WO2023020036A1 - 一种基于回声状态网络的冗余机械臂追踪控制方法 - Google Patents

一种基于回声状态网络的冗余机械臂追踪控制方法 Download PDF

Info

Publication number
WO2023020036A1
WO2023020036A1 PCT/CN2022/092420 CN2022092420W WO2023020036A1 WO 2023020036 A1 WO2023020036 A1 WO 2023020036A1 CN 2022092420 W CN2022092420 W CN 2022092420W WO 2023020036 A1 WO2023020036 A1 WO 2023020036A1
Authority
WO
WIPO (PCT)
Prior art keywords
manipulator
moment
joint angle
end effector
cerebellum
Prior art date
Application number
PCT/CN2022/092420
Other languages
English (en)
French (fr)
Inventor
谭宁
余鹏
Original Assignee
中山大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中山大学 filed Critical 中山大学
Publication of WO2023020036A1 publication Critical patent/WO2023020036A1/zh

Links

Images

Classifications

    • BPERFORMING OPERATIONS; TRANSPORTING
    • B25HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25JMANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00Programme-controlled manipulators
    • B25J9/16Programme controls
    • B25J9/1602Programme controls characterised by the control system, structure, architecture
    • B25J9/161Hardware, e.g. neural networks, fuzzy logic, interfaces, processor
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B25HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25JMANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00Programme-controlled manipulators
    • B25J9/16Programme controls
    • B25J9/1612Programme controls characterised by the hand, wrist, grip control
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B25HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25JMANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00Programme-controlled manipulators
    • B25J9/16Programme controls
    • B25J9/1628Programme controls characterised by the control loop
    • B25J9/163Programme controls characterised by the control loop learning, adaptive, model based, rule based expert control
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B25HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25JMANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00Programme-controlled manipulators
    • B25J9/16Programme controls
    • B25J9/1628Programme controls characterised by the control loop
    • B25J9/1643Programme controls characterised by the control loop redundant control
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B25HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25JMANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00Programme-controlled manipulators
    • B25J9/16Programme controls
    • B25J9/1656Programme controls characterised by programming, planning systems for manipulators
    • B25J9/1661Programme controls characterised by programming, planning systems for manipulators characterised by task planning, object-oriented languages

Definitions

  • the invention relates to the field of manipulator tracking control, in particular to a redundant manipulator tracking control method based on an echo state network.
  • the manipulator When the manipulator has more degrees of freedom to drive (such as the number of joint angles of the manipulator) than is required to complete the task, the manipulator can be called a redundant manipulator. Due to their extra degrees of freedom, redundant manipulators have higher flexibility and safety, so they have broad potential applications in various fields. However, the additional degrees of freedom and structural features of this type of manipulator also bring challenges to the control of the manipulator.
  • a redundant manipulator tracking control method based on the echo state network is proposed to solve the existing redundant manipulator control method based on the pseudo-inverse of the Jacobian matrix, which needs to know the accurate manipulator model information (such as the length of the connecting rod), so that Calculate the value of the Jacobian matrix, and the existence of various uncertain factors will affect the accuracy of the model information, thereby affecting the accuracy of the control scheme.
  • the object of the present invention is to provide a redundant manipulator tracking control method based on the echo state network, which does not need to know the model information of the redundant manipulator, only needs to use the position information of the end effector as feedback, and does not need The speed and acceleration information of the end effector, and has high tracking control accuracy.
  • the first technical solution adopted by the present invention is: a redundant manipulator tracking control method based on echo state network, comprising the following steps:
  • the cerebellum-like model based on the echo state network calculates the target trajectory compensation information y(n+1) according to the input current moment joint angle variation ⁇ and the next moment target position p d (n+1) of the end effector of the mechanical arm;
  • Steps S2-S7 are repeated until the tracking control of the target trajectory is completed.
  • step of collecting the data of the manipulator and training the feedforward neural network controller to obtain the trained feedforward neural network also specifically includes:
  • the cerebellum-like model based on the echo state network includes afferent units, granule cells, Purkinje cells and efferent units.
  • the cerebellum-like model based on the echo state network calculates the target trajectory compensation information y (n+1 ) This step specifically includes:
  • update x(n+1) formula for updating the state of the granular cells at the n+1 moment is as follows:
  • Win represents the input connection weight between the afferent unit and the granule cell
  • W represents the internal Connection weight between nodes.
  • tanh represents the hyperbolic tangent activation function
  • W out represents the output connection weight between the Purkinje cell and the efferent unit
  • p(n+1) represents the state of the Purkinje cell at the n+1 moment .
  • the step of measuring the current position p a (n+1) of the end effector of the manipulator and updating the output weight W out of the cerebellum-like model of the acoustic state network specifically includes:
  • the output weight W out of the cerebellum-like model is updated according to the change amount ⁇ W out of the output weight.
  • T represents vector transpose
  • the present invention integrates the cerebellar-like controller and the data-driven learning method through the present invention, so that the scheme can cope with the influence of various uncertain factors to a certain extent without knowing the model parameter information of the manipulator, thereby effectively Complete the tracking control task of the redundant robotic arm.
  • Fig. 1 is a flow chart of steps of a redundant manipulator tracking control method based on an echo state network in the present invention
  • Fig. 2 is the control schematic diagram of the specific embodiment of the present invention.
  • Fig. 3 is a schematic diagram of a feedforward neural network training process of a specific embodiment of the present invention.
  • Fig. 4 is a schematic diagram of the target trajectory in the task space and the actual trajectory (first task cycle) of the end effector of the mechanical arm according to a specific embodiment of the present invention
  • Fig. 5 is a schematic diagram of the target trajectory in the task space and the actual trajectory (the tenth task period) of the end effector of the mechanical arm in a specific embodiment of the present invention
  • Fig. 6 shows the error variation of controlling the redundant robot arm to complete ten cycles of trajectory tracking tasks according to a specific embodiment of the present invention.
  • the present invention provides a kind of redundant manipulator tracking control method based on echo state network, comprises the following steps:
  • Both the input weight matrix W in and the internal connection weight matrix W are randomly initialized to a value between [-1, 1], and remain unchanged during the working process of the model.
  • the manipulator is controlled to move randomly, the data of the manipulator is collected, and a three-layer feed-forward neural network controller is trained.
  • the controller of the manipulator is an approximate inverse kinematics solver, and the controller is used to control the redundant manipulator There will be a large error when , so it is necessary to use a cerebellum-like model to improve the accuracy of the controller.
  • the cerebellum-like model based on the echo state network calculates the target trajectory compensation information y(n+1) according to the input current moment joint angle variation ⁇ and the next moment target position p d (n+1) of the end effector of the mechanical arm;
  • the measurement positions and joint angles are measured using sensor devices.
  • Steps S2-S7 are repeated until the tracking control of the target trajectory is completed.
  • the step of collecting mechanical arm data and training the feedforward neural network controller to obtain the trained feedforward neural network specifically includes:
  • the cerebellum-like model based on the echo state network includes an afferent unit Granulocytes Purkinje cells and outgoing unit
  • the afferent layer includes K units
  • the granular cells include N units
  • the Purkinje cells include N units
  • the efferent layer includes L units
  • t represents time.
  • connection weights between nodes inside granular cells Parallel fibers between granule cells and Purkinje cells, and output connection weights between Purkinje cells and efferent cells
  • the state of all granular cells is initialized as During the working process of the cerebellum-like model, the state update formula of each unit of the granular cell at the n+1 moment is as follows:
  • Win represents the input connection weight between the afferent unit and the granule cell
  • W represents the connection weight between the internal nodes of the granule cell
  • the output activation function of the cell usually using the sigmoid activation function:
  • Both the input weight W in and the internal connection weight W are randomly initialized to a value between [-1, 1], and remain unchanged during the working process of the model. It is worth noting that, according to existing theories, in order to ensure that the model can work normally, it is usually necessary to scale the internal connection weight W. Assume that the inner join weights are randomly initialized as Then the scaled internal connection weight matrix is where the scaling factor ⁇ needs to be slightly smaller than can, representation matrix the spectral radius. In the working process of the cerebellum-like model, what needs to be trained is the weight W out of the output layer. First, W out needs to be initialized as a matrix with all zero values, and then the output weights need to be trained. The training methods of W out can be divided into offline training and online training.
  • the cerebellum-like model based on the echo state network calculates the target trajectory compensation according to the input current moment joint angle variation ⁇ and the next moment target position p d (n+1) of the end effector of the mechanical arm
  • the step of information y(n+1) specifically includes:
  • the splicing formula is as follows:
  • the step of measuring the current position p a (n+1) of the end effector of the mechanical arm and updating the output weight W out of the cerebellum-like model of the acoustic state network specifically includes:
  • T represents vector transpose
  • the output weight W out of the cerebellum-like model is updated according to the change amount ⁇ W out of the output weight, as follows:
  • the redundant manipulator is controlled to continuously complete ten cycles of trajectory tracking tasks, and the tracking results of the first cycle and the tenth cycle are shown in Figure 4 and Figure 5 .
  • the first cycle the initial position of the end effector of the manipulator has a certain distance from the target trajectory, and it can be seen that the actual moving trajectory of the end effector quickly converges to the target trajectory.
  • the tenth cycle it can be seen that the actual moving trajectory of the end effector is almost consistent with the target trajectory.
  • the variation of the tracking error in each mission period is shown in Fig. 6. It can be seen that in the tenth mission period, the order of magnitude of the average tracking error is 10 -4 meters.
  • the above results reflect the validity of the scheme proposed by the present invention.
  • a redundant manipulator tracking control device based on echo state network A redundant manipulator tracking control device based on echo state network:
  • At least one memory for storing at least one program
  • the at least one processor When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned echo state network-based redundant manipulator tracking control method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Robotics (AREA)
  • Mechanical Engineering (AREA)
  • Automation & Control Theory (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Physics & Mathematics (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Orthopedic Medicine & Surgery (AREA)
  • Feedback Control In General (AREA)

Abstract

一种基于回声状态网络的冗余机械臂追踪控制方法,包括:初始化机械臂关节角度和基于回声状态网络的类小脑模型的参数,并输入目标轨迹信息;训练前馈神经网络控制器;根据输入的当前时刻关节角变化量和机械臂末端执行器下一时刻目标位置计算目标轨迹补偿信息;计算下一时刻的期望轨迹;测量机械臂末端执行器位置和机械臂关节角度,并输入至训练完成的前馈神经网络,输出下一时刻关节角度;根据下一时刻关节角度控制机械臂运动并追踪目标轨迹;更新基于回声状态网络的类小脑模型的输出权重。该方法不需要知道冗余机械臂的模型信息即可实现追踪控制,且具备较高的追踪控制精确度,可广泛应用于机械臂追踪控制领域。

Description

一种基于回声状态网络的冗余机械臂追踪控制方法 技术领域
本发明涉及机械臂追踪控制领域,尤其涉及一种基于回声状态网络的冗余机械臂追踪控制方法。
背景技术
当机械臂的驱动自由度(如机械臂的关节角个数)多于完成任务所需要的自由度时,机械臂可以被称为冗余机械臂。由于其额外的自由度,冗余机械臂具备更高的灵活性和安全性,因此他们在各个领域中有着广阔的潜在应用前景。然而该类机械臂的额外自由度和结构特点也给机械臂的控制带来了挑战。
技术问题
提出一种基于回声状态网络的冗余机械臂追踪控制方法,来解决已有的冗余机械臂控制基于雅可比矩阵伪逆的方法需要知道准确的机械臂模型信息(如连杆长度),从而计算雅可比矩阵的值,而各种不确定性因素的存在会影响模型信息的准确性,从而影响控制方案的准确性的问题。
技术解决方案
为了解决上述问题,本发明的目的是提供一种基于回声状态网络的冗余机械臂追踪控制方法,不需要知道冗余机械臂的模型信息,仅需将末端执行器位置信息作为反馈,不需要末端执行器的速度和加速度信息,并且具备较高的追踪控制精确度。
本发明所采用的第一技术方案是:一种基于回声状态网络的冗余机械臂追踪控制方法,包括以下步骤:
S1、初始化机械臂关节角度θ(0)和基于回声状态网络的类小脑模型的参数,并输入目标轨迹信息p d
S2、采集机械臂数据并训练前馈神经网络控制器,得到训练完成的前馈神经网络;
S3、基于回声状态网络的类小脑模型根据输入的当前时刻关节角变化量Δθ和机械臂末端执行器下一时刻目标位置p d(n+1)计算目标轨迹补偿信息y(n+1);
S4、根据目标轨迹补偿信息y(n+1)计算下一时刻的期望轨迹
Figure PCTCN2022092420-appb-000001
并输入至训练完成的前馈神经网络;
S5、测量机械臂末端执行器位置p a(n)和机械臂关节角度θ(n),并输入至训练完成的前馈 神经网络,输出下一时刻关节角度θ(n+1);
S6、根据下一时刻关节角度θ(n+1)控制机械臂运动并追踪目标轨迹;
S7、测量机械臂末端执行器当前位置p a(n+1)并更新声状态网络的类小脑模型的输出权重W out
S8、循环步骤S2~S7直至完成目标轨迹的追踪控制。
进一步,所述采集机械臂数据并训练前馈神经网络控制器,得到训练完成的前馈神经网络这一步骤,其具体还包括,其具体还包括:
S21、记录第k个时刻的关节角度θ(k)和机械臂末端执行器位置p a(k);
S22、控制机械臂随机运动并得到关节角度变化量Δθ;
S23、更新第k+1个时刻的关节角度θ(k)+1)=θ(k)+Δθ,并使用第k+1个时刻关节角θ(k+1)控制机械臂运动,测量末端执行器位置p a(k+1);
S24、存储一组训练数据,其中输入为[p a(k);θ(k);p a(k+1)],训练的输出为θ(k+1),修改计时变量k=k+1;
S25、循环步骤S21~S24直至k大于预设阈值,得到训练集;
S26、以第k个时刻的关节角度θ(k)、第k个时刻的末端执行器实际位置p a(k)以及第k+1个时刻的末端执行器位置p a(k+1)为输入,以第k+1个时刻的机械臂关节角度θ(k+1)为输出,基于训练集训练前馈神经网络控制器,得到训练完成的前馈神经网络。
进一步,所述基于回声状态网络的类小脑模型包括传入单元、粒状细胞、浦肯野细胞和传出单元。
进一步,所述基于回声状态网络的类小脑模型根据输入的当前时刻关节角变化量Δθ和机械臂末端执行器下一时刻目标位置p d(n+1)计算目标轨迹补偿信息y(n+1)这一步骤,其具体包括:
S31、将第n个时刻的机械臂关节角变化量Δθ(n)和机械臂末端执行器第n+1个时刻目标位置p d(n+1)拼接,作为类小脑模型的传入单元;
S32、更新粒状细胞在第n+1个时刻的状态x(n+1);
S33、通过平行纤维将粒状细胞在第n+1个时刻的状态传递给浦肯野细胞;
S34、计算传出单元在第n+1个时刻的状态y(n+1),得到第n+1个时刻的目标轨迹补偿量。
进一步,所述更新粒状细胞在第n+1个时刻的状态的更新x(n+1)公式如下:
x(n+1)=f(W inu(n+1)+Wx(n)),
上式中,f(·)=[f 1,f 2,…,f N] T表示粒状细胞的输出激活函数,W in表示传入单元和粒状细胞之间输入连接权重,W表示粒状细胞内部节点之间连接权重。
进一步,所述计算传出单元在第n+1个时刻的状态y(n+1)的计算公式如下:
y(n+1)=tanh(W outp(n+1)).
上式中,tanh表示双曲正切激活函数,W out表示浦肯野细胞和传出单元之间的输出连接权重,p(n+1)表示浦肯野细胞在第n+1个时刻的状态。
进一步,所述测量机械臂末端执行器当前位置p a(n+1)并更新声状态网络的类小脑模型的输出权重W out这一步骤,其具体包括:
测量机械臂末端执行器当前位置p a(n+1)并根据目标位置p d(n+1)和末端执行器当前位置p a(n+1)计算末端执行器的位置误差e;
根据位置误差e计算类小脑模型输出权重的变化量ΔW out
根据输出权重的变化量ΔW out对类小脑模型的输出权重W out进行更新。
进一步,所述计算类小脑模型输出权重的变化量ΔW out的计算公式:
ΔW out=ep T(n+1)
上式中,T表示向量转置。
有益效果
本发明通过本发明将类小脑控制器与数据驱动的学习方法进行融合,使得本方案能够不需要知道机械臂的模型参数信息,在一定程度上应对各种不确定性因素的影响,从而有效地完成冗余机械臂的追踪控制任务。
附图说明
图1是本发明一种基于回声状态网络的冗余机械臂追踪控制方法的步骤流程图;
图2是本发明具体实施例的控制示意图;
图3是本发明具体实施例前馈神经网络训练流程示意图;
图4是本发明具体实施例任务空间中的目标轨迹和机械臂末端执行器的实际轨迹(第一个任务周期)示意图;
图5是本发明具体实施例任务空间中的目标轨迹和机械臂末端执行器的实际轨迹(第十个任务周期)示意图;
图6是本发明具体实施例控制冗余机械臂完成十个周期的轨迹追踪任务的误差变化。
本发明的实施方式
下面结合附图和具体实施例对本发明做进一步的详细说明。对于以下实施例中的步骤编号,其仅为了便于阐述说明而设置,对步骤之间的顺序不做任何限定,实施例中的各步骤的执行顺序均可根据本领域技术人员的理解来进行适应性调整。
参照图1和图2,本发明提供了一种基于回声状态网络的冗余机械臂追踪控制方法,包括以下步骤:
S1、初始化机械臂关节角度θ(0)和基于回声状态网络的类小脑模型的参数,并输入目标轨迹信息p d
具体地,以一个6自由度冗余机械臂为例,任务空间的维度为3,给定机械臂的初始关节角度θ(0)=[1.675;2.843;-3.216;4.187;-1.71;0.77]弧度。然后对类小脑模型进行初始化,其中粒状细胞的单元数N=400,传入单元和传出单元的数量随具体的输入输出维度而确定,粒状细胞的初始状态向量x(0)均设为0。输入权重矩阵W in和内部连接权重矩阵W都是随机初始化为[-1,1]之间的数值,并且在模型工作过程中保持不变。假设内部连接权重被随机初始化为
Figure PCTCN2022092420-appb-000002
则缩放后的内部连接权重矩阵为
Figure PCTCN2022092420-appb-000003
其中缩放因子α需稍小于
Figure PCTCN2022092420-appb-000004
即可,
Figure PCTCN2022092420-appb-000005
表示矩阵
Figure PCTCN2022092420-appb-000006
的谱半径。输出权重矩阵W out的值全部被初始化为0,输出权重矩阵更新的学习率设为η=0.001。最后,在任务空间中定义目标轨迹的数学表达式
Figure PCTCN2022092420-appb-000007
其中ι=0.1米表示目标轨迹的半径,T d=20秒表示追踪任务的周期,其中时间采样间隔设为τ=0.05秒。
S2、采集机械臂数据并训练前馈神经网络控制器,得到训练完成的前馈神经网络;
具体地,控制机械臂随机地运动,采集机械臂数据,训练一个三层的前馈神经网络控制器,该机械臂控制器为近似的逆运动学求解器,使用该控制器控制冗余机械臂时会存在较大的误差,因此需要使用一个类小脑模型来改善控制器的精确度。
S3、基于回声状态网络的类小脑模型根据输入的当前时刻关节角度变化量Δθ和机械臂末端执行器下一时刻目标位置p d(n+1)计算目标轨迹补偿信息y(n+1);
具体地,当前时刻即第n个时刻,当n为0的时候,关节角变化量Δθ被初始化为零向量,当n大于0的时候,关节角变化量的计算方式为Δθ=θ(n)-θ(n-1),其中θ(n)和θ(n-1)分别表示第n个时刻和第n-1个时刻的前馈神经网络的输出。
S4、根据目标轨迹补偿信息y(n+1)计算下一时刻的期望轨迹
Figure PCTCN2022092420-appb-000008
并输入至训练完成的前馈神经网络;
S5、测量机械臂末端执行器位置p a(n)和机械臂关节角度θ(n),并输入至训练完成的前馈神经网络,输出下一时刻关节角度θ(n+1);
具体地,测量位置和关节角度采用传感器设备测量。
S6、根据下一时刻关节角度θ(n+1)控制机械臂运动并追踪目标轨迹;
S7、测量机械臂末端执行器当前位置p a(n+1)并更新声状态网络的类小脑模型的输出权重W out
S8、循环步骤S2~S7直至完成目标轨迹的追踪控制。
进一步作为本方法的优选实施例,所参照图3,述采集机械臂数据并训练前馈神经网络控制器,得到训练完成的前馈神经网络这一步骤,其具体还包括:
S21、记录第k个时刻的关节角度θ(k)和机械臂末端执行器位置p a(k);
S22、控制机械臂随机运动并得到关节角度变化量Δθ;
S23、更新第k+1个时刻的关节角度θ(k+1)=θ(k)+Δθ,并使用第k+1个时刻关节角θ(k+1)控制机械臂运动,测量末端执行器位置p a(k+1);
S24、存储一组训练数据,其中输入为[p a(k);θ(k);p a(k+1)],训练的输出为θ(k+1),修改计时变量k=k+1;
S25、循环步骤S21~S24直至k大于预设阈值,得到训练集;
S26、以第k个时刻的关节角度θ(k)、第k个时刻的末端执行器实际位置p a(k)以及第k+1个时刻的末端执行器位置p a(k+1)为输入,以第k+1个时刻的机械臂关节角度θ(k+1)为输出,基于训练集训练前馈神经网络控制器,得到训练完成的前馈神经网络。
进一步作为本方法的优选实施例,所述基于回声状态网络的类小脑模型包括传入单元
Figure PCTCN2022092420-appb-000009
粒状细胞
Figure PCTCN2022092420-appb-000010
浦肯野细胞
Figure PCTCN2022092420-appb-000011
和传出单元
Figure PCTCN2022092420-appb-000012
其中传入层包括K个单元,粒状细胞包括N个单元,浦肯野细胞包括N个单元,传出层包括L个单元,t表示时刻。在传入单元和粒状细胞之间存在输入连接权重
Figure PCTCN2022092420-appb-000013
在粒状细胞内部节点之间存在连接权重
Figure PCTCN2022092420-appb-000014
在粒状细胞和浦肯野细胞之间存在平行纤维,而在浦肯野细胞和传出单元间则存在输出连接权重
Figure PCTCN2022092420-appb-000015
首先,所有粒状细胞的状态被初始化为
Figure PCTCN2022092420-appb-000016
类小脑模型工作过程中,粒状细胞的各个单元在第n+1个时刻的状态更新公式如下:
x(n+1)=f(W inu(n+1)+Wx(n)),
上式中,W in表示传入单元和粒状细胞之间输入连接权重,W表示粒状细胞内部节点之间连接权重,f(·)=[f 1,f 2,…,f N] T表示粒状细胞的输出激活函数,通常采用sigmoid激活函数:
Figure PCTCN2022092420-appb-000017
浦肯野细胞则通过平行纤维与粒状细胞相连,所以浦肯野细胞在第n+1个时刻的状态为p(n+1)=x(n+1)。而传出单元的状态的计算方式为y(n+1)=tanh(W outp(n+1)),其中tanh(·)为双曲正切激活函数,其定义为:
Figure PCTCN2022092420-appb-000018
输入权重W in和内部连接权重W都是随机初始化为[-1,1]之间的数值,并且在模型工作过程中保持不变。值得注意的是,根据已有的理论,为了保证模型能正常工作,通常需要对内部连接权重W进行缩放。假设内部连接权重被随机初始化为
Figure PCTCN2022092420-appb-000019
则缩放后的内部连接权重矩阵为
Figure PCTCN2022092420-appb-000020
其中缩放因子α需稍小于
Figure PCTCN2022092420-appb-000021
即可,
Figure PCTCN2022092420-appb-000022
表示矩阵
Figure PCTCN2022092420-appb-000023
的谱半径。在类小脑模型工作过程中,需要训练的是输出层权重W out,首先需要将W out初始化为值全部为零的矩阵,然后对输出权重进行训练。W out的训练方式可分为离线训练和在线训练两种方式,其中离线训练是指在n个连续的时刻给定输入信号U=[u(1),u(2),…,u(n)],然后收集每个时刻对应的粒状细胞状态X=[x(1),x(2),…,x(n)]和期望的输出Y d=[y d(1),y d(2),…,y d(n)],最后使用最小二乘法计算出满足一下要求的W out
进一步作为本方法优选实施例,所述基于回声状态网络的类小脑模型根据输入的当前时刻关节角变化量Δθ和机械臂末端执行器下一时刻目标位置p d(n+1)计算目标轨迹补偿信息y(n+1)这一步骤,其具体包括:
S31、将第n个时刻的机械臂关节角变化量Δθ(n)和机械臂末端执行器第n+1个时刻目标位置p d(n+1)拼接,作为类小脑模型的传入单元;
具体地,拼接公式如下:
u(n+1)=[Δθ(n);p d(n+1)].
S32、更新粒状细胞在第n+1个时刻的状态x(n+1);
S33、通过平行纤维将粒状细胞在第n+1个时刻的状态传递给浦肯野细胞;
S34、计算传出单元在第n+1个时刻的状态y(n+1),得到第n+1个时刻的目标轨迹补偿量。
进一步作为本方法优选实施例,所述测量机械臂末端执行器当前位置p a(n+1)并更新声状态网络的类小脑模型的输出权重W out这一步骤,其具体包括:
测量机械臂末端执行器当前位置p a(n+1)并根据目标位置p d(n+1)和末端执行器当前位置p a(n+1)计算末端执行器的位置误差e,如下式:
e=p d(n+1)-p a(n+1).
根据位置误差e计算类小脑模型输出权重的变化量ΔW out,如下式:
ΔW out=ep T(n+1)
上式中,T表示向量转置。
根据输出权重的变化量ΔW out对类小脑模型的输出权重W out进行更新,如下式:
W out=W out+ηΔW out
基于以上技术方案,控制冗余机械臂连续完成十个周期的轨迹追踪任务,其中第一个周期和第十个周期的追踪结果如图4和图5所示。在第一个周期中,机械臂末端执行器的初始位置与目标轨迹有一定的距离,可以看出末端执行器的实际移动轨迹很快收敛到目标轨迹上。在第十个周期中,可以看到末端执行器的实际移动轨迹与目标轨迹几乎一致。在这十个任务周期中,每一个任务周期的追踪误差的变化如图6所示,可以看出,在第十个任务周期的时候,追踪误差平均值的数量级为10 -4米。以上结果体现了本发明所提出的方案的有效性。
一种基于回声状态网络的冗余机械臂追踪控制装置:
至少一个处理器;
至少一个存储器,用于存储至少一个程序;
当所述至少一个程序被所述至少一个处理器执行,使得所述至少一个处理器实现如上所述一种基于回声状态网络的冗余机械臂追踪控制方法。
上述方法实施例中的内容均适用于本装置实施例中,本装置实施例所具体实现的功能与上述方法实施例相同,并且达到的有益效果与上述方法实施例所达到的有益效果也相同。
以上是对本发明的较佳实施进行了具体说明,但本发明创造并不限于所述实施例,熟悉本领域的技术人员在不违背本发明精神的前提下还可做作出种种的等同变形或替换,这些等同的变形或替换均包含在本申请权利要求所限定的范围内。

Claims (8)

  1. 一种基于回声状态网络的冗余机械臂追踪控制方法,其特征在于,包括以下步骤:
    S1、初始化机械臂关节角度θ(0)和基于回声状态网络的类小脑模型的参数,并输入目标轨迹信息p d
    S2、采集机械臂数据并训练前馈神经网络控制器,得到训练完成的前馈神经网络;
    S3、基于回声状态网络的类小脑模型根据输入的当前时刻关节角度变化量Δθ和机械臂末端执行器下一时刻目标位置p d(n+1)计算目标轨迹补偿信息y(n+1);
    S4、根据目标轨迹补偿信息y(n+1)计算下一时刻的期望轨迹
    Figure PCTCN2022092420-appb-100001
    并输入至训练完成的前馈神经网络;
    S5、测量机械臂末端执行器位置p a(n)和机械臂关节角度θ(n),并输入至训练完成的前馈神经网络,输出下一时刻关节角度θ(n+1);
    S6、根据下一时刻关节角度θ(n+1)控制机械臂运动并追踪目标轨迹;
    S7、测量机械臂末端执行器当前位置p a(n+1)并更新声状态网络的类小脑模型的输出权重W out
    S8、循环步骤S2~S7直至完成目标轨迹的追踪控制。
  2. 根据权利要求1所述一种基于回声状态网络的冗余机械臂追踪控制方法,其特征在于,所述采集机械臂数据并训练前馈神经网络控制器,得到训练完成的前馈神经网络这一步骤,其具体还包括:
    S21、记录第k个时刻的关节角度θ(k)和机械臂末端执行器位置p a(k);
    S22、控制机械臂随机运动并得到关节角度变化量Δθ;
    S23、更新第k+1个时刻的关节角度θ(k+1)=θ(k)+Δθ,并使用第k+1个时刻关节角θ(k+1)控制机械臂运动,测量末端执行器位置p a(k+1);
    S24、存储一组训练数据,其中输入为[p a(k);θ(k);p a(k+1)],训练的输出为θ(k+1),修改计时变量k=k+1;
    S25、循环步骤S21~S24直至k大于预设阈值,得到训练集;
    S26、以第k个时刻的关节角度θ(k)、第k个时刻的末端执行器实际位置p a(k)以及第k+1个时刻的末端执行器位置p a(k+1)为输入,以第k+1个时刻的机械臂关节角度θ(k+1)为输出,基于训练集训练前馈神经网络控制器,得到训练完成的前馈神经网络。
  3. 根据权利要求2所述一种基于回声状态网络的冗余机械臂跟踪控制方法,其特征在于,所述基于回声状态网络的类小脑模型包括传入单元u(t)、粒状细胞x(t)、浦肯野细胞p(t)和传出单元y(t)。
  4. 根据权利要求3所述一种基于回声状态网络的冗余机械臂追踪控制方法,其特征在于,所述基于回声状态网络的类小脑模型根据输入的当前时刻关节角度变化量Δθ和机械臂末端执行器下一时刻目标位置p d(n+1)计算目标轨迹补偿信息y(n+1)这一步骤,其具体包括:
    S31、将第n个时刻的机械臂关节角变化量Δθ(n)和机械臂末端执行器第n+1个时刻目标位置p d(n+1)拼接,作为类小脑模型的传入单元;
    S32、更新粒状细胞在第n+1个时刻的状态x(n+1);
    S33、通过平行纤维将粒状细胞在第n+1个时刻的状态传递给浦肯野细胞;
    S34、计算传出单元在第n+1个时刻的状态y(n+1),得到第n+1个时刻的目标轨迹补偿量。
  5. 根据权利要求4所述一种基于回声状态网络的冗余机械臂追踪控制方法,其特征在于,所述更新粒状细胞在第n+1个时刻的状态x(n+1)的更新公式如下:
    x(n+1)=f(W inu(n+1)+Wx(n)),
    上式中,f(·)=[f 1,f 2,…,f N] T表示粒状细胞的输出激活函数,W in表示传入单元和粒状细胞之间输入连接权重,W表示粒状细胞内部节点之间连接权重。
  6. 根据权利要求5所述一种基于回声状态网络的冗余机械臂追踪控制方法,其特征在于,所述计算传出单元在第n+1个时刻的状态y(n+1)的计算公式如下:
    y(n+1)=tanh(W outp(n+1)).
    上式中,tanh表示双曲正切激活函数,W out表示浦肯野细胞和传出单元之间的输出连接权重,p(n+1)表示浦肯野细胞在第n+1个时刻的状态。
  7. 根据权利要求6所述一种基于回声状态网络的冗余机械臂追踪控制方法,其特征在于,所述测量机械臂末端执行器当前位置p a(n+1)并更新声状态网络的类小脑模型的输出权重W out这一步骤,其具体包括:
    测量机械臂末端执行器当前位置p a(n+1)并根据目标位置p d(n+1)和末端执行器当前位置p a(n+1)计算末端执行器的位置误差e;
    根据位置误差e计算类小脑模型输出权重的变化量ΔW out
    根据输出权重的变化量ΔW out对类小脑模型的输出权重W out进行更新。
  8. 根据权利要求7所述一种基于回声状态网络的冗余机械臂追踪控制方法,其特征在于,所述计算类小脑模型输出权重的变化量ΔW out的计算公式:
    ΔW out=ep T(n+1)
    上式中,T表示向量转置。
PCT/CN2022/092420 2021-08-18 2022-05-12 一种基于回声状态网络的冗余机械臂追踪控制方法 WO2023020036A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110946820.XA CN113650014B (zh) 2021-08-18 2021-08-18 一种基于回声状态网络的冗余机械臂追踪控制方法
CN202110946820.X 2021-08-18

Publications (1)

Publication Number Publication Date
WO2023020036A1 true WO2023020036A1 (zh) 2023-02-23

Family

ID=78480771

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/092420 WO2023020036A1 (zh) 2021-08-18 2022-05-12 一种基于回声状态网络的冗余机械臂追踪控制方法

Country Status (2)

Country Link
CN (1) CN113650014B (zh)
WO (1) WO2023020036A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113650014B (zh) * 2021-08-18 2022-05-17 中山大学 一种基于回声状态网络的冗余机械臂追踪控制方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107263481A (zh) * 2017-07-27 2017-10-20 青岛格莱瑞智能控制技术有限公司 一种多自由度机器人的类脑学习控制方法
CN107972031A (zh) * 2017-11-10 2018-05-01 浙江科技学院 一种冗余机械臂可重复运动的初始位置定位方法
CN110977992A (zh) * 2020-01-02 2020-04-10 中山大学 一种面向机械臂的无运动学模型轨迹跟踪方法及一种机械臂系统
US20210034959A1 (en) * 2017-03-22 2021-02-04 Larsx Continuously learning and optimizing artificial intelligence (ai) adaptive neural network (ann) computer modeling methods and systems
CN112894812A (zh) * 2021-01-21 2021-06-04 中山大学 一种面向机械臂的视觉伺服轨迹跟踪控制方法及系统
CN113650014A (zh) * 2021-08-18 2021-11-16 中山大学 一种基于回声状态网络的冗余机械臂追踪控制方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10686327B2 (en) * 2018-04-13 2020-06-16 Honeywell International Inc. Energy storage controller
CN109159151B (zh) * 2018-10-23 2021-12-10 北京无线电测量研究所 一种机械臂空间轨迹跟踪动态补偿方法和系统
CN112605996B (zh) * 2020-12-16 2021-12-24 中山大学 一种面向冗余机械臂的无模型碰撞避免控制方法
CN112706165A (zh) * 2020-12-22 2021-04-27 中山大学 一种面向轮式移动机械臂的跟踪控制方法及系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210034959A1 (en) * 2017-03-22 2021-02-04 Larsx Continuously learning and optimizing artificial intelligence (ai) adaptive neural network (ann) computer modeling methods and systems
CN107263481A (zh) * 2017-07-27 2017-10-20 青岛格莱瑞智能控制技术有限公司 一种多自由度机器人的类脑学习控制方法
CN107972031A (zh) * 2017-11-10 2018-05-01 浙江科技学院 一种冗余机械臂可重复运动的初始位置定位方法
CN110977992A (zh) * 2020-01-02 2020-04-10 中山大学 一种面向机械臂的无运动学模型轨迹跟踪方法及一种机械臂系统
CN112894812A (zh) * 2021-01-21 2021-06-04 中山大学 一种面向机械臂的视觉伺服轨迹跟踪控制方法及系统
CN113650014A (zh) * 2021-08-18 2021-11-16 中山大学 一种基于回声状态网络的冗余机械臂追踪控制方法

Also Published As

Publication number Publication date
CN113650014A (zh) 2021-11-16
CN113650014B (zh) 2022-05-17

Similar Documents

Publication Publication Date Title
CN112571420B (zh) 一种未知参数下的双功能模型预测控制方法
CN108555914B (zh) 一种基于腱驱动灵巧手的dnn神经网络自适应控制方法
CN109352656B (zh) 一种具有时变输出约束的多关节机械臂控制方法
CN109445274B (zh) 一种柔性空间机械臂振动控制方法及系统
CN116460860B (zh) 一种基于模型的机器人离线强化学习控制方法
WO2023020036A1 (zh) 一种基于回声状态网络的冗余机械臂追踪控制方法
CN115990888B (zh) 一种具有死区和时变约束功能的机械臂控制方法
CN114942593A (zh) 一种基于干扰观测器补偿的机械臂自适应滑模控制方法
CN114211503B (zh) 基于视觉反馈的绳驱柔性机器人轨迹控制方法及系统
CN110682290B (zh) 一种基于动量观测器的闭环机械臂系统碰撞检测方法
CN114310851B (zh) 一种机器人免力矩传感器的拖动示教方法
CN112223276B (zh) 基于自适应神经网络滑模控制的多关节机器人控制方法
CN111872933B (zh) 一种基于改进二次型迭代学习控制的scara机器人轨迹跟踪控制方法
CN116638507A (zh) 一种自适应阻抗控制与预测控制相结合的遥操作控制方法
Cui et al. Experiment on impedance adaptation of under-actuated gripper using tactile array under unknown environment
CN107894709A (zh) 基于自适应评价网络冗余机器人视觉伺服控制
CN113219841B (zh) 基于自适应鲁棒的水下多关节液压机械臂非线性控制方法
CN116300468A (zh) 一种基于神经网络类机理建模的导弹制导控制一体化方法
CN111158238B (zh) 一种基于粒子群算法的力反馈设备动力学参数估计算法
CN113325716A (zh) 基于扩张观测器的水下液压机械臂非线性鲁棒控制方法
Shaoming et al. Research on trajectory tracking control of multiple degree of freedom manipulator
Heyu et al. Impedance control method with reinforcement learning for dual-arm robot installing slabstone
Wang et al. Adaptive Impedance Control in Uncertain Environment for Uncertain Manipulator
Song et al. Neural network model based control of a flexible link manipulator
CN113325711B (zh) 一种用于柔性机械臂预定精度定位的智能控制方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22857353

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE