基于强化学习的三指灵巧手抓取方法研究  

Research on Three-finger Dexterous Hand Grasping Method Based on Reinforcement Learning

作　　者：杨扬[1] 韦君玉 YANG Yang;WEI Junyu(Col.of Information,North China Univ.of Tech.,100144,Beijing,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《北方工业大学学报》2020年第2期99-104,共6页Journal of North China University of Technology

摘　　要：为了实现对圆锥等特殊物品的抓取,本文提出了一种基于强化学习的三指灵巧手机器人抓取方法.本文使用DenseNet-121网络实现了图像的分类,使用ImageNet数据集进行预训练,以解决数据集较少时易出现过拟合的问题;并基于强化学习中的Q-Learning算法,在完全卷积网络中进行端到端的训练,学习视觉运动策略,最后在少量数据集上测试算法.实验结果表明,使用三指灵巧手能够有效抓取圆锥、圆球等物体;同时,该方法训练生成的模型有效建立了像素图像数据与机器人执行运动抓取动作之间的映射关系,根据这种映射关系选出的最优Q值,提高了抓取成功率.In order to grasp special objects, this paper proposes a three-finger dexterous hand robot grasping method based on reinforcement learning. This article uses DenseNet-121 to achieve image classification, pre-training on ImageNet to solve the problem of overfitting when the data set is small;and based on the Q-Learning algorithm in reinforcement learning, end-to-end training is performed in a fully convolutional network to learn visual motion strategies, and finally the algorithm is tested on a small number of data sets. The experimental results show that the use of three-finger dexterous hands can effectively grasp cones and balls. At the same time, the model generated by this method effectively establishes the mapping relationship between pixel image data and the robot’s motion grabbing action. The optimal Q value selected according to this mapping relationship improves the success rate of grabbing.

关 键 词：三指灵巧手 强化学习 完全卷积网络 

分 类 号：TP249[自动化与计算机技术—检测技术与自动化装置]