2018/09/14 by Diego Rodríguez, Corbin Cogswell, Rodriguez, Diego +6
Computer Science · Engineering · #FOS: Computer and information sciences #Human Pose and Action Recognition #Reinforcement Learning in Robotics #Robot Manipulation and Learning #Robotics (cs.RO) #Robotics and Sensor-Based Localization
paper · pdf · doi:10.48550/arxiv.1809.05353
openalex publication_date 2018/09/14 · openalex created_date 2022/08/03 · openalex updated_date 2026/07/28
Robots acting in open environments need to be able to handle novel objects.\nBased on the observation that objects within a category are often similar in\ntheir shapes and usage, we propose an approach for transferring grasping skills\nfrom known instances to novel instances of an object category. Correspondences\nbetween the instances are established by means of a non-rigid registration\nmethod that combines the Coherent Point Drift approach with subspace methods.\nThe known object instances are modeled using a canonical shape and a\ntransformation which deforms it to match the instance shape. The principle axes\nof variation of these deformations define a low-dimensional latent space. New\ninstances can be generated through interpolation and extrapolation in this\nshape space. For inferring the shape parameters of an unknown instance, an\nenergy function expressed in terms of the latent variables is minimized. Due to\nthe class-level knowledge of the object, our method is able to complete novel\nshapes from partial views. Control poses for generating grasping motions are\ntransferred efficiently to novel instances by the estimated non-rigid\ntransformation.\n