For an uncalibrated camera and unknown target's 3D model, this paper proposes a model-free visual servoing method for robot positioning. By introducing the state space, the state and observation equations are established on the robot “vision space -movement space” nonlinear mapping of the Jacobian matrix, and the neural network unite Kalman filtering algorithm is presented, while the neural network is used to compensate the approximation error of the dynamic system and paremeter estmation error, for the optimal prediction of the Jacobian with the least mean square error. Then a model-free visual servoing control scheme is further constructed using the Lyapunov stability criterion, which can, avoid the problems of camera calibration and target modeling. Finally, the positioning comparision experiments condlucted for the “eye-in-hand” six degrees of freedom robot show that the trajectories of the image features are smooth and stabilized in the field of view of the canera, and the movement of the robot end-effector is stable with no vibration and retreat in Cartesian space, and the end-efector positioning accuracy is within 10 pixels.