Vision-based control via navigation functions

Contemporary commercial robots generally lack external sensing and can operate in only the most controlled environments (e.g. an assembly line). The environments are crafted to enable industrial robots to achieve tasks in a repeatable fashion. Recent advances in computing and sensing have dramatically improved robot sensory intelligence, but there remain tremendous algorithmic challenges to transforming visual sensation into meaningful actions. Vision-based control provides a rich source of such algorithms.; A classic problem in vision-based control is to position a fully actuated rigid body by reducing the “perceived” error on the image plane of a computer vision system. Most visual servoing systems to date suffer from several drawbacks. For example they are kinematic, and hence low performance. Moreover, they cannot keep features from self-occluding or leaving the field-of-view, and hence are not safe. Finally, convergence from a large set of initial conditions is seldom guaranteed.; This dissertation presents a framework to solve the problem of visually servoing a fully actuated rigid body which affords several enhancements over the state-of-the-art, by recourse to a special class of artificial potential functions called navigation functions. NF-based visual servos dynamically converge to a goal from any visible initial position while maintaining full view of all the feature points during transients. The solution incorporates three components: (1) A model space for the “occlusion-free” workspace. (2) A change of coordinates from image to model coordinates. (3) A navigation function for the model space. These three components are assembled to construct a global, occlusion-free, second order “mass-spring-damper” type of controller.; This thesis presents three specific applications of the proposed framework. First, we demonstrate an image-based controller for a 3DOF planar rigid body with three point features in view of a 1D pinhole projection camera. Second, we present an imaged-based controller for a 3DOF spatial robot (the Buehgler Arm) endowed with an oriented feature point in view of a spatial perspective projection camera. Finally, a 6DOF fully actuated rigid body with a planar array of features is controlled via a tasked-based visual servo. Experimental results for the Buehgler Arm and a 6DOF commercial robot validate the practical effectiveness of the algorithms.