Aitor Aldoma from the Technische Universität Wien (TUW) spent his internship at Willow Garage working on a 3D feature called Clustered Viewpoint Feature Histogram (CVFH) useful for object recognition and pose estimation of rigid objects. The feature was integrated into PCL and the PR2 grasping pipeline to allow grasping of objects in table top scenes (see video).
CVFH is a semi-global feature that can deal with partial occlusions, noise or segmentation artefacts. Moreover, it can be learned on synthetic models of objects represented as CAD models and yet perform very well recognising objects seen with a depth sensor like the Microsoft Kinect mounted on top of the PR2. Because of the invariance of CVFH about the roll of the camera, the Camera's Roll Histogram (CRH) has been proposed to solve this final degree of freedom to provide a full 6DOF pose of the object in the scene.
Being able to learn on CAD models has several advantages. Besides simplifying the training stage as there is no need for calibrated systems, there are several grasp simulators that given a CAD model of the PR2 gripper and a CAD model of an object, can compute off-line several grasp hypothesis for the object. Once the PR2 recognises any of this objects on a real scene and their pose is fully estimated, the grasps learned off-line can be used to grasp the real object.