CN117930663B - Quadruped Robot Motion Control System Based on Eight-element Neural Network - Google Patents

Abstract

The present invention discloses a quadruped robot motion control system based on an eight-element neural network, and belongs to the field of quadruped robot motion control. The system includes: a signal control module, which is used to generate gait control parameters; a signal generation module, which is used to receive the gait control parameters generated by the signal control module, and input them into the eight-element neural network, respectively solve the ordinary differential equations of eight neurons, and obtain gait rhythm signals; a signal post-processing module, which is used to receive the gait rhythm signals solved in the signal generation module, and convert them into displacement signals corresponding to the actuation of a total of eight joints on the four legs of the quadruped robot. The present invention designs the rhythmicity of the gait of the central pattern generator and the phase relationship of the hip and knee joints based on the principle of symmetry, realizes low computing power and high reliability control of multiple joints of the quadruped robot, and improves the mobility and environmental adaptability of the quadruped robot under five gaits.

Description

Quadruped Robot Motion Control System Based on Eight-element Neural Network

Technical Field

The invention belongs to the field of quadruped robot motion control, and in particular relates to a quadruped robot motion control system based on an octal neural network.

Background technique

In recent years, compared with wheeled robots, legged robots have shown better mobility, terrain adaptability and stability in unstructured environments, and are suitable for complex tasks such as military exploration and disaster search and rescue. Among them, quadruped robots have shown better mobility and stability among all legged robots. It can not only perform tasks flexibly in various complex terrain environments, but also walk stably in unstructured terrains such as mud and grass.

At present, the common methods in the field of quadruped robot control are model-based control, learning-based control and central pattern generator control. The model-based control method is a classic control method in the field of robotics, and is often used for precise control of large quadruped robots, such as the Cheetah robot and the Big Dog robot. This control method requires precise modeling and motion planning of the robot model and environment, and has disadvantages such as complex calculations and poor environmental adaptability. Learning-based control is a method that uses methods such as reinforcement learning to design and improve control strategies. It can achieve precise control of robots in complex, nonlinear or uncertain environments, but this method is highly dependent on data and training environments, and still faces huge challenges when migrating models to real environments.

Central Pattern Generator (CPG) is a small neural network formed by a series of interacting neuron models or oscillators coupled together. It has been widely shown to exist in the central nervous system of vertebrates and ganglia of invertebrates. CPG can generate basic signals for rhythmic behaviors (such as breathing or movement) without sensory feedback, but sensory feedback is required to regulate CPG signals. In the field of robotic control, CPG is widely used to control the movement and behavior of various robotic systems. For example: salamander robots used to study salamander swimming and walking, motion control of bipedal quadrupedal and hexapod robots combined with sensory feedback, motion control of pneumatic quadrupedal robots without electronics, and combined with reinforcement learning to control the motion of quadrupedal robots and soft snake-like robots.

Compared with the other two methods, the biggest advantage of CPG is its simple structure and high computational efficiency. It only needs to calculate the ordinary differential equations of the power system to achieve coordinated rhythmic motion control of multiple joints of the robot. Not only that, CPG also has inherent stability and adaptability.

At present, most CPG networks for quadruped motion control are usually four-neuron networks. It has a simple structure, is easy to integrate into motion controllers, and has good compatibility with sensors and other learning algorithms. But it also has some disadvantages. First, the number of gait rhythms that can be generated is limited. Common gaits of quadrupeds include: walk, trot, pace, bound, pronk, jump, etc., and the phase relationship is shown in a~f in Figure 1. Most four-neuron networks can only achieve no more than three gait types (generally walk, trot, and bound). Secondly, the number of controllable joints is limited (generally four). At present, the most common design of quadruped robots is that each leg has three joints: hip abduction-adduction, hip flexion-extension, and knee flexion-extension. However, in all four-neuron or even eight-neuron CPGs, at most only four signals can be generated to control the phase relationship between the four legs. A mapping method is usually required to map the signal of a single neuron into two joint position signals to achieve the phase relationship between the knee and hip joints. From a biological perspective, the signals that control multiple joints are likely to be generated by specific neurons, which requires a more complex network architecture.

Designing a more complex network architecture than the four-neuron network does not weaken its advantages, and the network with additional neurons can also contain more symmetries, thereby generating more gait types. In addition, the mechanism of achieving coordinated control of the hip and knee joints through the intrinsic characteristics of the network can deepen the understanding of the gait control mechanism of quadrupeds, which will also help promote related research in robotics and biology.

One feasible method for designing CPG networks is the principle of symmetry. If the network is symmetrical, the constraints imposed by the symmetry on the dynamics usually lead to neuronal synchronization or phase locking. The rhythmic nature of the gait here is essentially a spatiotemporal symmetry of time domain signals between joints. The symmetry of the network structure can constrain the symmetry of CPG neuron signals. However, how to specifically design the rhythmicity of the CPG gait and the phase relationship of the hip and knee joints to achieve low-computing power and high-maneuverability control of multiple joints of a quadruped robot, and then achieve multiple gaits, thereby improving the maneuverability and environmental adaptability of the robot system, is still lacking in the current existing technology. Efficient solutions.

Summary of the invention

The purpose of the present invention is to solve the problem of limited number of degrees of freedom and types of rhythmic signals of quadruped robot control in the prior art, and to provide a quadruped robot motion control system based on an octal neural network to achieve eight-degree-of-freedom coordinated control of the hip and knee joints of the quadruped robot, while generating five types of gait rhythmic signals, thereby improving the environmental adaptability of the system.

The specific technical solutions adopted by the present invention are as follows:

A quadruped robot motion control system based on an eight-element neural network, comprising:

A signal control module for generating gait control parameters;

The signal generation module is used to receive the gait control parameters generated by the signal regulation module and input them into the eight-element neural network, solve the ordinary differential equations of the eight neurons respectively, and obtain the gait rhythm signal; the eight-element neural network is composed of a first unidirectional coupling network layer and a second unidirectional coupling network layer, each of which has a four-fold rotational symmetry; the first unidirectional coupling network layer is composed of a first neuron, a third neuron, a second neuron, and a fourth neuron connected end to end in sequence to form a unidirectional coupling ring network, and the second unidirectional coupling network layer is composed of a fifth neuron, an eighth neuron, a sixth neuron, and a seventh neuron connected end to end in sequence to form a unidirectional coupling ring network, and the coupling direction of the neurons in the first unidirectional coupling network layer is consistent with that of the neurons in the second unidirectional coupling network layer. The coupling directions of neurons in the network layer are opposite; the first unidirectional coupling network layer and the second unidirectional coupling network layer are grouped in pairs, and bidirectional coupling is formed between the two unidirectional coupling network layers, and the four groups of neurons correspond to each other to control the four legs of the quadruped robot; and in each group of neurons, the neurons located in the first unidirectional coupling network layer are used to control the hip joint, and the neurons located in the second unidirectional coupling network layer are used to control the knee joint; in the ordinary differential equation corresponding to the eight-element neural network, the driving signal of each neuron is determined by three parts, namely, the gait control parameters generated by the signal regulation module, the coupling effect of neurons from the same unidirectional coupling network layer, and the coupling effect of neurons from another unidirectional coupling network layer;

The signal post-processing module is used to receive the gait rhythm signal solved in the signal generation module and convert it into a displacement signal corresponding to the eight joints on the four legs of the quadruped robot.

Preferably, the gait forms that can be controlled by the quadruped robot include walking, trotting, strolling, jumping and leaping.

Preferably, in the octal neural network, the first neuron and the fifth neuron form a group, which are respectively used to control the hip joint and knee joint of the left hind leg, the second neuron and the sixth neuron form a group, which are respectively used to control the hip joint and knee joint of the right hind leg, the third neuron and the seventh neuron form a group, which are respectively used to control the hip joint and knee joint of the right front leg, and the fourth neuron and the eighth neuron form a group, which are respectively used to control the hip joint and knee joint of the left front leg.

Preferably, in the eight-element neural network, the couplings between neurons are all inhibitory couplings. Further, the first neuron inhibits the third neuron, the third neuron inhibits the second neuron, the second neuron inhibits the fourth neuron, the fourth neuron inhibits the first neuron, the fifth neuron inhibits the eighth neuron, the eighth neuron inhibits the sixth neuron, the sixth neuron inhibits the seventh neuron, the seventh neuron inhibits the fifth neuron, the first neuron and the fifth neuron are bidirectionally inhibited, the second neuron and the sixth neuron are bidirectionally inhibited, the third neuron and the seventh neuron are bidirectionally inhibited, and the fourth neuron and the eighth neuron are a group of bidirectionally inhibited.

Preferably, in the eight-element neural network, the ordinary differential equation of any neuron i adopts the Stein neuron model, wherein the driving signal of any neuron i is ,in , , are all gait control parameters generated by the signal control module, t is time, and are the same-layer weight hyperparameters and different-layer weight hyperparameters, respectively. X is the sum of the membrane potentials of the neurons that inhibit neuron i in the one-way coupled network layer where the current neuron i is located. Y is the sum of the membrane potentials of the neurons that inhibit neuron i in the one-way coupled network layer that does not contain the current neuron i.

Preferably, when solving ordinary differential equations, the same-layer weight hyperparameters of the ordinary differential equations corresponding to the eight neurons are set to the same, and the absolute values of the different-layer weight hyperparameters of the ordinary differential equations corresponding to the neurons in the second unidirectional coupled network layer are greater than the absolute values of the different-layer weight hyperparameters of the ordinary differential equations corresponding to the neurons in the first unidirectional coupled network layer.

Preferably, the same-layer weight hyperparameters of the ordinary differential equations corresponding to the eight neurons are all set to -0.15.

Preferably, the heterogeneous layer weight hyperparameters of the ordinary differential equations corresponding to the neurons in the first unidirectional coupled network layer are all set to -0.1, and the heterogeneous layer weight hyperparameters of the ordinary differential equations corresponding to the neurons in the second unidirectional coupled network layer are all set to -0.6.

Preferably, the ordinary differential equation is numerically solved by a fourth-order Runge-Kutta method.

Preferably, the gait rhythm signal of each neuron solved in the signal generation module is a time domain signal with a numerical range between 0 and 1, and the signal post-processing module converts the time domain signal into a joint position signal or a voltage signal for controlling the actuation of the corresponding joint.

Compared with the prior art, the present invention has the following beneficial effects:

1. Compared with the common CPG quaternary network architecture, the network architecture proposed in the present invention can control more degrees of freedom of the quadruped robot, achieve coordinated control between the hip and knee joints, and maintain phase locking between the hip and knee joints.

2. The network architecture of the present invention can generate more types of gait rhythm signals, making the quadruped robot more adaptable to the environment.

3. Compared with the computational scale and computing power required by model-based methods or learning-based methods, the bionic motion control architecture based on the CPG octal D ₄ symmetry network proposed in the present invention is computationally efficient and simpler. It only needs to set the parameters of the signal control layer and then numerically solve a set of ordinary differential equations to obtain the rhythmic signal of gait.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the gait phase relationship of a quadruped robot;

FIG2 is a schematic diagram of a three-layer bionic motion control architecture used in the present invention;

FIG3 is a schematic diagram of an eight-element neural network architecture used in the present invention;

Figure 4 shows the simulation results in an example of the present invention, where (a)-(e) are the numerical simulation results of the output state quantities of the two-layer network and the phase diagrams of the hip and knee joint neurons under five gaits, and (f) is the stability test result after the disturbance is injected.

Detailed ways

In order to make the above-mentioned purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation mode of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined accordingly without conflicting with each other.

In the description of the present invention, it is to be understood that when an element is considered to be "connected" to another element, it may be directly connected to the other element or indirectly connected, that is, there are intermediate elements. On the contrary, when an element is said to be "directly" connected to another element, there are no intermediate elements.

In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.

In a preferred embodiment of the present invention, a quadruped robot motion control system based on an eight-element neural network is provided, which can generate different gait rhythm signals and maintain the phase lock of the hip and knee joints by constructing an eight-neuron central pattern generator (CPG) network, thereby realizing low computing power and high maneuverability control of the quadruped robot. Moreover, the common four-element network generally generates three gait rhythm signals. The present invention designs an eight-element network composed of eight neurons (i.e., the first neuron to the eighth neuron) based on the principle of symmetry, which can generate five gait rhythm signals, thereby improving the environmental adaptability of the system.

As shown in Figure 2, the quadruped robot motion control system adopts a three-layer bionic motion control architecture, namely the signal regulation layer, the signal generation layer and the signal post-processing layer. The three-layer architecture can be implemented through corresponding functional modules, which are described in detail below.

The first layer of control architecture is the signal control module, which is used to generate gait control parameters. In the embodiment of the present invention, the signal control module can imitate the way the mesencephalic locomotor region (MLR) of the biological hypothalamus generates gait command signals, and adjusts the subsequent drive signals by generating control parameters for regulating CPG. , to regulate the gait rhythmic signals of the CPG network.

The second-level control architecture is the signal generation module, which receives the gait control parameters generated by the signal control module and inputs them into the eight-element neural network. It solves the ordinary differential equations of the eight neurons separately to obtain the gait rhythm signal.

The third-level control architecture is the signal post-processing module, which is used to receive the gait rhythm signal solved in the signal generation module and convert it into a displacement signal corresponding to the eight joints on the four legs of the quadruped robot.

In an embodiment of the present invention, eight Stein neurons can be selected to construct a CPG eight-element symmetric neural network, that is, as a biological neural circuit, it accepts modulation parameters from a signal control module to generate rhythmic signals of different gaits and gait switching. Since the eight-element neural network is the core of the present invention to realize the motion control of a quadruped robot, in order to facilitate the understanding of the principles and advantages of the network, the design ideas of the eight-element neural network are described in detail through two design links below. In addition, in order to facilitate subsequent descriptions, the first neuron, the second neuron, the third neuron, the fourth neuron, the fifth neuron, the sixth neuron, the seventh neuron, and the eighth neuron in the eight-element neural network are named neuron 1, neuron 2, neuron 3, neuron 4, neuron 5, neuron 6, neuron 7, and neuron 8, respectively, but it can be understood that the numbering and naming of neurons does not essentially affect the implementation effect of the technical solution.

Design stage 1: Design the global symmetry of gait rhythm

Since the present invention needs to design a CPG network architecture for quadruped robot motion control, it can realize walking, trotting, pacing, bounding and pronk gaits. In a network with Z ₄ symmetry (i.e., four-fold cyclic symmetry, also known as four-fold periodic symmetry), the spatiotemporal symmetry of trotting and bounding gaits is always conjugated, which means that trotting and bounding gaits cannot coexist in a single Z ₄ network. Therefore, the present invention designs the global symmetry of the network as D ₄ , which is a four-element network with bidirectional coupling, indicating that the network has four-fold rotational symmetry (also known as four-sided symmetry or square symmetry). In this network, four neurons 1, neuron 2, neuron 3 and neuron 4 are used to control the left hind leg (LH), right hind leg (RH), right front leg (RF) and left front leg (LF) of the quadruped robot respectively. The generators of the D ₄ group are ω= (1324) and κ= (13)(24). The group elements of D ₄ can be expressed as:

In the symmetry group D ₄ of the above-mentioned bidirectional coupled four-element network, the cyclic quotient subgroup and isotropic subgroup of the five gaits required by the present invention are found. Table 1 summarizes the subgroup selection for all gaits, where _xi represents the state of the neuron and lists the phase relationship of other neurons relative to neuron No. 1 in different gaits. It is thus proved that the bidirectional coupled four-neuron network with D ₄ symmetry meets the spatiotemporal symmetry requirements of all desired gaits. Therefore, the global symmetry of the eight-element CPG network architecture that needs to be designed is set to D ₄ symmetry.

Table 1 Subgroup selection of all gaits

Design stage 2: Designing the local symmetry of hip-knee phase locking

In order to achieve coordinated control of the hip and knee joints of the quadruped robot and maintain the corresponding phase relationship, it is necessary to further solve two problems in design link 2 based on design link 1:

First, the bidirectionally coupled four-element network designed in the above link is expanded to an eight-element network, and the D ₄ symmetry of the network must be maintained.

Second, the local symmetry of neurons in the hip and knee joints is designed to achieve a locked phase between the hip and knee joints.

Therefore, based on the above bidirectionally coupled quaternary D ₄ network, a single neuron is split into a group of two neurons to increase the number of neurons in the network to 8. In each neuron group, two neurons establish a bidirectional connection to ensure that the local symmetry is Z ₂ while ensuring the global symmetry. In the extended eight-neuron network, the generator corresponding to the local symmetry is λ =(15)(26)(37)(48), so the generator of the eight-neuron network is:

Finally, the eight-neuron network obtained after the expansion of the present invention is shown in Figure 3. The eight-element neural network is divided into two layers, the upper layer of neurons is used to control the hip joint, which is called the first unidirectional coupling network layer, and the lower layer of neurons is used to control the knee joint, which is called the second unidirectional coupling network layer. The first unidirectional coupling network layer of the upper network layer and the second unidirectional coupling network layer of the lower network layer both have _D4 symmetry. The first unidirectional coupling network layer is formed by the first neuron, the third neuron, the second neuron, and the fourth neuron connected end to end to form a unidirectional coupling ring network, that is, the first neuron points to the third neuron, the third neuron points to the second neuron, the second neuron points to the fourth neuron, and the fourth neuron points to the first neuron. Similarly, the second unidirectional coupling network layer is formed by the fifth neuron, the eighth neuron, the sixth neuron, and the seventh neuron connected end to end to form a unidirectional coupling ring network, that is, the fifth neuron points to the eighth neuron, the eighth neuron points to the sixth neuron, the sixth neuron points to the seventh neuron, and the seventh neuron points to the fifth neuron. As a result, the coupling direction of the neurons in the first unidirectional coupling network layer is opposite to the coupling direction of the neurons in the second unidirectional coupling network layer. In addition, in the eight-element neural network, two neurons form a group between the first unidirectional coupling network layer and the second unidirectional coupling network layer, and a group of two neurons comes from different unidirectional coupling network layers, and the two neurons are bidirectionally connected, forming a bidirectional coupling between the two layers of unidirectional coupling network layers to maintain the equivalence of the neurons corresponding to the hip joint and the neurons corresponding to the knee joint. The four groups of neurons correspond one by one to control the four legs of the quadruped robot; and in each group of neurons, the neurons located in the first unidirectional coupling network layer are used to control the hip joint, and the neurons located in the second unidirectional coupling network layer are used to control the knee joint. Thus, the eight neurons of the network can be controlled independently. In an embodiment of the present invention, there is a relationship between the numbers of each group of neurons between the upper and lower unidirectional coupling network layers, that is, the neuron i in the first unidirectional coupling network layer and the neuron i+4 in the second unidirectional coupling network layer constitute a group of bidirectionally coupled neurons. In the eight-element neural network, neurons 1 and 5 form a group, which are used to control the hip joint and knee joint of the left hind leg, respectively; neurons 2 and 6 form a group, which are used to control the hip joint and knee joint of the right hind leg, respectively; neurons 3 and 7 form a group, which are used to control the hip joint and knee joint of the right front leg, respectively; neurons 4 and 8 form a group, which are used to control the hip joint and knee joint of the left front leg, respectively. In actual implementation, the bidirectional coupling between neurons in a group can also be obtained by adding self-coupling to each neuron in the four-neuron network, so that the entire network still maintains D ₄ symmetry.

Referring to FIG. 3, in the above-mentioned eight-element neural network, the coupling between neurons adopts inhibitory coupling, wherein neuron 1 inhibits neuron 3, neuron 3 inhibits neuron 2, neuron 2 inhibits neuron 4, neuron 4 inhibits neuron 1, neuron 5 inhibits neuron 8, neuron 8 inhibits neuron 6, neuron 6 inhibits neuron 7, neuron 7 inhibits neuron 5, there is bidirectional inhibition between neuron 1 and neuron 5, bidirectional inhibition between neuron 2 and neuron 6, bidirectional inhibition between neuron 3 and neuron 7, and bidirectional inhibition between neuron 4 and neuron 8 as a group. There are four types of coupling in the eight-element neural network:

The first type: coupling in the top hip neurons (corresponding to the subsequent parameter α)

The second type: coupling in the underlying knee neurons (corresponding to the subsequent parameter β)

The third type: coupling from the top layer to the bottom layer (corresponding to the subsequent parameter γ)

The fourth type: coupling from the bottom layer to the top layer (corresponding to the subsequent parameter δ)

In order to maintain the _D4 global symmetry, α and β should be equal, so that the neurons in the same layer are equivalent. γ and δ should also be equal in theory to keep a pair of knee-hip neurons equivalent. Therefore, the global symmetry of the network is formed by α and β, and the local symmetry is formed by γ and δ. However, the present invention can further optimize and adjust γ and δ in the future to meet the requirements of quadruped robot motion control.

In addition, in CPG, which is a small neural network formed by a group of interacting coupled connections, the connection between neurons is called a synapse. In order to realize the motion control of neurons, a set of neuron models described by ordinary differential equations are required to model the behavior of neurons, and a coupling matrix is used to describe the connection between neurons. In the above-mentioned eight-element neural network, due to the existence of different coupling situations, the driving signal of each neuron in the ordinary differential equation corresponding to the eight-element neural network is determined by three parts, namely, the gait control parameters generated by the signal control module, the coupling effect of neurons from the same unidirectional coupling network layer, and the coupling effect of neurons from another unidirectional coupling network layer. Therefore, different neuron couplings can be taken into account in the ordinary differential equations, thereby realizing the eight-degree-of-freedom coordinated control of the hip and knee joints of the quadruped robot, and realizing five gait rhythm signals at the same time.

In the embodiment of the present invention, taking any neuron i in the octal neural network as an example, i=1, 2, ..., 8, the construction and optimization method of the neuron model is generally demonstrated.

In the eight-element neural network, the ordinary differential equation of any neuron i adopts the Stein neuron model. The Stein neuron model belongs to the prior art, and the equation formula can be expressed as:

in, Indicates The membrane potential of a neuron, , , are all the state quantities to be solved in the equation, but only the state quantities The output of the neuron is then input into the subsequent signal post-processing module. , , They are , , The first-order derivative with respect to time t. In a two-layer network with eight neurons, since the eight neurons are located in two unidirectionally coupled network layers, some parameters need to be expressed by superscript and To distinguish, they correspond to the first unidirectional coupling network layer that controls the hip joint and the second unidirectional coupling network layer that controls the knee joint. is the rate constant affecting the neuron frequency, and Represent the parameters corresponding to the neurons in the first unidirectional coupled network layer and the second unidirectional coupled network layer respectively , is the driving signal of the neuron, which can also be divided into and . is an adaptive constant (a hyperparameter) that determines the degree of adaptability of the neuron. and is the rate constant for the cumulative conversion of sodium ions (a hyperparameter).

In the network, nodes are represented as neuron models, and the connections between nodes can be understood as the mutual coupling between neurons. In the eight-element neural network of the present invention, due to the four different coupling conditions between neurons, the coupling effect from other neurons in the above Stein neuron model should also be included in the driving signal of its ordinary differential equation. Therefore, the present invention needs to modify and transform the ordinary differential equation of a single Stein neuron to construct the control equation of the eight-element network.

In the embodiment of the present invention, the ordinary differential equation is based on the Stein neuron model, and the driving signal of any neuron i is modified to ,in , , are all gait control parameters generated by the signal control module, t is time, and are the same-layer weight hyperparameters and different-layer weight hyperparameters, respectively; X is the sum of the membrane potentials of the neurons that inhibit neuron i in the one-way coupled network layer where the current neuron i is located; Y is the sum of the membrane potentials of the neurons that inhibit neuron i in the one-way coupled network layer that does not contain the current neuron i.

For ease of description and easier understanding, the aforementioned superscript and To distinguish the driving signal equation of neuron i in the upper and lower unidirectional coupled network layers, the same-layer weight hyperparameter in the first unidirectional coupled network layer is and the different layer weight hyperparameters Recorded as and , the same-layer weight hyperparameters in the second unidirectional coupling network layer and the different layer weight hyperparameters Recorded as and Therefore, the driving signal equation is It can be further expressed as:

in (The corresponding two-layer networks are ) is the amplitude parameter of the entire driving signal, (The corresponding two-layer networks are )and (The corresponding two-layer networks are ) is the amplitude and frequency of the driving signal in the gait control parameters from the signal control module. The summation term and is the coupling effect from neurons in the same layer, and refers to the coupling effect from neurons in other layers, where It is used to indicate whether neuron j has a unidirectional coupling effect on neuron i. If yes, then , if otherwise Therefore, according to the coupling relationship between the eight neurons in the eight-element neural network shown in Figure 3, an 8×8 coupling matrix can be constructed as shown in Table 2 To record the one-way coupling relationship between neurons.

Table 2 Coupling matrix of eight-element neural network

The parameters in the above equations (3) and (4) are and The gait of the quadruped robot ultimately controlled by the eight-element neural network can be adjusted. These parameters, namely the aforementioned gait control parameters, can be generated by the signal control module and transmitted to the eight-element neural network.

In addition, in order to ensure the eight-element network architecture Global symmetry: when solving ordinary differential equations, the weight hyperparameters of the same layer of the ordinary differential equations corresponding to the eight neurons are set to the same, that is, α=β is satisfied, and the absolute value of the weight hyperparameter of the different layers of the ordinary differential equations corresponding to the neurons in the second one-way coupled network layer is greater than the absolute value of the weight hyperparameter of the different layers of the ordinary differential equations corresponding to the neurons in the first one-way coupled network layer, that is, .

In the embodiment of the present invention, the model parameters of the network are preferably set as shown in Table 3. The coupling parameters α and β are equal, so they are both set to −0.15, and the negative sign indicates the inhibitory coupling between neurons. For local symmetry, if , which will cause a pair of knee-hip neurons to produce Symmetry. Even if the other control parameters of this pair of neurons are not strictly equal, since they are located in different layers with opposite coupling loops, they still bring about a phase lock of approximately 1/2 cycle. Therefore, in the eight-element neural network of this embodiment, and are set to −0.6 and −0.1 respectively. This corresponds to the biological hypothesis that the upper network controls the gait of the entire octet network, and the lower network follows the signal generated by the upper network. It will break the overall situation Symmetry, because it causes a pair of knee-hip neurons to be not completely equivalent, so in this embodiment and Not the same.

Table 3 Parameter categories of eight-element network

In the embodiment of the present invention, the ordinary differential equations corresponding to the above eight neurons can be solved by any feasible numerical solution method.

Therefore, the motion control process of the quadruped robot according to the present invention can be described as follows: by adjusting the gait control parameters ( and ) is set, so that the eight-element neural network CPG in the signal generation module generates the corresponding gait signal, which can be sent to the signal post-processing module for processing. Generally speaking, the gait rhythm signals of each neuron solved in the signal generation module are time domain signals with a numerical range between 0 and 1, and the control signal actually executed by the system is a joint angle signal or a voltage signal. Therefore, the signal processing layer is required to map and process the rhythm signal of the signal generation layer and send it. Therefore, the time domain signal can be converted into a joint position signal or a voltage signal that controls the actuation of the corresponding joint through the signal post-processing module. This process is a mature technology in robot control and will not be described in detail.

In order to demonstrate the advantages of the present invention over the prior art, the technical effects that can be achieved by the present invention are demonstrated below through specific preferred examples.

In a preferred example of the present invention, the numerical simulation of the eight-element neural network is written in the Python 3.8 environment, and the ordinary differential equation of the network is calculated by the fourth-order Runge-Kutta method numerical solution method. The initial state parameters of the eight-element neural network and the control parameters corresponding to the five gaits generated by the eight-element network are shown in Tables 4 and 5, respectively, and five gaits can be generated: walking, trotting, gliding, jumping and leaping.

Table 4 Initial state of the eight-element neural network

Table 5 Control parameters of the eight-element neural network for five gaits

Finally, the numerical simulation results of the eight-element network executing five gaits are shown in Figure 4. (a)-(e) respectively plot the output of the CPG signal of the knee-hip neurons under the five gaits of the eight-element neural network and the phase diagram between the knee-hip neurons. The phase diagram between the knee-hip neurons can prove that the eight-element network can produce stable knee-hip neuron phase locking, and (f) shows the test results after disturbance injection, indicating that the present invention has high stability and anti-interference. In addition, the application of the eight-element neural network to the simulation of a quadruped robot also proves the feasibility of the technical solution.

In summary, compared with the common four-element network, the eight-element neural network proposed in the present invention can realize more types of gaits and coordinated control of the hip and knee joints, so that the motion control system of the quadruped robot has better environmental adaptability and maneuverability. Moreover, unlike the common model-based control and learning-based control methods of quadruped robots, the eight-element neural network architecture proposed in the present invention has the advantages of simple structure and small amount of calculation, and different gait types can be generated by adjusting the model parameters of the signal control layer.

Finally, it should be noted that in each embodiment provided by the present application, different modules in the quadruped robot motion control system based on the eight-element neural network can essentially be executed by a computer program. Moreover, in each embodiment provided by the present application, the division of modules in the system is only a logical function division. There may be other division methods in actual implementation, such as multiple modules can be combined or integrated together, and a module can also be split. The programs corresponding to these modules can be stored in a computer-readable storage medium in the form of software function modules written in various programming languages, and sold or used as independent products. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium, including several instructions to enable one or more computer devices (which can be personal computers, servers, or network devices, etc.) to execute all or part of the modules described in each embodiment of the present invention, thereby realizing the overall system function. Moreover, it can be understood that the above-mentioned storage medium may include a random access memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk storage. At the same time, the storage medium can also be a U disk, a mobile hard disk, a magnetic disk or an optical disk, etc., which can store program codes. The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims (9)

1. Four-foot robot motion control system based on eight-element neural network, characterized by comprising:

The signal regulation and control module is used for generating gait control parameters;

The signal generation module is used for receiving the gait control parameters generated by the signal regulation and control module, inputting the gait control parameters into the eight-element neural network, and respectively solving ordinary differential equations of the eight neurons to obtain gait rhythm signals; the eight-element neural network consists of a first unidirectional coupling network layer and a second unidirectional coupling network layer which have quadruple rotational symmetry respectively; the first unidirectional coupling network layer is formed by sequentially connecting a first neuron, a third neuron, a second neuron and a fourth neuron end to form a unidirectional coupling annular network, the second unidirectional coupling network layer is formed by sequentially connecting a fifth neuron, an eighth neuron, a sixth neuron and a seventh neuron end to form a unidirectional coupling annular network, and the coupling direction of the neurons in the first unidirectional coupling network layer is opposite to the coupling direction of the neurons in the second unidirectional coupling network layer; two-by-two neurons are used as a group between the first unidirectional coupling network layer and the second unidirectional coupling network layer, bidirectional coupling is formed between the two unidirectional coupling network layers, and four groups of neurons respectively control four legs of the four-legged robot in a one-to-one correspondence manner; and in each group of neurons, the neurons in the first unidirectional coupling network layer are used for controlling the hip joint, and the neurons in the second unidirectional coupling network layer are used for controlling the knee joint; in the ordinary differential equation corresponding to the eight-element neural network, the driving signal of each neuron is determined by three parts, namely gait control parameters generated by a signal regulation and control module, the coupling effect of the neurons from the same unidirectional coupling network layer and the coupling effect of the neurons from another unidirectional coupling network layer; in the eight-neuron neural network, a first neuron and a fifth neuron are in a group and are respectively used for controlling the hip joint and the knee joint of the left rear leg, a second neuron and a sixth neuron are in a group and are respectively used for controlling the hip joint and the knee joint of the right rear leg, a third neuron and a seventh neuron are in a group and are respectively used for controlling the hip joint and the knee joint of the right front leg, and a fourth neuron and an eighth neuron are in a group and are respectively used for controlling the hip joint and the knee joint of the left front leg;

the signal post-processing module is used for receiving the gait rhythm signal obtained by solving in the signal generating module and converting the gait rhythm signal into a displacement signal which correspondingly controls eight joints on four legs of the four-legged robot to actuate.

2. The eight-element neural network based motion control system of a quadruped robot of claim 1, wherein the gait pattern controllable by the quadruped robot includes walking, jogging, running, jumping and leaping.

3. The eight-neuron neural network based motion control system of a four-foot robot of claim 1, wherein the eight-neuron neural network uses inhibitory coupling for all couplings between neurons.

4. The eight-neuron-network-based quadruped robot motion control system according to claim 3, wherein the ordinary differential equation of any one neuron i in the eight-neuron-network adopts a Stein neuron model, and the driving signal of any one neuron i is thatWherein/>、/>、/>Are gait control parameters generated by the signal regulation and control module,/>For time,/>And/>The method comprises the steps of respectively obtaining the same-layer weight super parameter and the different-layer weight super parameter, wherein X is the membrane potential sum of neurons which form inhibition on the neurons i in a unidirectional coupling network layer where the current neurons i are located, and Y is the membrane potential sum of neurons which form inhibition on the neurons i in the unidirectional coupling network layer without the current neurons i.

5. The eight-element neural network-based motion control system of a four-foot robot of claim 4, wherein when the ordinary differential equation is solved, the same-layer weight super-parameters of the ordinary differential equations corresponding to the eight neurons are all set to be the same, and the different-layer weight super-parameters of the ordinary differential equations corresponding to the neurons in the second unidirectional coupling network layer have absolute values larger than those of the ordinary differential equations corresponding to the neurons in the first unidirectional coupling network layer.

6. The eight-neuron neural network based motion control system of a four-foot robot of claim 5, wherein the same-layer weight hyper-parameters of the ordinary differential equations for the eight neurons are all set to-0.15.

7. The eight-element neural network based motion control system of a four-foot robot of claim 5, wherein the outlier weight super parameters of the ordinary differential equations corresponding to the neurons in the first unidirectional coupling network layer are all set to-0.1, and the outlier weight super parameters of the ordinary differential equations corresponding to the neurons in the second unidirectional coupling network layer are all set to-0.6.

8. The eight-element neural network-based motion control system of a four-foot robot of claim 1, wherein the ordinary differential equation is numerically solved by a fourth-order longgrid tower method.

9. The eight-element neural network-based motion control system of the quadruped robot according to claim 1, wherein the gait rhythm signal of each neuron solved by the signal generation module is a time domain signal with a numerical range between 0 and 1, and the signal post-processing module converts the time domain signal into a joint position signal or a voltage signal for controlling the corresponding joint to actuate.