Dynamic Eye-in-Hand Visual Servoing with Neural-Adaptive Backstepping

Abstract

This thesis investigates eye-in-hand visual servoing, where a camera on the robot arm provides information for the motor-control feedback loop. Current methods use a dual-loop strategy, where the outer-loop uses the visual servo error to compute desired joint velocities, while an inner-loop accomplishes the tracking. Since it is difficult to establish global stability with this strategy, this thesis instead investigates backstepping control. This provides a guarantee of uniformly ultimately bounded signals and explicitly accounts for the coupling between outer and inner loops. First a method with knowledge of the feature Jacobian is developed, then it is extended to an adaptive method that uses supervisory estimates of the feature Jacobian to maintain stable adaptation. The methods are further extended to the visual servo control of n-link robots, multiple features using a switched controller, and visual tracking. Non-linearities in the system are approximated using the computationally-efficient Cerebellar Model Articulation Controller neural network.