Grasp the Graph (GtG): A Super Light Graph-RL Framework for Robotic Grasping

Abstract

This paper introduces Grasp the Graph, an innovative framework addressing challenges posed by processing point clouds with conventional neural networks for robotic grasping. Such challenges often lead to the conversion of point clouds to multi-view images or voxel grids, requiring large networks with millions of parameters. Grasp the Graph formulates robotic grasping as a one-step reinforcement learning problem and leverages graph representation and graph neural networks to efficiently process geometric relationships in point clouds. The lightweight design with only 67k parameters facilitates swift training and strong generalization to novel objects, offering the capability to propose multiple valid grasping candidates. Empirical testing demonstrates the effectiveness of the proposed approach, achieving a 90.4% success rate on test objects while being trained with only 10 objects. Even with a three-camera setup during training, reducing the cameras during evaluation minimally impacts performance, resulting in less than a 3% decrease. <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>