基于多约束条件的机器人抓取策略学习方法  

A learning method of robotic grasping strategy based on multi-constraint conditions

作　　者：崔涛 李凤鸣[1] 宋锐[1] 李贻斌[1] CUI Tao;LI Feng-ming;SONG Rui;LI Yi-bin(School of Control Science and Engineering,Shandong University,Ji’nan 250061,China)

机构地区：[1]山东大学控制科学与工程学院,济南250061

出　　处：《控制与决策》2022年第6期1445-1452,共8页Control and Decision

基　　金：广东省重点领域研发计划项目(2020B090925001);国家自然科学基金面上项目(61973196)。

摘　　要：针对机器人在多类别物体不同任务下的抓取决策问题,提出基于多约束条件的抓取策略学习方法.该方法以抓取对象特征和抓取任务属性为机器人抓取策略约束,通过映射人类抓取习惯规划抓取模式,并采用物体方向包围盒(OBB)建立机器人抓取规则,建立多约束条件的抓取模型.利用深度径向基(DRBF)网络模型结合减聚类算法(SCM)实现抓取策略的学习,两种算法的结合旨在提高学习鲁棒性与精确性.搭建以Refiex 1型灵巧手和AUBO六自由度机械臂组成的实验平台,对多类别物体进行抓取实验.实验结果表明,所提出方法使机器人有效学习到对多物体不同任务的最优抓取策略,具有良好的抓取决策能力.Aiming at the problem of robot grasping decision under different tasks of multiple objects,a learning method of the grasping strategy based on multiple constraints is proposed,in which,the characteristics of grasping objects and the attributes of grasping tasks are taken as the multiple constraints of the a robot grasping strategy.Furthermore,the method uses human grasping habit to map to robot grasp types,and the grasping rules of a robot are established by using the object bounding box(OBB).The fetching model with multiple constraints is established.Then,the radial basis function RBF network model is combined with the de-clustering algorithm SCM to realize the grasping strategy learning.The combination of the two algorithms aims to improve the robustness and accuracy of learning.Using the AUBO six-degree-of-freedom robotic arm with Refiex 1 dexterity hand,experiments are conducted to grasp objects with different shapes and multiple tasks.Experimental results show that the proposed method enables the robot to effectively learn the optimal grasping strategy for different tasks of multiple objects and has good grasping decision-making ability.

关 键 词：机器人抓取策略 抓取决策 多约束条件 神经网络 方向包围盒 灵巧手 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]