Calibration of the arc-welding robot by neural network

Based on the analysis of the robot calibration methods, this paper presents the neural network calibration method and calibrates an arc-welding robot with two approaches. The first one requires just the nominal model of the robot to be calibrated. This is a peculiarity of the proposed method. It reduces the pose errors to 1/5 of initial values. The second variant combines a BP network with an already calibrated parametrical model of the robot. This is a high performance solution. The presence of the neural network permits the compensation of several effects, even those not considered by the parametrical model. Calibration results are compared with those obtained by traditional parametric methodologies. Simulation results show that this method improves the calibration effect further and achieve better calibration effect.