Robot Motion Control Method, Device and Medium

Abstract

The present invention discloses a robot motion control method, device and medium. The method includes: determining the target point position, target motion path and target mode of the robot; performing point-to-point motion control on the robot according to the target point position, target motion path and target mode, and based on a pre-trained controller, where the controller is modeled based on the width learning algorithm and the controller model is trained. The present invention avoids the complex process of repeated parameter tuning and reduces the professional requirements for users.

Description

Technical Field

[0001] The present invention relates to the technical field of robot control, and particularly to a robot motion control method, device and medium. Background Art

[0002] Intelligent robots are now widely used in many fields including engineering manufacturing, medical treatment, and micro-operations. In common applications of intelligent robots, robots often need to complete the motion of complex paths in a specific posture. For example, in a polishing operation, the robot needs to ensure a specific posture and complete a variety of complex motions to achieve the required smoothness of the workpiece surface. At the same time, the robot sometimes also needs to avoid some positions that do not need to be polished, which involves path planning and posture control. This control process of knowing the positions of the known target point and the starting point and controlling the robot to move from the starting point to the target point under specific conditions is called point-to-point motion control.

[0003] In some existing point-to-point motion control methods, one method is to use a theoretically calculated method based on pre-programming for point-to-point motion control of the robotic arm. This method designs the speeds and accelerations of the robot at multiple characteristic positions in the path using complex calculation formulas according to the pre-known path, thus achieving smooth point-to-point control and reducing the vibration during the motion of the robotic arm. However, this method requires repeated calculations for different paths and is more suitable for pipeline work with single and repeated paths.

[0004] Another method is to use a closed-loop control system to achieve autonomous navigation of a mobile robot in a dynamic environment. This method enables the robot to autonomously achieve human-like navigation behaviors, improving the safety and efficiency of robot navigation. However, this method is based on closed-loop control, involves complex motion modeling and controller parameter adjustment processes, and this scheme also requires an additional design of a path planning algorithm outside of motion control, increasing the complexity of the overall design and the professional requirements for users. Summary of the Invention

[0005] The main objective of the embodiments of the present invention is to provide a robot motion control method, device and medium, aiming to avoid complex repeated parameter adjustment processes in robot motion control and reduce the professional requirements for users.

[0006] To achieve the above objective, an embodiment of the present invention provides a robot motion control method, and the method includes:

[0007] Determine the target point position, target motion path and target mode of the robot;

[0008] According to the target point position, target motion path, and target mode, and based on a pre-trained controller, perform point-to-point motion control on the robot, where the controller is modeled based on the width learning algorithm and the controller model is trained.

[0009] Optionally, the step of performing point-to-point motion control on the robot according to the target point position, target motion path, and target mode, and based on a pre-trained controller includes:

[0010] Calculate the target control rate according to the target motion path and target mode and based on a pre-trained controller;

[0011] Execute the target control rate through a low-order system and calculate the current position and attitude of the robot;

[0012] Repeat the above steps to calculate the position and attitude of the robot at the next time point until the target point position;

[0013] Based on the calculated position and attitude of the robot, perform point-to-point motion control on the robot.

[0014] Optionally, before the step of determining the target point position, target motion path, and target mode of the robot, it further includes:

[0015] Perform controller modeling based on the width learning algorithm and train the controller model to obtain a trained controller.

[0016] Optionally, the control system of the robot includes: a high-order dynamic control system and a low-order control system. The step of performing controller modeling based on the width learning algorithm and training the controller model to obtain a trained controller includes:

[0017] Teach and sample the motion process of the robot to obtain teaching data;

[0018] Use the width learning algorithm through the high-order dynamic control system to model the controller to obtain a controller model, and use the extreme learning machine to fit the nonlinear relationship of the low-order control system, and integrate to obtain the sample control rate of the control system;

[0019] Analyze and derive the system stability constraints based on the sample control rate;

[0020] Import the teaching data and system stability constraints into the controller model, and train the controller model based on the width learning algorithm, the sample control rate, and combined with the solution of the optimization function to obtain the final parameters of the controller, and obtain a trained controller.

[0021] Optionally, the step of teaching and sampling the motion process of the robot to obtain teaching data includes:

[0022] Based on different satisfactory performance indicators, use multiple adjusted sample controllers to control the robot to complete the same motion path and mode;

[0023] Record the position, speed, attitude, and magnetic field data of the robot at each sampling time;

[0024] Repeat the above two steps for different motion paths and modes to collect teaching data.

[0025] Optionally, the step of using the width learning algorithm to model the controller through the high-order dynamic control system to obtain a controller model, fitting the nonlinear relationship of the low-order control system using the extreme learning machine, and integrating to obtain the sample control rate of the control system includes:

[0026] Use the width learning algorithm through the high-order dynamic control system to model the controller to obtain a controller model, and calculate the high-order system control rate;

[0027] Fit the nonlinear relationship of the low-order control system using the extreme learning machine to obtain the low-order system control rate;

[0028] Combine the high-order system control rate and the low-order system control rate to form the sample control rate of the complete control system.

[0029] Optionally, in the step of importing the teaching data and system stability constraints into the controller model, and training the controller model based on the width learning algorithm, the sample control rate, and combined with the optimization function to solve, to obtain the final parameters of each controller and get a trained controller, it includes:

[0030] Solve the constraint conditions of the controller parameters through the Lyapunov theory in the width learning algorithm.

[0031] Optionally, the controller model satisfies the following conditions for the high-order dynamic system: under the control of the control rate, the robot needs to reach the specified position while satisfying the specified attitude; during the point-to-point motion process, the motion of the robot needs to satisfy a specific motion trajectory or a specific motion attitude.

[0032] Optionally, the method further includes:

[0033] Update the controller model by adding new teaching data in a real-time online update manner to obtain a controller that can handle different requirements.

[0034] An embodiment of the present invention also provides a robot control device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the robot motion control method described above is implemented.

[0035] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the robot motion control method described above is implemented.

[0036] A robot motion control method, device, and medium provided by an embodiment of the present invention determine the target point position, target motion path, and target mode of the robot; and perform point-to-point motion control on the robot according to the target point position, target motion path, and target mode, and based on a pre-trained controller. The controller is modeled based on the width learning algorithm and the controller model is trained. The solution of the present invention uses the width learning-based algorithm. On the one hand, it solves the drawback of the pre-programming method that requires re-design for different paths. On the other hand, it enables the user to manually control the robot to move according to the requirements when facing different needs, thereby collecting teaching data, and a controller that can handle different needs can be updated through real-time online updates. Accordingly, the method of the present invention avoids the complex process of repeated parameter adjustment, reduces the professional requirements for users, and improves the accuracy of motion control.

[0037] Compared with the prior art, the width learning algorithm used in the present invention has the characteristic of supporting online model update. Therefore, for newly added teaching data, the present invention does not need to re-train the controller, but directly obtains a controller that has both the characteristics of the original controller and the characteristics of the newly added teaching data by adding nodes. This method is very suitable for solving the variability of micro-robot motion control and motion paths, and can avoid frequent adjustment of the controller. While ensuring a certain accuracy, it has the advantages of being fast, simple, and supporting incremental online model update. Therefore, when applying the width learning method to the problem of micro-robot point-to-point motion control, new training data can be added in real time and quickly, thus avoiding frequent re-training of the controller. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the functional modules of the device to which the robot motion control device of the present invention belongs;

[0039] Figure 2 It is a schematic flowchart of an embodiment of the robot motion control method of the present invention;

[0040] Figure 3 It is a schematic flowchart of another embodiment of the robot motion control method of the present invention;

[0041] Figure 4 This is the overall system diagram including high-order and low-order systems in the embodiments of the present invention;

[0042] Figure 5 This is a detailed flowchart showing the process of modeling a controller based on the width learning algorithm and training the controller model to obtain a trained controller in the embodiments of the robot motion control method of the present invention;

[0043] Figure 6 This is a motion control diagram based on the width learning algorithm in the embodiments of the present invention;

[0044] Figure 7 This is a schematic diagram of the functional modules of an embodiment of the robot motion control device of the present invention.

[0045] The realization, functional features, and advantages of the object of the present invention will be further described in conjunction with the embodiments with reference to the accompanying drawings. Detailed implementation manners

[0046] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] The main solution of the embodiments of the present invention is as follows: By determining the target point position, target motion path, and target mode of the robot; according to the target point position, target motion path, and target mode, and based on a pre-trained controller, perform point-to-point motion control on the robot, and the controller is obtained by modeling a controller based on the width learning algorithm and training the controller model. The solution of the present invention uses an algorithm based on width learning. On the one hand, it solves the disadvantage that the pre-programming method needs to be redesigned for different paths. On the other hand, it also enables the user to only need to manually control the robot to move according to the requirements when facing different needs, so as to collect teaching data, and a controller that can handle different needs can be updated through real-time online updating. Accordingly, the method of the present invention avoids the complex process of repeated parameter adjustment and reduces the professional requirements for users.

[0048] The embodiments of the present invention consider that: The existing point-to-point motion control methods can be roughly divided into three categories. Among them:

[0049] The first category of methods is to control the robot to perform motion control on a specific path through pre-programming. On the one hand, this method is difficult to make real-time corrections and adjustments. On the other hand, it is impossible to achieve motion control of complex paths because there is often a strong non-linear relationship between the motion control instructions and the robot's motion behavior, and relying solely on the pre-programming method is likely to produce large errors when the path is complex.

[0050] The second type of method is achieved through classical control techniques, such as PID, fuzzy, sliding mode control, etc. Using this method to achieve the point-to-point motion control of a robot often requires prior modeling of the nonlinear relationship between the robot's motion and the actuator. Therefore, this type of method often requires a complex process to determine the controller parameters. At the same time, the controller constructed based on this method is very sensitive to the dynamic modeling of the robot's motion and the actuator, which further increases the difficulty of determining the controller parameters. On the other hand, using this method also requires the design of a separate trajectory tracking algorithm for the point-to-point motion trajectory at the same time, which undoubtedly also increases the complexity of the robot's point-to-point motion control. Finally, for the field of micro-robot control involved in the present invention, the desired state value of the robot changes frequently, so the controller parameters also need to be adjusted frequently accordingly, which is difficult to achieve for users without professional knowledge of control systems.

[0051] The third type of method is to solve the trajectory tracking problem through teaching learning. The teaching learning method provides a solution for non-professionals to adjust the controller, but the existing teaching learning methods applied to the trajectory control of micro-robots still have the following problems. First, some learning algorithms, such as the Extreme Learning Machine (ELM), are difficult to directly add new teaching data after the training process is completed. For problems such as the control of micro-robots that require frequent adjustment of control parameters and control paths, each time new teaching data is added, the controller needs to be retrained. Second, the use of the Broad Learning System (BLS) can achieve the addition of new training data without retraining, but this method has not been applied to the field of micro-robot point-to-point motion control yet.

[0052] Therefore, the embodiment of the present invention proposes a solution. By sampling a variety of different point-to-point motion trajectories of human teaching, a data-driven control model is established to control the micro-robot to perform point-to-point motion that meets specific conditions (such as trajectory or posture). The present invention first introduces the broad learning algorithm into the micro-robot point-to-point motion control, and simultaneously performs modeling based on the learning algorithm for both high-order systems and low-order systems. This method has a relatively flexible structure, and while ensuring a certain accuracy, it has the advantages of being fast, simple, and supporting incremental online model updates. Therefore, when applying the broad learning algorithm to the micro-robot point-to-point motion control problem, new training data can be added quickly and in real time, thus avoiding frequent retraining of the controller.

[0053] The present invention mainly solves two difficulties in the point-to-point motion control of micro-robots. (1) For a given target point, the robot needs to reach the target point while meeting specific pose requirements. (2) During the process of the robot reaching the target point, it needs to pass through a special path (such as obstacle avoidance) or meet specific poses. While solving the above two problems, the method of the present invention also needs to have sufficient generalization performance, so that for different micro-robots and actuators, the control rate sufficient to solve the above two difficulties can be obtained through a learning algorithm.

[0054] The controller design method proposed by the present invention does not require professional control or programming knowledge. Users only need to teach the robot motion control to obtain a controller with teaching performance, which is suitable for industrial people.

[0055] In addition, the point-to-point motion control method of the robot proposed by the present invention is a general imitation learning method, which is applicable to any point-to-point motion control problem.

[0056] Specifically, referring to Figure 1 , Figure 1 is a schematic diagram of the functional modules of the device to which the robot motion control device of the present invention belongs. The robot motion control device can be a device independent of the device and capable of data processing, and it can be carried on the device in the form of hardware or software. The device can be an intelligent mobile terminal with data processing functions such as a mobile phone or a tablet computer, or a fixed terminal device or a server with data processing functions, etc.

[0057] In this embodiment, the device to which the robot motion control device belongs at least includes an output module 110, a processor 120, a memory 130, and a communication module 140.

[0058] The memory 130 stores an operating system and a robot motion control program; the output module 110 can be a display screen, etc. The communication module 140 can include a WIFI module, a mobile communication module, a Bluetooth module, etc., and communicates with external devices or servers through the communication module 140.

[0059] Among them, when the robot motion control program in the memory 130 is executed by the processor, the following steps are implemented:

[0060] Determine the target point position, target motion path, and target mode of the robot;

[0061] According to the target point position, target motion path, and target mode, and based on a pre-trained controller, perform point-to-point motion control on the robot, and the controller is modeled based on a width learning algorithm and the controller model is trained.

[0062] Further, when the robot motion control program in the memory 130 is executed by the processor, the following steps are also implemented:

[0063] Calculate a target control rate according to the target motion path and target mode and based on a pre-trained controller;

[0064] Execute the target control rate through a low-order system, and calculate the current position and attitude of the robot;

[0065] Repeat the above steps to calculate the position and attitude of the robot at the next time point until the target point position;

[0066] Perform point-to-point motion control on the robot based on the calculated position and attitude of the robot.

[0067] Further, when the robot motion control program in the memory 130 is executed by the processor, the following steps are also implemented:

[0068] Perform controller modeling based on the width learning algorithm and train the controller model to obtain a trained controller.

[0069] Further, when the robot motion control program in the memory 130 is executed by the processor, the following steps are also implemented:

[0070] Teach and sample the motion process of the robot to obtain teaching data;

[0071] Use the width learning algorithm to model the controller through a high-order dynamic control system to obtain a controller model, and use the extreme learning machine to fit the nonlinear relationship of the low-order control system, and integrate to obtain the sample control rate of the control system;

[0072] Analyze and deduce the system stability constraint based on the sample control rate;

[0073] Import the teaching data and system stability constraint into the controller model, and train the controller model based on the width learning algorithm, the sample control rate and combined with the solution of the optimization function to obtain the final parameters of the controller, and obtain a trained controller.

[0074] Further, when the robot motion control program in the memory 130 is executed by the processor, the following steps are also implemented:

[0075] Based on different satisfactory performance indicators, use multiple adjusted sample controllers to control the robot to complete the same motion path and mode;

[0076] Record the position, speed, attitude and magnetic field data of the robot at each sampling time;

[0077] Repeat the above two steps for different motion paths and modalities to collect the teaching data.

[0078] Further, when the robot motion control program in the memory 130 is executed by the processor, the following steps are further implemented:

[0079] Model the controller using a width learning algorithm through a high-order dynamic control system to obtain a controller model, and calculate the high-order system control rate;

[0080] Fit the non-linear relationship of the low-order control system using an extreme learning machine to obtain the low-order system control rate;

[0081] Combine the high-order system control rate and the low-order system control rate to form the sample control rate of the complete control system.

[0082] Further, when the robot motion control program in the memory 130 is executed by the processor, the following steps are further implemented:

[0083] Update the controller model by adding new teaching data in a real-time online update manner to obtain a controller that can meet different requirements.

[0084] In this embodiment, through the above solution, specifically, the target point position, target motion path, and target modality of the robot are determined; according to the target point position, target motion path, and target modality, and based on a pre-trained controller, point-to-point motion control of the robot is performed, and the controller is obtained by modeling a controller based on a width learning algorithm and training the controller model. The solution of the present invention uses an algorithm based on width learning. On the one hand, it solves the disadvantage that the pre-programming method needs to be redesigned for different paths. On the other hand, it also enables the user to only need to manually control the robot to move according to the requirements when facing different requirements, thereby collecting teaching data, and a controller that can meet different requirements can be updated through a real-time online update method. Accordingly, the method of the present invention avoids the complex repeated parameter adjustment process and reduces the professional requirements for users.

[0085] Compared with the prior art, the width learning algorithm used in the present invention has the characteristic of supporting online model update. Therefore, for newly added teaching data, the present invention does not need to retrain the controller, but directly obtains a controller that simultaneously has the characteristics of the original controller and the newly added teaching data by adding nodes. This method is very suitable for solving the variability of micro-robot motion control and motion paths, and can avoid frequent adjustment of the controller. While ensuring a certain accuracy, it has the advantages of being fast, simple, and supporting incremental online model update. Therefore, when applying the width learning method to the micro-robot point-to-point motion control problem, new training data can be added in real time and quickly, thereby avoiding frequent retraining of the controller.

[0086] Based on the above device architecture but not limited to the above architecture, the method embodiments of the present invention are proposed.

[0087] The execution subject of the method in this embodiment can be a robot motion control device, or a robot motion control device or server. In this embodiment, a robot motion control device is taken as an example. The robot motion control device can be integrated on terminal devices such as smart phones and tablets with data processing functions, and can also be integrated on fixed terminals or servers.

[0088] Referring to Figure 2 , Figure 2 is a schematic flow chart of the first embodiment of the robot motion control method of the present invention. The robot motion control method in this embodiment is mainly applied to the point-to-point motion control of a micro robot. The method includes:

[0089] Step S101, determine the target point position, target motion path and target mode of the robot;

[0090] In this embodiment, the controller model is pre-modeled and the controller model is trained based on the width learning algorithm to obtain the final parameters of the controller. The controller obtained after training is used to perform point-to-point motion control on the robot currently.

[0091] Among them, the sample data used for controller model training is the pre-collected teaching data, which includes different motion paths and modes completed by the robot, and further may include: the position, speed, attitude and magnetic field data of the robot at different sampling points when completing different motion paths and modes.

[0092] Among them, as an implementation manner, the mode may include a motion mode or a motion method, as well as an attitude (such as the angle between the robot and the horizontal line during motion), speed, acceleration, etc.

[0093] Among them, for the controller completed through model training, in subsequent applications, only the target point position, target motion path and target mode of the robot need to be input into the controller, the current target control rate is calculated, and then, under the calculated target control rate, the current point-to-point motion of the robot is controlled.

[0094] In this embodiment, the determination method of the target point position, target motion path and target mode of the robot can be selected or set according to the current actual needs, and this embodiment does not make specific limitations on this.

[0095] Among them, in this embodiment, a variety of different point-to-point motion trajectories of human teaching are sampled, so as to establish a data-driven control model to control the micro robot to perform point-to-point motion that meets specific conditions (such as trajectory or attitude).

[0096] In this embodiment, the final parameters of the controller are obtained by pre-modeling the controller and training the controller model based on the width learning algorithm. The width learning algorithm is introduced into the point-to-point motion control of micro-robots for the first time. This method has a relatively flexible structure and has the advantages of being fast, concise, and supporting incremental online model update while ensuring a certain accuracy. Therefore, when applying the width learning algorithm to the point-to-point motion control problem of micro-robots, new training data can be added in real time and quickly, thus avoiding frequent retraining of the controller and the complex process of repeated parameter tuning, and reducing the professional requirements for users.

[0097] Step S102: According to the target motion path and target mode, and based on the pre-trained controller, perform point-to-point motion control on the robot. The controller is obtained by modeling the controller based on the width learning algorithm and training the controller model.

[0098] This embodiment trains corresponding controllers for different motion trajectories and modes.

[0099] After determining the target point position, target motion path, and target mode of the robot, input the target point position, target motion path, and target mode of the robot into the corresponding trained controller, calculate the current target control rate of the robot, and then perform point-to-point motion control on the current robot under the calculated target control rate.

[0100] Specifically, as an implementation manner, the step of performing point-to-point motion control on the robot according to the target point position, target motion path, and target mode and based on the pre-trained controller may include:

[0101] Calculate the target control rate according to the target motion path and target mode and based on the pre-trained controller;

[0102] Execute the target control rate through a low-order system and calculate the current position and attitude of the robot;

[0103] Repeat the above two steps to calculate the position and attitude of the robot at the next time point until the target point position;

[0104] Perform point-to-point motion control on the robot based on the calculated position and attitude of the robot.

[0105] Through the solution of this embodiment, the target point position, target motion path, and target mode of the robot are determined; according to the target point position, target motion path, and target mode, and based on a pre-trained controller, point-to-point motion control of the robot is performed. The controller is modeled based on the width learning algorithm and the controller model is trained. The solution of the present invention uses the width learning-based algorithm. On the one hand, it solves the disadvantage of the pre-programming method that requires re-design for different paths. On the other hand, it enables the user to only manually control the robot to move according to the requirements when facing different needs, so as to collect teaching data, and a controller that can handle different needs can be updated through real-time online update. Accordingly, the method of the present invention avoids the complex process of repeated parameter adjustment and reduces the professional requirements for users.

[0106] Compared with the prior art, the width learning algorithm used in the present invention has the characteristic of supporting online model update. Therefore, for newly added teaching data, the present invention does not need to re-train the controller, but directly obtains a controller that has both the characteristics of the original controller and the characteristics of the newly added teaching data by adding nodes. This method is very suitable for solving the variability of the motion control and motion path of micro-robots and can avoid frequent adjustment of the controller. While ensuring a certain accuracy, it has the advantages of being fast, simple, and supporting incremental online model update. Therefore, when applying the width learning method to the point-to-point motion control problem of micro-robots, new training data can be added in real time and quickly, thus avoiding frequent re-training of the controller.

[0107] As Figure 3 shown, another embodiment of the present invention proposes a robot motion control method. Based on the above Figure 2 shown embodiment, in this embodiment, before step S101 of determining the target point position, target motion path, and target mode of the robot, it further includes:

[0108] Step S100, perform controller modeling based on the width learning algorithm and train the controller model to obtain a trained controller.

[0109] In this embodiment, mainly through the pre-collected teaching data, combined with performing controller modeling based on the width learning algorithm for high-order systems and low-order systems and training the controller model to obtain a trained controller.

[0110] As described above, the teaching learning method provides a solution for non-professionals to adjust the controller. However, the existing teaching learning methods applied to the trajectory control of micro-robots still have the following problems. First, for some learning algorithms, such as the extreme learning machine, it is difficult to directly add new teaching data after the training process is completed. For problems like the micro-robot control problem that require frequent adjustment of control parameters and control paths, each time new teaching data is added, the controller needs to be retrained. Second, the width learning algorithm can be used to add new training data without retraining, but this method has not been applied to the field of micro-robot point-to-point motion control yet.

[0111] Therefore, an embodiment of the present invention proposes a solution. By sampling a variety of different point-to-point motion trajectories of human teaching, a data-driven control model is established to control the micro-robot to perform point-to-point motion that meets specific conditions (such as trajectory or posture).

[0112] The point-to-point motion control involved in the solution of this embodiment simultaneously considers the high-order dynamic control system (hereinafter referred to as the high-order system) and the low-order control system.

[0113] Therefore, as an implementation manner, the control system of the robot in this embodiment includes: a high-order dynamic control system and a low-order control system, as Figure 4 shown, Figure 4 is the overall system diagram including the high-order system and the low-order system.

[0114] Among them, the high-order system is a dynamic control system, and the low-order system is an actuator control system.

[0115] In the high-order system, a controller model is built based on the width learning algorithm, and the controller model is trained by combining the collected teaching data and the high-order motion control algorithm to calculate parameters such as speed or acceleration required in the control rate equation. Among them, special limiting conditions for enhancing algorithm performance are introduced in the width learning algorithm. At the same time, the non-linear relationship of the low-order system is fitted using the extreme learning machine. The control rate simultaneously considers the high-order system and the low-order system, and combines the two to form a complete control rate.

[0116] In the low-order system, the parameters input by the high-order system are command-converted to obtain corresponding forces, torques or other parameters, which are executed by the robot actuator, and the current position and posture of the robot are measured by measuring sensors. Finally, the point-to-point motion control of the robot is completed.

[0117] The technical solution proposed in this embodiment for controller modeling based on the width learning algorithm and training the controller model to achieve robot trajectory tracking control mainly includes the following steps: Step 1, teaching data preparation, that is, teaching and sampling the movement process of the robot. Step 2, controller training, that is, using the width learning algorithm to model the controller, and at the same time using the extreme learning machine to fit the nonlinear relationship of the low-order system, and combining the two to form a complete control rate. Step 3, analyze and derive the system stability constraints. Step 4, import the teaching data and stability constraints into the controller model, and perform learning algorithm training to obtain the final parameters of each controller. Subsequently, the trained controller can be applied to the point-to-point motion control system of the micro robot.

[0118] Specifically, as Figure 5 shown, the steps of modeling the controller based on the width learning algorithm and training the controller model to obtain a trained controller in this embodiment may include:

[0119] Step S1001, teach and sample the movement process of the robot to obtain teaching data;

[0120] Specifically, as an implementation manner, for the collection of teaching data, first, based on different satisfactory performance indicators, use multiple well-adjusted sample controllers to control the robot to complete the same motion path and mode.

[0121] Then, record the position, speed, attitude and magnetic field data of the robot at each sampling time and preprocess these data at the same time.

[0122] Then, repeat the above two steps for different motion paths and modes to collect teaching data.

[0123] Step S1002, use the width learning algorithm to model the controller through a high-order dynamic control system to obtain a controller model, and use the extreme learning machine to fit the nonlinear relationship of the low-order control system, and integrate to obtain the sample control rate of the control system;

[0124] Specifically, as an implementation manner, use the width learning algorithm to model the controller through a high-order system to obtain a controller model, and calculate the high-order system control rate. At the same time, use the extreme learning machine to fit the nonlinear relationship of the low-order control system to obtain the low-order system control rate;

[0125] Combine the high-order system control rate and the low-order system control rate to form the sample control rate of the complete control system.

[0126] Step S1003, analyze and derive the system stability constraints based on the sample control rate;

[0127] Step S1004: Import the teaching data and system stability constraints into the controller model, and train the controller model based on the width learning algorithm, the sample control rate, and the solution combined with the optimization function to obtain the final parameters of each part of the controller, and obtain a trained controller.

[0128] Among them, repeat the above steps S1001 - S1004 for different motion trajectories and modes, so as to obtain controllers corresponding to different motion trajectories and modes.

[0129] Among them, the controller model satisfies the following conditions for high-order dynamic systems: under the control of the control rate, the robot needs to meet the specified posture while reaching the specified position; during the point-to-point motion process, the motion of the robot needs to meet a specific motion trajectory or a specific motion posture.

[0130] In addition, in the width learning algorithm for training the controller model, the constraint conditions for solving the controller parameters are obtained through the Lyapunov theory.

[0131] The technical solution of this embodiment is a point-to-point motion control method based on the width learning algorithm, and the method can be applied to a micro-robot system. This control rate takes into account both high-order systems and low-order systems, and has many advantages of width learning, such as simple structure and no need to retrain for new teaching data. In addition, the Lyapunov theory is cleverly combined in the width learning algorithm to solve the constraint conditions of the controller parameters, and finally a point-to-point motion controller is obtained that not only has the characteristics of the teaching data, but also has strong generalization and error convergence performance.

[0132] The following combines Figure 4 and Figure 6 , and elaborates in detail the algorithm principle of modeling the controller based on the width learning algorithm and training the controller model to obtain a trained controller in this embodiment. Among them, Figure 6 is a schematic diagram of motion control based on the width learning algorithm.

[0133] As Figure 6 shown, the motion control algorithm based on the width learning algorithm in this embodiment includes a controller training part and a closed-loop control system for controlling the robot using the controller.

[0134] The algorithm principle involved in this embodiment includes:

[0135] (1) Modeling for the point-to-point motion control problem

[0136] As Figure 4As shown, the point-to-point motion control involved in this embodiment simultaneously considers a high-order dynamic control system (hereinafter referred to as the high-order system) and a low-order micro-robot control system (hereinafter referred to as the low-order system).

[0137] For the high-order system, the point-to-point motion control can be decomposed into two parts: (1) Under the control of the control rate, the robot needs to reach the specified position while satisfying the specified posture. (2) During the point-to-point motion process, the motion of the robot needs to meet specific conditions, for example, satisfying a specific motion trajectory or a specific motion posture. For the above two requirements, the model of high-order motion control can be described as follows:

[0138]

[0139]

[0140] Among them, ζ and θ respectively represent the position and posture of the robot's motion, and respectively represent the velocity and acceleration of the position and posture. f1(·), f2(·), f3(·) represent three different nonlinear models. In the point-to-point motion control problem, the only end point is defined as Therefore, when the initial position of the robot is not 0, the control instruction should be appropriate velocity and acceleration to make the robot move and stop at the target position.

[0141] Among them, the low-order robot control system applicable to point-to-point motion control can include various different types. In this embodiment, the micro-robot magnetic control system is taken as an example.

[0142] The motion control of the micro-robot includes two aspects:

[0143] One is the external magnetic field model. For the commonly used oscillating magnetic field in magnetic control, the magnetic field model can be defined by φ B , n P Two parameters. Among them, φ B represents the included angle of magnetic field oscillation, and n p represents the direction vector of the central axis of magnetic field oscillation.

[0144] The other is the response of the micro-robot itself to the external magnetic field. This part is mainly determined by the size, dimensions and magnetization of the micro-robot itself. Therefore, once the micro-robot is manufactured, the control response of this part will be fixed.

[0145] In summary, the control rate for the low-order micro-robot magnetic control system can be expressed as:

[0146]

[0147]

[0148] where \(u\) represents the control rate of the lower-order system, \(f(\cdot)\) represents the nonlinear equation of the lower-order system, and \(s(\cdot)\) represents the nonlinear relationship between the control rate and the final velocity determined by the characteristics (size, magnetization, etc.) of the robot itself.

[0149] (2) For the design of the control rate based on the width learning algorithm

[0150] Combined with Figure 6 As shown, in this embodiment, a method based on the broad learning system (BLS) is adopted to design the control algorithm for point-to-point motion. The system input is designed as \(c = \zeta\) T , \(\theta\) T T , based on the width learning algorithm and the motion control model, the control rate equations for both the higher-order and lower-order systems are as follows:

[0151]

[0152] Further define:

[0153]

[0154] h j = (r j z j + j )( )

[0155] N = J( )

[0156] where \(\varphi(\cdot)\) and \(g(\cdot)\) are activation functions. In this embodiment, the sigmoid function is selected as the activation function. Therefore, the activation function can be expressed as:

[0157]

[0158]

[0159] Through the above control rate equations and related definitions, \(Z = [z_1, z_2, \ldots, z\) N is the feature node of the width learning neural network, \(H = [h_1, h_2, \ldots, h\) J is the enhanced node, \(\omega\) i , k j are the output weights of the feature node and the enhanced node respectively, and \(a\) i , r j , b i , p j are randomly generated parameters, which are the input weights and biases of the feature node and the enhanced node respectively. Since the input of this system is \(c = \zeta\)​T , θ T T , a can be i split into corresponding to ζ and θ respectively. Accordingly, the control rate equation can be further rewritten as:

[0160]

[0161] The acquisition methods of the parameters in Equation (11) are as follows. a i , r j , b i , p j are randomly generated parameters, and ω i , k j can be obtained through the training process. Before training, teaching data needs to be obtained. For different low-order systems, methods such as manual control or non-linear controllers can be used to control the robot to move along the required route. The same movement route will be repeated multiple times, and finally multiple groups of input data c o and output data u o can be collected. The superscript o is used to represent the teaching data. After obtaining enough teaching data, the training process can be completed through the following optimization function:

[0162]

[0163] Finally, using common computing software, such as Matlab, to solve the above optimization problem, ω i , k j can be obtained.

[0164] (3) For the stability analysis of the control system

[0165] First of all, assuming that the low-order system will not suddenly get out of control, the core of the design focuses on the stability analysis of the high-order system control rate. In order to use the control rate equation of Equation (11), the stability of the system itself needs to be guaranteed. The common Lyapunov stability method can be selected. Based on this method, the state quantity of the system c = [ζ; θ] is defined, and this state quantity will globally asymptotically stabilize at c * =[ζ * ; θ * =0. The continuous and continuously differentiable Lyapunov candidate function (LCF), V(c), should satisfy:

[0166]

[0167]

[0168] V(c * ) = 0 (13c) ​

[0169]

[0170] According to Lyapunov theory, the LCF satisfying Equation (13) can be designed, and the stability limit of the system can be deduced accordingly. The LCF is designed as follows:

[0171]

[0172] Next, by taking the derivative of the LCF, we can obtain:

[0173]

[0174] Combining Equation (5) and Equation (4), we can further deduce:

[0175]

[0176] Substituting Equation (15) into Equation (14), we can obtain:

[0177]

[0178] Next, select the extreme learning machine (ELM) algorithm to model s(·):

[0179]

[0180] Substituting Equation (17) into Equation (14) and using the mean value theorem to rewrite it, we can obtain:

[0181]

[0182] where φ(0) = 0, φ′(λ l ) is or the gradient of. Next, substituting the control rate in Equation (5) into Equation (18), we can obtain:

[0183]

[0184] Based on the mean value theorem and defining φ(0) = g(0) = 0, further rewriting Equation (19) can obtain:

[0185]

[0186] where φ′(κ i ), g′(η j ) are respectively (or and η j ∈(0, r j z j + p j )(or ηj ∈(r j z j +p j , 0)). The gradient of. To simplify Equation (20), the following assignments are made to some of the parameters therein:

[0187] n l = 0 (21a)

[0188] b i = 0, or b j = 0 (21b)

[0189] p j = 0 (21c)

[0190] Substituting Equation (21) into Equation (20), the simplified LCF can be obtained:

[0191]

[0192] According to Equation (13b), the loose constraint conditions for the stability to hold can be derived:

[0193] n l = 0 (23a)

[0194] b i = 0, or b j = 0 (23b)

[0195] p j = 0 (23c)

[0196]

[0197]

[0198] Here, the symbol <0 indicates that the matrix is a negative definite matrix. Since the slope of the activation function (Equations (9, 10)) selected in this embodiment is always greater than zero, and L = N = J. Accordingly, the strict constraint conditions for the stability to hold can be derived:

[0199] n l = 0 (24a)

[0200] b i = 0 (24b)

[0201] p j = 0 (24c)

[0202]

[0203]

[0204] Based on the above derivations, the design of a stable controller and learning algorithm can be expressed as an optimization problem:

[0205]

[0206] where \(i, j, l = 1, 2, \ldots, N\), and the following constraints are satisfied:

[0207] n l = 0 (26a)

[0208] b i = 0 (26b)

[0209] p j = 0 (26c)

[0210]

[0211]

[0212] The above optimization problem can be solved during the training process by using the fmincon function in Matlab.

[0213] Through the above scheme, in this embodiment, a controller based on the width learning algorithm is designed, and both high-order systems and low-order systems are considered. At the same time, modeling of high-order systems and low-order systems based on the learning algorithm is carried out. This method has a relatively flexible structure, and while ensuring a certain accuracy, it has the advantages of being fast, concise, and supporting incremental online model updates. Therefore, when applying the width learning method to the point-to-point motion control problem of a micro-robot, new training data can be added in real time and quickly, thus avoiding frequent retraining of the controller.

[0214] In addition, the method of this embodiment takes into account the stability problem. Through strict stability calculations, the limiting conditions for ensuring control stability are obtained. Moreover, by learning the motion characteristics of a micro-robot in various different paths and postures, the results obtained by the learning algorithm have generalization ability, and motion control can be completed for different robots and point-to-point motion control requirements.

[0215] Furthermore, the method further includes:

[0216] Updating the controller model by adding new teaching data in a real-time online update manner to obtain a controller that can meet different requirements. Therefore, the controller designed by this scheme has the ability of online update and can add new teaching data without retraining.

[0217] In addition, through simulation calculations and experimental tests in this embodiment of the invention, it is proved that the control algorithm is stable and effective.

[0218] Compared with the prior art, the method proposed in this embodiment has the following advantages:

[0219] 1. By sampling a variety of different point-to-point motion trajectories of human teaching, a data-driven control model is established to control the micro-robot to perform point-to-point motion that meets specific conditions (such as trajectory or posture).

[0220] 2. The embodiment of the present invention solves two difficulties in the point-to-point motion control of micro-robots. (1) For a given target point, the robot needs to meet specific posture requirements while reaching the target point. (2) During the process of the robot reaching the target point, it needs to pass through a special path (such as obstacle avoidance) or meet specific postures. While solving the above two problems, the method of the present invention also has sufficient generalization performance, so that for different micro-robots and drivers, a control rate sufficient to solve the above two difficulties can be obtained through a learning algorithm.

[0221] 3. The controller design method proposed in the embodiment of the present invention is applicable to any point-to-point motion control problem. Without professional control or programming knowledge, users only need to teach the robot motion control to obtain a controller with teaching performance, which is suitable for industrial people.

[0222] 4. The width learning algorithm used in the embodiment of the present invention has the characteristic of supporting online model update. Therefore, for newly added teaching data, the present invention does not need to retrain the controller, but directly obtains a controller that has the characteristics of the original controller and the characteristics of the newly added teaching data by adding nodes. This method is very suitable for solving the variability of micro-robot motion control and motion paths, and can avoid frequent adjustment of the controller.

[0223] Therefore, the embodiment of the present invention uses an algorithm based on width learning. On the one hand, it solves the disadvantage of the pre-programming method that needs to be redesigned for different paths. On the other hand, it also enables users to only need to manually control the robot to move according to the requirements when facing different needs, thereby collecting teaching data, and can update a controller that can handle different needs through real-time online update. Accordingly, this method avoids the complex process of repeated parameter adjustment and reduces the professional requirements for users.

[0224] Combined Figure 6 As shown, in practical applications, the motion control method based on the width learning algorithm proposed in the embodiment of the present invention can be adopted. Combining the above formulas, the specific implementation steps include:

[0225] Step 1: Teaching data preparation process

[0226] 1.1, Teaching data collection. Based on different satisfactory performance indicators, use multiple well - tuned sample controllers to control the micro - robot to complete the same motion path and mode.

[0227] 1.2, Record the position, speed, attitude, and magnetic field data at each sampling time As inputs and outputs, and pre - process the data simultaneously.

[0228] 1.3, Repeat the above two steps for different motion paths and modes to collect teaching data.

[0229] Step 2: Motion controller training process

[0230] 2.1, Use a set of teaching data for one motion path and mode to train the controller. Solve the function s(·) based on the extreme learning machine by fitting to obtain h l ,m l ,n l .

[0231] 2.2, Solve the constrained optimization problems in equations (25) and (26). Among them, h l ,m l ,n l is known, and solve for w i ,a i ,k j ,r j ;

[0232] 2.3, Based on b i = 0,p j = 0, solve the optimization problem of equation (25).

[0233] 2.4, Determine the parameters of the control law shown in equation (5).

[0234] 2.5, Repeat the above steps (2.1)-(2.4) for different motion trajectories and modes to obtain controllers corresponding to different motion trajectories and modes. Refer to Figure 6 the controller training part in

[0235] Step 3: Use appropriate motion trajectories and modes to complete point - to - point motion control.

[0236] 3.1, Define the target point position and select appropriate motion paths and modes.

[0237] 3.2, For the selected motion path and mode and based on f(ζ,θ) trained in equation (3), calculate the control law of equation (6).

[0238] 3.3. The control rate of the above steps is executed by the lower-order system, and the current position and posture of the robot are measured by the measurement sensor.

[0239] 3.4. Return to (3.1) of step 3, and repeat the process to calculate the state at the next time point. Finally, the point-to-point motion control of the robot is completed, referring to Figure 6 the closed-loop control system part in

[0240] In addition, referring to Figure 6 the node layout method in the controller training part shown, the width learning algorithm used in this embodiment has the characteristic of supporting online model update. For the newly added teaching data, it is not necessary to retrain the controller, but a controller with the characteristics of both the original controller and the newly added teaching data can be directly obtained by adding nodes. This method is very suitable for solving the variability of the motion control and motion path of the micro robot, and can avoid frequent adjustment of the controller. While ensuring a certain accuracy, it has the advantages of fast, simple, and supporting incremental online model update. Therefore, when applying the width learning method to the point-to-point motion control problem of the micro robot, new training data can be added in real time and quickly, thus avoiding frequent retraining of the controller.

[0241] In addition, as Figure 7 shown, an embodiment of the present invention also proposes a robot motion control device, and the device includes:

[0242] A determination module, configured to determine the target point position, target motion path, and target mode of the robot;

[0243] A control module, configured to perform point-to-point motion control on the robot according to the target point position, target motion path, and target mode, and based on a pre-trained controller, where the controller is modeled based on the width learning algorithm and the controller model is trained.

[0244] For the principle and implementation process of realizing robot motion control in this embodiment, please refer to the above embodiments, and details are not described herein again.

[0245] In addition, an embodiment of the present invention also proposes a robot control device, and the robot control device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the robot motion control method in the above embodiment is implemented.

[0246] Since all the technical solutions of the foregoing embodiments are adopted when the robot motion control program is executed by the processor, it has at least all the beneficial effects brought by all the technical solutions of the foregoing embodiments, and details are not described herein one by one.

[0247] In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the robot motion control method described in the above embodiment is implemented.

[0248] Since all the technical solutions of the foregoing embodiments are adopted when the robot motion control program is executed by the processor, it has at least all the beneficial effects brought by all the technical solutions of the foregoing embodiments, which will not be elaborated herein one by one.

[0249] A robot motion control method, device and medium provided by an embodiment of the present invention determine the target point position, target motion path and target mode of the robot; according to the target point position, target motion path and target mode, and based on a pre-trained controller, perform point-to-point motion control on the robot. The controller is modeled based on the width learning algorithm and the controller model is trained. The solution of the present invention uses the width learning-based algorithm. On the one hand, it solves the disadvantage that the pre-programming method needs to be redesigned for different paths. On the other hand, it enables the user to manually control the robot to move according to the requirements when facing different requirements, so as to collect teaching data, and a controller that can handle different requirements can be updated through real-time online update. Accordingly, the method of the present invention avoids the complex process of repeated parameter adjustment and reduces the professional requirements for users.

[0250] Compared with the prior art, the width learning algorithm used in the present invention has the characteristic of supporting online model update. Therefore, for newly added teaching data, the present invention does not need to retrain the controller, but directly obtains a controller that has both the characteristics of the original controller and the characteristics of the newly added teaching data by adding nodes. This method is very suitable for solving the variability of the motion control and motion path of micro-robots and can avoid frequent adjustment of the controller. While ensuring a certain accuracy, it has the advantages of fast, simple and supporting incremental online model update. Therefore, when applying the width learning method to the point-to-point motion control problem of micro-robots, new training data can be added in real time and quickly, thus avoiding frequent retraining of the controller.

[0251] It should be noted that in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or method. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article or method including the element.

[0252] The serial numbers of the embodiments of the present invention above are only for description and do not represent the superiority or inferiority of the embodiments.

[0253] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium as described above (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, a computer, a server, a controlled terminal, or a network device, etc.) to execute the methods of each embodiment of the present invention.

[0254] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the contents of the specification and drawings of the present invention, or directly or indirectly applied to other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A robot motion control method, characterized in that, The method includes: Determining the target point position, target motion path, and target mode of the robot; Based on the target point position, target motion path, and target mode, and using a pre-trained controller, performing point-to-point motion control on the robot, where the controller is modeled based on the width learning algorithm and the controller model is trained; Before the step of determining the target point position, target motion path, and target mode of the robot, it further includes: Modeling the controller based on the width learning algorithm and training the controller model to obtain a trained controller; The control system of the robot includes: a high-order dynamic control system and a low-order control system. The step of modeling the controller based on the width learning algorithm and training the controller model to obtain a trained controller includes: Teaching and sampling the motion process of the robot to obtain teaching data; Using the width learning algorithm through the high-order dynamic control system to model the controller to obtain a controller model, fitting the non-linear relationship of the low-order control system using the extreme learning machine, and integrating to obtain the sample control rate of the control system; Analyzing and deriving the system stability constraints based on the sample control rate; Importing the teaching data and system stability constraints into the controller model, and training the controller model based on the width learning algorithm, the sample control rate, and combined with the optimization function solution to obtain the final parameters of the controller, and obtaining a trained controller.

2. The method according to claim 1, characterized in that, The step of performing point-to-point motion control on the robot based on the target point position, target motion path, and target mode, and using a pre-trained controller includes: Calculating the target control rate based on the target motion path and target mode and using a pre-trained controller; Executing the target control rate and calculating the current position and attitude of the robot; Repeating the above steps to calculate the position and attitude of the robot at the next time point until the target point position; Based on the calculated position and attitude of the robot, performing point-to-point motion control on the robot.

3. The method according to claim 1, characterized in that, The step of teaching and sampling the motion process of the robot to obtain teaching data includes: Based on different satisfactory performance indicators, using multiple adjusted sample controllers to control the robot to complete the same motion path and mode; Recording the position, speed, attitude, and magnetic field data of the robot at each sampling time; Repeating the above two steps for different motion paths and modes to collect teaching data.

4. The method according to claim 1, characterized in that, The step of using the width learning algorithm through the high-order dynamic control system to model the controller to obtain a controller model, fitting the non-linear relationship of the low-order control system using the extreme learning machine, and integrating to obtain the sample control rate of the control system includes: Using the width learning algorithm through the high-order dynamic control system to model the controller to obtain a controller model, and calculating the high-order system control rate; Fitting the non-linear relationship of the low-order control system using the extreme learning machine to obtain the low-order system control rate; Combining the high-order system control rate and the low-order system control rate to form the sample control rate of the complete control system.

5. The method according to claim 1, characterized in that, The step of importing the teaching data and system stability constraints into the controller model and training the controller model based on the width learning algorithm, the sample control rate, and the solution of the combined optimization function to obtain the final parameters of the controller and obtain the trained controller includes: In the width learning algorithm, the constraint conditions of the controller parameters are solved through the Lyapunov theory.

6. The method according to claim 1, characterized in that, The controller model satisfies the following conditions for the high-order dynamic system: under the control of the control rate, the robot needs to satisfy the specified posture while reaching the specified position; during the point-to-point movement, the movement of the robot needs to satisfy a specific movement trajectory or a specific movement posture.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: By means of real-time online update, new teaching data is added to update the controller model to obtain a controller that can meet different requirements.

8. A robot control device, characterized in that, The robot control device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, it implements the robot motion control method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium. When the computer program is executed by the processor, it implements the robot motion control method according to any one of claims 1-7.

Images