Homography-based uncalibrated visual servoing with neural-network-assisted robust filtering scheme and adaptive servo gain

In this paper, a homography-based uncalibrated visual servo system with neural-network-assisted robust filtering scheme and adaptive servo gain is presented. This system employs a homography-based task function which is robust to image defects. A neural-network-assisted robust filtering method which combines the new form of smooth variable structure filter (SVSF) with a radial basis function (RBF) neural network is proposed to estimate the total Jacobian between task function and robot joints. The RBF neural network in this filtering method plays the role as a corrector to further improve the accuracy and compensate the interference caused by the measurement errors of image features. The controller that directly controls the robot joints based on the estimated total Jacobian is designed for achieving the robustness to robot parameters errors. By adopting this filtering scheme, the visual servo system shows better accuracy and convincing anti-interference ability. In addition, a novel Q-learning strategy is introduced for this homography-based system to make adaptive adjustment for the servo gain. This adaptive gain enables the system to achieve a faster convergence speed while ensuring the accuracy. Several simulations and experiments have been carried out to verify the performance of the proposed system.