Shape recognition and vision-based robot control by shape morphing

The article presents a unified approach for object recognition and recognition based control of robotic interactions with objects. The approach is based on morphing the shape of objects. The morph of one object to another is treated as a deformation and is defined such that it can be quantified using a physics based model. This quantification serves as a dissimilarity measure and is used for shape recognition. By storing the views that represent the grasp positions for different object, the proposed framework is used to identify the desired end position a robot needs to attain for interacting with the given object. Furthermore, the images synthesized during the morph define a view based trajectory, starting with the view defining the initial robot-object orientation to the view defining the desired robot-object orientation. We propose a technique where these synthetic images are used to control the motion of a PUMA 560 eye-in-hand manipulator to execute alignment and grasping tasks. The proposed approach does not require complete calibration information, obviates manual feature selection and correspondence, can provide smooth trajectories, and needs a single image for each object (in a 4-DOF formulation). Potential applications range from recognition and positioning with respect to partially occluded or deformable objects as well as planning robotic grasping based on human demonstration.