The invention provides a robot walking control method based on deep reinforcement learning. The method comprises: setting a reward mechanism; constructing a multi-expert neural network; wherein the multi-expert neural network comprises a top-layer network and at least one bottom-layer network; training a top-layer network by using the reward mechanism and the collected sample data; and in the operation process of the robot, the top-layer network fusiung the output of the bottom-layer network according to the training result and the data measured by the robot in real time, outputting a controlinstruction according to the fusion result, sending the control instruction to the robot and controlling motors of joints in the robot. According to the method, continuous switching can be carried outamong different expert networks, the diversity of combination is increased, and the defects of asymmetry and non-naturalness of movement gaits of an existing robot are overcome. According to the method, the walking gait obtained by human motion capture is used as a reference object for training and learning, so that the multi-expert neural network can be converged to a periodic symmetrical walking strategy which is natural like a human body more quickly.

    本发明提供的基于深度强化学习的机器人行走控制方法,设置奖励机制;构建多专家神经网络;所述多专家神经网络包括一个顶层网络和至少一个底层网络;利用所述奖励机制和采集的样本数据对顶层网络进行训练;在机器人运行过程中,顶层网络根据训练结果和机器人实时测量到的数据对底层网络的输出进行融合,并根据融合结果输出控制指令,将所述控制指令发送给机器人,控制机器人中关节的电机。该方法可以在不同的专家网络之间进行连续切换,增加了组合的多样性,改善现有机器人运动步态非对称、非自然的缺陷。该方法将人类动作捕捉获得的行走步态作为训练学习的参考对象,能让多专家神经网络更快地收敛到像人一样自然的周期性对称行走策略。


    Zugriff

    Download


    Exportieren, teilen und zitieren



    Titel :

    Robot walking control method and system based on deep reinforcement learning and medium


    Weitere Titelangaben:

    基于深度强化学习的机器人行走控制方法、系统及介质


    Beteiligte:
    YANG CHUANYU (Autor:in) / PU CAN (Autor:in)

    Erscheinungsdatum :

    2020-08-25


    Medientyp :

    Patent


    Format :

    Elektronische Ressource


    Sprache :

    Chinesisch


    Klassifikation :

    IPC:    G05B Steuer- oder Regelsysteme allgemein , CONTROL OR REGULATING SYSTEMS IN GENERAL / B62D MOTOR VEHICLES , Motorfahrzeuge



    Walking Motion Generation of Bipedal Robot Based on Planar Covariation Using Deep Reinforcement Learning

    Yamano, Junsei / Kurokawa, Masaki / Sakai, Yuki et al. | Springer Verlag | 2024


    ROBOT MOTION PARAMETER ADAPTIVE CONTROL METHOD AND SYSTEM BASED ON DEEP REINFORCEMENT LEARNING

    REN LIANG / WANG CHUNLEI / YANG YA et al. | Europäisches Patentamt | 2022

    Freier Zugriff

    Walking system of tracked robot, walking system control method and storage medium

    LI QINGFENG / SU XIAOJIE / MA TIEDONG | Europäisches Patentamt | 2023

    Freier Zugriff

    Biped walking robot control method and biped walking robot control system

    MORIDAIRA TOMOHISA | Europäisches Patentamt | 2016

    Freier Zugriff

    Planet wheel, walking robot, control method of walking robot and readable storage medium

    JIA ZHENZHONG / ZHU ZHENG / WANG WENHUI et al. | Europäisches Patentamt | 2023

    Freier Zugriff