Method and device for controlling a space robot, electronic equipment and storage medium

Abstract

The present application relates to the technical field of space manipulator control, and particularly relates to a space manipulator control method and device, electronic equipment and storage medium. The method comprises: obtaining a target manipulator state of a target assembly task to be executed by a space manipulator; inputting the target manipulator state into a pre-trained operation control model to output a target manipulator action corresponding to the target manipulator state, so as to control the movement of the space manipulator by using the target manipulator action. The technical scheme of the present application can accelerate the judgment and decision-making process of the space manipulator, so as to consume less time for judgment and decision-making.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of space manipulator control, and in particular to a space manipulator control method and device, an electronic device, and a storage medium. BACKGROUND

[0002] In the field of aerospace, space manipulators are often assigned the task of performing precise assembly.

[0003] In related technologies, in order to improve assembly efficiency, space manipulators usually automatically perform assembly tasks according to pre-set program logic. However, this approach requires obstacles to be avoided, and using traditional path planning methods requires a long time to make decisions.

[0004] Therefore, there is an urgent need to provide a space manipulator control method, device, electronic device, and storage medium to solve the above technical problems. SUMMARY

[0005] The present application describes a space manipulator control method, device, electronic device, and storage medium that can accelerate the decision-making process of a space manipulator to consume less time for decision-making.

[0006] According to a first aspect, the present application provides a space manipulator control method, comprising:

[0007] obtaining a target manipulator state of a target assembly task to be performed by a space manipulator;

[0008] inputting the target manipulator state into a pre-trained operation control model to output a target manipulator action corresponding to the target manipulator state, and using the target manipulator action to control the movement of the space manipulator.

[0009] According to a second aspect, the present application provides a space manipulator control device, comprising:

[0010] an obtaining unit configured to obtain a target manipulator state of a target assembly task to be performed by a space manipulator;

[0011] a control unit configured to input the target manipulator state into a pre-trained operation control model to output a target manipulator action corresponding to the target manipulator state, and use the target manipulator action to control the movement of the space manipulator.

[0012] According to a third aspect, the present application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method of the first aspect.

[0013] According to a fourth aspect, the application provides a computer readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method of the first aspect.

[0014] According to the space mechanical arm control method, device, electronic equipment and storage medium provided by the application, the target mechanical arm state of the target assembly task to be performed by the space mechanical arm is acquired, and the target mechanical arm state is input into the pre-trained operation control model, and a target mechanical arm action corresponding to the target mechanical arm state is output, so that the target mechanical arm action can be used to quickly control the movement of the space mechanical arm, so as to better complete the assembly task. Therefore, the above technical solution can accelerate the judgment and decision-making process of the space mechanical arm, so as to consume less time for judgment and decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0016] Figure 1 A flowchart of a space mechanical arm control method according to an embodiment is shown;

[0017] Figure 2 A schematic block diagram of a space mechanical arm control device according to an embodiment is shown;

[0018] Figure 3 Error test results of a space mechanical arm control method according to an embodiment are shown. DETAILED DESCRIPTION

[0019] The schemes provided by the application will be described below in combination with the drawings.

[0020] Figure 1 A flowchart of a space mechanical arm control method according to an embodiment is shown. It can be understood that the method can be executed by any device, equipment, platform, device cluster with computing and processing capabilities. As shown, the method includes: Figure 1

[0021] Step 100, acquiring the target mechanical arm state of the target assembly task to be performed by the space mechanical arm;

[0022] ​In step 102, the target robot state is input into the pre-trained operation control model, and a target robot action corresponding to the target robot state is output, so as to control the space robot motion by using the target robot action.

[0023] In the embodiment, by obtaining the target robot state of the target assembly task to be performed by the space robot, and inputting the target robot state into the pre-trained operation control model, a target robot action corresponding to the target robot state is output, so that the target robot action can be quickly used to control the space robot motion, so as to better complete the assembly task. Therefore, the above technical solution can accelerate the judgment and decision-making process of the space robot, so as to consume less time for judgment and decision-making.

[0024] In the field of aerospace, the widespread use of teleoperation devices is crucial for the successful execution of space missions. In the extreme space environment, factors such as microgravity, strong radiation, and extreme temperature differences pose unique challenges to extravehicular tasks, and using a teleoperation handle to control a space robot becomes an indispensable technology. The microgravity environment in space makes the operation of the space robot more complex than on Earth, so through the teleoperation handle, the operator can adjust the motion of the space robot in real time through a highly sensitive control system, ensuring accurate and stable operation in the microgravity environment. In the space environment, equipment may be damaged by factors such as micro-meteors and cosmic dust. Using a teleoperation handle to control a space robot, the operator can remotely perform emergency repairs and maintenance work, replace damaged parts, and ensure that the spacecraft operates normally during the mission.

[0025] One inevitable problem in teleoperation is the active-passive switching problem of space robot task execution. With the trend of technological change and development in today's world, automatic assembly of components in tasks not only reduces the trouble of operators, but also saves labor costs, standardizes task execution processes, and reduces human operation errors. However, such automated control processes require high precision, and the design of program logic can greatly affect whether the task can be completed normally.

[0026] The robot arm can imitate and perform similar actions by observing and learning the actions of humans or other robot arms. This learning method can make the robot arm more adaptable and flexible, and can handle different tasks and environments. For complex tasks, it is difficult to give explicit problem specifications or reward functions by manual design. In the process of robot arm movement, if obstacles are encountered or computer decision-making is required, traditional path planning requires a long time and more complex decision-making. The above technical solution trains the robot arm using human expert demonstration data by utilizing prior knowledge. In this way, the robot arm can move according to a predetermined route in the face of complex situations, reducing the demand for computing power and time consumption. In real-world tasks, human experience data is often needed to speed up the decision-making process.

[0027] As a preferred embodiment, the robot arm state includes joint angle, joint speed, joint acceleration, end force and end position.

[0028] As a preferred embodiment, the operation control model is trained by the following method:

[0029] Obtain standard action data of the spatial robot arm when performing an assembly task, and extract sample robot arm states and sample robot arm actions from the standard action data;

[0030] Input the sample robot arm states and sample robot arm actions into the preset neural network for training to obtain an intermediate transition model;

[0031] Determine the correction value corresponding to each time point based on the task completion degree of the spatial robot arm at different time points;

[0032] Based on the vector composed of all correction values, the weights of the intermediate transition model are corrected to obtain the operation control model.

[0033] In this embodiment, since the state space of the spatial robot arm is large, it may lead to slow learning or even unable to learn. By obtaining the correction value corresponding to each time point based on the task completion degree of the spatial robot arm at different time points, the vector composed of these correction values is used to continuously learn and correct the weights of the intermediate transition model, so that a more accurate operation control model can be obtained.

[0034] As a preferred embodiment, the step of "determining the correction value corresponding to each time point based on the task completion degree of the spatial robot arm at different time points" can specifically include:

[0035] The task completion degree of the spatial robot arm at different time points is determined by the following formula:

[0036]

[0037] wherein p t is the task completion degree of the space manipulator at time t;

[0038] The correction value corresponding to each time is determined by the following formula:

[0039]

[0040] wherein R t+1 is the correction value corresponding to time t+1, p t-1 is the task completion degree of the space manipulator at time t-1, a is a preset proportion coefficient, b is a preset activity range of the end of the manipulator, and ΔD is the spatial distance between the end of the manipulator and the target position.

[0041] According to the specific assembly task, the setting rule of the correction value can be determined as follows:

[0042] 1) When the manipulator moves along the direction of the demonstration trajectory data, the correction value is positive; otherwise, if the manipulator moves in the opposite direction of the reference direction, the correction value is negative;

[0043] 2) The smaller the spatial distance between the end of the manipulator and the target position, the larger the positive correction value; when the distance exceeds a certain range, the correction value is negative, and the farther the distance, the smaller the correction value;

[0044] 3) When the manipulator completes the task, the maximum positive correction value can be obtained; otherwise, when the corresponding task is not completed, the minimum negative correction value is obtained.

[0045] As a preferred embodiment, the neural network is a back propagation neural network.

[0046] The back propagation neural network is a machine learning algorithm, which can effectively establish the complex and nonlinear internal relationship between physical quantities through processing or analyzing a large amount of original data, and then realize accurate prediction of the target problem, and has significant advantages and good practicability in data mining.

[0047] The training process of the back propagation neural network includes two parts: first, the working signal is propagated from the input layer to the output layer through the hidden layer, completing the forward propagation, and in this process, the weights and thresholds between layers are unchanged; then, the actual output is compared with the expected output, that is, the sample value, and the error signal is propagated from the output layer to the input layer through the hidden layer, completing the reverse propagation, and in this process, the weights and thresholds are corrected in the opposite direction of the gradient of the error signal, and through repeated adjustment of the weights and thresholds, the actual output and the expected output are as close as possible.

[0048] The initial weight and threshold of the back propagation neural network are randomly generated, and the iteration update is a small correction to the initial value. Therefore, the back propagation neural network is very sensitive to the initial solution, and the algorithm has unstable factors. In this regard, the inventor creatively finds that the network can be initialized with a fixed weight matrix.

[0049] In addition, the back propagation algorithm uses the gradient descent method to determine the search direction. Since there are flat regions and multiple minimum points in the search space, the algorithm is prone to slow convergence speed or falling into local minimum. In view of this problem, the inventor creatively finds that the particle swarm optimization algorithm or the genetic algorithm with excellent global search performance can be considered for optimization. The particle swarm optimization algorithm or the genetic algorithm has strong adaptability, does not need to use gradient and other problem information in the iteration process, and does not have the problem of flat region. In addition, the selection operator can eliminate the model falling into the local minimum. There are multiple individuals in a particle swarm in the particle swarm optimization algorithm or the genetic algorithm, which perform global search in parallel to ensure that the optimal model can be finally found. Further, the inventor creatively finds that the particle swarm optimization algorithm combined with the genetic algorithm can achieve better optimization effect for optimizing the initial weight and threshold of the back propagation neural network.

[0050] As a preferred embodiment, the intermediate transition model is trained by the following method:

[0051] The network structure of the back propagation neural network is determined; wherein the network structure includes the weight and the threshold;

[0052] The weight and the threshold of the back propagation neural network are encoded to obtain an initial particle swarm; wherein the weight and the threshold are both taken as particles of the initial particle swarm;

[0053] The algorithm parameters of the genetic algorithm are set; wherein the algorithm parameters include: the particle swarm size, the maximum evolution number, the crossover probability, the mutation probability, the learning rate, the velocity range and the position range of the particle;

[0054] The velocity and the position of each particle in the initial particle swarm are initialized, and the fitness value of each particle in the initial particle swarm is calculated;

[0055] The velocity and the position of each particle in the initial particle swarm are updated, and the fitness value of each updated particle is calculated;

[0056] Based on the fitness value of each updated particle, the particles of the initial particle swarm are selected, crossed and mutated to obtain a next generation particle swarm. The fitness value calculation, selection, crossover and mutation of the particles of each generation particle swarm are repeatedly executed until the iteration of the maximum evolution number is completed, and the global optimal particle with the highest fitness value is output.

[0057] Decoding the globally optimal particle to obtain optimal weights and optimal thresholds;

[0058] Assigning the optimal weights and the optimal thresholds to the back propagation neural network to obtain a target back propagation neural network;

[0059] Training the target back propagation neural network using the sample robot arm state and the sample robot arm action to obtain an intermediate transition model.

[0060] As a preferred embodiment, the step of "training the target back propagation neural network using the sample robot arm state and the sample robot arm action to obtain an intermediate transition model" can specifically include:

[0061] In the process of training the target back propagation neural network using the sample robot arm state and the sample robot arm action, the learning rate of the target back propagation neural network is adjusted according to the following formula:

[0062]

[0063] wherein η is the learning rate, β is a constant between 0 and 1, γ is a constant between 1 and 2, n is the current training number of the network, n is a finite positive integer, d ik is the kth expected output value of the ith sample, a ik is the kth actual network output value of the ith sample, mse(n) represents the mean square error between the actual network output value and the expected output value, and N is the number of samples.

[0064] When the mean square error of the target back propagation neural network is less than the preset target mean square error or the training number of the target back propagation neural network reaches the preset maximum training number, the training of the target back propagation neural network is completed, and the intermediate transition model is obtained.

[0065] In this embodiment, on the basis of optimizing the weights and thresholds of the back propagation neural network using the particle swarm algorithm combined with the genetic algorithm, the adaptive learning rate is additionally added to the back propagation neural network, which further promotes the efficient learning and rapid convergence of the network and effectively improves the defects of the ordinary back propagation neural network, such as slow convergence speed and easy to fall into local minimum.

[0066] According to the above rules, the neural network based on genetic algorithm is designed to complete the planning of the motion trajectory of the mechanical arm. The reasons for selecting the genetic algorithm are as follows: strong global search capability, suitable for various problems; the value of each point in the solution space can be fully utilized, especially suitable for complex function optimization problems such as non-linear, non-convex and multi-peak; prior knowledge can be added to guide the search and improve the search efficiency; the genetic algorithm is used to realize the weight optimization, and the training set is used to train the neural network. The performance of the neural network is verified by using the test set, and the error test results of the neural network trained by using the back propagation neural network are as shown in the following figure: Figure 3 Figure 3 The horizontal coordinate in the figure represents the i-th test result, and the vertical coordinate represents the error e. The calculation formula of the error is set as follows:

[0067]

[0068] In the formula, x, y and z respectively represent the spatial position coordinates of the next step state of the mechanical arm corresponding to the standard action data, x', y' and z' represent the next step spatial position of the mechanical arm end calculated by the network; f and f' respectively represent the end force at the next time of the standard action data and the end force predicted by the network.

[0069] In addition, time delay inevitably occurs in the process of long-distance signal transmission. Compared with the time delay in the traditional control system, the signal transmission of the teleoperation system not only has delay in the control loop, but also has delay in the feedback loop. The delay of the whole system is the round-trip communication delay. The existence of the delay reduces the transparency and the sense of presence of the teleoperation system, brings great inconvenience to the operation of the operator, and even affects the stability of the teleoperation system. The passivity control method can ensure the stability under any time delay, which provides a simple and robust tool for analyzing nonlinear systems. The method allows connection to other systems and maintains the stability of the whole system. The wave variable method based on the passivity theory is widely used in teleoperation systems and has been studied by many scholars in recent years. The method converts the power variable into the wave variable for transmission through wave transformation. A small amount of parameters are adjusted to ensure the passivity of the system, thereby improving the robustness of the system. Wave transformation changes the transmission value of the speed, increases the position error, and reduces the transparency of teleoperation.

[0070] Therefore, the present application proposes a wave variable position correction method based on force feedback. The method designs an error accumulator to compensate the speed signal of the slave end according to the position error, so as to reduce the position error of the master and slave ends. The method uses a correction coefficient positively related to the position error to buffer the speed of the position change, and designs the correction coefficient to ensure the stability of the teleoperation system. Simulation proves that the proposed wave variable method reduces the synchronization position error of the master and slave ends and improves the tracking performance of the teleoperation system. ​

[0071] In addition, virtual reality technology has been introduced into various industries in recent years due to its unique sense of presence and space. A remote operation system can establish a virtual environment space at the slave end by introducing virtual reality technology, which can effectively predict the subsequent state of the actual operation space, improve the operator's sense of presence, and avoid safety accidents. At the same time, by directly controlling the virtual environment, the operator can control the slave end robot with the remote operation hand controller, which can be changed from the master end to the virtual environment and then to the slave end. This can bypass the operation delay problem and provide a new idea for the development of remote operation technology.

[0072] In the current use of remote operation control, the operator needs to analyze, find and judge problems existing in task execution through various sensors (force sensors, speed and acceleration sensors, etc.) and real-time images captured by cameras and handle them in a timely manner. However, due to the large number of parameters transmitted by various sensors, it is difficult to quickly analyze the problems occurring in the remote operation environment through data changes. Moreover, the information transmitted to the operator by the camera is not comprehensive. When the camera transmits complex information in three-dimensional space to a two-dimensional control interface, it ignores a lot of three-dimensional depth information, which will undoubtedly affect the operator's manual operation. Therefore, virtual reality technology can be used to assist remote operation to solve the above problems.

[0073] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order other than that described in the embodiments and still achieve desirable results. In addition, the processes depicted in the figures do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous or possible.

[0074] According to another aspect, embodiments of the present application provide a control device of a space robot arm. Figure 2 A schematic block diagram of a control device of a space robot arm according to an embodiment is shown. It can be understood that the device can be implemented by any device, equipment, platform and cluster of equipment with computing and processing capabilities. As shown, the device includes an acquisition unit 200 and a control unit 202. The main functions of each component unit are as follows: Figure 2 The acquisition unit 200 is configured to acquire a target robot arm state of a target assembly task to be executed by the space robot arm;

[0075]

[0076] ​The control unit 202 is configured to input the target robot arm state into a pre-trained operation control model, and output a target robot arm action corresponding to the target robot arm state, so as to control the motion of the space robot arm by using the target robot arm action.

[0077] As a preferred implementation, the robot arm state includes joint angle, joint speed, joint acceleration, end force and end position.

[0078] As a preferred implementation, the operation control model is trained by the following method:

[0079] Obtain standard action data of the space robot arm when performing an assembly task, and extract sample robot arm states and sample robot arm actions from the standard action data.

[0080] Input the sample robot arm states and the sample robot arm actions into a preset neural network for training to obtain an intermediate transition model.

[0081] Determine a correction value corresponding to each time point based on the task completion degree of the space robot arm at different time points.

[0082] Correct the weight of the intermediate transition model based on a vector composed of all the correction values to obtain the operation control model.

[0083] As a preferred implementation, the determination of the correction value corresponding to each time point based on the task completion degree of the space robot arm at different time points includes:

[0084] The task completion degree of the space robot arm at different time points is determined by the following formula:

[0085]

[0086] In the formula, p t is the task completion degree of the space robot arm at time t;

[0087] The correction value corresponding to each time point is determined by the following formula:

[0088]

[0089] In the formula, R t+1 is the correction value corresponding to time t+1, p t-1 is the task completion degree of the space robot arm at time t-1, a is a preset proportion coefficient, b is a preset activity range of the robot arm end, and ΔD is the spatial distance between the robot arm end and the target position.

[0090] As a preferred implementation, the neural network is a back propagation neural network.

[0091] As a preferred embodiment, the intermediate transition model is trained by the following way:

[0092] determining a network structure of a back propagation neural network; wherein the network structure comprises weights and thresholds;

[0093] encoding the weights and the thresholds of the back propagation neural network to obtain an initial particle swarm; wherein the weights and the thresholds are both taken as particles of the initial particle swarm;

[0094] setting algorithm parameters of a genetic algorithm; wherein the algorithm parameters comprise: a particle swarm size, a maximum evolution number, a crossover probability, a mutation probability, a learning rate, a velocity range and a position range of the particles;

[0095] initializing the velocity and the position of each particle in the initial particle swarm, and calculating a fitness value of each particle in the initial particle swarm;

[0096] updating the velocity and the position of each particle in the initial particle swarm, and calculating a fitness value of each particle after updating;

[0097] based on the fitness value of each particle after updating, selecting, crossing and mutating the particles in the initial particle swarm to obtain a next generation particle swarm, and cyclically performing the fitness value calculation, selection, crossing and mutation of the particles in each generation particle swarm until the iteration of the maximum evolution number is completed, and outputting a global optimal particle with the highest fitness value;

[0098] decoding the global optimal particle to obtain optimal weights and optimal thresholds;

[0099] assigning the optimal weights and the optimal thresholds to the back propagation neural network to obtain a target back propagation neural network;

[0100] training the target back propagation neural network by using sample robot arm states and sample robot arm actions to obtain an intermediate transition model.

[0101] As a preferred embodiment, the training of the target back propagation neural network by using sample robot arm states and sample robot arm actions to obtain an intermediate transition model comprises:

[0102] in the process of training the target back propagation neural network by using sample robot arm states and sample robot arm actions, adjusting the learning rate of the target back propagation neural network according to the following formula:

[0103]

[0104] wherein η is a learning rate, β is a constant between 0 and 1, γ is a constant between 1 and 2, n is a current training number of the network, n is a finite positive integer, d ik is a kth expected output value of the ith sample, a ik is a kth actual network output value of the ith sample, mse(n) represents a mean square error between the actual network output value and the expected output value, and N is a sample number.

[0105] When the mean square error of the target back propagation neural network is less than a preset target mean square error or a training number of the target back propagation neural network reaches a preset maximum training number, the training of the target back propagation neural network is completed, and an intermediate transition model is obtained.

[0106] According to an embodiment of still another aspect, an electronic device is also provided, which includes a memory and a processor, the memory has stored executable code, and the processor implements the executable code to implement the method as Figure 1 described above.

[0107] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment mainly explains the difference from other embodiments. Especially, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the related parts can be referred to the part of the method embodiments.

[0108] Those skilled in the art should be aware that, in one or more examples described above, the functions described with reference to the present application can be implemented in hardware, software, firmware or any combination thereof. When implemented in software, the functions can be stored in a computer readable medium or transmitted as one or more instructions or codes on a computer readable medium.

[0109] The above detailed description of the specific implementation of the present application further explains the purpose, technical solution and beneficial effects of the present application. It should be understood that the above detailed description is only the specific implementation of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solution of the present application should be included in the protection scope of the present application.

Claims

1. A method for controlling a space robot arm, characterized in that: include: Obtain the target manipulator state of the target assembly task to be performed by the space manipulator; Inputting the target manipulator state into a pre-trained operation control model, and outputting a target manipulator action corresponding to the target manipulator state, so as to control the motion of the spatial manipulator using the target manipulator action; The state of the robot arm includes joint angle, joint velocity, joint acceleration, end force and end position; The operation control model is trained in the following way: Acquiring standard motion data of the space manipulator when performing an assembly task, and extracting sample manipulator arm states and sample manipulator arm motions from the standard motion data; Inputting the sample robotic arm state and the sample robotic arm motion into a preset neural network for training to obtain an intermediate transition model; Based on the mission completion degree of the space manipulator at different times, determine the correction value corresponding to each moment; Based on the vector composed of all the correction values, the weight of the intermediate transition model is corrected to obtain an operation control model; The step of determining the correction value corresponding to each moment based on the mission completion degree of the space manipulator at different moments includes: The mission completion degree of the space manipulator at different times is determined by the following formula: Where p t is the mission completion degree of the space manipulator at time t; The correction value corresponding to each moment is determined by the following formula: Where R t+1 is the correction value corresponding to time t+1, p t-1 is the task completion degree of the space manipulator at time t-1, a is the preset proportional coefficient, b is the preset range of motion of the end of the manipulator, and ΔD is the spatial distance between the end of the manipulator and the target position; The neural network is a back propagation neural network; The intermediate transition model is trained in the following way: Determining a network structure of a back propagation neural network; wherein the network structure includes weights and thresholds; Encoding the weights and thresholds of the back propagation neural network to obtain an initial particle swarm; wherein the weights and thresholds are both used as particles of the initial particle swarm; Setting algorithm parameters of the genetic algorithm; wherein the algorithm parameters include: particle swarm size, maximum number of evolutions, crossover probability, mutation probability, learning rate, particle speed range and position range; Initializing the speed and position of each particle in the initial particle swarm, and calculating the fitness value of each particle in the initial particle swarm; The speed and position of each particle in the initial particle swarm are updated, and the fitness value of each updated particle is calculated; Based on the updated fitness value of each particle, the particles of the initial particle swarm are selected, crossed and mutated to obtain the next generation particle swarm, and the fitness value calculation, selection, crossover and mutation of the particles of each generation particle swarm are cyclically performed until the maximum number of evolutionary iterations is completed, and the global optimal particle with the highest fitness value is output; Decoding the global optimal particle to obtain an optimal weight and an optimal threshold; Assigning the optimal weight and the optimal threshold to the back propagation neural network to obtain a target back propagation neural network; The target back propagation neural network is trained using sample robotic arm states and sample robotic arm actions to obtain an intermediate transition model.

2. The method according to claim 1, characterized in that The target back propagation neural network is trained using the sample manipulator state and the sample manipulator action to obtain an intermediate transition model, including: In the process of training the target back propagation neural network using the sample manipulator state and the sample manipulator action, the learning rate of the target back propagation neural network is adjusted according to the following formula: Where η is the learning rate, β is a constant between 0 and 1, γ is a constant between 1 and 2, n is the current number of training times of the network, n is a finite positive integer, d ik is the kth expected output value of the i-th sample, a ik is the kth actual network output value of the i-th sample, mse(n) represents the mean square error between the actual network output value and the expected output value, and N is the number of samples; When the mean square error of the target back propagation neural network is less than a preset target mean square error or the number of training times of the target back propagation neural network reaches a preset maximum number of training times, the training of the target back propagation neural network is completed and an intermediate transition model is obtained.

3. A control device for a space robot arm, characterized in that: The method for performing any one of claims 1 to 2 comprises: An acquisition unit is configured to acquire a target manipulator state of a target assembly task to be performed by the space manipulator; The control unit is configured to input the target manipulator state into a pre-trained operation control model, and output a target manipulator action corresponding to the target manipulator state, so as to use the target manipulator action to control the movement of the spatial manipulator.

4. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 2 is implemented.

5. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 2.

Images