Black-box modeling of a 2-DOF manipulator in the image plane using recurrent neurofuzzy networks

Visual servoing, using the visual measurements direct in the control loop, is a problem that in recent years has grown in interest. One of the main problems involved in these systems is that, while the robot manipulator has a well known model and identification methods have been available, the vision system introduces a nonlinear transformation and modifies the dynamics as seen in the image plane. In this work we present a black-box modeling of a 2-DOF planar robot in the image plane using recurrent neural networks with output feedback. The input for the identification is the voltage fed into the actuators, and the output is the angle of each joint as seen in the image plane. The learning law is inspired by adaptive observer theory, and proven to be convergent in the parameters and stable in the Lyapunov sense. Simulation and experimental results are shown in order to validate the presented modeling, using only input-output data and no knowledge on the manipulator dynamics, forward kinematics and camera model.