Inheritance-based network generation operating system

Abstract

The present application relates to the field of operating system, particularly to a kind of inherited network generation operating system.Application is controlled to graphic user interface and only configured posture control model power robot.The graphic user interface at least includes input module, network module, model generation determination module and output control module icon.Each time, draw connection between two icon corresponding input, output mark, controller controls background corresponding function module to establish communication connection and transmit preset data.The present application establishes control link between the background function module required by power robot posture control model and front-end graphic interface icon and graphic drawing operation, and the user draws connection simply and intuitively to graphic interface graph according to model training logic, generates posture control model training control graph.The background starts corresponding function module, and data is transmitted according to input-output connection line, realizes posture control model training operation, so that model training construction is more intuitive and simple, and easy to use.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of operating systems, in particular to an inheritance type network generation operating system. BACKGROUND

[0002] The control model of a powered robot refers to a set of mathematical models and algorithms used to control the movement and operation of the robot, which ensures that the robot can perform actions according to predetermined targets, such as control models that can be used to control the joints or wheels of the robot and maintain stability and efficiency in dynamic environments.

[0003] Generally, in order to make the powered robot better maintain the target posture in the corresponding use scene, a posture control model that conforms to the current use scene is configured for the robot. In the prior art, when training the posture control model of the powered robot, multiple steps and techniques are usually involved to ensure that the robot can accurately perform the expected action and maintain a stable posture. This process usually includes key links such as data collection, model selection, training method and evaluation. However, these links usually need to be adjusted through the configuration of the background program to build a complete model training process. The technical ability of the operator is required to be high, and the operation is more complex and cumbersome, which is not convenient to use. SUMMARY

[0004] To solve one of the above technical problems, the technical solution adopted by the present application is:

[0005] According to one aspect of the present application, an inheritance type network generation operating system is provided, which is applied to a controller and a powered robot. The controller has a graphical user interface, which at least includes: an input module icon, a network module icon, a model generation determination module icon and an output control module icon. Only a posture control model is configured in the powered robot.

[0006] Whenever a connection line is drawn between the input label and the output label corresponding to the two icons, the controller controls the communication connection between the function modules corresponding to the two icons in the background to transmit preset data.

[0007] When the connection between the input module icon, the network module icon, the model generation determination module icon and the output control module icon is completed, the controller controls the function modules corresponding to each icon to perform the following steps:

[0008] The network module corresponding to the network model icon is configured to start the posture control model of the power robot; the posture control model is configured to generate a posture adjustment instruction of the power robot according to an input state parameter; the state parameter includes current posture data of the power robot; the posture control model includes an input layer, an output layer and at least one hidden layer, and full connection is arranged between two adjacent layers;

[0009] The output control module corresponding to the output control module icon is configured to adjust the posture of the power robot according to the posture adjustment instruction.

[0010] The input module corresponding to the input module icon is configured to obtain the current posture data of the power robot and input the current posture data into the posture control model and the model generation determination module corresponding to the model generation determination module icon respectively.

[0011] The model generation determination module corresponding to the model generation determination module icon is configured to perform model parameter solidification processing according to the current posture data in each training period, so as to update the model weight of the posture control model.

[0012] The model parameter solidification processing includes:

[0013] Integrating the difference between the current posture data and the adjacent historical posture data to generate an adjustment posture coefficient of the posture adjustment instruction corresponding to the current posture data;

[0014] If the absolute value of the difference between the adjustment posture coefficient corresponding to the current posture data and a target posture coefficient is greater than or equal to a preset validity threshold, the model training invalidity coefficient is accumulated by one;

[0015] If the absolute value of the difference between the adjustment posture coefficient corresponding to the current posture data and the target posture coefficient is less than the preset validity threshold, the model training validity coefficient is accumulated by one;

[0016] When the model solidification node corresponding to the current training period is reached, if the ratio between the model training validity coefficient increment and the model training invalidity coefficient increment in the training period is greater than a preset proportion threshold, the model weight in the posture control model is updated to the model weight at the current model solidification node.

[0017] The present application has at least one of the following beneficial effects:

[0018] The various background function modules required by the posture control model of the power robot are respectively associated with the icons in the front-end graphical user interface and the dependent graphical drawing operation to establish relevant control links, so that the user can generate a training control graph of a certain posture control model by performing a relatively simple and intuitive drawing connection on the corresponding graph in the graphical user interface according to the logic of model training. The corresponding background will start the function modules corresponding to the graphs, and transmit data in different modules according to the input-output connection lines between the graphs to realize the training operation of the posture control model in the background. Therefore, the construction work of model training can be made more intuitive and simple, and is convenient to use.

[0019] Meanwhile, in the model generation determination module of the present application, whether the posture of the power robot is more close to the target posture after being adjusted by the posture adjustment instruction output by the posture control model each time in the training period is determined to determine whether the current training optimization direction of the model is effective. Then, whether the posture control model is converging to the target state in the current training period is determined by the proportion between the number of invalid optimization in the whole training period, and the model weight parameter is fixed. Therefore, in each subsequent period, the model is continuously trained on the basis of the model with updated parameters obtained in the previous training period, which can make the model gradually converge to the target state quickly, and the training efficiency is higher. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 A schematic diagram of a graphical user interface in an inheritance type network generation operating system is provided for the embodiments of the present application.

[0022] Figure 2 An interface schematic diagram of a posture control model after updating the model weight after a training period is completed in the posture control model training process using the inheritance type network generation operating system is provided for the embodiments of the present application.

[0023] Figure 3 An interface schematic diagram of a posture control model after the posture control model training using the inheritance type network generation operating system is completed is provided for the embodiments of the present application.

[0024] Figure 4 A flowchart of an inheritance type network training method is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort fall within the protection scope of the present application.

[0026] As a possible embodiment of the present application, as shown in Figure 1 , a hereditary network generation operating system is provided, which is applied to a controller and a power robot, the controller has a graphical user interface, as shown in Figure 1 , the graphical user interface at least includes: an input module icon (i.e. an angle sensor corresponding icon), a network module icon (i.e. a computing network corresponding icon), a model generation determination module icon (i.e. a function corresponding icon) and an output control module icon (i.e. a motor control corresponding icon). The power robot is only configured with a posture control model.

[0027] Specifically, in the embodiment, the power robot refers to those robots that can move or operate autonomously, have certain intelligent control ability, and are usually driven by electric motors, hydraulic systems or pneumatic systems, etc. Such as industrial robots, unmanned aerial vehicles, robot dogs, humanoid robots and service robots, etc.

[0028] Whenever a connection line is drawn between the input label and the output label corresponding to the two icons respectively, the controller controls the function modules corresponding to the two icons in the background to establish a communication connection to transmit preset data.

[0029] Of course, in the graphical user interface of the operating system, the configuration function of the function module corresponding to the icon in the background can also be started by clicking an input module icon, and then adaptive configuration operation is performed on the function module. For example, the configuration function of the input module can be started by double-clicking the input module icon, and then the number of output label graphics is set according to the need to configure which posture data to collect input, so that the obtained data can be sent to multiple modules respectively.

[0030] When the connection between the input module icon, the network module icon, the model generation determination module icon and the output control module icon is completed, the controller controls the function modules corresponding to the icons to perform the following steps:

[0031] The network module corresponding to the network model icon is configured to start a posture control model of the powered robot. The posture control model is configured to generate a posture adjustment instruction of the powered robot according to an input state parameter. The state parameter includes current posture data of the powered robot. The posture control model includes an input layer, an output layer, and at least one hidden layer, and full connection is provided between two adjacent layers.

[0032] Specifically, the posture control model is a deep learning neural network model. The number of nodes of the input layer and the output layer in the posture control model can be determined according to the type and number of posture data to be collected and the number of movement joints of the powered robot to be controlled. Taking a quadcopter unmanned aerial vehicle as the powered robot, and taking the posture control model as a model used to maintain the unmanned aerial vehicle in a horizontal and centered posture, and taking the posture data including a pitch angle and a roll angle as an example, the number of nodes of the input layer is 2, and the number of nodes of the output layer is 4. The number of layers of the hidden layer can be one, two, or three. Since the posture control logic of the powered robot is relatively simple, a good control of the posture of the powered robot can be achieved by using a plurality of AND or NAND strategies. Correspondingly, in the neural learning network, a hidden layer within three layers can completely express the combination of a plurality of AND or NAND strategies, and thus the number of layers of the hidden layer and the number of nodes of each layer can be adaptively set according to needs in this embodiment.

[0033] The output control module corresponding to the output control module icon is configured to adjust the posture of the powered robot according to the posture adjustment instruction.

[0034] The input module corresponding to the input module icon is configured to acquire current posture data of the powered robot and input the current posture data into the posture control model and the model generation determination module corresponding to the model generation determination module icon.

[0035] The current posture data of the powered robot is acquired by:

[0036] According to a preset data acquisition interval, the current posture data of the powered robot is acquired according to information received by the powered robot through ROS2 publishing at each acquisition time. In each preset data acquisition interval, the powered robot performs at least one posture adjustment according to the posture adjustment instruction.

[0037] In the current powered robot, the topic mechanism in ROS2 (Robot Operating System 2) is basically used to publish the current posture value of the unmanned aerial vehicle. Therefore, the input module in this embodiment can convert the received posture information into data in a corresponding input format according to the corresponding acquisition time.

[0038] Generally, the power robot needs to re-acquire the adjusted posture information after each adjustment according to the posture adjustment instruction, and then input the model generation determination module and the posture control model again to perform the next training adjustment. Therefore, in the embodiment, the preset data acquisition interval is generally the time interval required for one complete posture adjustment. At the same time, a training period needs to include multiple posture adjustments, so the length of the training period is greater than the length of the preset data acquisition interval.

[0039] In addition, the posture input module can also acquire data of other influencing factors in addition to the current posture data. For example, in the graphical user interface, the input module icons can include posture input module icons and environmental influencing factor input module icons.

[0040] The input module corresponding to the posture input module icon is used to acquire the current posture data of the power robot, and input the model generation determination module corresponding to the model generation determination module icon and the posture control model respectively.

[0041] The input module corresponding to the environmental influencing factor input module icon is used to acquire environmental data that affects the posture in the current environment of the power robot, and input the model generation determination module corresponding to the model generation determination module icon and the posture control model respectively. For example, the environmental data can be wind speed.

[0042] The model generation determination module corresponding to the model generation determination module icon is used to perform model parameter solidification processing according to the current posture data in each training period, so as to update the model weight of the posture control model.

[0043] The model parameter solidification processing includes:

[0044] Integrating the difference between the current posture data and the adjacent historical posture data to generate the adjustment posture coefficient of the posture adjustment instruction corresponding to the current posture data.

[0045] If the absolute value of the difference between the adjustment posture coefficient corresponding to the current posture data and the target posture coefficient is greater than or equal to the preset validity threshold, the model training invalidity coefficient is accumulated by one.

[0046] If the absolute value of the difference between the adjustment posture coefficient corresponding to the current posture data and the target posture coefficient is less than the preset validity threshold, the model training validity coefficient is accumulated by one.

[0047] When the model solidification node corresponding to the current training period is reached, if the ratio between the model training validity coefficient increment and the model training invalidity coefficient increment in the training period is greater than the preset proportion threshold, the model weight in the posture control model is updated to the model weight at the current model solidification node.

[0048] When reaching the model solidification node corresponding to the current training period, if the ratio between the model training effective coefficient increment and the model training ineffective coefficient increment in the training period is less than or equal to the preset proportion threshold, the model weight in the attitude control model is randomly changed.

[0049] If the attitude control effect of the current attitude control model reaches the preset condition, the model parameter solidification processing is stopped, and the current model weight is configured to the local attitude control model of the power robot.

[0050] The model generation determination module is a key functional module in the embodiment. Whether the attitude adjustment instruction output by the model is effective is determined by the attitude change trend of the current power robot after each attitude adjustment, whether the trend is close to the target attitude. Whether the model converges effectively to the target state in the current training period is determined by whether the ratio between the effective adjustment times and the ineffective adjustment times in each training period reaches the threshold.

[0051] If yes, the weight parameters of the model are updated and saved for further optimization in the subsequent training period based on the current state, so as to improve the training speed of the model by inheriting the training results of the previous period.

[0052] If no, the weight parameters of the model are randomly changed so that the model jumps into another feasible region for optimization.

[0053] Specifically, the power robot is a quadcopter, the attitude control model is used to maintain the quadcopter in a horizontal attitude, the attitude data includes the pitch angle, the preset data collection interval is 20 ms, the target pitch angle is 0°, and the preset proportion threshold is 0.8. Taking a preset effectiveness threshold of 30° as an example, the model parameter solidification processing in a training period (1S) is described:

[0054] The current collected pitch angle is A t , the last adjacent historical pitch angle is A t-1 , and A t The adjustment attitude coefficient Pt corresponding to the attitude adjustment instruction satisfies the following conditions:

[0055] ;

[0056] Wherein, b is the current data collection times.

[0057] Therefore, the above integral operation can represent the deviation of the current attitude of the quadcopter from 0° after each attitude adjustment. Therefore, if the absolute value of Pt is greater than 30°, it is considered that the current adjustment deviates from the target too much, and the adjustment strategy is invalid, and the model training ineffective coefficient is accumulated by one. Otherwise, the model training effective coefficient is accumulated by one.

[0058] Finally, there are 50 times of whether the effective determination in the whole training cycle, if the proportion of the number of times of the whole is more than 0.8, or the ratio of the number of times of the effective and the number of times of the invalid is greater than 4, the model weight in the posture control model is updated to the model weight when the current model is solidified. Otherwise, the model weight in the posture control model is randomly changed.

[0059] In this embodiment, the various background function modules required for the posture control model of the power robot are respectively associated with the icons and the corresponding graphical drawing operations in the front-end graphical user interface to establish relevant control links, so that the user can generate a training control graph of a certain posture control model according to the logic of model training, and perform relatively simple and intuitive drawing connections on the corresponding graphs in the graphical user interface. The corresponding background will start the function modules corresponding to the graphs, and transmit data in different modules according to the input-output connection lines between the graphs, so as to realize the training operation of the posture control model in the background. Therefore, the construction work of model training can be made more intuitive and simple, and is convenient to use.

[0060] Meanwhile, in this embodiment, the model generation determination module automatically determines whether the current training optimization direction of the model is effective according to whether the posture of the power robot after being adjusted by the posture adjustment instruction output by the posture control model each time in the training cycle tends to the target posture. Then, whether the posture control model is converging to the target state in the current training cycle is determined by the proportion between the number of invalid optimization times in the whole training cycle, and the model weight parameters are fixed. Therefore, in each subsequent cycle, the model is gradually and quickly converging to the target state based on the model with updated parameters obtained in the previous training cycle, and the training efficiency is higher.

[0061] In this embodiment, the use of the inherited network generation operation system in this embodiment is also provided to train and generate a posture control network model of a self-adjusting gimbal robot (i.e., a camera stabilizer). The gimbal robot controls the swing of the support through four driving motors below to change the posture of the camera above. The corresponding posture control model is used to ensure that the camera above is in a horizontal state. The input layer of the model has 6 input nodes corresponding to the X-direction acceleration, Y-direction acceleration, Z-direction acceleration, roll angle, pitch angle, and yaw angle. The output layer has 4 corresponding output control nodes controlling the four motors, and a hidden layer.

[0062] In the model training generation process using the system in this embodiment, as shown in FIG. 12, the model training process is as follows: Figure 2As shown in FIG. 6, it is a posture control model after a certain training period in the training process, and the content displayed in the lower left corner of the frame is the model weight recorded at each time of fixing the model weight. As shown in FIG. 7, it is a posture control model after the training is completed. Figure 3 As shown in FIG. 8, it is a posture control model after the training is completed.

[0063] As another embodiment of the present application, as shown in FIG. 9, a method for training an inherited network is also provided. The method can further train a power robot equipped with a mature original posture control model, and can make the new model inherit the ability of the original posture control model and further optimize and improve the posture control ability of the power robot in a new use scenario without performing two model fusion superposition. Figure 4

[0064] Specifically, the method comprises the following steps:

[0065] S100: A posture control optimization model of a power robot is newly created. The posture control optimization model (equivalent to the posture control model in the previous embodiment) is used to generate a posture adjustment instruction of the power robot according to an input state parameter. The state parameter comprises current posture data of the power robot. The posture control optimization model comprises an input layer, an output layer and at least one hidden layer, and is fully connected between two adjacent layers. The power robot has been configured with an original posture control model. The original posture control model is a posture control model generated by the power robot to meet a first use task requirement.

[0066] In the embodiment, the original posture control model and the newly created posture control optimization model are different. The difference is not only the difference in model ability, but also the difference in model type. For example, the original posture control model is a PID control model, and the posture control optimization model is a deep learning neural network model.

[0067] The original posture control model is usually a general control model trained by the manufacturer when the power robot is shipped. The first use task requirement to which the model is usually targeted in the training process is different from the second use task requirement to which the posture control optimization model is targeted in the embodiment.

[0068] Taking a humanoid power robot as an example, the first use task requirement is that the power robot can stand up again after falling down in different use scenarios, and the second use task requirement is that the power robot can stand up again after falling down in different use scenarios while carrying a certain load.

[0069] ​S200: input the acquired current attitude data of the power robot into an attitude control optimization model and a training determination module, respectively. The attitude control optimization model is a newly built neural learning network model to be trained for the power robot to meet a second use task requirement. The attitude influencing factors of the power robot in the first use task requirement and the second use task requirement are different.

[0070] S300: the training determination module is configured to execute model parameter solidification processing according to the current attitude data in each training cycle to update the model weight of the attitude control optimization model.

[0071] The model parameter solidification processing includes:

[0072] S301: integrate the difference between the current attitude data and the adjacent historical attitude data to generate an adjustment attitude coefficient of the attitude adjustment instruction corresponding to the current attitude data.

[0073] S302: if the absolute value of the difference between the adjustment attitude coefficient corresponding to the current attitude data and the target attitude coefficient is greater than or equal to a preset validity threshold, the model training invalid coefficient is accumulated by one.

[0074] S303: if the absolute value of the difference between the adjustment attitude coefficient corresponding to the current attitude data and the target attitude coefficient is less than the preset validity threshold, the model training valid coefficient is accumulated by one.

[0075] S304: when reaching the model solidification node corresponding to the current training cycle, if the ratio between the model training valid coefficient increment and the model training invalid coefficient increment in the training cycle is greater than a preset proportion threshold, the model weight in the attitude control optimization model is updated to the model weight at the current model solidification node.

[0076] S305: when reaching the model solidification node corresponding to the current training cycle, if the ratio between the model training valid coefficient increment and the model training invalid coefficient increment in the training cycle is less than or equal to the preset proportion threshold, the model weight in the attitude control optimization model is randomly changed.

[0077] In the embodiment, the power robot trains the newly constructed posture control optimization model to adapt to a new use scenario. The training of the posture control optimization model is based on the regulation and control instructions output by the model and the adjusted posture information of the power robot, which are used as the basis for judging whether the current adjustment of the model is effective to drive the model to converge to the target state. In this process, the original posture control model deployed in the power robot is also running in parallel, that is, the adjusted posture information of the power robot contains the adjustment information sent by the original posture control model to the power robot. Therefore, when training the newly constructed posture control optimization model, the training data (i.e., posture information) contains the adjustment information sent by the original posture control model to the power robot and the compensation adjustment information output by the newly constructed posture control optimization model. Correspondingly, the posture adjustment instructions output by the trained posture control optimization model take into account the compensation information of the posture adjustment instructions output by the original posture control model, so that the current posture control optimization model and the original posture control model do not need to be fused and superimposed, and the posture optimization adjustment of the power robot in adapting to the new use scenario can be further realized on the basis of the original posture control model. The method consumes less time, is more efficient, and does not have the problem of incompatibility between models.

[0078] In addition, in the embodiment, the training determination module determines whether the current training optimization direction of the model is effective according to whether the posture of the power robot is more close to the target posture after being adjusted by the posture adjustment instructions output by the posture control optimization model each time in the training period. Then, the proportion between the number of invalid optimization times in the entire training period is used to determine whether the posture control optimization model is converging to the target state in the current training period, and the model weight parameters are fixed. Therefore, in each subsequent period, the model is trained based on the model with updated parameters obtained in the previous training period, which can make the model gradually converge to the target state quickly and improve the training efficiency.

[0079] In addition, although the steps of the method in the present disclosure are described in a specific order in the drawings, this does not require or imply that the steps must be performed in this specific order or that all the steps shown must be performed to achieve the desired results. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step, and / or one step can be divided into multiple steps, etc.

[0080] Those skilled in the art can clearly understand the example embodiments described herein through the above description of the example embodiments that the example embodiments described herein can be implemented by software or by software in combination with necessary hardware. Therefore, the technical solutions according to the embodiments of the present disclosure can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, or the like) or on a network, and includes a plurality of instructions to enable a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, or the like) to execute the method according to the embodiments of the present disclosure.

[0081] The above merely describes specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements within the technical scope disclosed by the present application can be easily conceived by those skilled in the art, and should be encompassed within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An inheritance network generating operating system, characterized by The system is applied to a controller and a powered robot, the controller has a graphical user interface, the graphical user interface at least includes: an input module icon, a network module icon, a model generation determination module icon and an output control module icon; only a posture control model is configured in the powered robot; Whenever an additional connection line is drawn between the input label and the output label corresponding to the two icons, the controller controls the functional modules corresponding to the two icons in the background to establish a communication connection to transmit preset data; When the connection drawing between the input module icon, the network module icon, the model generation determination module icon and the output control module icon is completed, the controller controls the functional modules corresponding to the icons to perform the following steps: The network model icon corresponds to a network module, which is used to start the posture control model corresponding to the powered robot; the posture control model is used to generate a posture adjustment instruction of the powered robot according to an input state parameter; the state parameter includes current posture data of the powered robot; the posture control model includes an input layer, an output layer and at least one hidden layer, and is fully connected between two adjacent layers; The output control module icon corresponds to an output control module, which is used to adjust the posture of the powered robot according to the posture adjustment instruction; The input module icon corresponds to an input module, which is used to obtain current posture data of the powered robot and input the current posture data into the posture control model and the model generation determination module icon corresponding to a model generation determination module; The model generation determination module icon corresponds to a model generation determination module, which is used to perform model parameter solidification processing according to the current posture data in each training period; and update the model weight of the posture control model; The model parameter solidification processing includes: Integrating the difference between the current posture data and the adjacent historical posture data to generate an adjustment posture coefficient of the posture adjustment instruction corresponding to the current posture data; If the absolute value of the difference between the adjustment posture coefficient corresponding to the current posture data and a target posture coefficient is greater than or equal to a preset validity threshold, the model training invalidity coefficient is accumulated by one; If the absolute value of the difference between the adjustment posture coefficient corresponding to the current posture data and the target posture coefficient is less than the preset validity threshold, the model training validity coefficient is accumulated by one; When the model solidification node corresponding to the current training period is reached, if the ratio between the model training validity coefficient increment and the model training invalidity coefficient increment in the training period is greater than a preset proportion threshold, the model weight in the posture control model is updated to the model weight at the current model solidification node.

2. The generational network generation operating system of claim 1, wherein, The current posture data of the powered robot includes: According to a preset data collection interval, the posture data of the powered robot corresponding to each collection time is obtained; in each preset data collection interval, the powered robot performs at least one posture adjustment according to the posture adjustment instruction.

3. The generational network generation operating system of claim 2, wherein, The current posture data of the powered robot is obtained according to the information received by the powered robot through ROS2 publishing at each collection time. ​ 4. The generational network generation operating system of claim 1, wherein, The posture control model is a deep learning neural network model.

5. The generational network generation operating system of claim 4, wherein, The number of hidden layers in the posture control model is one, two or three.

6. The generational network generation operating system of claim 1, wherein, The power robot is a drone, and the current posture data includes a pitch angle and / or a roll angle.

7. The generational network generation operating system of claim 1, wherein, When reaching the model solidification node corresponding to the current training cycle, the model parameter solidification process further includes: If, during the training cycle, the ratio of the model training effective coefficient increment to the model training invalid coefficient increment is less than or equal to a preset ratio threshold, the model weights in the posture control model are randomly changed.

8. The generational network generation operating system of claim 1, wherein, The model generation and determination module is further configured to perform the following steps: If the posture control effect of the current posture control model reaches the preset condition, the model parameter solidification process is stopped, and the current model weight is configured into the local posture control model of the power robot.

9. The generational network generation operating system of claim 1, wherein, The input module icons include a posture input module icon and an environmental influence factor input module icon; The input module corresponding to the posture input module icon is used to obtain the current posture data of the power robot and input it into the posture control model and the model generation determination module corresponding to the model generation determination module icon respectively; The input module corresponding to the environmental influencing factor input module icon is used to obtain environmental data that affects the posture of the power robot in the current environment, and input it into the posture control model and the model generation judgment module corresponding to the model generation judgment module icon respectively.

10. The generational network generation operating system of claim 9, wherein, The environmental data includes wind speed.

Images