6DoF Bimanual-GraspNet: A Deep Learning Approach for Bimanual Grasping

Abstract

This study proposes a deep learning architecture that generate bimanual grasp for dual arm robot system. In Chapter 1, backgrounds and related works of robotic grasping are described to introduce the idea of bimanual grasp generation. To generate grasp candidates, current methods using data-driven manner utilizing grasp dataset collected from simulator with analytic calculations. In Chapter 2, the proposed methods, namely 6DoF Bimanual-GraspNet, with suggested loss, called permutation invariant min distance loss, is formulated. Before explain about proposed model, introduce about the motivation and the toy problem. 6DoF Bimanual-GraspNet extracts representative feature by PointNet++ encoder and generate grasp candidates using modified transformer decoder similar to DETR decoder. Chapter 3 covers experimental setup for evaluating the model. We discover how the used dataset consists of and explain about the simulator setup how to evaluate the proposed model. In Chapter4, bimanual grasping result in both point cloud and simulator comparison with other method is presented. The experiment demonstrates that the proposed method outperforms other method. Performance on grasp success rate in simulator of ours is 81.6% on seen objects and 57.9% on unseen objects while randomly sampled from single arm grasp candidates success only 50.6% on seen objects and 47.3% on unseen objects. In Chapter 5, the proposed model with high quality are demonstrated with visualization. Experimental results suggested that high quality grasps is reasonable for bimanual grasping scenarios which consider grasp stability, dexterity and torque optimization. Also, this method does not require additional module such as grasp sampler model which need for grasp classification model for example dual-pointnetGPD which is proposed previous works. This make our model more robust and have more generalizability in real-world scenarios.