A robot compliant shaft-hole assembly method based on width learning model

Abstract

The present invention provides a robot compliant shaft-hole assembly method based on a width learning model, the method comprising: obtaining contact information between the end effector and the environment through a torque sensor; initializing and training the width learning model in simulation training to obtain an initial hole position search model, and dynamically generalizing the initial hole position search model through incremental learning to update the latest hole position search model; based on the contact information and the end position of the robot, hole position search is performed through the latest hole position search model to obtain the optimal hole position; based on the optimal hole position, the robot is controlled to perform assembly operations through an anthropomorphic compliant control algorithm to complete the shaft-hole assembly task. Through this application, the robot can automatically adjust its own parameters under different environments and force feedback conditions, and the robot has the ability to interact with the environment in anthropomorphic compliance, which is conducive to improving the success rate of shaft-hole assembly tasks.

Description

Technical Field

[0001] The invention relates to the technical field of robot shaft hole compliant assembly, and in particular to a robot compliant shaft hole assembly method based on a width learning model. Background Art

[0002] With the development of industrial robot technology, robot shaft-hole assembly technology has been widely used in many actual production processes such as aviation, automobile, foundry industry, and electronic 3C fields. Shaft-hole assembly is a typical assembly task. Due to the existence of hole-finding positioning errors, it is difficult to achieve accurate assembly of shaft holes, and assembly gaps and large assembly stresses are prone to occur. At present, there is still a long way to go to achieve true "machine replacement". In addition, due to the complexity and variety of assembly tasks and the limitations of hardware conditions such as sensors, how to enable robots to have intelligent and flexible assembly capabilities is still one of the problems that need to be solved in the current field of robot assembly.

[0003] In the process of interaction between robots and the environment, compliant control is an effective way to achieve the expected interactive behavior and obtain good interactive performance. Because the expected interactive behavior will change with the change of task scenarios, active compliant control with variable structure and high flexibility has been more widely used.

[0004] At present, the main methods of active compliant control are force or position hybrid control, impedance control and admittance control. The force or position hybrid control method ignores the dynamic coupling between the robot and the environment, and its control effect depends on the modeling accuracy of the robot dynamics and the environment. When the modeling accuracy of the robot dynamics or the environment is insufficient, this method is likely to cause unstable interaction between the robot and the environment; the impedance control method is difficult to provide hard behavior suitable for interaction with the soft environment. Affected by the unmodeled friction force, the position accuracy of the manipulator using impedance control is poor when it is in free motion and interacting with the soft environment, but it can interact stably with the hard environment; the admittance control method is difficult to provide soft behavior suitable for interaction with the hard environment, and the manipulator using admittance control has poor robustness when interacting with the hard environment. Summary of the invention

[0005] Based on this, the purpose of the present invention is to provide a robot compliant shaft hole assembly method based on a width learning model to address the deficiencies in the prior art.

[0006] To achieve the above object, the present invention provides a robot compliant shaft hole assembly method based on a width learning model, the method comprising:

[0007] The contact information between the end effector and the environment is obtained through the torque sensor;

[0008] Initializing the width learning model in the simulation training to obtain an initial hole position search model, and dynamically generalizing the initial hole position search model through incremental learning to update the latest hole position search model;

[0009] Based on the contact information and the end position of the robot, hole position search is performed through the latest hole position search model to obtain the optimal hole position;

[0010] Based on the optimal hole position, the robot is controlled to perform assembly operations through a humanoid compliant control algorithm to complete the shaft hole assembly task.

[0011] The beneficial effects of the present invention are as follows: an initial hole position search model is obtained by initializing and training the width learning model in simulation training, the initial hole position search model is dynamically generalized through incremental learning to update the latest hole position search model, and the hole position search is performed based on the contact information between the end effector and the environment through the latest hole position search model to obtain the optimal hole position, which is beneficial to improving the error accumulation problem of relying on the visual system during the shaft hole alignment process, and then efficiently and accurately searching for the hole position, and then based on the optimal hole position, the robot is controlled by an anthropomorphic compliant control algorithm to perform assembly operations, so that the robot can automatically adjust its own parameters under different environments and force feedback conditions, and the robot has the ability to interact with the environment in anthropomorphic compliance, which is beneficial to improving the success rate of shaft hole assembly tasks.

[0012] Preferably, the contact information includes a generalized contact force, and the expression of the generalized contact force is as follows:

[0013]

[0014] in, is the generalized contact force, f and τ are the measured contact force and torque respectively, x, y, z are the corresponding coordinate axes, is the transpose of the matrix. When ensuring the consistency of the generalized contact force data collected in the simulation, the indirect force controller is applied along the z-axis of the tool coordinate system as follows:

[0015]

[0016] in, For force control gain, and are the tool velocity and the desired contact force along the z-axis, respectively, Set to a fixed value.

[0017] Preferably, before initializing the width learning model in the simulation training, the method further includes:

[0018] The simulation robot collects training data in a pre-designed grid space covering the hole boundary, and the training data is used to perform initialization training on the width learning model.

[0019] Preferably, the step of dynamically generalizing the initial hole position search model through incremental learning comprises:

[0020] An initial starting point is randomly generated, and based on the initial starting point, hole position searches are performed N times on the initial hole position search model, wherein each hole position search operation outputs a corresponding output label, and the network weights of the current hole position search model are updated online based on the corresponding output label, and the current hole position search model is the hole position search model obtained after the previous hole position search.

[0021] Preferably, the step of controlling the robot to perform assembly operations by using a humanoid compliant control algorithm comprises:

[0022] The motion trajectory, damping, stiffness and feedforward force parameters of the robot are uniformly encoded, and the damping, stiffness and feedforward force parameters of the robot are adaptively updated through an online adaptive update law. The updated damping, stiffness and feedforward force parameters of the robot are optimally controlled by minimizing the total cost function of interactive control and motion tracking errors.

[0023] Preferably, the step of uniformly encoding the motion trajectory, damping, stiffness and feedforward force parameters of the robot comprises:

[0024] The robot's motion trajectory is encoded through dynamic motion primitives, where the dynamic motion primitive equations in the encoding process are as follows:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030] in, is the motion position, For speed, and are the initial position and target position of the motion trajectory, s is the phase variable evolved by the canonical system, and are the preset coefficients for dynamic motion primitives, is the system forcing term obtained by linear combination of m nonlinear radial basis functions, is the torque, for The derivative of for The derivative of is the derivative of s, is the number of nonlinear radial basis functions, is the radial basis kernel function, is the number of radial basis kernel functions, and are the width and center parameters of the radial basis function, is the motion weight parameter, is the motion basis vector, whose elements can be expressed as:

[0031] .

[0032] Preferably, the expression for dynamically adjusting the damping, stiffness and feedforward force parameters of the robot through the online adaptive update law is as follows:

[0033]

[0034]

[0035]

[0036] in, , and Respectively represent the weights of damping, stiffness and feedforward force at time t, and The time period is initialized to a zero matrix of appropriate dimension, , , are the weight updates of damping, stiffness and feedforward force respectively, , , They are The weights of damping, stiffness and feedforward force at each moment, , and are learning rates, , and For the forgetting factor, , , are the matrices of damping, stiffness and feedforward force basis vectors, for The transpose of is the position error matrix e The transpose of is the parameter period.

[0037] Preferably, the expression of the total cost function for minimizing the interaction control and motion tracking errors is as follows:

[0038]

[0039]

[0040]

[0041] Where M is a fixed inertia and is designed to be a symmetric positive definite matrix, is the sliding error, , and are all symmetric positive definite matrices, is the total cost function, is the minimum motion error cost function, is the minimum control cost function, is the parameter period, , , are the update matrices of damping, stiffness and feedforward force weights, respectively. ∈( , ), is the time variable.

[0042] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flow chart of a robot compliant shaft hole assembly method based on a width learning model provided in an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the principles of step S102 and step S103 provided in an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of the principle of step S104 provided in an embodiment of the present invention.

[0046] The following specific implementation manner will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0048] Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can also be applied to other similar scenarios based on these drawings without creative work. In addition, it can also be understood that although the efforts made in this development process may be complicated and lengthy, for ordinary technicians in this field related to the content disclosed in this application, some changes in design, manufacturing or production based on the technical content disclosed in this application are just conventional technical means, and should not be understood as insufficient content disclosed in this application.

[0049] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0050] Unless otherwise defined, the technical terms or scientific terms involved in this application should be understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the" and the like involved in this application do not indicate a quantity limitation, and may indicate the singular or plural. The terms "include", "comprise", "have" and any of their variations involved in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0051] The robot compliant shaft-hole assembly method based on width learning model in the embodiment of the present invention aims to improve the accuracy and compliance of the automated assembly of shaft holes, wherein the assembly method is implemented by a robot, which includes a 6-DOF robotic arm, a 6-dimensional force or torque sensor, an end effector and a controller, the 6-dimensional force or torque sensor is used to detect generalized contact force, the 6-DOF robotic arm is used to perform assembly actions, the end effector is used to perform shaft hole insertion actions, the controller is used to control the 6-DOF robotic arm to perform hole search and shaft insertion operations, and the controller can achieve compliant interaction with the environment based on the generalized contact force by online adaptive adjustment of damping, stiffness and feedforward force parameters.

[0052] See also Figure 1 , is a flow chart of a robot compliant shaft hole assembly method based on a width learning model in an embodiment of the present invention, the method comprising the following steps:

[0053] Step S101, obtaining contact information between the end effector and the environment through a torque sensor;

[0054] The contact information includes the generalized contact force between the end effector and the environment.

[0055] Step S102, initializing the width learning model in simulation training to obtain an initial hole position search model, and dynamically generalizing the initial hole position search model through incremental learning to update the latest hole position search model;

[0056] Among them, the width learning model is initialized and trained in the simulation training, so that the trained initial hole position search model has the hole position search ability, and the incremental learning of input samples, feature mapping and enhanced nodes in the width direction of the initial hole position search model is performed through the incremental learning framework, so that there is no need to retrain the existing structure and parameters, and only the parameters of the newly added part need to be calculated, which greatly shortens the training time and can ensure a certain accuracy. The network is expanded by online updating, that is, in view of the continuous updating of training input samples, the width neural network can be dynamically expanded to make the network reflect the current input in real time, so as to realize dynamic generalization of the initial hole position search model, so as to update the latest hole position search model, such as Figure 2 shown.

[0057] Step S103, based on the contact information and the end position of the robot, and using the latest hole position search model, a hole position search is performed to obtain an optimal hole position;

[0058] Wherein, based on the end position of the robot and the generalized contact force as the perception input data, the hole position search is performed through the latest hole position search model.

[0059] Step S104, based on the optimal hole position, the robot is controlled to perform assembly operations through a humanoid compliant control algorithm to complete the shaft hole assembly task.

[0060] The step of controlling the robot to perform assembly operations by using a humanoid compliant control algorithm comprises:

[0061] The motion trajectory, damping, stiffness and feedforward force parameters of the robot are uniformly encoded, and the damping, stiffness and feedforward force parameters of the robot are adaptively updated through an online adaptive update law. The updated damping, stiffness and feedforward force parameters of the robot are optimally controlled by minimizing the total cost function of interactive control and motion tracking errors.

[0062] It should be noted that a humanoid compliant controller with position control as the inner loop is designed to uniformly encode the motion trajectory, damping, stiffness and feedforward force parameters of the robot. The damping, stiffness and feedforward force parameters are dynamically adjusted through the online adaptive update law, so that the robot has humanoid compliant interaction capabilities during the assembly process. Figure 3 shown.

[0063] Through the above steps, the width learning model is initialized and trained in the simulation training to obtain the initial hole search model, and the initial hole search model is dynamically generalized through incremental learning to update the latest hole search model. Based on the generalized contact force between the end effector and the environment and the end position of the robot, the hole search is performed through the latest hole search model to obtain the optimal hole position, which is beneficial to improve the error accumulation problem of relying on the visual system during the shaft hole alignment process, and then the hole position is efficiently and accurately searched. Then, based on the optimal hole position, the robot is controlled to perform assembly operations through the anthropomorphic compliant control algorithm, so that the robot can automatically adjust its own parameters under different environments and force feedback conditions. The robot has the ability to interact with the environment in anthropomorphic compliance, which is beneficial to improving the success rate of shaft hole assembly tasks.

[0064] In some embodiments, the contact information includes a generalized contact force, and the expression of the generalized contact force is as follows:

[0065]

[0066] in, is the generalized contact force, f and τ are the measured contact force and torque respectively, x, y, z are the corresponding coordinate axes, is the transpose of the matrix. In order to ensure the consistency of the generalized contact force data collected in the simulation, the indirect force controller is applied along the z-axis of the tool coordinate system as follows:

[0067]

[0068] in, For force control gain, and are the tool velocity and the desired contact force along the z-axis, respectively, Set to a fixed value, the output of the latest hole search model is the position offset between the shaft and the hole , They are the position offset between the shaft holes on the x-axis and y-axis respectively.

[0069] In some embodiments, before initializing the width learning model in the simulation training, the method further includes:

[0070] The simulation robot collects training data in a pre-designed grid space covering the hole boundary, and the training data is used to initialize the training of the width learning model.

[0071] The data sampling points in the grid space are evenly distributed at a certain interval.

[0072] In some embodiments, the step of initializing the width learning model in the simulation training to obtain the initial hole position search model is specifically as follows: for the width learning model, firstly, an input data set of the width learning model is given , N is the number of samples in the data set, M is the dimension of each sample, is a set of real numbers, then the i-th group of feature maps can be represented as the transformation of the input after linear mapping and activation function, and the i-th group of feature maps It is expressed as:

[0073]

[0074] The jth group of enhanced nodes It can be expressed as:

[0075]

[0076] in, is a feature layer consisting of b groups of feature maps, and are all predefined nonlinear activation functions. and are the weights of the linear mapping corresponding to the feature map and the enhancement node, and are the corresponding biases. These weights and biases are randomly generated and remain unchanged during network training. The connection matrix of the feature map and enhancement node of the width learning model is defined as , then the output of the width learning model can be expressed as:

[0077]

[0078]

[0079]

[0080] in, To enhance the connection matrix of the node, is the output of the width learning model, Learn the network weights of the model for the width, To enhance the weight of the node, is the bias of the enhancement node.

[0081] It should be noted that in order to improve the search efficiency, the connection matrix of the width learning model is obtained by the Ridge Regression method, and the Moore-Penrose generalized inverse matrix of the connection matrix A is obtained. , the network weight is:

[0082]

[0083]

[0084] in, is the constraint on the sum of squared weights, is the identity matrix, is the transpose of A.

[0085] It should be noted that given the new input sample And the corresponding label , the new feature map And the newly added enhanced nodes They are:

[0086]

[0087]

[0088] For Corresponding to the newly added feature map, the connection matrix can be updated as , and its pseudo-inverse update algorithm is as follows:

[0089]

[0090] in:

[0091]

[0092]

[0093] The updating formula of the network weights of the width learning model is:

[0094]

[0095] in, To add enhanced node connection matrix, for The transpose of is the weight of the b-th enhanced node, is the bias of the b-th enhancement node, , U, and L are all auxiliary variables constructed. , They are , the transpose of U, for The transpose of The labels output by the initial hole position search model.

[0096] In some embodiments, the step of dynamically generalizing the initial hole position search model through incremental learning includes:

[0097] An initial starting point is randomly generated, and based on the initial starting point, hole position searches are performed N times on the initial hole position search model, wherein each hole position search operation outputs a corresponding output label, and the network weights of the current hole position search model are updated online based on the corresponding output label, and the current hole position search model is the hole position search model obtained after the previous hole position search.

[0098] The robot starts from a randomly generated initial point and successfully searches for a hole after k attempts. Each attempt collects a set of generalized contact forces as the input of the initial hole search model, and obtains the position offset output to adjust the position of the axis. The k sets of input and output label pairs of the initial hole search model are ,in, The kth group output It can correctly represent the offset between the shaft position and the actual position of the hole, and obtain the generalized contact force input corresponding to the i-th group The correct label pair for:

[0099]

[0100] Among them, for an initial point, the initial hole position search model can obtain k groups of new input and output label pairs , , and update the network weights online.

[0101] When the robot repeatedly starts from different initial points and continuously attempts to search for hole positions, the initial hole position search model is continuously updated, and its prediction error will gradually decrease with the increase in the number of attempts, which means that the basic hole position search skills learned in the simulation are transferred to the real environment and gradually generalized into more advanced skills through incremental learning, thereby adapting to and completing complex real tasks without the need for model retraining, complex modeling under contact conditions, and time-consuming data collection processes in the real environment.

[0102] It should be noted that the hole search strategy based on incremental learning uses the robot's end position and generalized contact force as perception inputs, and the learned width learning model (i.e., the initial hole search model) as the basic model for hole search. It is generalized into a complex strategy through incremental learning, enabling the robot to quickly search for hole locations in real scenarios, avoiding complex modeling under contact conditions, time-consuming real data collection, and model retraining processes.

[0103] In some embodiments, the step of uniformly encoding the motion trajectory, damping, stiffness and feedforward force parameters of the robot comprises:

[0104] The robot's motion trajectory is encoded through dynamic motion primitives, where the dynamic motion primitive equations in the encoding process are as follows:

[0105]

[0106]

[0107]

[0108]

[0109]

[0110] in, is the motion position, For speed, and are the initial position and target position of the motion trajectory, s is the phase variable evolved by the canonical system, and are the preset coefficients for dynamic motion primitives, is the system forcing term obtained by linear combination of m nonlinear radial basis functions, is the torque, for The derivative of for The derivative of is the derivative of s, is the number of nonlinear radial basis functions, is the radial basis kernel function, is the number of radial basis kernel functions, and are the width and center parameters of the radial basis function, is the motion weight parameter, is the motion basis vector, whose elements can be expressed as:

[0111] .

[0112] It should be noted that given a reference trajectory, a regression algorithm, such as local weighted regression, can be used to calculate the motion weight parameters Based on the human motion learning mechanism, the humanoid compliant controller is designed as follows:

[0113]

[0114] in, is the generalized contact force, For reference position and virtual equilibrium position The error between Generated by the DMP (Dynamic Movement Primitives) system, for Assuming that the robot's position servo accuracy is high enough, it can be considered that the actual position x of the robot end is consistent with the reference position Equal, that is M is a fixed inertia and is designed to be a symmetric positive definite matrix. C is a symmetric positive definite matrix whose minimum eigenvalue satisfies the following relationship to ensure system stability:

[0115]

[0116] is the sliding error, for The derivative of is the eigenvalue function of C, is the eigenvalue of the positive definite matrix C, the sliding error It is expressed as follows:

[0117]

[0118] in, is the preset coefficient. , and They are the damping, stiffness and feedforward force that need to be adaptive during the interaction process, and are further expressed in parameter space as:

[0119]

[0120] is the damping weight matrix The transpose of is the stiffness weight matrix The transpose of is the feedforward force weight matrix The transpose of , , are matrices composed of damping, stiffness and feedforward force basis vectors, respectively. is the compliance parameter matrix, expressed as follows:

[0121]

[0122]

[0123] As you can see, In addition, the definition includes the compliant basis vectors Matrix, vector ,vector and vector as follows:

[0124]

[0125]

[0126] in, Compliant Basis Vectors The transpose of , the compliant basis vector It can be expressed as:

[0127]

[0128]

[0129] represents the i-th compliant basis vector, is the i-th kernel function, and the generalized contact force is obtained by using Taylor expansion It is expressed as follows:

[0130]

[0131] in, , and are all unknown Taylor expansion coefficients. Due to the characteristics of the environment, it can be assumed that these coefficients are periodic:

[0132]

[0133] and are all time variables, , , They represent the contact force, stiffness and damping coefficient at time t respectively. Given the generalized contact force , the goal of the designed controller is to learn the ideal damping corresponding to the minimum interactive control cost , stiffness and feedforward force , that is, the following relationship is satisfied:

[0134]

[0135] The errors between the actual and ideal values ​​of damping, stiffness and feedforward force can be expressed in parameter space as:

[0136]

[0137]

[0138]

[0139] in, , is the time variable, for The transpose of the sliding error matrix at time , for The generalized contact force matrix at time , , , They are The ideal damping, stiffness and feedforward force at each moment, for The position error at the moment, for The derivative of is the error between the actual value and the ideal value of damping, is the error between the actual value and the ideal value of stiffness, is the error between the actual value and the ideal value of the feedforward force, , , are the transpose of the update matrices for damping, stiffness, and feedforward force weights, respectively.

[0140] In some embodiments, the expression for dynamically adjusting the damping, stiffness and feedforward force parameters of the robot through the online adaptive update law is as follows:

[0141]

[0142]

[0143]

[0144] in, , and Respectively represent the weights of damping, stiffness and feedforward force at time t, and The time period is initialized to a zero matrix of appropriate dimension, , , are the weight updates of damping, stiffness and feedforward force respectively, , , They are The weights of damping, stiffness and feedforward force at each moment, , and are learning rates, , and For the forgetting factor, , , are the matrices of damping, stiffness and feedforward force basis vectors, for The transpose of is the position error matrix e The transpose of is the parameter period.

[0145] In some embodiments, the expression of the total cost function for minimizing the interaction control and motion tracking errors is as follows:

[0146]

[0147]

[0148] The expression of the total cost function that minimizes the interaction control and motion tracking errors is as follows:

[0149]

[0150]

[0151]

[0152] Where M is a fixed inertia and is designed to be a symmetric positive definite matrix, is the sliding error, , and are all symmetric positive definite matrices, is the total cost function, is the minimum motion error cost function, is the minimum control cost function, is the parameter period, , , are the update matrices of damping, stiffness and feedforward force weights, respectively. ∈( , ), is a time variable, where M is a fixed inertia and is designed to be a symmetric positive definite matrix, is the sliding error, , and are all symmetric positive definite matrices, is the total cost function, is the minimum motion error cost function, is the minimum control cost function, is the parameter period, , , are the update matrices of damping, stiffness and feedforward force weights, respectively. ∈( , ), is the time variable.

[0153] It should be noted that the above implementation process is only to illustrate the feasibility of the present application, but this does not mean that the robot flexible shaft hole assembly method based on the width learning model of the present application has only the above-mentioned only implementation process. On the contrary, as long as the robot flexible shaft hole assembly method based on the width learning model of the present application can be implemented, it can be included in the feasible implementation plan of the present application.

[0154] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0155] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A robot compliant shaft hole assembly method based on a width learning model, characterized in that: The method comprises: The contact information between the end effector and the environment is obtained through the torque sensor; Initializing the width learning model in the simulation training to obtain an initial hole position search model, and dynamically generalizing the initial hole position search model through incremental learning to update the latest hole position search model; Based on the contact information and the end position of the robot, hole position search is performed through the latest hole position search model to obtain the optimal hole position; Based on the optimal hole position, the robot is controlled to perform assembly operations through a humanoid compliant control algorithm to complete the shaft hole assembly task; Before initializing the width learning model in the simulation training, the method further includes: Collecting training data in a pre-designed grid space covering the hole boundary by a simulation robot, wherein the training data is used to initialize the width learning model; The step of dynamically generalizing the initial hole position search model through incremental learning comprises: An initial starting point is randomly generated, and based on the initial starting point, hole position searches are performed N times on the initial hole position search model, wherein each hole position search operation outputs a corresponding output label, and the network weights of the current hole position search model are updated online based on the corresponding output label, and the current hole position search model is the hole position search model obtained after the previous hole position search; The step of initializing the width learning model in the simulation training to obtain the initial hole position search model includes: Transforming the feature map in the width learning model through linear mapping and activation function to obtain the transformed feature map; Obtaining the latest enhancement node in the width learning model through a predefined nonlinear activation function and based on the transformed feature map calculation; Acquire a connection matrix of feature maps and enhanced nodes in the width learning model by a ridge regression method, and obtain a network weight of the enhanced node based on the connection matrix; An initial hole position search model is obtained based on the network weights, the latest enhanced nodes and the transformed feature map.

2. The robot compliant shaft hole assembly method based on the width learning model according to claim 1 is characterized in that: The contact information includes a generalized contact force, and the expression of the generalized contact force is as follows: in, is the generalized contact force, f and τ are the measured contact force and torque respectively, x, y, z are the corresponding coordinate axes, is the transpose of the matrix. While maintaining the consistency of the generalized contact force data collected in the simulation, the indirect force controller is applied along the z-axis of the tool coordinate system as follows: in, For force control gain, and are the tool velocity and the desired contact force along the z-axis, respectively, Set to a fixed value.

3. The robot compliant shaft hole assembly method based on the width learning model according to claim 1 is characterized in that: The step of controlling the robot to perform assembly operations by using a humanoid compliant control algorithm comprises: The motion trajectory, damping, stiffness and feedforward force parameters of the robot are uniformly encoded, and the damping, stiffness and feedforward force parameters of the robot are adaptively updated through an online adaptive update law. The updated damping, stiffness and feedforward force parameters of the robot are optimally controlled by minimizing the total cost function of interactive control and motion tracking errors.

4. The robot compliant shaft hole assembly method based on the width learning model according to claim 3 is characterized in that: The step of uniformly encoding the motion trajectory, damping, stiffness and feedforward force parameters of the robot comprises: The robot's motion trajectory is encoded through dynamic motion primitives, where the dynamic motion primitive equations in the encoding process are as follows: in, is the motion position, For speed, and are the initial position and target position of the motion trajectory, s is the phase variable evolved by the canonical system, and are the preset coefficients for dynamic motion primitives, is the system forcing term obtained by linear combination of m nonlinear radial basis functions, is the torque, for The derivative of for The derivative of is the derivative of s, is the number of nonlinear radial basis functions, is the radial basis kernel function, is the number of radial basis kernel functions, and are the width and center parameters of the radial basis function, is the motion weight parameter, is the motion basis vector, whose elements can be expressed as: 。 5. The robot compliant shaft hole assembly method based on width learning model according to claim 3 is characterized in that: The expression for dynamically adjusting the damping, stiffness and feedforward force parameters of the robot through the online adaptive update law is as follows: in, , and Respectively represent the weights of damping, stiffness and feedforward force at time t, and The time period is initialized to a zero matrix of appropriate dimension, , , are the weight updates of damping, stiffness and feedforward force respectively, , , They are The weights of damping, stiffness and feedforward force at each moment, , and are learning rates, , and For the forgetting factor, , , are the matrices of damping, stiffness and feedforward force basis vectors, for The transpose of is the position error matrix e The transpose of is the parameter period.

6. The robot compliant shaft hole assembly method based on width learning model according to claim 3 is characterized in that: The expression of the total cost function that minimizes the interaction control and motion tracking errors is as follows: Where M is a fixed inertia and is designed to be a symmetric positive definite matrix, is the sliding error, , and are all symmetric positive definite matrices, is the total cost function, is the minimum motion error cost function, is the minimum control cost function, is the parameter period, , , are the update matrices of damping, stiffness and feedforward force weights, respectively. ∈( , ), is the time variable.

Images