A Learning Approach for Robotic Grasp Selection in Open-Ended Domains

Enabling a robot to grasp unknown objects is still an ongoing challenge in robotics. The main problem is to find an appropriate grasp configuration including the position and orientation of the arm relative to the object and fingers configuration. One approach is to recognize an appropriate grasp pose for an object based on one or more grasp demonstrations of the same or other objects. We focus on familiar objects, i.e. unknown objects sharing some features with known objects. The underlying assumption in grasping familiar objects is that, where they are similar to known ones, they may be grasped in a similar way. However finding an object representation and a similarity metric are still the main challenge to transfer grasp experiences to new objects. In this paper, we present an interactive object view recognition approach and a similarity metric to grasp familiar objects. Object view recognition is incrementally capable of recognizing object view labels. The grasp pose learning approach learns a grasp template for a recognized object view using local and global visual features of a demonstrated grasp. In grasp pose recognition, a similarity measure based on Mahalanobis distance is used for grasp template matching. The experimental results reveal the high reliability of the developed template matching approach. We also demonstrate how the proposed grasp learning system can incrementally improve its performance in grasping familiar objects.