Robotic grasp detection toward unknown objects using 3D edge detection and Gaussian mixture model for clustering candidates

Abstract

We propose a novel grasp detection method for arbitrary shape objects. This method is applicable to multijoint manipulators with a two-finger hand and an RGB-D hand-eye camera. When a robot tries to grasp an unknown object with a camera as the only sensor, it must choose a position to grasp without mass distribution information. Moreover, even if the robot can find some candidates of locations for grasping, some of them will be false due to noise. The proposed method chooses an appropriate candidate by variational Bayesian clustering. In the clustering, the candidates are classified based on their closeness. Then the method chooses one existing at the center of the largest cluster since it may be uninfluenced by noise. This set of procedures does not require training unlike learning methods. The proposed method is evaluated with an actual robot. Though the method is a heuristic, it can choose suitable grasping locations from piles of objects.