摘要
传统仿生机器狗步态规划需构建动力学模型,对此提出一种基于DQN的仿生机器狗步态规划方法。结合DQN算法的基本原理,将DQN算法与仿生机器狗控制器结合,进而大量训练仿生机器狗的DQN控制器,设置DQN的参数和计算DQN的奖励函数,并在Webhots仿真环境中对四足机器狗进行训练,得到四足仿生机器狗运动的翻滚角、俯仰角以及损失函数的变化。结果表明,通过DQN训练得到的仿生机器狗的翻滚角、俯仰角和损失函数都能趋于稳定,说明机器狗能快速平稳地完成步态运动,证实了DQN控制器的有效性,可为四足机器狗的步态生成提供新的方案。
Aiming at the problem that the traditional bionic robot dog gait planning needs to build a dynamic model,a bionic robot dog gait planning method based on DQN is proposed.DQN algorithm is combined with bionic robot dog controller,and then a large number data are used to train the DQN controllers of bionic robot dog.The parameters of DQN are set and the reward function of DQN is calculated.The quadruped robot dog is trained in Webhots simulation environment,and the changes of roll angle,pitch angle and loss function of quadruped bionic robot dog are obtained.The results show that the roll angle,pitch angle and loss function of the bionic robot dog trained by DQN gradually tend to be stable,which shows that the robot dog can move quickly and smoothly,proves the effectiveness of DQN controller,and can provide a new reference for the gait generation of quadruped robot dog.
作者
苗俊杰
郭猛
邹捷
王爽
MIAO Junjie;GUO Meng;ZOU Jie;WANG Shuang(Hebei Electric Power Corporation,Shijiazhuang 050021,China;State Grid Baoding Power Supply Company,Baoding 071051,China)
出处
《微型电脑应用》
2023年第3期47-51,共5页
Microcomputer Applications
基金
国网河北省电力有限公司科技项目(kj2020-014)。
关键词
仿生机器狗
步态规划
强化学习
DQN算法
bionic robot dog
gait planning
strengthen learning
DQN algorithm
