Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp\n Synthesis Approach

Abstract

This paper presents a real-time, object-independent grasp synthesis method\nwhich can be used for closed-loop grasping. Our proposed Generative Grasping\nConvolutional Neural Network (GG-CNN) predicts the quality and pose of grasps\nat every pixel. This one-to-one mapping from a depth image overcomes\nlimitations of current deep-learning grasping techniques by avoiding discrete\nsampling of grasp candidates and long computation times. Additionally, our\nGG-CNN is orders of magnitude smaller while detecting stable grasps with\nequivalent performance to current state-of-the-art techniques. The light-weight\nand single-pass generative nature of our GG-CNN allows for closed-loop control\nat up to 50Hz, enabling accurate grasping in non-static environments where\nobjects move and in the presence of robot control inaccuracies. In our\nreal-world tests, we achieve an 83% grasp success rate on a set of previously\nunseen objects with adversarial geometry and 88% on a set of household objects\nthat are moved during the grasp attempt. We also achieve 81% accuracy when\ngrasping in dynamic clutter.\n