摘要
触觉在机器人抓取过程中扮演着重要的角色,但在大多数强化学习任务中,触觉仅被用于拓展状态空间,其提供的位置和压力等信息很少被完全利用.针对该问题,同时受内在奖励机制启发,首先设计了一种“倒T”形传感器阵列布局;然后基于这种传感器阵列提出了新的内在激励方法,该方法根据机械臂末端与物体接触位置的不同,给予不同的重视程度,鼓励智能体以更有效的姿态来夹取物体;最后将该方法在仿真环境中进行测试,结果表明该方法在夹取椭球和圆球物体任务中收敛速度比最新的基准方法平均提高了约20%.
Although play an important role in the process of the robot grasping,haptics is only used to extend the state space,and the information provided by it,such as position and pressure,is rarely fully utilized in most reinforcement tasks.Inspired by the intrinsic reward mechanism,an intrinsic incentive method based on the“inverted T”array sensor is proposed.According to the position where the end effector of the robot touches the object,the method gives degrees of importance,and encourages the agent to achieve the goal with a more effective posture.Finally,the method was tested in the simulation environment,and the results showed that the speed of convergence of the method in the task gripping ellipsoid objects was about 20%faster than the latest benchmark method.
作者
宋相兵
季玉龙
俎文强
何扬
杨红雨
SONG Xiang-Bing;JI Yu-Long;ZU Wen-Qiang;HE Yang;YANG Hong-Yu(National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610065,China;School of Aeronautics and Astronautics,Sichuan University,Chengdu 610065,China;College of Computer Science,Sichuan University,Chengdu 610065,China)
出处
《四川大学学报(自然科学版)》
CAS
CSCD
北大核心
2022年第3期53-62,共10页
Journal of Sichuan University(Natural Science Edition)
基金
国家自然科学基金重点项目(U20A20161)。
关键词
深度强化学习
机械臂
抓取
触觉
内在奖励
Deep reinforcement learning
Robot arm
Grasping
Haptics
Intrinsic reward