of manipulator. As shown in [4] and [5], RBFNN was applied to the parameter estimation of the robot with uncertain parameters, and RBFNN was trained based on the trajectory tracking error to ensure the stability of the robot system. The Lyapunov function proved that the position tracking error would converge to the specified boundary within the wired time. The simulation results show that the control method based on RBFNN parameter identification has a better stability. However, in traditional RBFNN, the parameters of neural nodes need to be designed in advance. In this case, the parameter estimation ability of neural network will be affected if the neural network parameters were not set properly, leading to the deterioration of the control effect of the manipulator. Researchers proposed incremental learning to improve traditional RBFNN method with adaptive node parameters designed [6].