Automated grasping of objects of unknown geometry a priori has applications in many industries such as clearing a mine shaft after blasting, agricultural applications such as fruit and vegetable handling, and many roles in the service industry such as fetching items for a handicapped individual. In these roles the system environment is highly unstructured, and the robot must be able to react to different types of objects needing to be grasped. In this thesis a vision guided robotic grasp planner for unstructured environments is presented. An evaluation method for robotic grasping involving two distinct sets of objects is also presented. Both the grasp planner and evaluation metric are evaluated by experimentation using an articulated robotic arm with an eye-in-hand video camera, line laser, and pneumatic gripper. Multiple grasping experiments were performed with the objects in random poses on a modified tabletop deemed the playfield that did not allow objects to rest flat.

The grasp planner focused on using a created model of the object from camera observations using silhouetting and line laser data. The object model and its computed convex hull were used to evaluate and select a single facet and point creating a grasping pair for the pneumatic gripper jaws. The grasp was attempted and then evaluated using a secondary camera and the developed evaluation method.

Grasp success rates ranged from 80.30% (Rectangular Block on playfield 137 attempts) to 97.69% (Hexagonal Nut 173 attempts), with a mean grasp computation time for the hexagonal nut of 0.57s.