基于深度确定性策略梯度算法的双轮机器人平衡控制研究  

Balance Control of Two-wheeled Robot Based on Deep Deterministic Policy Gradient

作　　者：刘胜祥 林群煦 杨智才 吴月玉 翟玉江 LIU Shengxiang;LIN Qunxu;YANG Zhicai;WU Yueyu;ZHAI Yujiang(School of Railway Tracks and Transportation,Wuyi University,Jiangmen 529020,China)

机构地区：[1]五邑大学轨道交通学院

出　　处：《机械工程师》2020年第3期142-144,共3页Mechanical Engineer

基　　金：国家自然基金(51307170);广东省自然科学基金(2018A030310046);五邑大学青年科研基金(2016zk02)

摘　　要：针对深度强化学习在双轮机器人中的应用问题,提出了一种基于深度确定性策略梯度算法的平衡控制方法。首先,该方法将双轮机器人作为智能体,读取状态信息并建立动作策略和奖惩机制;其次,智能体依据算法给出的随机动作指令执行,执行完动作之后反馈状态信息,算法根据反馈的状态信息给出新动作;最后,通过多回合训练使双轮机器人保持平衡。实验结果表明,通过多回合训练之后,双轮机器人的姿态角度波动范围在-4°~4°以内,双轮机器人的平稳性得到了显著改善,实现了机器人的平衡控制行为学习,证明了该方法的有效性。Aiming at the application of deep reinforcement learning in two-wheeled robot,a balance control method based on Deep Deterministic Policy Gradient is proposed.Firstly,the two-wheeled robot is used as the agent to read the state information and establish the action strategy and reward and punishment mechanism.Then DDPG gives the random action and gives feedback state information after the agent performs the action.DDPG gives the new action according to the feedback state information.Compared with the traditional PID balance control,this algorithm does not need complex dynamic modeling.Finally,the two-wheeled robot is balanced by multi-round training.The experimental results show that the stability of the two-wheeled robot is obviously improved after multi-round training,and the learning of the balanced control behavior of the robot is realized,which proves the effectiveness of the proposed method.

关 键 词：双轮机器人 平衡控制 策略梯度 动作策略 奖惩机制 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]