Learning-based approaches to grasping have become a popular alternative to geometric and model-based planning over the past decade. In particular grasp learning has shown to generalize well to previously unseen objects where only partial-view visual information is available. More recently, researchers have looked to capitalize on the success of deep neural networks to improve grasp learning. Broadly speaking deep neural network methods for grasp learning can be split into two approaches: predicting grasp success for an image patch associated with a gripper configuration and directly predicting a grasp configuration from an image or image patch using regression. While these deep learning approaches have shown impressive performance for parallel jaw grippers relatively little work has focused on the more difficult problem of multi-fingered grasping. We believe two primary difficulties restrict the use of deep learning for multi-fingered grasping (1) the input representation used for grasp configurations in neural networks and (2) the reliance on external planners for generating candidate grasps.
展开▼