摘要
在机器人自主抓取领域,由于抓取对象的大小形状以及分布状态的随机性,仅靠单一的抓取操作完成对工作区域内物体的抓取是十分困难的,而推动和抓取动作的结合可以降低抓取环境的复杂性,通过推动操作可以改变抓取对象的分布以便于更好的抓取。但是推动动作的添加同时也会产生一些无效的推动,会降低模型的学习效率。在基于深度Q网络(deep Q-network,DQN)的视觉推动抓取(visual pushing for grasping,VPG)模型的基础上,提出了一种可供性方案用于简化机器人动作规划空间的搜索复杂度,加快机器人抓取的学习进程。通过减少在任何给定情况下可用的行动数量来实现更快的计划,有助于从数据中更高效和精确地学习模型。最后通过在V-rep仿真平台上的仿真场景验证了所提方法的有效性。
In the field of robotic autonomous grasping,due to the randomness of the sizes and shapes and the distribution states of the objects,it is very difficult to complete the grasping of the object in the working area by a single grasping operation.However,the combination of pushing and grasping actions can greatly decrease the complexity of the grasping environment,and change the distribution of the objects through the pushing operation for better grasping.But the addition of pushing actions will also produce some ineffective pushing,and will reduce the learning efficiency of the model.In this paper,we propose an affordance method based on the visual pushing for grasping(VPG)model using deep Q-network(DQN)to simplify the search complexity of the robot action planning space and accelerate the learning process of robot grasping.This approach can achieve faster planning by reducing the number of actions available in any given situation,helping to learn the model more efficiently and precisely from the data.Finally,the effectiveness of the proposed method is verified through simulation scenarios on the V-rep simulation platform.
作者
温凯
李东年
陈成军
赵正旭
WEN Kai;LI Dongnian;CHEN Chengjun;ZHAO Zhengxu(School of Mechanical and Automotive Engineering,Qingdao University of Technology,Qingdao 266520,China)
出处
《组合机床与自动化加工技术》
北大核心
2024年第11期34-37,43,共5页
Modular Machine Tool & Automatic Manufacturing Technique
基金
山东省自然科学基金项目(ZR2022ME203)。
关键词
机器人抓取
可供性
深度Q网络
深度强化学习
robotic grasping
affordance
deep Q-network
deep reinforcement learning
