A skill learning method and system considering robot end posture

Abstract

The application discloses a skill learning method and system considering the posture of a robot end, and the method comprises the following steps: acquiring a plurality of sets of teaching data, each set of teaching data comprising the Cartesian coordinates of the robot end and the posture of the robot end; aligning the teaching data by using a DTW algorithm based on the plurality of sets of teaching data to obtain teaching samples; obtaining an ideal teaching trajectory by clustering and regression of a GMM-GMR model based on the teaching samples; encoding the ideal teaching trajectory by using a DMP model to obtain a task trajectory; and driving the robot to complete the reproduction or generalization of the skill based on the task trajectory. The skill learning method provided by the application considers the posture of the robot end in the skill reproduction process, regards Euler angles as teaching data, expands a two-dimensional plane task to a three-dimensional space, and enables the robot to more accurately reproduce the skill and generalize the skill in multiple dimensions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of robot skill learning, and particularly relates to a skill learning method and system considering robot end posture. BACKGROUND

[0002] The transformation of manufacturing industry has higher and higher requirements on the intelligence level of robots, and existing industrial robots based on fixed programming have gradually been difficult to meet current production and manufacturing requirements, and it is necessary to develop more flexible production systems with self-adaptive ability. The human-robot skill transfer technology is a kind of effective strategy to solve the above problems, which is developed on the basis of teaching programming, abstracts the working ability of robots or humans as a kind of "skill", and transfers the skill of humans to robots after modeling and generalization, so as to realize efficient programming and motion planning of robots.

[0003] The human-robot skill transfer technology has made great progress after forty years of development, but some problems have not been solved. The main defects are as follows: (1) the traditional programming based on the programmer requires that the operator has a certain programming foundation, and the programming cost is high. (2) It is limited to two-dimensional plane and does not consider the robot end posture in three-dimensional space. (3) The intelligence of robots in the whole production link is low, the generalization ability of robot skills is low, and the robot can only complete a specific task once programmed, and cannot change according to the environment. SUMMARY

[0004] The purpose of the application is to overcome the deficiencies in the prior art, provide a skill learning method and system considering robot end posture, and expand two-dimensional plane tasks to three-dimensional space, so that the robot can more accurately reproduce skills and generalize skills in multiple dimensions.

[0005] The application provides the following technical solutions:

[0006] In a first aspect, a skill learning method considering robot end posture is provided, including: acquiring multiple groups of teaching data, each group of teaching data including Cartesian coordinates of the robot end and robot end posture; aligning the teaching data through a DTW algorithm based on multiple groups of the teaching data to obtain a teaching sample; obtaining an ideal teaching trajectory through clustering and regression of a GMM-GMR model based on the teaching sample; encoding the ideal teaching trajectory through a DMP model to obtain a task trajectory; and driving the robot to complete skill reproduction or generalization based on the task trajectory.

[0007] As an optional technical solution of the present invention, the step of aligning the teaching data based on multiple sets of teaching data using the DTW algorithm includes: using one set of sequences from the multiple sets of teaching data as a template time series; compressing or stretching other sequences using the DTW algorithm until a shortest regularized path is found that minimizes the cumulative distance between corresponding points of the template time series and other sequences; and obtaining the teaching sample after aligning all sequences.

[0008] As an optional technical solution of the present invention, the step of obtaining the ideal teaching trajectory based on the teaching samples through clustering and regression using the GMM-GMR model includes:

[0009] Clustering of the teaching samples using the GMM model includes:

[0010] Each sample in the teaching samples is a state-action pair. ,in, This represents the robot's state data. This represents the robot's motion data. T Indicates shared ownership T A sequence, t Indicates the first t A sequence;

[0011] The teaching samples are input into the GMM model to obtain the probability density function. , is represented as:

[0012] (1)

[0013] in, K Indicates that there are a total of K A Gaussian function, k Indicates the first k A Gaussian function, N This represents the probability density function of a Gaussian distribution. Indicates the first k The weights of a Gaussian function, Indicates the first k A Gaussian function covariance matrix, Indicates the first k The mean of a Gaussian function;

[0014] A set of optimal GMM parameters was obtained using the EM algorithm. , is represented as:

[0015] (2)

[0016] Based on equation (2), the log-likelihood maximized by the GMM model is obtained. , is represented as:

[0017] (3)

[0018] wherein, denotes the variance of the k th Gaussian function.

[0019] As an optional technical solution of the present application, the ideal teaching trajectory is obtained based on the teaching sample through clustering and regression of the GMM-GMR model, and further comprising:

[0020] The behavior action of the robot is obtained through the GMR model, which comprises:

[0021] For a given state space and a conditional probability distribution of the action space is expressed as:

[0022] (4)

[0023] wherein, denotes the mean of the fitted Gaussian function, denotes the variance of the fitted Gaussian function;

[0024] The expectation of the action space is expressed as:

[0025] (5)

[0026] is expressed as: (6)

[0027] wherein, denotes the variant of the robot behavior action.

[0028] As an optional technical solution of the present application, the three dimensions of the Cartesian coordinates of the robot end are respectively expressed as , , ; the three dimensions of the robot end pose are respectively expressed as roll , pitch , yaw .

[0029] As an optional technical solution of the present application, the ideal teaching trajectory is encoded through the DMP model, comprising:

[0030] The ideal teaching trajectory is encoded through the DMP model, comprising: , , , roll , pitch ​、 yaw A set of models are built in six dimensions, and then the six DMP models are synchronized using the same regularization system, and the ideal teaching trajectory is coded into the weight values of the Gaussian basis functions in the nonlinear forcing term that can record the trajectory shape, wherein the DMP formula is expressed as:

[0031] (7)

[0032] wherein, y denotes a task trajectory, and denote the first derivative and the second derivative of y respectively, g denotes an ideal desired trajectory, and denote constants;

[0033] f denotes a nonlinear forcing term for learning the ideal desired trajectory shape, which is expressed as:

[0034] (8)

[0035] wherein, x denotes a phase term, denotes the k th Gaussian basis function, denotes a starting point;

[0036] The Gaussian basis function is expressed as:

[0037] (9)

[0038] wherein, and denote the center and the width of the k th Gaussian function respectively.

[0039] As an optional technical solution of the present application, based on formula (8), a set of weight terms are needed to be found to minimize the error between the task trajectory and the ideal desired trajectory, at this time, it is the most ideal state, and the nonlinear forcing term at this time is the nonlinear forcing term under the most ideal state, which is expressed as: (10)

[0040] wherein, denotes the second derivative of the trajectory under the most ideal state;

[0041] Based on formula (10), the weight term is expressed as:

[0042] (11)

[0043] wherein, represents a nonlinear forcing term fitted by a Gaussian basis function , represents a time-dependent variable, represents the minimum of the sum .

[0044] As an optional technical solution of the present application, the driving the robot to complete the skill reproduction or generalization based on the task trajectory comprises: inputting different starting points and ending points into the DMP model, driving the nonlinear forcing term through the phase term of the regular system f , the DMP model generates a task trajectory with the same action shape as the ideal teaching trajectory, and generalizes the joint angle values of each joint of the robot, thereby controlling the skill reproduction and generalization of the robot.

[0045] In a second aspect, a compliant control system for robot skill reproduction is provided, comprising:

[0046] An acquisition module is configured to acquire a plurality of sets of teaching data, each set of teaching data comprising Cartesian coordinates of the robot end and a robot end pose;

[0047] An alignment module is configured to align the teaching data based on the plurality of sets of teaching data by using a DTW algorithm to obtain a teaching sample;

[0048] A conversion module is configured to obtain an ideal teaching trajectory by clustering and regression of a GMM-GMR model based on the teaching sample;

[0049] An encoding module is configured to encode the ideal teaching trajectory by using a DMP model to obtain a task trajectory;

[0050] A driving module is configured to drive the robot to complete skill reproduction or generalization based on the task trajectory.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] The skill learning method considering the robot end pose provided by the present application considers the end pose in the robot skill reproduction process, uses Euler angles as teaching data, expands a two-dimensional plane task to a three-dimensional space, and enables the robot to more accurately reproduce skills and generalize skills in multiple dimensions. BRIEF DESCRIPTION OF DRAWINGS

[0053] figure 1 is a structural schematic diagram of the skill learning method in the embodiment of the present application;

[0054] figure 2 is a logical diagram of the DMP model in the embodiment of the application. DETAILED DESCRIPTION

[0055] The application will be further described below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application.

[0056] Embodiment 1

[0057] The embodiment provides a skill learning method considering a robot end posture, which can transfer the skill of a human to a robot in a human-robot physical interaction (dragging demonstration) manner, and the robot can realize accurate skill reproduction and skill generalization in multiple dimensions.

[0058] Before the dragging demonstration is performed, a robot experiment platform needs to be built. In this embodiment, the robot experiment platform is composed of a collaborative robot arm and a computer, and is demonstrated by an operator. The collaborative robot arm is an xMateER3 Pro manufactured by ROKAE, and each joint of the robot is loaded with a joint torque sensor to support the zero force dragging mode. The computer needs to be loaded with MATLAB and Network Assistant. The computer and the robot need to be under the same local area network, and can be connected by a local area network or a network cable. The two communicate through the TCP protocol, and the Network Assistant is responsible for the data transmission and reception.

[0059] As shown in figure 1 , the skill learning method considering the robot end posture specifically includes the following steps:

[0060] Step 1: Obtain a plurality of sets of demonstration data, each set of demonstration data including Cartesian coordinates of the robot end and a robot end posture.

[0061] Specifically, the demonstrator completes the demonstration by demonstrating the process of performing a certain task, and the demonstration mode can be realized through various human-computer interaction interfaces. The robot is in a zero force dragging mode, and the demonstrator can freely drag the robot to complete the task. In the demonstration process, the Cartesian coordinates of the robot are recorded, and the three dimensions are respectively represented as , , . At the same time, the robot end posture is recorded, and the three dimensions are respectively represented as roll , pitch , yaw , and the data acquisition process does not need to consider the absolute consistency of the sampling points.

[0062] Step 2: Align the demonstration data by using a DTW algorithm based on the plurality of sets of demonstration data, to obtain a demonstration sample.

[0063] The teaching data in the embodiment, i.e., the teaching trajectory, includes six time sequences, respectively , , , roll , pitch , yaw time sequences in six dimensions.

[0064] The DTW (Dynamic Time Warping) algorithm aligns another template time sequence by compressing or lengthening an unknown time sequence, in which the unknown time sequence is twisted or bent so as to align its features with the template time sequence. It includes:

[0065] Step 2.1, taking a sequence in a group of the teaching data as a template time sequence.

[0066] Specifically, since each teaching trajectory is almost the same, only the length of the trajectory is different, generally a teaching trajectory with better teaching effect is selected as the template time sequence.

[0067] Step 2.2, compressing or lengthening other sequences by the DTW algorithm until a shortest regular path is found to minimize the cumulative distance between the corresponding points of the template time sequence and other sequences.

[0068] Step 2.3, obtaining the teaching sample after aligning all sequences.

[0069] In the embodiment, each sample in the teaching sample is a state-action pair , wherein represents the state data of the robot, represents the action data of the robot, T represents a total of T sequences, t represents the i-th sequence. t

[0070] Step 3: based on the teaching sample, an ideal teaching trajectory is obtained by clustering and regression of the GMM-GMR model. Specifically, it includes the following steps:

[0071] Step 3.1, clustering the teaching sample by the GMM model (Gaussian Mixture Model). It includes:

[0072] Step 3.1.1, inputting the teaching sample into the GMM model to obtain the probability density function , denoted as:

[0073] (1)

[0074] ​wherein, K denotes the total number of Gaussian functions in the GMM model, K k denotes the i-th Gaussian function, k N denotes the probability density of the Gaussian distribution, denotes the weight of the i-th Gaussian function, k denotes the covariance matrix of the i-th Gaussian function, k denotes the mean of the i-th Gaussian function. k

[0075] Step 3.1.2, obtaining a set of optimal GMM parameters by EM algorithm denoted as:

[0076] (2)

[0077] Based on equation (2), the log-likelihood of the GMM model is maximized denoted as:

[0078] (3)

[0079] wherein, denotes the variance of the i-th Gaussian function. k

[0080] Step 3.2, obtaining the behavior action of the robot by GMR model (Gaussian Mixture Regression), i.e., finding the mapping relationship from the state space to the action space . It includes:

[0081] For a given state space and the conditional probability distribution of the action space denoted as:

[0082] (4)

[0083] wherein, denotes the mean of the fitted Gaussian function, denotes the variance of the fitted Gaussian function.

[0084] Further, reflects how the state space affects the expected value of the action space , reflects how the state space affects the variance of the action space . ​​​​​​​

[0085] behavioral space is expressed as:

[0086] (5)

[0087] is expressed as: (6)

[0088] wherein, represents a variant of a robot behavior action. Further, defines the general form of a behavior, whereas depicts the range of variation or uncertainty allowed when performing the action.

[0089] Step 4: Encode the ideal demonstration trajectory by a DMP model to obtain a task trajectory.

[0090] Specifically, a set of modeling is performed for the ideal demonstration trajectory by the DMP model (dynamic motion primitive) in , , , roll , pitch , yaw six dimensions, and then the six DMP models are synchronized by using the same regular system to encode the ideal demonstration trajectory into the weight values of the Gaussian basis functions in the nonlinear forcing term capable of recording the trajectory shape.

[0091] As shown in figure 2 , after a series of operations (integration is represented in the figure) S , the task trajectory y finally converges on the ideal demonstration trajectory g . Specifically, it includes:

[0092] Step 4.1, introduce a nonlinear forcing term.

[0093] The DMP formula is expressed as: (7)

[0094] wherein, y represents a task trajectory, and represent the first-order derivative and the second-order derivative of y respectively, g represents an ideal desired trajectory, and represent constants.

[0095] f represents a nonlinear forcing term used to learn the shape of the ideal desired trajectory, which is expressed as:

[0096] (8)

[0097] wherein, x denotes the phase term, denotes the k th Gaussian kernel, denotes the starting point.

[0098] The Gaussian kernel is denoted as:

[0099] (9)

[0100] wherein, and denote the center and width of the k th Gaussian kernel, respectively.

[0101] Step 4.2, based on equation (8), a set of weight terms is needed to minimize the error between the task trajectory and the ideal desired trajectory, which is the most ideal state, and the nonlinear forcing term at this time is the nonlinear forcing term at the most ideal state , denoted as:

[0102] (10)

[0103] wherein, denotes the second derivative of the trajectory at the most ideal state.

[0104] Based on equation (10), the weight term is denoted as:

[0105] (11)

[0106] wherein, denotes the nonlinear forcing term obtained by fitting the Gaussian kernel , denotes the variable related to time, denotes the .

[0107] Step 5: based on the task trajectory, drive the robot to complete the reproduction or generalization of skills.

[0108] Specifically, different starting points and endpoints are input into the DMP model, and the nonlinear forcing term fThe DMP model generates a task trajectory with the same action shape as the ideal teaching trajectory, and generalizes the joint angle values of each joint of the robot, controls the reproduction and generalization of the robot skill. That is, the data generated by the DMP model is sent to a robot inverse kinematics solver to generate corresponding joint variables, which are sent to a robot controller to drive the robot to complete trajectory reproduction or generalization. Embodiment 2

[0109] The embodiment provides a skill learning system considering a robot end posture, comprising:

[0110] An acquisition module is configured to acquire a plurality of sets of teaching data, each set of teaching data comprising Cartesian coordinates of the robot end and a robot end posture.

[0111] An alignment module is configured to align the teaching data by using a DTW algorithm based on the plurality of sets of teaching data, to obtain teaching samples.

[0112] A conversion module is configured to obtain an ideal teaching trajectory by clustering and regression of a GMM-GMR model based on the teaching samples.

[0113] An encoding module is configured to encode the ideal teaching trajectory by using a DMP model, to obtain a task trajectory.

[0114] A driving module is configured to drive the robot to complete the reproduction or generalization of the skill based on the task trajectory.

[0115] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] The present application is described with reference to 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 flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing apparatus to produce a machine, so that the instructions executed by the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. figure 1 The functions specified in one flow or multiple flows and / or blocks figure 1 The functions specified in one flow or multiple flows and / or blocks

[0117] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the figure 1 function specified in the flow or flows and / or blocks figure 1 of the block or blocks.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing the figure 1 function specified in the flow or flows and / or blocks figure 1 of the block or blocks.

[0119] The above description is only preferred embodiments of the present application. It should be pointed out that, for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present application, and these improvements and modifications should also be considered as the protection scope of the present application.

Claims

1. A skill learning method considering robot end-effector posture, characterized in that, include: Acquire multiple sets of teaching data, each set of teaching data including the Cartesian coordinates and robot end-effector pose; Based on multiple sets of teaching data, the teaching data are aligned using the DTW algorithm to obtain teaching samples; Based on the teaching samples, the ideal teaching trajectory is obtained through clustering and regression using the GMM-GMR model; The ideal teaching trajectory is encoded using the DMP model to obtain the task trajectory; Based on the task trajectory, the robot is driven to reproduce or generalize the skill; The step of aligning the teaching data based on multiple sets of teaching data using the DTW algorithm includes: One sequence from the multiple sets of teaching data is used as a template time series; The DTW algorithm is used to compress or stretch other sequences until a shortest regular path is found that minimizes the cumulative distance between corresponding points of the template time series and other sequences. The teaching sample is obtained by aligning all sequences. The process of obtaining the ideal teaching trajectory based on the teaching samples through clustering and regression using the GMM-GMR model includes: Clustering of the teaching samples using the GMM model includes: Each sample in the teaching samples is a state-action pair. ,in, This represents the robot's state data. This represents the robot's motion data. T Indicates shared ownership T A sequence, t Indicates the first t A sequence; The teaching samples are input into the GMM model to obtain the probability density function. , is represented as: （1） in, K Indicates that there are a total of K A Gaussian function, k Indicates the first k A Gaussian function, N This represents the probability density function of a Gaussian distribution. Indicates the first k The weights of a Gaussian function, Indicates the first k A Gaussian function covariance matrix, Indicates the first k The mean of a Gaussian function; A set of optimal GMM parameters was obtained using the EM algorithm. , is represented as: （2） Based on equation (2), the log-likelihood maximized by the GMM model is obtained. , is represented as: （3） in, Indicates the first k The variance of a Gaussian function; The step of obtaining the ideal teaching trajectory based on the teaching samples through clustering and regression using the GMM-GMR model further includes: The robot's behavior is obtained through the GMR model, including: For a given state space and behavioral space conditional probability distribution Represented as: （4） in, This represents the mean of the fitted Gaussian function. This represents the variance of the fitted Gaussian function; Behavioral Space The expectation is expressed as: （5） Represented as: (6) in, This refers to variations of robot behavior.

2. The skill learning method considering robot end-effector posture according to claim 1, characterized in that, The three dimensions of the Cartesian coordinates of the robot's end effector are respectively represented as follows: , , The three dimensions of the robot's end-effector posture are respectively represented as follows: roll , Pitch , yaw .

3. The skill learning method considering robot end-effector posture according to claim 2, characterized in that: The encoding of the ideal teaching trajectory using the DMP model includes: The DMP model is used to define the ideal teaching trajectory. , , , roll , Pitch , yaw A set of models is created across six dimensions, and then the six DMP models are synchronized using the same regularization system. The ideal teaching trajectory is encoded as the weights of the Gaussian function in a nonlinear forcing term that records the trajectory shape, where the DMP formula is expressed as: （7） in, y Indicates the task trajectory. and They represent y The first and second derivatives, g Represents the ideal expected trajectory. and Represents a constant; f The nonlinear forcing term used to learn the shape of the desired trajectory is expressed as: （8） in, x Indicates the phase term, Indicates the first k Gaussian functions, Indicates the starting point; Gaussian function Represented as: （9） in, and They represent the first k The center and width of a Gaussian function.

4. The skill learning method considering robot end-effector posture according to claim 3, characterized in that, Based on equation (8), a set of weight terms needs to be found. The goal is to minimize the error between the task trajectory and the desired trajectory; this is the ideal state, and the nonlinear forcing term in this state is the same as the nonlinear forcing term in the ideal state. , is represented as: （10） in, This represents the second derivative of the trajectory under the most ideal condition; Based on equation (10), the weight term Represented as: （11） in, Indicates the use of Gaussian functions The nonlinear forcing term obtained from the fitting, Representing variables related to time, Indicates taking the minimum sum .

5. The skill learning method considering robot end-effector posture according to claim 1, characterized in that, The step of driving the robot to reproduce or generalize skills based on the task trajectory includes: inputting different start and end points into the DMP model, and driving nonlinear forcing terms through the phase term of the regularized system. f The DMP model generates a task trajectory with the same shape as the ideal teaching trajectory, and generalizes the joint angle values ​​of each joint of the robot to control the reproduction and generalization of robot skills.

6. A skill learning system considering robot end-effector posture, characterized in that, A skill learning method for implementing any one of claims 1-5, taking into account the robot's end-effector posture, comprises: The acquisition module is used to acquire multiple sets of teaching data, each set of teaching data including the Cartesian coordinates of the robot end effector and the robot end effector posture; The alignment module is used to align the teaching data based on multiple sets of teaching data using the DTW algorithm to obtain teaching samples; The conversion module is used to obtain the ideal teaching trajectory based on the teaching samples through clustering and regression of the GMM-GMR model; The encoding module is used to encode the ideal teaching trajectory using the DMP model to obtain the task trajectory; The driving module is used to drive the robot to reproduce or generalize skills based on the task trajectory.

Images