Improving Robot Precision Positioning Using a Neural Network Based on Levenberg Marquardt-APSO Algorithm

Abstract

This paper proposes a robot calibration method that uses an extended Kalman filter (EKF) and a neural network based on Levenberg-Marquardt combined accelerated particle swarm optimization (LMAPSO) to improve the accuracy of the robot's absolute position. After the EKF optimizes all geometric parameters, the robot position still contains non-geometric errors due to joint clearance, gear backlash, and link deflection that are impossible to model. Therefore, an artificial neural network model (ANN) is designed to compensate for these un-modeled errors. The Levenberg-Marquardt combined accelerated particle swarm optimization (LMAPSO) provides a robust optimization search algorithm to optimize the weight and bias of the neural network based on the training set. An experiment on a five-bar parallel robot shows that geometric and non-geometric calibration reduced the maximum absolute position error from (1.548 to 0.045) mm. The experimental results demonstrate the proposed calibration method's effectiveness with the robot's absolute position accuracy improving by 98%.