Dynamic Eye-in-Hand Visual Servoing using Adaptive Backstepping with a Supervisory Feature Jacobian

This paper develops two backstepping controllers for the eye-in-hand image-based visual servoing setpoint regulation problem. The first uses knowledge of the feature Jacobian matrix to implement a resolved velocity control in the first step of backstepping. The second provides an adaptive modification to the first, with initial estimates of the feature Jacobian supervising the training of an adaptive feature Jacobian matrix. Experiments demonstrate both methods outperform a classical inner/outer loop control strategy. The backstepping controllers show smoother image trajectories, require less control effort, and have less steady-state error. With the proposed supervisory parameter update, the adaptive method performs well irrespective of the accuracy of the feature Jacobian estimate.