【作者】 刘明；

【导师】 李满天；

【作者基本信息】 哈尔滨工业大学 ， 机械工程， 2022， 博士

【摘要】 具有类似足式生物那样“从不会走到会走”的稳定节律性运动控制能力一直是四足机器人研究领域关注的热点之一。现代神经生理学研究表明,足式动物具有上述能力主要是由其中枢环路网络中各个神经的突触在外界信息不断评价、引导下发育产生的。因此,如何建立具有类似足式生物神经突触发育机理的中枢环路网络模型成为实现四足机器人从不会走到会走的关键之一。然而,当前对中枢环路网络的建模研究由于忽略了中枢环路自身的发育特性而较难赋予四足机器人“从不会走到会走”的能力。针对当前四足机器人对节律性运动能力的需求,以及中枢环路网络模型研究存在的不足,本文借鉴生物运动控制神经系统(中枢环路)初期的发育过程以及突触电信号振动特性,结合突触物理特性对神经细胞信号传递的作用以及评价对中枢环路网络修正发育的作用,建立了神经网络突触发育式模型,通过四足机器人实验证明了本文所建立的模型、方法的准确性与有效性。首先针对当前神经细胞电信号振动模型存在的神经细胞微观与宏观特性相脱离问题,从钾、钠离子浓度波动入手,利用神经细胞电信号振动的宏观特性,建立了神经细胞电信号振动模型,进而结合神经细胞电信号电生理实验的特性分析,对所建立神经细胞电信号振动模型的稳定性、周期解存在性、驰豫振动以及混沌等特性进行了研究,获得了相应的参数范围,并进行了所建模型与环节软体动物运动神经振动响应对比实验,实验结果表明所建立的神经细胞电信号振动模型能够很好的描述生物运动神经细胞的电信号振动特性。其次从生物神经发育角度入手,结合突触发育所具有的方向性与随机性,在突触发育特性与随机动力学的物理等价性分析与突触发育特性分析的基础上,建立了突触发育模型,该模型包括:基于引力场的突触发育方向引导模型、突触发育方程、突触随机转向模型以及突触发育修正模型,并开展了突触发育参数分析,获得了突触发育步长和突触随机转角系数对突触发育过程的影响特性。再次通过对拉尔提出的突触等效电缆模型进行简化,建立了突触路径传递函数,通过分析传递函数特性,获得了传递函数的时间常数,在四足生物单腿运动特性以及腿间协调关系分析的基础上,融合运动速度、占空比以及相位差,建立了突触发育式网络评价算法,并进行了突触发育式网络仿真实验,实验结果表明了所建立的模型与算法的准确性与有效性。最后利用仿真平台结合发育式神经网络发育、评价与优化方法对机器人进行运动控制实验,验证了发育式神经网络模型与网络评价算法能够实现四足机器人从不会走到会走的运动控制过程。在此基础上,利用半月形腿式四足机器人平台,进一步验证了发育式神经网络模型与网络评价算法具有实现四足机器人从不会走到会走的能力。进而利用液压驱动四足机器人平台进行了与Matsuoka中枢环路模型控制行走对比实验,验证了所建立的发育式网络对液压驱动四足机器人平台运动控制有更好的效果。更多还原

【Abstract】 It has always been one of the hotspots in the research field of quadruped robots to enable quadruped robots to have the controllability of stable rhythmic motion,similar to the process of “never walk to walk” of quadrupeds.Modern neurophysiological research has shown that the process of “never walk to gain rhythmic gait motion ability” of quadrupeds is mainly caused by the development of the synapses of each nerve in the central loop network under the continuous evaluation and guidance of external information.Therefore,the establishment of a central loop network model,similar to the mechanism of synaptic development of quadrupeds has become one of the keys to realize the goal of “never walks to can walk” of quadruped robots.However,the current modeling research on the central loop network find it difficult to give a quadruped robot the ability of “never walk to walk”,resulting of the ignorance of the developmental characteristics of the central loop.Modern biological behavior studies have found that the rhythmic gait motion controllability of quadrupeds is not innate,but gradually formed in a period of time after birth.For example,the small gazelle cannot even stand when it is just born,but can stand and perform small-range static gait motion after tens of minutes,obtaining rhythmic motion ability in a few hours.The process of "never walks to walk" of this kind of quadrupeds is mainly caused by the development of the synapses of each nerve in the central loop network under the continuous evaluation and guidance of external information,the detail of which is: the random development of nerve synapses→the initial formation of the central loop network→generation of random small-range motions→evaluation of each synapse in the network→guidance of the corrective development of each synapse in the network→continuously expansion of the range of motion→until the central loop network meets the needs of rhythmic motion control.Targeting at the current quadruped robot’s need to realize the rhythmic gait motion controllability of "never walks to walk" and the deficiencies in the res earch of the central loop network model,this article draws on the initial developmental process of biological motion control nervous system(central loop network)and the vibration characteristics of synaptic electrical signals,combined with the effect o f synaptic physical characteristics on signal transmission of nerve cell and the evaluation of the effect on the corrective development of the central loop network,establishing a neural network synapse fertility model.In addition,experiments through a quadruped robot are implemented to test the accuracy and effectiveness of the model and method established in this article.First,focusing on the problem of the separation of the microscopic and macroscopic characteristics of nerve cells in the current neu ral cell electrical signal vibration model,this article starts from the concentration fluctuation of potassium and sodium ion,using the macroscopic characteristics of nerve cell electrical signal vibration,establishing a neural cell electrical signal vi bration model.Then combined with the characteristic analysis of the electrophysiological experiment of the nerve cell electrical signal,the stability,the existence of the periodic solution,the relaxation vibration and the chaos of the established nerve cell electrical signal vibration model are studied,as well as the corresponding parameter range is obtained.Moreover,a comparative experiment on the vibration response of the built model and the motor nerve of an annelid mollusk is carried out.The exp erimental results show that the established nerve cell electrical signal vibration model can well describe the electrical signal vibration characteristics of biological motor nerve cells.Secondly,starting from the perspective of biological neurodevelopme nt,combining the directionality and randomness of synaptic development,based on the analysis of the physical equivalence of synaptic development characteristics and stochastic dynamics and the analysis of synaptic development characteristics,the developmental model of synapse is established,including: synaptic growth direction guidance model based on gravitational field,synaptic growth equation,synaptic random steering model,and synaptic corrective development model.Then,synaptic growth simulation experiments were carried out,whose results were analyzed.From the analysis,it is shown that the established synaptic breeding model is correct and effective.Once again,by simplifying the synaptic equivalent cable model proposed by Lahr,the synaptic path transfer function is established.By analyzing the characteristics of the transfer function,the time constant of the transfer function is obtained.Based on the analysis of the duty ratio,movement speed,and coordinated relationship of legs of quadrupeds,with the fusion of movement speed,duty ratio and phase difference,a synaptic development network evaluation algorithm is established.A synaptic development network simulation experiment is carried out,whose results show the accuracy and effectiveness of established model and algorithm.Finally,in order to verify the effectiveness of the above established models and algorithms,the quadruped robot simulation platform was used to develop and optimize the developmental neural network,on which robot motion control experiment was carried out to verify that the developmental neural network model and network evaluation algorithm can realize the motion control process of “never walk to walk” of the quadruped robot.On this basis,by using the half-moon-legged quadruped robot platform,it is further verified that the developmental neural network model and network evaluation algorithm are able to realize that the process of “never walk to know how to walk” of quadruped robot.Then,the hydraulically driven quadruped robot platform was used to carry out a comparative experiment on control walking,compared with the Matsuoka central loop model,which verifies that the established developmental network has a better effect on the motion control of the hydraulically driven quadruped robot platform.更多还原