基于深度Q网络的仿人机器人步态优化  

Gait Optimization of Humanoid Robot Based on Deep Q Network

作　　者：袁雯 刘惠义[1] YUAN Wen;LIU Hui-yi(College of Computer and Information, Hohai University, Nanjing 210098, China)

机构地区：[1]河海大学计算机与信息学院,江苏南京210098

出　　处：《计算机与现代化》2019年第4期47-51,58,共6页Computer and Modernization

基　　金：江苏省水利厅科技计划项目(2017003ZB)

摘　　要：为实现仿人机器人快速稳定的行走,在满足有效参数组合的条件下,提出一种基于深度强化学习的步行参数训练算法以优化机器人步态。首先,从环境中捕获机器人步态模型参数作为DQN的输入;然后,用DQN来拟合机器人行走产生的状态-动作值函数;最后,通过动作选择策略选择当前机器人执行的步态动作,同时产生奖励函数达到更新DQN的目的。选择NAO仿真机器人为实验对象,在Robo Cup3D仿真平台上进行实验,结果证明在此算法下,NAO仿人机器人可以获得稳定的双足步行。In order to realize the fast and stable walking of humanoid robot, and under the condition that the effective parameter combination is satisfied, a walking parameter training algorithm based on deep reinforcement learning is proposed to optimize the gait of humanoid robot. First of all, we capture the robot gait model parameters from the environment as the input of DQN. And then, DQN is used to fit the robot state-action value function. At last, by action selection strategy, we choose the gait of a robot to perform current action, at the same time produce reward function to achieve the aim of updating DQN. By selecting NAO robot as the experimental object and conducting experiments on the RoboCup3D simulation platform, the results show that using this algorithm, NAO robot can achieve stable bipedal walking.

关 键 词：仿人机器人 深度强化学习 DQN 步态优化 ROBOCUP3D 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]