Robot teaching obstacle avoidance method based on neural network and dynamic system model

Abstract

The application discloses a robot teaching obstacle avoidance method based on a neural network and a dynamic system model, and comprises the following steps: obtaining a demonstration data set; encoding a current robot state into a high-dimensional feature vector; extracting a nonlinear feature component from the coded feature, performing weighted summation on the feature and adding a quadratic term to construct an energy function; constructing a decision function by multiplying the state of the robot and environmental information by corresponding weights to obtain the output of the decision function; inputting the output of the decision function into a multi-task hybrid neural network to generate a parameter vector, constructing a lower triangular matrix based on the parameter vector and constructing a modulation matrix through cholosky decomposition; constructing a loss function and obtaining optimal model parameters through a nonlinear programming method; and taking the output of the neural dynamic system model as a robot control strategy to complete a specified task. The application solves the bottleneck problems of time dependence and dynamic obstacle avoidance by introducing a dynamic system model and a neural network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot control, and particularly relates to a robot teaching obstacle avoidance method based on a neural network and a dynamic system model. BACKGROUND

[0002] As the production environment of robots is becoming more and more complex and the interference is becoming more and more frequent, it is difficult for robots to master human skills through offline programming to meet the needs in unstructured environments, and therefore it is necessary to use teaching programming to enable robots to master human skills, such as drag teaching.

[0003] Through teaching, robots can quickly master human skills, but many current robot teaching technologies are time-dependent and cannot perform autonomous obstacle avoidance or return to the task state autonomously after obstacle avoidance. If an interference or obstacle appears in the robot assembly space due to a sudden factor or accident in an actual production scene, the robot will be damaged due to collision, and even interfere with the operation of the production line, causing huge economic losses if the conventional teaching method does not have obstacle avoidance capability.

[0004] For example, patent No. CN119077728A, a robot autonomous learning and adaptive trajectory planning method for human-robot collaboration, uses dynamic primitives to learn human teaching samples to obtain a generalized trajectory, uses a particle swarm algorithm to quickly search for a local obstacle avoidance trajectory based on the generalized trajectory, and combines the generalized trajectory and the obstacle avoidance trajectory obtained by the particle swarm algorithm to achieve task completion and obstacle avoidance integration. However, this method is based on a dynamic primitive teaching algorithm with high time dependence, and interruption during operation caused by interference will result in task failure.

[0005] For example, patent No. CN109702744A, a robot imitation learning method based on a dynamic system model, uses a dynamic system model to learn human teaching samples and generate a generalized trajectory, uses control theory to construct a Gaussian mixture model as a nonlinear dynamic system model, and adds a stability constraint condition to ensure that the model converges stably to the task target point, thereby ensuring that the task can be completed. However, this method uses a dynamic system teaching algorithm with low time dependence, but does not detect and avoid obstacles in the environment.

[0006] Therefore, the limitations of mainstream robot teaching schemes, such as the time-dependent dynamic primitive (DMP) method, in unstructured environments are particularly evident. For example, when the robot is interrupted due to sudden interference (such as temporary obstacles or sensor noise), the DMP method needs to re-plan the time parameters to resume the task, which not only increases the computational complexity, but also may cause the robot to lose control due to time synchronization problems, resulting in serious losses. In addition, although the traditional dynamic system method guarantees the convergence of the target through stability constraints, its obstacle avoidance function usually relies on the assumption of a static environment and cannot respond to real-time changes in the position of dynamic obstacles. SUMMARY

[0007] In order to overcome the defects and deficiencies existing in the prior art, the present application provides a robot teaching obstacle avoidance method based on a neural network and a dynamic system model. The present application learns human demonstration trajectories based on a nonlinear dynamic system model. A modulation matrix combined with a decision function is designed in the dynamic system model to simultaneously learn the task characteristics and obstacle avoidance characteristics in the demonstration. An energy function containing a neural network is designed to ensure that the trajectory has a unique task target point. Constraint conditions are designed to ensure that the model can stably converge to guarantee that all trajectories generalized can reach the task target point. The parameter learning problem in the model is converted into an optimization problem containing constraints to obtain the optimal model parameters. The task starting point and obstacle information are input into the dynamic system model to obtain the generalized trajectory to guide the robot system to complete the task. The present application solves the bottleneck problems of time dependence and dynamic obstacle avoidance by introducing a dynamic system model and a neural network.

[0008] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0009] The present application provides a robot teaching obstacle avoidance system based on a neural network and a dynamic system model, comprising: a demonstration data acquisition module, a decision function construction module, a modulation matrix construction module, an energy function construction module, and a control strategy output module.

[0010] The demonstration data acquisition module is used to collect human demonstration data and obstacle information to generate a demonstration data set.

[0011] The decision function construction module is used to fuse state and environmental information to construct a decision function, dynamically allocate the weights of task tracking and obstacle avoidance behaviors using the decision function, and generate an adaptive control strategy according to the robot state and environmental information.

[0012] The modulation matrix construction module is used to construct a modulation matrix through a multi-task hybrid neural network. The modulation matrix dynamically adjusts the direction and amplitude of the robot's motion, and learns to reproduce complex trajectory characteristics and obstacle avoidance behaviors.

[0013] The energy function construction module is configured to extract nonlinear feature components from the coded features of the robot state through a neural network, to sum the features with weighting and add a quadratic term, to construct an energy function, and to make the trajectory converge to a unique target point based on the energy function.

[0014] The control strategy output module is configured to construct a loss function, to obtain optimal parameters of the neural dynamic system model through a nonlinear programming method, and to use the output of the neural dynamic system model as a robot control strategy to complete a specified task.

[0015] The application further provides a robot teaching obstacle avoidance method based on a neural network and a dynamic system model, which comprises the robot teaching obstacle avoidance system based on the neural network and the dynamic system model and the following steps.

[0016] Human demonstration data and obstacle information are collected to generate a demonstration data set.

[0017] The current robot state is coded into a high-dimensional feature vector.

[0018] Nonlinear feature components are extracted from the coded features through a neural network, the features are summed with weighting and a quadratic term is added to construct an energy function.

[0019] A decision function is constructed by fusing the state and environmental information, the state and environmental information of the robot are multiplied by corresponding weights to obtain the output of the decision function.

[0020] The output of the decision function is input into a multi-task hybrid neural network to generate a parameter vector, a lower triangular matrix is constructed based on the parameter vector and a modulation matrix is constructed through cholosky decomposition.

[0021] A loss function is constructed to convert a parameter learning problem into a constrained optimization problem, and optimal model parameters are obtained through a nonlinear programming method.

[0022] The output of the neural dynamic system model is used as a robot control strategy to complete a specified task.

[0023] As a preferred technical solution, human demonstration data and obstacle information are collected to generate a demonstration data set, which is specifically represented as:

[0024]

[0025] wherein x t,n is the position state of the robot at time t in the nth demonstration, is a velocity state, and x obj is obstacle information in the environment.

[0026] As a preferred technical solution, the nonlinear feature component is extracted from the encoded feature through a neural network, the features are weighted and summed and a quadratic term is added to construct an energy function, which specifically includes:

[0027] The nonlinear feature component is extracted from the encoded feature, which is represented as:

[0028]

[0029] wherein a k and b k are learnable feature parameters, sigma is an activation function, c(x) represents the encoded feature, and g k (x) represents the nonlinear feature component.

[0030] All nonlinear feature components g k (x) are weighted and summed through a weight vector omega to obtain a neural network part P1(x):

[0031]

[0032] A quadratic term P2(x) is constructed:

[0033] P2(x) = kappa xTx (kappa > 0)

[0034] wherein kappa is a positive number.

[0035] The energy function is represented as:

[0036] V(x) = P1(x) - P1(0) + P2(x).

[0037] As a preferred technical solution, the encoded feature c(x) is specifically represented as:

[0038]

[0039] wherein epsilon > 0 is an adjustment parameter, and x represents the state of the robot.

[0040] As a preferred technical solution, the parameter constraint of the energy function is:

[0041]

[0042] wherein omega represents the weight vector.

[0043] As a preferred technical solution, the state and environmental information of the robot are multiplied by corresponding weights to obtain the output of the decision function, which specifically includes:

[0044] The distance between the robot and the obstacle is calculated, and the weights of the control task tracking parameter alpha1 and the obstacle avoidance parameter alpha2 are dynamically allocated through a Sigmoid function.

[0045] The state x of the robot and the environment information z are multiplied by the corresponding weights α1 and α2 respectively to obtain the output of the decision function, which is specifically represented as:

[0046] D(x,z)=[α1x,α2z]。

[0047] As a preferred technical solution, the environment information z is represented as:

[0048]

[0049] Wherein, x represents the current state of the robot, x obj represents the center position of the obstacle, represents the speed of the robot, represents the gradient of the energy function.

[0050] As a preferred technical solution, the lower triangular matrix is constructed based on the parameter vector and the modulation matrix is constructed through cholosky decomposition, which is specifically represented as:

[0051]

[0052] Wherein, D(x,z) represents the output of the decision function, ∈>0 is a regulating parameter, and I is a unit matrix.

[0053] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0054] (1) The present application realizes the robot teaching and obstacle avoidance function by introducing a dynamic system model, which is different from other time-dependent teaching methods, and can continue to complete the task after being interrupted or disturbed, thereby enhancing the anti-interference ability of the system.

[0055] (2) The present application extracts task tracking features and obstacle avoidance features from human demonstration through the decision function module, so that when the robot generalizes autonomously, it can realize task tracking and obstacle avoidance functions according to whether it encounters obstacles, respectively.

[0056] (3) The present application constructs an energy function and a modulation matrix through a neural network, the energy function ensures that the dynamic system can have a unique minimum point, and guarantees that the task can be completed at the target position, and the modulation matrix ensures the stability of the system, so that the robot can accurately and stably converge to the minimum point, that is, the task target point. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a schematic diagram of the overall implementation framework of the robot teaching and obstacle avoidance method based on neural network and dynamic system model of the present application;

[0058] Figure 2A flowchart for constructing a decision function of the present application is shown in the figure;

[0059] Figure 3 A flowchart for constructing a modulation matrix of the present application is shown in the figure;

[0060] Figure 4 A flowchart for constructing an energy function of the present application is shown in the figure;

[0061] Figure 5 A comparison diagram of the obstacle avoidance performance of the present application and other methods is shown in the figure. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0063] Example 1

[0064] The present embodiment provides a robot teaching obstacle avoidance system based on neural network and dynamic system model, comprising a demonstration data acquisition module, a decision function construction module, a modulation matrix construction module and an energy function construction module, and a control strategy output module.

[0065] The demonstration data acquisition module collects human demonstration data and obstacle information to generate a demonstration data set.

[0066] The decision function construction module is used to distinguish the task trajectory tracking and obstacle avoidance trajectory in the human demonstration data set, and dynamically adjusts the behavior of the system between trajectory tracking and obstacle avoidance tasks by combining the current state of the robot and the environmental information. The decision function consists of two parts: one part is the current state information of the robot, and the other part is the environmental information, including the relative position, velocity of the robot and the obstacle, and the gradient of the energy function. By introducing a smooth switching mechanism, the decision function can dynamically adjust the weight according to the distance between the robot and the obstacle, ensuring that the system prioritizes obstacle avoidance behavior when approaching the obstacle, and focuses on the trajectory tracking task when moving away from the obstacle.

[0067] The role of the decision function mainly reflects in two aspects: one is to enhance the obstacle avoidance ability of the system, so that the robot can autonomously avoid obstacles in complex environments; the other is to improve the generalization performance of the system, so that the robot can adapt to different obstacle layouts and initial conditions, thereby efficiently completing tasks in various scenarios.

[0068] The modulation matrix construction module is configured to construct a modulation matrix by a multi-task hybrid neural network, an input of which combines a current state of the robot, environmental obstacle information, and a gradient of the energy function, so as to comprehensively perceive a task demand and an environmental constraint. The modulation matrix dynamically adjusts a direction and a magnitude of a motion of the robot, so that the robot can accurately learn and reproduce a complex trajectory feature and an obstacle avoidance behavior. Specifically, the modulation matrix extracts a trajectory shape feature from the demonstration data and fuses the obstacle distance information to generate an adaptive motion correction strategy based on a nonlinear fitting capability of the neural network. The modulation matrix has advantages of flexibility and robustness. On one hand, the modulation matrix can automatically learn local details of the trajectory by training of the neural network, while capturing a global influence of the obstacle on the trajectory. On the other hand, a symmetric positive definite structure design of the modulation matrix ensures stability of the dynamic system, so as to avoid deviation of the trajectory caused by local disturbance.

[0069] The energy function construction module provides a stable motion guiding mechanism for the system by constructing an energy function NEUM which is continuously differentiable, radially unbounded, and has a unique global minimum point. The energy function comprises two parts: one part is a nonlinear feature item extracted based on the neural network, which is used to capture local details of the complex trajectory; and the other part is a quadratic term, which is used to ensure radial unboundedness of the energy function, so as to ensure that the system can still converge when far away from the target. The energy function has advantages of flexibility and robustness. By the nonlinear fitting capability of the neural network, the energy function can accurately model shape features of the complex trajectory, while the quadratic term ensures global stability of the system. In addition, the unique minimum point design of the energy function avoids a local minimum value problem, so that the system can always converge to the target state.

[0070] The control strategy output module takes an output of the neural dynamic system model as a robot control strategy to complete a specified task.

[0071] Embodiment 2

[0072] As shown in Figure 1 , the embodiment provides a robot teaching obstacle avoidance method based on a neural network and a dynamic system model, comprising the following steps:

[0073] Step 1: collecting human demonstration data and obstacle information, a human expert completes a robot task by teaching to generate a demonstration data set wherein x t,n is a position state of the robot at time t in the nth demonstration, is a velocity state, and x obj is obstacle information in the environment;

[0074] Step 2: state coding and feature expansion, coding the current robot state x into a high-dimensional feature vector c(x) to enhance nonlinear expression capability;

[0075] Step 3: Constructing the energy function, first extract nonlinear feature components from the encoded features c(x) through a neural network where a k and b k are feature parameters that need to be trained and learned, and σ is the activation function, then the feature weighted sum and added quadratic term, construct the energy function V(x);

[0076] As Figure 4 shown, first receive the trajectory data set of human demonstration which contains the state of the robot x and its corresponding dynamic change

[0077] The input state x is processed by the state encoding module c(x) to expand the nonlinear features:

[0078]

[0079] where ∈>0 is an adjustment parameter to enhance the nonlinear representation ability of the state.

[0080] Feature extraction and weighted combination: the encoded state c(x) is input to the feature extraction module to generate a set of nonlinear feature components g k (x):

[0081]

[0082] where a k and b k are learnable feature parameters, and σ is the activation function (such as the hyperbolic tangent function), and all feature components g k (x) are weighted and summed through the weight vector ω to form the neural network part P1(x):

[0083]

[0084] Energy function structure integration: the energy function V(x) is composed of two parts, the neural network part P1(x) that captures the nonlinear features of complex trajectories, and the quadratic term P2(x) that ensures the radial unboundedness of the energy function, where P2(x) is defined as:

[0085] P2(x) = κxTx (κ > 0)

[0086] κ is a small positive number, and the final energy function form is:

[0087] V(x) = P1(x) - P1(0) + P2(x)

[0088] where P1(0) is used for normalization to ensure that the energy is minimized at the target point x = 0;

[0089] Energy function parameter constraint: during the learning optimization process of the energy function parameters, the parameters in the energy function must satisfy the following conditions to ensure that the system can converge to the target position:

[0090]

[0091] Parameter optimization: this embodiment converts the parameter learning problem of the model into an optimization problem with additional constraints, which can be solved by optimization methods and nonlinear programming methods. The objective function is set as:

[0092]

[0093] Step 4: fuse state and environment information to construct decision function, calculate distance between robot and obstacle:

[0094] d = ||x-x obj ||.

[0095] Dynamically allocate the weights of control task tracking parameter α1 and obstacle avoidance parameter α2 through Sigmoid function;

[0096] As shown in Figure 2 , the input of the decision function includes the current state of the robot x, the center position of the obstacle x obj , the speed of the robot , and the gradient of the energy function These information together constitute the environment information z, which is used for subsequent decision-making process;

[0097] Distance calculation: calculate the distance between the current position of the robot and the center of the obstacle d = ||x-x obj ||, which is used to determine whether the robot is close to the obstacle, so as to decide whether to switch to obstacle avoidance mode;

[0098] Weight calculation: calculate two weight coefficients α1 and α2 through Sigmoid function. Among them, α1 represents the weight of task tracking, and α2 represents the weight of obstacle avoidance. The calculation of these two weights is based on distance d and a safety threshold β, which ensures that the obstacle avoidance weight gradually increases when approaching the obstacle, and the task tracking weight dominates when moving away from the obstacle;

[0099] Decision function construction: multiply the state of the robot x and the environment information z by the corresponding weights α1 and α2 respectively to get the output of the decision function. Specifically, the output of the decision function is D(x,z) = [α1x,α2z], where

[0100] Step 5: Generating the modulation matrix, the decision function output D(x, z) is input into the multi-task hybrid neural network Net(·) to generate the parameter vector n, and the lower triangular matrix is constructed by n and the modulation matrix M is constructed by cholosky decomposition;

[0101] As shown in Figure 3 , the input of the modulation matrix includes the output of the decision function , which contains the current state x of the robot and the environment information z;

[0102] The core of the modulation matrix is a multi-task hybrid neural network (Multi-task Hybrid Neural Network), denoted as The task of this neural network is to learn the features of the task trajectory and obstacle avoidance behavior from the input data. The design of the neural network selects a three-layer fully connected neural network, and the number of neurons in each hidden layer is 80. The calculation process of each layer of the network is as follows:

[0103] Input layer to hidden layer 1, input data D(x, z) will be transmitted to the first layer of the network. For the calculation of the first layer, the network will perform the following operations:

[0104] h1=σ(W1D(x,z)+b1)

[0105] Where W1 is the first layer weight matrix, the number of neurons is 80, b1 is the first layer bias vector, σ(·) is the activation function, and tanh function is used as the activation function;

[0106] Similarly, the calculation process from hidden layer 1 to hidden layer 2 is as follows:

[0107] h2=σ(W2h1+b2)

[0108] Where W2 is the second weight matrix, the number of neurons in this layer is also 80, b2 is the second layer bias vector, and h2 is the output of the second layer, which is also the input of the output layer;

[0109] The calculation process from hidden layer 2 to output layer is as follows:

[0110] n=W3h2+b3

[0111] Where W3 is the output layer weight matrix, b3 is the output layer bias vector, and n is the final output of the network;

[0112] Output of the neural network: The output of the neural network is a vector n, whose dimension is Where n is the dimension of the state space. This vector contains the parameters required by the modulation matrix;

[0113] Construction of modulation matrix: according to the output vector n of the neural network, a lower triangular matrix N is constructed, specifically, the elements N of the matrix N ij is generated by the following way:

[0114] If i>j or i=j, then N ij = nk, where k is the corresponding index in the vector n;

[0115] If i<j, then N ij =0;

[0116] Positive definite of modulation matrix: in order to ensure that the modulation matrix M is positive definite (which is a key condition for system stability), the idea of cholesky decomposition is adopted for design, specifically, the modulation matrix M is generated by the following way:

[0117]

[0118] Wherein, ∈ is a small positive number, I is a unit matrix, this step ensures that the modulation matrix M is symmetric and positive definite, so as to meet the stability requirements of the system;

[0119] Step 6: construct the loss function, convert the parameter learning problem into a constrained optimization problem, and obtain the optimal model parameters by nonlinear programming method;

[0120] Step 7: the output of the neural dynamic system model is used as the robot control strategy to complete the specified task.

[0121] As Figure 5 shown, and combined with the following table 1, the obstacle avoidance performance of the present application is compared with that of other methods;

[0122] Table 1 comparison of obstacle avoidance performance of the present application and other methods

[0123]

[0124] Compared with the existing obstacle avoidance algorithm based on DMP, the method of the present application is more stable and has higher obstacle avoidance ability. By calculating the root mean square error of the two methods, the method of the present application better preserves the characteristics of the original trajectory, realizes an error of 9.73 cm, while the method based on DMP reduces the trajectory accuracy in the obstacle avoidance process, resulting in an error of 10.65 cm. The present application can avoid obstacles for multiple different obstacle positions, while the method based on DMP will collide at some obstacles; on the other hand, the method of the present application can demonstrate multiple trajectories, and learn the characteristics of multiple trajectories, while the method based on DMP can only learn a single trajectory, and cannot learn the characteristics of multiple expert demonstrations.

[0125] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications, etc. made without departing from the spirit and principles of the present application should be equivalent replacement manners and should be included in the protection scope of the present application.

Claims

1. A robot teach-and-avoid obstacle system based on neural networks and dynamic system models, characterized in that, include: The demonstration module includes a data acquisition module, a decision function construction module, a modulation matrix construction module, an energy function construction module, and a control strategy output module. The demonstration data acquisition module is used to collect human demonstration data and obstacle information to generate a demonstration dataset. The decision function construction module is used to integrate state and environmental information to construct a decision function, dynamically allocate the weights of task tracking and obstacle avoidance behaviors using the decision function, and generate an adaptive control strategy based on the robot's state and environmental information. The modulation matrix construction module is used to construct a modulation matrix through a multi-task hybrid neural network. The modulation matrix dynamically adjusts the direction and amplitude of the robot's motion and learns to reproduce complex trajectory features and obstacle avoidance behavior. The energy function construction module is used to extract nonlinear feature components from the features encoded by the robot state through a neural network, sum the features by weight and add a quadratic term to construct an energy function, and make the trajectory converge to a unique target point based on the energy function. The control strategy output module is used to construct a loss function, obtain the optimal parameters of the neural dynamic system model through nonlinear programming, and use the output of the neural dynamic system model as the robot control strategy to complete the specified task.

2. A robot teach-and-avoid-obstacle method based on a neural network and a dynamic system model, characterized by, The robot teaching obstacle avoidance system based on neural networks and dynamic system models as described in claim 1 includes the following steps: Collect human demonstration data and obstacle information to generate a demonstration dataset; Encode the current robot state into a high-dimensional feature vector; The nonlinear feature components are extracted from the encoded features using a neural network. The features are then weighted, summed, and a quadratic term is added to construct an energy function. The decision function is constructed by fusing state and environmental information. The robot's state and environmental information are multiplied by their respective weights to obtain the output of the decision function. The output of the decision function is input into a multi-task hybrid neural network to generate a parameter vector. A lower triangular matrix is ​​constructed based on the parameter vector and a modulation matrix is ​​constructed through Cholesky decomposition. By constructing a loss function, the parameter learning problem is transformed into a constrained optimization problem, and the optimal model parameters are obtained through nonlinear programming. The output of the neural dynamic system model is used as the robot control strategy to complete the specified task.

3. The robot teach-and-avoid obstacle method based on neural network and dynamic system model according to claim 2, characterized in that, Collect human demonstration data and obstacle information to generate a demonstration dataset, specifically as follows: where x t,n is the position state of the robot at time t in the n-th demonstration, is the velocity state, x obj is the obstacle information in the environment.

4. The robot teach-and-avoid obstacle method based on neural network and dynamic system model according to claim 2, characterized in that, The nonlinear feature components are extracted from the encoded features using a neural network. The features are then weighted, summed, and a quadratic term is added to construct an energy function, which specifically includes: The nonlinear feature components are extracted from the encoded features and represented as follows: where a k and b k are learnable feature parameters, σ is an activation function, c(x) represents the encoded features, and g k (x) represents a non-linear feature component; all nonlinear feature components g k (x) weighted summation by the weight vector ω to obtain the neural network part P1(x): Construct the quadratic term P2(x): P2(x) = Kx T x (K > 0) Where κ is a positive number; The energy function is expressed as: V(x) = P1(x) - P1(0) + P2(x).

5. The robot teach-and-avoid obstacle method based on neural network and dynamic system model according to claim 4, characterized in that, The encoded feature c(x) is specifically represented as follows: Where ∈>0 is the adjustment parameter, and x represents the robot state.

6. The robot teach-and-avoid obstacle method based on neural network and dynamic system model according to claim 4, characterized in that, The parameter constraints of the energy function are: Where ω represents the weight vector.

7. The robot teach-and-avoid method based on neural network and dynamic system model according to claim 2, characterized in that, A decision function is constructed by fusing state and environmental information. The robot's state and environmental information are multiplied by their respective weights to obtain the output of the decision function, which specifically includes: Calculate the distance between the robot and obstacles, and dynamically allocate the weights of the tracking parameter α1 and obstacle avoidance parameter α2 using the Sigmoid function; The state x and the environment information z of the robot are multiplied by the corresponding weights α1 and α2, respectively, to obtain the output of the decision function, which is specifically represented as: D(x,z)=[α1x,α2z].

8. The robot teach-and-avoid obstacle method based on neural network and dynamic system model according to claim 2, characterized in that, The environment information z is represented as: where x represents the current state of the robot, x obj represents the center position of the obstacle, represents the velocity of the robot, represents the gradient of the energy function.

9. The robot teach-and-avoid method based on neural network and dynamic system model according to claim 2, wherein, A lower triangular matrix is constructed based on the parameter vector, and a modulation matrix is constructed through cholosky decomposition, which is specifically represented as: where D(x,z) represents the output of the decision function, ∈>0 is a regulation parameter, and I is a unit matrix.

Images