Adaptive Fuzzy Gaussian Mixture Models for Shape Approximation in Robot Grasping

Abstract

Robotic grasping has always been a challenging task for both service and industrial robots. The ability of grasp planning for novel objects is necessary for a robot to autonomously perform grasps under unknown environments. In this work, we consider the task of grasp planning for a parallel gripper to grasp a novel object, given an RGB image and its corresponding depth image taken from a single view. In this paper, we show that this problem can be simplified by modeling a novel object as a set of simple shape primitives, such as ellipses. We adopt fuzzy Gaussian mixture models (GMMs) for novel objects’ shape approximation. With the obtained GMM, we decompose the object into several ellipses, while each ellipse is corresponding to a grasping rectangle. After comparing the grasp quality among these rectangles, we will obtain the most proper part for a gripper to grasp. Extensive experiments on a real robotic platform demonstrate that our algorithm assists the robot to grasp a variety of novel objects with good grasp quality and computational efficiency.