In order to implement the automatic visual guidance of 6- DOF robots in intelligent welding system,a new control model of uncalibrated visual servoing is proposed.In this model,multiple support vector regression (SVR) machines are used to estimate the Jacobian matrix of images,and the nonlinear mapping between the image feature and the robot joint angle is constructed.Thus,an uncalibrated visual servoing of welding robots can be carried out.The authors also put forward an image Jacobian matrix expression with Gaussian kernel and conduct visual guidance experiments of a welding robot with both eye- in- hand and eye- to- hand camera configurations by using the SVR- Jaco-bian estimator automatically trained under the control of computer.The results indicate that the visual guidance of welding robots using the SVR- Jacobian estimator converges at the desired goal and is superior to the conventional Broyden- estimator because it helps to acquire high dynamic response quality.