A machine learning based method and system for fine control of a robot arm

Abstract

The application relates to the technical field of mechanical arms and discloses a mechanical arm fine control method and system based on machine learning, which comprises the following steps: determining a first grabbing strategy or a second grabbing strategy based on the relationship between material parameters and a historical material set; when the first grabbing strategy is determined, determining the grabbing strength of the mechanical arm based on the historical material set; when the second grabbing strategy is determined, determining the grabbing strength of the mechanical arm based on the clustering result of the historical material set and a grabbing strength model; determining the minimum moving load based on a moving load curve diagram and determining the grabbing load of a driving motor; judging whether the grabbing strength needs to be adjusted according to the load relationship between the minimum moving load and the grabbing load; determining the grabbing strength factor based on the minimum moving load and the grabbing load; and completing grabbing according to the adjusted grabbing strength. The first grabbing strategy, the second grabbing strategy and the grabbing strength factor ensure the stability of the grabbing process.

Description

Technical Field

[0001] The present invention relates to the technical field of robotic arms, and in particular to a method and system for fine control of a robotic arm based on machine learning. Background Art

[0002] With the development of technology, robots are widely used in various automated production lines, such as assembly, handling, and sorting. Their high speed, high precision, and high repeatability improve production efficiency and quality. However, in specific grasping operations, the gripping force of the robot arm's end effector—the gripper—is usually set by the operator's experience. This setting is easily affected by individual differences. Moreover, the set gripping force cannot be dynamically adjusted according to actual conditions, and the lack of a feedback adjustment mechanism leads to grasping failures and slippage during the grasping process, resulting in insufficient stability of the robot arm and high debugging costs.

[0003] Therefore, it is necessary to design a machine learning-based robotic arm fine control method and system to solve the problems existing in current technology. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for fine control of a robotic arm based on machine learning, aiming to solve the problem that the grasping force is usually set by the operator's experience, the setting of human experience is easily affected by individual differences, and the set grasping force cannot be dynamically adjusted according to actual conditions, resulting in insufficient stability of the robotic arm and high debugging costs.

[0005] In one aspect, the present invention proposes a method for fine control of a robotic arm based on machine learning, comprising:

[0006] Determine material parameters of the material to be grasped, and determine a first grasping strategy or a second grasping strategy based on a relationship between the material parameters and a historical material set;

[0007] When the first grasping strategy is determined, the grasping force of the robot arm is determined based on the historical material set; when the second grasping strategy is determined, the grasping force of the robot arm is determined based on the clustering result of the historical material set and the grasping force model;

[0008] Obtaining the moving load of the drive motor during the moving time period, analyzing all moving loads, determining an effective moving load and a moving load curve based on the analysis results, determining a minimum moving load based on the moving load curve, grasping the material to be grasped according to the grasping force, determining the grasping load of the drive motor, and determining whether to adjust the grasping force based on the load relationship between the minimum moving load and the grasping load;

[0009] When it is determined that the grasping force needs to be adjusted, a grasping force factor is determined based on the minimum movement load and the grasping load, and the grasping force is adjusted according to the grasping force factor, and grasping is completed according to the adjusted grasping force.

[0010] Furthermore, when determining the first grabbing strategy or the second grabbing strategy based on the relationship between the material parameters and the historical material set, it includes:

[0011] The historical material set includes a plurality of historical material parameters and a plurality of historical grasping forces, and each historical material parameter corresponds to a historical grasping force;

[0012] When the historical material set contains historical material parameters that are equal to the material parameters, determining that the robotic arm adopts the first grasping strategy;

[0013] When there are no historical material parameters equal to the material parameters in the historical material set, it is determined that the robotic arm adopts the second grasping strategy.

[0014] Furthermore, when the first grasping strategy is determined, the grasping force of the robot arm is determined based on the historical material set, including:

[0015] When there is a historical material parameter in the historical material set that is equal to the material parameter, and the historical material parameter is unique, the historical grasping force corresponding to the historical material parameter is determined as the grasping force;

[0016] When there is a historical material parameter equal to the material parameter in the historical material set, and the historical material parameter is not unique, the average of the historical grasping forces corresponding to the historical material parameters is determined as the grasping force.

[0017] Furthermore, when the second grasping strategy is determined, the grasping force of the robot arm is determined based on the clustering result of the historical material set and the grasping force model, including:

[0018] The material parameters and the historical material set are used as the set to be clustered, the expected number of clusters k is determined to be 2, and the parameters of the Gaussian distribution are initialized. The probability that each data in the set to be clustered belongs to each Gaussian distribution is determined and the responsibility value is determined. Based on the responsibility value, a cluster set corresponding to the material parameters is determined, and the mean of the historical grasping strength in the cluster set is used as the cluster grasping strength;

[0019] Dividing the historical material set into a training set and a test set, preselecting a gradient boosting tree model, and training the gradient boosting tree model based on the training set, testing the trained gradient boosting tree model based on the test set, and determining the model grasping strength based on the test results and the material parameters;

[0020] The grasping force of the robotic arm is determined based on the cluster grasping force and the model grasping force.

[0021] Furthermore, when the gradient boosting tree model is trained according to the training set, the trained gradient boosting tree model is tested according to the test set, and the model grasping strength is determined based on the test results and the material parameters, the method includes:

[0022] If the F1 score of the currently trained gradient boosted tree model is less than the F1 score of the previously trained gradient boosted tree model, cross-validation is used in combination with gridding to reduce the overfitting of the model, and training is continued until the F1 score of the currently trained gradient boosted tree model is greater than or equal to the F1 score of the previously trained gradient boosted tree model.

[0023] If the F1 score of the currently trained gradient boosting tree model is greater than or equal to the F1 score of the previously trained gradient boosting tree model, then stop training and determine the currently trained gradient boosting tree model as the grasping force model;

[0024] Substituting the material parameters into the grasping force model to determine the model grasping force.

[0025] Furthermore, when determining the grasping force of the robotic arm based on the cluster grasping force and the model grasping force, the method includes:

[0026] When the cluster grasping force and the model grasping force are equal, the cluster grasping force or the model grasping force is determined as the grasping force;

[0027] When the cluster grasping strength and the model grasping strength are equal, the average of the cluster grasping strength and the model grasping strength is determined as the grasping strength.

[0028] Furthermore, when analyzing all the moving loads, determining the effective moving load and the moving load curve based on the analysis results, and determining the minimum moving load based on the moving load curve, the method includes:

[0029] Convert all moving loads into load coordinate points, where the X-axis coordinate value of the load coordinate point represents the unit moving time period, and the Y-axis coordinate value of the load coordinate point represents the moving load;

[0030] A load coordinate system is established based on all load coordinate points, and all load coordinate points are fitted to determine the load fitting curve, and the load coordinate points that are not fitted are deleted. The mobile load curve diagram is determined based on the deletion result, and the load coordinate points on the mobile load curve diagram are used as the effective mobile load, and the minimum value of all effective mobile loads is determined as the minimum mobile load.

[0031] Furthermore, when determining whether to adjust the gripping force according to the load relationship between the minimum moving load and the gripping load, the method includes:

[0032] When the minimum moving load is greater than or equal to the grabbing load, it is determined that the grabbing force should be adjusted;

[0033] When the minimum moving load is less than the grasping load, it is determined that the grasping force is not adjusted, and grasping is completed with the grasping force.

[0034] Furthermore, when determining to adjust the gripping force, determining a gripping force factor based on the minimum moving load and the gripping load, and adjusting the gripping force according to the gripping force factor includes:

[0035] Obtaining a load ratio Q of the minimum moving load and the grabbing load;

[0036] When 1≤Q<1.5, the first grasping force factor is determined as the grasping force factor;

[0037] When 1.5≤Q<1.7, the second grasping force factor is determined as the grasping force factor;

[0038] When 1.7≤Q, the third grasping force factor is determined as the grasping force factor;

[0039] The third grasping force factor is greater than the second grasping force factor, the second grasping force factor is greater than the first grasping force factor, and the first grasping force factor is greater than 1;

[0040] The grasping force is directly proportional to the grasping force factor.

[0041] Compared with the prior art, the beneficial effects of the present invention are: dynamically determining the grasping strategy according to the relationship between material parameters and historical material sets, directly calling the historical material sets to determine the grasping force based on the grasping experience of the historical material sets, thereby ensuring the efficiency of the grasping operation, and for material parameters that are not fully matched, analyzing the historical material sets based on machine learning and clustering algorithms to determine the grasping force. The dual-strategy mechanism of experience reuse, clustering algorithm and machine learning not only avoids the blindness and uncertainty of human experience settings, but also ensures the stability of the grasping force that can be determined according to actual conditions. The minimum moving load is determined according to the established moving load curve diagram. During the grasping process, the real-time collected grasping load and the minimum moving load form a closed-loop feedback, thereby dynamically adjusting the grasping force, and then constructing a complete closed loop of grasping material identification, grasping strategy determination, real-time load monitoring and adjustment of grasping force, thereby ensuring the reliability and stability of fine control of the robotic arm.

[0042] On the other hand, the present application also provides a machine learning-based fine control system for a robotic arm, which is used to apply the above-mentioned machine learning-based fine control method for a robotic arm, including:

[0043] a collection and analysis unit configured to determine material parameters of a material to be grasped, and determine a first grasping strategy or a second grasping strategy based on a relationship between the material parameters and a historical material set;

[0044] a first processing unit configured to, when determining the first grasping strategy, determine the grasping force of the robotic arm based on the historical material set, and when determining the second grasping strategy, determine the grasping force of the robotic arm based on the clustering result of the historical material set and a grasping force model;

[0045] a second processing unit configured to obtain a moving load of the drive motor during a moving time period, analyze all moving loads, determine an effective moving load and a moving load curve based on the analysis results, determine a minimum moving load based on the moving load curve, grasp the material to be grasped according to the grasping force, determine the grasping load of the drive motor, and determine whether to adjust the grasping force based on a load relationship between the minimum moving load and the grasping load;

[0046] The control adjustment unit is configured to determine a grasping force factor based on the minimum moving load and the grasping load when determining to adjust the grasping force, adjust the grasping force according to the grasping force factor, and complete the grasping according to the adjusted grasping force.

[0047] It is understandable that the above-mentioned machine learning-based robotic arm fine control method and system have the same beneficial effects and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0049] Figure 1 A flowchart of a method for fine control of a robotic arm based on machine learning provided by an embodiment of the present invention;

[0050] Figure 2 A functional block diagram of a machine learning-based robotic arm precision control system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features described in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0052] In some embodiments of this application, see Figure 1 As shown, a method for fine control of a robotic arm based on machine learning includes:

[0053] S100: Determine material parameters of a material to be grasped, and determine a first grasping strategy or a second grasping strategy based on a relationship between the material parameters and a historical material set.

[0054] S200: When the first grasping strategy is determined, the grasping force of the robot arm is determined based on the historical material set; when the second grasping strategy is determined, the grasping force of the robot arm is determined based on the clustering result of the historical material set and the grasping force model.

[0055] S300: Obtain the moving load of the driving motor during the moving time period, and analyze all moving loads, determine the effective moving load and the moving load curve based on the analysis results, determine the minimum moving load based on the moving load curve, grab the material to be grabbed according to the grabbing force, and determine the grabbing load of the driving motor, and determine whether to adjust the grabbing force based on the load relationship between the minimum moving load and the grabbing load.

[0056] S400: When it is determined that the grasping force needs to be adjusted, a grasping force factor is determined based on the minimum moving load and the grasping load, and the grasping force is adjusted according to the grasping force factor, and the grasping is completed according to the adjusted grasping force.

[0057] Specifically, the material parameters of the material to be grasped are determined. The material parameters include data such as the weight, shape, size and material of the material to be grasped. These data are key indicators reflecting the material grasping characteristics. The material parameters are compared with the historical material set. The historical material set includes various types of data when grasping different materials in the past, including material parameters and corresponding grasping data, etc. Through comparison, it is judged whether the material to be grasped matches the existing historical grasping experience, and the first grasping strategy or the second grasping strategy is dynamically determined. When the first grasping strategy is determined, it indicates that the material parameters of the material to be grasped are relatively matched with the historical grasping experience. The grasping force of the robot arm is determined based on the historical material set, and the mature grasping scheme in the historical experience is used to achieve a quick and effective determination of the grasping force. When the second grasping strategy is determined, it indicates that there are certain differences between the material parameters of the material to be grasped and the historical grasping experience, which is a relatively special or complex situation. The grasping force of the robot arm is determined based on the clustering results of the historical material set and the grasping force model. The clustering result is obtained by clustering the historical material set, and the historical experience and material parameters with similar grasping characteristics are classified into one category. The grasping force model is obtained by training the historical material set using a machine learning algorithm, so that the grasping force model can output prediction results based on the material parameters. The grasping force of the robot arm is determined based on the clustering results and the model output prediction results, so that the robot arm can perform refined analysis according to different materials to be grasped, thereby ensuring the stability of the grasping force and avoiding the uncertainty and differences set by human experience.

[0058] It can be understood that in the actual grasping process, the drive motor drives the robotic arm to move accordingly, and the moving load of the drive motor in this moving time period is obtained. The moving load reflects the load of the robotic arm during the movement process. All moving loads are analyzed to determine the effective moving load and draw a moving load curve. The moving load curve can intuitively show the changing trend of the moving load over time, from which the minimum moving load can be determined. The minimum moving load is the load required by the robotic arm during stable movement, and is an important indicator to ensure the smooth operation of the robotic arm. When grabbing the material to be grabbed according to the grabbing force, the driving motor will generate a grabbing load. The grabbing load reflects the actual load of the robot arm when grabbing the material to be grabbed under the current grabbing force. The relationship between the grabbing load and the minimum moving load indicates whether the current grabbing force may cause grabbing failure or slipping. If a large deviation is found between the current grabbing load and the minimum moving load, the grabbing force factor is determined according to the load relationship between the minimum moving load and the grabbing load, and the current grabbing force is adjusted, thereby avoiding the uncertainty of setting the grabbing force based on human experience, realizing dynamic adjustment of the grabbing force, and enhancing the adaptability and stability of the robot arm. During the grabbing process, by obtaining the moving load and grabbing load of the driving motor, the working status of the robot arm and the grabbing situation of the material to be grabbed can be perceived in real time, and the grabbing force can be dynamically adjusted according to the actual situation, forming an effective feedback adjustment mechanism, reducing the occurrence of grabbing failure and slipping, and improving the stability of the robot arm under complex working conditions.

[0059] In some embodiments of the present application, when determining the first grasping strategy or the second grasping strategy based on the relationship between material parameters and historical material sets, it includes: the historical material set includes several historical material parameters and several historical grasping forces, and each historical material parameter corresponds to a historical grasping force. When there are historical material parameters equal to the material parameters in the historical material set, the robotic arm is determined to be the first grasping strategy; when there are no historical material parameters equal to the material parameters in the historical material set, the robotic arm is determined to be the second grasping strategy.

[0060] In some embodiments of the present application, when the first grasping strategy is determined, the grasping force of the robotic arm is determined based on the historical material set, including: when there is a historical material parameter equal to the material parameter in the historical material set, and the historical material parameter is unique, then the historical grasping force corresponding to the historical material parameter is determined as the grasping force; when there is a historical material parameter equal to the material parameter in the historical material set, and the historical material parameter is not unique, then the average of the historical grasping forces corresponding to each historical material parameter is determined as the grasping force.

[0061] Specifically, by judging the degree of match between the current material parameters and the historical material parameters, when historical material parameters equal to the material parameters are found, the robot arm is determined to be the first grasping strategy, and the historical data is directly used to reuse the historical grasping force, thereby ensuring the reliability and consistency of the grasping operation. When there are multiple historical material parameters equal to the material parameters, the average of the historical grasping forces corresponding to these historical material parameters is determined as the grasping force, thereby balancing the influence between the data and reducing the grasping failure rate. For the situation where the current material parameters do not match the historical data, the robot arm is determined to be the second grasping strategy to cope with changes in different material parameters. Through data-driven automated adjustment, the dependence on human experience and intuition is reduced, and the uncertainty and grasping risk brought by human judgment are reduced, thereby improving the automation level and stability of the grasping process. By comprehensively utilizing the historical material set, it can learn and optimize from historical experience, so that the robot arm can finely determine the grasping force and continuously improve the accuracy and efficiency of grasping.

[0062] In some embodiments of the present application, when the second grasping strategy is determined, the grasping force of the robotic arm is determined based on the clustering results of the historical material set and the grasping force model, including: taking the material parameters and the historical material set as the set to be clustered, determining the expected number of clusters k as 2, and initializing the parameters of the Gaussian distribution, determining the probability that each data in the set to be clustered belongs to each Gaussian distribution and determining the responsibility value, determining the cluster set corresponding to the material parameter based on the responsibility value, and taking the mean of the historical grasping force in the cluster set as the cluster grasping force, dividing the historical material set into a training set and a test set, pre-selecting a gradient boosting tree model, and training the gradient boosting tree model according to the training set, testing the trained gradient boosting tree model according to the test set, determining the model grasping force based on the test results and the material parameters, and determining the grasping force of the robotic arm based on the cluster grasping force and the model grasping force.

[0063] Specifically, a clustering algorithm analyzes the historical material set to identify the cluster set that most closely matches the current grasping conditions. This improves the accuracy of cluster grasping force determination and ensures adaptive cluster grasping force determination under varying material parameters, thereby maintaining the stability and efficiency of the grasping process. This reduces reliance on human experience and judgment, thereby reducing human error during operation and improving the automation and reliability of the overall grasping process. The historical material set is divided into a training set and a test set, typically in a 7:3 ratio, to enhance the model's adaptability and generalization to the data. The training set is used to train a gradient boosting tree model, while the test set is used to evaluate the trained model's performance. The gradient boosting tree model is used to construct a mapping between historical material parameters and historical grasping force, thereby capturing complex relationships in the data. While single clustering methods are susceptible to outliers in historical data, the gradient boosting tree method balances potential biases in clustering results. This dual determination mechanism reduces the risk of misjudgment, ensures the stability of grasping force, and avoids the uncertainty and variability inherent in human experience.

[0064] In some embodiments of the present application, when training a gradient boosting tree model according to a training set, testing the trained gradient boosting tree model according to a test set, and determining the model grasping strength based on the test results and material parameters, it includes: if the F1 score of the currently trained gradient boosting tree model is less than the F1 score of the previously trained gradient boosting tree model, cross-validation is used in combination with grid reduction to reduce the result of model overfitting, and training is continued until the F1 score of the currently trained gradient boosting tree model is greater than or equal to the F1 score of the previously trained gradient boosting tree model; if the F1 score of the currently trained gradient boosting tree model is greater than or equal to the F1 score of the previously trained gradient boosting tree model, training is stopped, and the currently trained gradient boosting tree model is determined as the grasping strength model, and the material parameters are substituted into the grasping strength model to determine the model grasping strength.

[0065] Specifically, during each training session, the model attempts to learn patterns and relationships in the data to improve its prediction or classification capabilities. After each training session, the model is tested using data from the test set. If the F1 score of the current training set is lower than the F1 score of the gradient boosting tree model after the previous training, it indicates that the model may have overfitted the details of the training set data and failed to learn the true distribution and general rules of the data. Such a model has poor generalization ability when facing new and unseen data, and thus cannot make effective predictions. Cross-validation is used in combination with grid search to reduce the results of model overfitting. Cross-validation divides the training set data into several parts to balance its stability and performance. Grid search reduces the results of model overfitting by exhaustively searching for parameter combinations in the parameter space, and continues training. If the F1 score of the gradient boosting tree model after the current training is greater than or equal to the F1 score of the gradient boosting tree model after the previous training, it indicates that the model has reached a satisfactory performance level. At this time, training is stopped, and the gradient boosting tree model after the current training is determined as the grasping strength model. At this time, the grasping strength model can stably approach the global optimal solution, thereby outputting the model prediction results, ensuring the stability and reliability of the model grasping strength.

[0066] In some embodiments of the present application, when determining the grasping force of the robotic arm based on the cluster grasping force and the model grasping force, it includes: when the cluster grasping force and the model grasping force are equal, the cluster grasping force or the model grasping force is determined as the grasping force; when the cluster grasping force and the model grasping force are equal, the average of the cluster grasping force and the model grasping force is determined as the grasping force.

[0067] Specifically, when the clustering result (clustering grasping force) is the same as the model prediction result (model grasping force), one of the two values ​​is directly used as the grasping force of the robotic arm. The two independent algorithms (clustering analysis based on historical experience and gradient boosting tree model based on mapping) obtain the same result, which essentially forms a double verification process, reducing the risk of errors caused by local optimal solutions in a single algorithm, thereby forming data complementarity and ensuring the stability and accuracy of the grasping force. When the two are not equal, it means that there is a certain difference between the historical clustering data and the data obtained from the grasping force model. By taking the average of the two as the grasping force, the deviation between the two can be balanced, reducing the risk caused by errors in a single algorithm and improving the stability and accuracy of the grasping.

[0068] In some embodiments of the present application, when analyzing all mobile loads, determining the effective mobile load and the mobile load curve diagram based on the analysis results, and determining the minimum mobile load based on the mobile load curve diagram, it includes: converting all mobile loads into load coordinate points, the X-axis coordinate value of the load coordinate point represents the unit mobile time period, and the Y-axis coordinate value of the load coordinate point represents the mobile load, establishing a load coordinate system based on all load coordinate points, and fitting all load coordinate points to determine the load fitting curve, and deleting the load coordinate points that are not fitted, determining the mobile load curve diagram based on the deletion results, and using the load coordinate points on the mobile load curve diagram as the effective mobile load, and determining the minimum value of all effective mobile loads as the minimum mobile load.

[0069] Specifically, the load coordinate system is represented as a rectangular coordinate system, or XY coordinate system. The X-axis coordinate value of a load coordinate point represents the unit movement time period. Assuming the movement time period is 30 seconds, the unit movement time period can be set to 1 second, 3 seconds, 5 seconds, etc. Specifically, an arithmetic progression is selected, representing the time interval between each moving load extraction, with a moving load extracted every two seconds. During the fitting process, a curve fitting method such as polynomial fitting, spline interpolation, or least squares can be used. During the fitting process, the drive motor load is acquired using torque or current sensors. These sensors have inherent accuracy limitations and noise interference, sometimes resulting in transient error readings that differ significantly from the actual load. The measurement environment (such as electromagnetic interference and temperature fluctuations) can also affect the sensor output, causing individual moving loads to deviate from the overall trend. These deviating load coordinate points are often referred to as "outliers" and have a significant impact on the determination of the moving load curve. Therefore, during the fitting process, unfitted load coordinate points (outliers) are deleted to ensure the stability and reliability of the minimum moving load, laying the foundation for subsequent judgment.

[0070] In some embodiments of the present application, when judging whether to adjust the grasping force based on the load relationship between the minimum moving load and the grasping load, it includes: when the minimum moving load is greater than or equal to the grasping load, it is determined to adjust the grasping force; when the minimum moving load is less than the grasping load, it is determined not to adjust the grasping force, and the grasping is completed with the grasping force.

[0071] In some embodiments of the present application, when determining to adjust the grasping force, the grasping force factor is determined based on the minimum moving load and the grasping load, and the grasping force is adjusted according to the grasping force factor, including: obtaining the load ratio Q of the minimum moving load and the grasping load, when 1≤Q<1.5, the first grasping force factor is determined as the grasping force factor, when 1.5≤Q<1.7, the second grasping force factor is determined as the grasping force factor, when 1.7≤Q, the third grasping force factor is determined as the grasping force factor, the third grasping force factor is greater than the second grasping force factor, the second grasping force factor is greater than the first grasping force factor, and the first grasping force factor is greater than 1, and the grasping force and the grasping force factor are in direct proportion.

[0072] Specifically, the minimum moving load is the load required by the robot arm during stable movement, and the grasping load reflects the actual load of the robot arm when grasping the material to be grasped under the current grasping force. In the grasping action of the robot arm, the grasping load of the drive motor will rise rapidly when it just contacts the material to be grasped, so as to maintain a high but relatively stable state during continuous clamping and movement. When the minimum moving load is greater than or equal to the grasping load, it indicates that the reaction force of the material to be grasped on the gripper is insufficient, which may cause unstable grasping and the risk of slipping or falling off. In this case, the grasping force is adjusted. Otherwise, it is considered that the current grasping force has met the grasping requirements and the grasping operation can be completed. The load ratio Q is constructed based on the load ratio Q. A three-level progressive adjustment mechanism is established. When 1≤Q<1.5, it indicates that the adjustment force is slightly insufficient. In this case, a smaller first gripping force factor (the first gripping force factor is greater than 1) is used for fine-tuning to avoid damage to the material to be gripped due to excessive adjustment. When 1.5≤Q<1.7, it indicates that the adjustment force is moderately insufficient. In this case, a larger second gripping force factor is enabled to quickly improve the stability of gripping. When Q≥1.7, it indicates that the adjustment force is seriously insufficient. In this case, the largest third gripping force factor is used to ensure stable gripping under special circumstances, thereby ensuring the stability and reliability of fine-tuning of small deviations and strong correction of large deviations for the refined control of the robot arm.

[0073] It can be understood that the grasping force is adjusted according to the grasping force factor. Assuming that the grasping force is F and the grasping force factor is S, the adjusted grasping force is determined to be F*S. When a larger grasping force is needed, as the grasping force factor increases, the uncertainty of human experience is avoided. While ensuring the reliability of grasping, the refinement and efficiency optimization of the robotic arm grasping are achieved.

[0074] To sum up, the beneficial effects of the present invention are: dynamically determining the grasping strategy according to the relationship between material parameters and historical material sets, directly calling the historical material set to determine the grasping force based on the grasping experience of the historical material set, ensuring the efficiency of the grasping operation, and analyzing the historical material set based on machine learning and clustering algorithms for material parameters that are not fully matched to determine the grasping force. The dual-strategy mechanism of experience reuse, clustering algorithm and machine learning not only avoids the blindness and uncertainty of human experience setting, but also ensures the stability of the grasping force that can be determined according to actual conditions. The minimum moving load is determined according to the established moving load curve diagram. During the grasping process, the real-time collected grasping load and minimum moving load form a closed-loop feedback, thereby dynamically adjusting the grasping force, and then constructing a complete closed loop of grasping material identification, grasping strategy determination, real-time load monitoring and adjustment of grasping force, ensuring the reliability and stability of fine control of the robotic arm.

[0075] In another preferred embodiment based on the above embodiment, refer to Figure 2 As shown, this embodiment provides a machine learning-based fine control system for a robotic arm, which is used to apply the above-mentioned machine learning-based fine control method for a robotic arm, including:

[0076] The acquisition and analysis unit is configured to determine material parameters of the material to be grasped, and determine the first grasping strategy or the second grasping strategy based on the relationship between the material parameters and the historical material set.

[0077] The first processing unit is configured to determine the grasping force of the robot arm based on the historical material set when the first grasping strategy is determined, and to determine the grasping force of the robot arm based on the clustering results of the historical material set and the grasping force model when the second grasping strategy is determined.

[0078] The second processing unit is configured to obtain the moving load of the driving motor during the moving time period and analyze all moving loads, determine the effective moving load and the moving load curve based on the analysis results, determine the minimum moving load based on the moving load curve, grab the material to be grabbed according to the grabbing force, and determine the grabbing load of the driving motor, and determine whether to adjust the grabbing force according to the load relationship between the minimum moving load and the grabbing load.

[0079] The control adjustment unit is configured to determine a grasping force factor based on the minimum moving load and the grasping load when it is determined to adjust the grasping force, adjust the grasping force according to the grasping force factor, and complete the grasping according to the adjusted grasping force.

[0080] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0082] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for fine control of a robotic arm based on machine learning, characterized in that: include: Determine material parameters of the material to be grasped, and determine a first grasping strategy or a second grasping strategy based on a relationship between the material parameters and a historical material set; When determining the first grabbing strategy or the second grabbing strategy based on the relationship between the material parameters and the historical material set, it includes: The historical material set includes a plurality of historical material parameters and a plurality of historical grasping forces, and each historical material parameter corresponds to a historical grasping force; When the historical material set contains historical material parameters that are equal to the material parameters, determining that the robotic arm adopts the first grasping strategy; When there is no historical material parameter equal to the material parameter in the historical material set, determining that the robotic arm adopts the second grasping strategy; When the first grasping strategy is determined, the grasping force of the robot arm is determined based on the historical material set; when the second grasping strategy is determined, the grasping force of the robot arm is determined based on the clustering result of the historical material set and the grasping force model; When the first grasping strategy is determined, the grasping force of the robot arm is determined based on the historical material set, including: When there is a historical material parameter in the historical material set that is equal to the material parameter, and the historical material parameter is unique, the historical grasping force corresponding to the historical material parameter is determined as the grasping force; When there is a historical material parameter in the historical material set that is equal to the material parameter, and the historical material parameter is not unique, the average of the historical grasping forces corresponding to each historical material parameter is determined as the grasping force; When the second grasping strategy is determined, the grasping force of the robot arm is determined based on the clustering result of the historical material set and the grasping force model, including: The material parameters and the historical material set are used as the set to be clustered, the expected number of clusters k is determined to be 2, and the parameters of the Gaussian distribution are initialized. The probability that each data in the set to be clustered belongs to each Gaussian distribution is determined and the responsibility value is determined. Based on the responsibility value, a cluster set corresponding to the material parameters is determined, and the mean of the historical grasping strength in the cluster set is used as the cluster grasping strength; Dividing the historical material set into a training set and a test set, preselecting a gradient boosting tree model, and training the gradient boosting tree model based on the training set, testing the trained gradient boosting tree model based on the test set, and determining the model grasping strength based on the test results and the material parameters; determining a grasping force of the robotic arm based on the cluster grasping force and the model grasping force; Obtaining the moving load of the drive motor during the moving time period, analyzing all moving loads, determining an effective moving load and a moving load curve based on the analysis results, determining a minimum moving load based on the moving load curve, grasping the material to be grasped according to the grasping force, determining the grasping load of the drive motor, and determining whether to adjust the grasping force based on the load relationship between the minimum moving load and the grasping load; When it is determined that the grasping force needs to be adjusted, a grasping force factor is determined based on the minimum movement load and the grasping load, and the grasping force is adjusted according to the grasping force factor, and grasping is completed according to the adjusted grasping force.

2. The method for fine control of a robotic arm based on machine learning according to claim 1, characterized in that: The method includes training the gradient boosting tree model according to the training set, testing the trained gradient boosting tree model according to the test set, and determining the model grasping strength based on the test results and the material parameters, including: If the F1 score of the currently trained gradient boosted tree model is less than the F1 score of the previously trained gradient boosted tree model, cross-validation is used in combination with gridding to reduce the overfitting of the model, and training is continued until the F1 score of the currently trained gradient boosted tree model is greater than or equal to the F1 score of the previously trained gradient boosted tree model. If the F1 score of the currently trained gradient boosting tree model is greater than or equal to the F1 score of the previously trained gradient boosting tree model, then stop training and determine the currently trained gradient boosting tree model as the grasping force model; Substituting the material parameters into the grasping force model to determine the model grasping force.

3. The method for fine control of a robotic arm based on machine learning according to claim 2, characterized in that: When determining the grasping force of the robotic arm based on the cluster grasping force and the model grasping force, the method includes: When the cluster grasping force and the model grasping force are equal, the cluster grasping force or the model grasping force is determined as the grasping force; When the cluster grasping strength and the model grasping strength are equal, the average of the cluster grasping strength and the model grasping strength is determined as the grasping strength.

4. The method for fine control of a robotic arm based on machine learning according to claim 3, characterized in that: When analyzing all moving loads, determining an effective moving load and a moving load curve diagram based on the analysis results, and determining a minimum moving load based on the moving load curve diagram, the method includes: Convert all moving loads into load coordinate points, where the X-axis coordinate value of the load coordinate point represents the unit moving time period, and the Y-axis coordinate value of the load coordinate point represents the moving load; A load coordinate system is established based on all load coordinate points, and all load coordinate points are fitted to determine the load fitting curve, and the load coordinate points that are not fitted are deleted. The mobile load curve diagram is determined based on the deletion result, and the load coordinate points on the mobile load curve diagram are used as the effective mobile load, and the minimum value of all effective mobile loads is determined as the minimum mobile load.

5. The method for fine control of a robotic arm based on machine learning according to claim 4, characterized in that: When determining whether to adjust the gripping force according to the load relationship between the minimum moving load and the gripping load, the method includes: When the minimum moving load is greater than or equal to the grabbing load, it is determined that the grabbing force should be adjusted; When the minimum moving load is less than the grasping load, it is determined that the grasping force is not adjusted, and grasping is completed with the grasping force.

6. The method for fine control of a robotic arm based on machine learning according to claim 5, characterized in that: When determining to adjust the gripping force, determining a gripping force factor based on the minimum movement load and the gripping load, and adjusting the gripping force according to the gripping force factor, the method includes: Obtaining a load ratio Q of the minimum moving load and the grabbing load; When 1≤Q<1.5, the first grasping force factor is determined as the grasping force factor; When 1.5≤Q<1.7, the second grasping force factor is determined as the grasping force factor; When 1.7≤Q, the third grasping force factor is determined as the grasping force factor; The third grasping force factor is greater than the second grasping force factor, the second grasping force factor is greater than the first grasping force factor, and the first grasping force factor is greater than 1; The grasping force is directly proportional to the grasping force factor.

7. A machine learning-based fine control system for a robotic arm, for applying the machine learning-based fine control method for a robotic arm according to any one of claims 1 to 6, characterized in that: include: a collection and analysis unit configured to determine material parameters of a material to be grasped, and determine a first grasping strategy or a second grasping strategy based on a relationship between the material parameters and a historical material set; a first processing unit configured to, when determining the first grasping strategy, determine the grasping force of the robotic arm based on the historical material set, and when determining the second grasping strategy, determine the grasping force of the robotic arm based on the clustering result of the historical material set and a grasping force model; a second processing unit configured to obtain a moving load of the drive motor during a moving time period, analyze all moving loads, determine an effective moving load and a moving load curve based on the analysis results, determine a minimum moving load based on the moving load curve, grasp the material to be grasped according to the grasping force, determine the grasping load of the drive motor, and determine whether to adjust the grasping force based on a load relationship between the minimum moving load and the grasping load; The control adjustment unit is configured to determine a grasping force factor based on the minimum moving load and the grasping load when determining to adjust the grasping force, adjust the grasping force according to the grasping force factor, and complete the grasping according to the adjusted grasping force.

Images