Object Retrieval Strategy in Unstructured Environments Using Active Vision and a Medium Complexity Gripper
Depth information from a mobile laser stripe scanner mounted on a PUMA560 robot is used for simple thresholding by z-height and region growing. Superquadric surfaces are then fit to the regions segmented. This data reduction to three axis parameters, three Euler angles and two squareness parameters allows grasp planning using the PENN hand medium complexity end effector. Additionally, in order to take into account the spatial relationships between objects, they are grouped according to a nearest neighbor measure by distance between centroids in the χ-γ plane and also by height. The convex hull of the groups is then computed using Graham's method. The convex hull object list permits objects with the best clearance for grasp to be identified, thus reducing the possibility of unwanted collisions during the enclosure phase. The geometric properties of the object are then used to determine whether an approach parallel or normal to the plane of support is necessary. This list of candidate grasps for the object is checked for intersections with the bounding boxes of neighboring objects and the finger trajectories. The most stable collision free grasp preshape which passes the intersection testing is chosen. If no grasp is collision free the next best object in terms of topology is chosen. Height clustering information is used to determine a baseline height for transporting objects in a collision free fashion. By combining these simple strategies of favoring objects at the exterior of groups and tall objects for initial grasping and removal, the chances for successful task completion are increased with minimal computational burden.