张森彦, 田国会, 张营, 刘小龙. 一种先验知识引导的基于二阶段渐进网络的自主抓取策略[J]. 机器人, 2020, 42(5): 513-524. DOI: 10.13973/j.cnki.robot.190623
引用本文: 张森彦, 田国会, 张营, 刘小龙. 一种先验知识引导的基于二阶段渐进网络的自主抓取策略[J]. 机器人, 2020, 42(5): 513-524. DOI: 10.13973/j.cnki.robot.190623
ZHANG Senyan, TIAN Guohui, ZHANG Ying, LIU Xiaolong. An Autonomous Grasping Strategy Based on Two-Stage Progressive Network Guidedby Prior Knowledge[J]. ROBOT, 2020, 42(5): 513-524. DOI: 10.13973/j.cnki.robot.190623
Citation: ZHANG Senyan, TIAN Guohui, ZHANG Ying, LIU Xiaolong. An Autonomous Grasping Strategy Based on Two-Stage Progressive Network Guidedby Prior Knowledge[J]. ROBOT, 2020, 42(5): 513-524. DOI: 10.13973/j.cnki.robot.190623

一种先验知识引导的基于二阶段渐进网络的自主抓取策略

An Autonomous Grasping Strategy Based on Two-Stage Progressive Network Guidedby Prior Knowledge

  • 摘要: 针对未知不规则物体在堆叠场景下的抓取任务,提出一种基于二阶段渐进网络(two-stage progressive network,TSPN)的自主抓取方法.首先利用端对端策略获取全局可抓性分布,然后基于采样评估策略确定最优抓取配置.将以上2种策略融合,使得TSPN的结构更加精简,显著减少了需评估样本的数量,能够在保证泛化能力的同时提升抓取效率.为了加快抓取模型学习进程,引入一种先验知识引导的自监督学习策略,并利用220种不规则物体进行抓取学习.在仿真和真实环境下分别进行实验,结果表明该抓取模型适用于多物体、堆叠物体、未知不规则物体、物体位姿随机等多种抓取场景,其抓取准确率和探测速度较其他基准方法有明显提升.整个学习过程历时10天,结果表明使用先验知识引导的学习策略能显著加快学习进程.

     

    Abstract: To grasp stacked objects with unknown and irregular shapes, an autonomous grasping method based on two-stage progressive network (TSPN) is proposed. Firstly, the global grasping distribution is acquired based on end-to-end strategy, and then the optimal grasping configuration is determined based on the sampling-evaluation strategy. By combining the above two strategies, the structure of TSPN is simplified, the number of samples to be evaluated is reduced significantly, and the generalization ability is guaranteed while the grasping efficiency is improved. In order to speed up the learning process of the grasping model, a self-supervised learning strategy guided by prior knowledge is introduced, and 220 irregular objects are used for grasping learning. Experiments are carried out in simulation and real environment respectively. The results show that the proposed grasping model is suitable for the grasping scenes with multiple objects, stacked objects, unknown irregular objects, and the objects with random position and pose. The grasping accuracy and detection speed are significantly improved compared with other reference methods. The whole learning process lasted for 10 days, and the results show that the learning strategy guided by prior knowledge can significantly speed up the learning process.

     

/

返回文章
返回