A scalable method and system for learning humanoid precision grasping skills in robots

Abstract

This invention discloses a scalable method and system for learning humanoid precision grasping skills in robots, involving the intersection of artificial intelligence and robotics. The method comprises: obtaining a robot's motion trajectory that imitates a human grasping; establishing a robot grasping skill model based on the motion trajectory that imitates the human grasping; learning and optimizing the parameters of the robot grasping skill model to obtain a grasping reference trajectory guided by the human's experience and knowledge; converting the grasping reference trajectory into a linear dynamic system and optimizing the parameters of the linear dynamic system to achieve global convergence of the robot's grasping skill; and reconstructing the robot grasping skill model using intermediate points to achieve rapid expansion of the robot's grasping skill. This invention can solve the problems of poor scalability and low accuracy in robot grasping path planning in unstructured environments.

Description

Technical Field

[0001] The present invention relates to the intersection of artificial intelligence and robotics, and in particular to an expandable method and system for learning humanoid precision grasping skills of robots. Background Art

[0002] Robot grasping skill learning is an important research direction at the intersection of artificial intelligence and robotics. It is the key to enabling robots to flexibly perform tasks in unstructured dynamic environments and has broad application prospects in fields such as intelligent manufacturing and logistics warehousing. Traditional robotic grasping systems are usually composed of modules such as object pose estimation based on visual perception, robotic grasping path planning, and motion control. However, the grasping path planning module faces many challenges, such as the need for dynamic adjustment to adapt to environmental changes, object pose requirements, stability, and production rhythm. Currently, this module is limited by reliance on manual experience, poor scalability, low precision, low efficiency, and high cost. It is difficult to meet the needs of object grasping in unstructured and complex scenarios, thus limiting the depth and breadth of robotic applications. Summary of the Invention

[0003] The present invention provides an expandable robot humanoid precision grasping skill learning method and system to solve the problems of poor scalability and low accuracy in the robot grasping path planning process in an unstructured environment.

[0004] To achieve the above objectives, the present invention provides the following technical solutions:

[0005] In a first aspect, the present invention provides a scalable method for learning a robot's humanoid precision grasping skills, comprising:

[0006] Obtain the robot's motion trajectory imitating the demonstrator's grasp;

[0007] Establishing a robot grasping skill model based on the motion trajectory of the grasping imitating demonstrator;

[0008] Learning and optimizing the parameters of the robot grasping skill model to obtain a grasping reference trajectory guided by the demonstrator's experience and knowledge;

[0009] Converting the grasping reference trajectory into a linear dynamic system and optimizing the parameters of the linear dynamic system to achieve global convergence of the robot's grasping skill;

[0010] The robot grasping skill model is reconstructed using the intermediate points to achieve rapid expansion of the robot grasping skill.

[0011] In a second aspect, the present invention provides an expandable robot humanoid precision grasping skill learning system, comprising:

[0012] A data acquisition module, which is used to obtain the motion trajectory of the robot imitating the demonstrator's grasping;

[0013] A model building module, which is used to build a robot grasping skill model based on the motion trajectory of the imitation demonstrator's grasping;

[0014] An experience guidance module, which is used to learn and optimize the parameters of the robot's grasping skill model to obtain a grasping reference trajectory guided by the demonstrator's experience knowledge;

[0015] A global convergence module, which is used to convert the grasping reference trajectory into a linear dynamic system and optimize the parameters of the linear dynamic system to achieve global convergence of the robot's grasping skill;

[0016] A rapid expansion module is used to reconstruct the robot grasping skill model using intermediate points to achieve rapid expansion of the robot grasping skill.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] By combining imitation learning with stability theory, this paper invents a scalable method for learning humanoid precision grasping skills in robots, which can quickly adapt to the deployment and application of robot grasping systems in unstructured environments.

[0019] (1) Based on the human-demonstrated motion trajectory (position, posture, and other information), a scalable humanoid grasping framework of "skill demonstration - skill learning - skill stability - skill generalization" was proposed to meet the needs of multi-task and cross-scenario grasping applications;

[0020] (2) A global convergent robot humanoid grasping skill model based on efficient kernel mapping was proposed to improve the accuracy of robot humanoid grasping in unstructured environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 Flowchart of a method for learning a robot's humanoid precision grasping skill in an embodiment of the present invention.

[0023] Figure 2 This is a logic diagram of a method for learning a robot's humanoid precision grasping skill in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0025] Example:

[0026] It should be noted that the terms "including" and "having" in the embodiments of the present invention and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or devices.

[0027] The present invention aims at the application requirements and existing problems of robot grasping system path planning in unstructured environments, and proposes a scalable robot humanoid precision grasping skill learning method by combining imitation learning and stability theory. The inventive concept of this method is to use human demonstration motion trajectory as a reference, introduce an efficient kernel mapping mechanism, and construct a multi-task and cross-scenario robot grasping "skill demonstration-skill learning-skill stability-skill generalization" method framework, which can quickly adapt to the deployment and application of robot grasping systems in unstructured environments.

[0028] First, under the robot gravity compensation mode, the demonstrator and the robot conduct physical interaction to realize multiple demonstrations and teaching of humanoid grasping skills. The motion trajectory of the humanoid grasping skill is obtained through data collection, data preprocessing and posture change methods. Then, a robot grasping skill modeling and learning method based on an efficient kernel mapping mechanism is constructed. ① Skill modeling: Based on the demonstration motion trajectory, the Gaussian process is introduced to obtain the experiential knowledge of human demonstration and construct a robot grasping skill model. ② Skill learning: After introducing the experiential knowledge of human demonstration, the parameters of the robot grasping skill model are learned and optimized, and finally a grasping reference trajectory guided by human experiential knowledge is obtained. Thirdly, the nonlinear reference trajectory is converted into a linear dynamic system, and the parameterized quadratic Lyapunov function is used to optimize the parameters of the linear system to achieve global convergence of the grasping skill and improve grasping accuracy. Finally, the skill model is reconstructed by introducing intermediate points in the robot grasping skill model based on the efficient kernel mapping mechanism, so as to achieve rapid expansion of grasping skills for different objects and across scenarios.

[0029] The present invention can acquire human-like grasping skills from multiple demonstrations and teaching processes of human-machine physical interaction, effectively improving the robot's ability to accurately and stably perform grasping tasks in unstructured environments, and has good generalization ability and scalability for complex task scenarios.

[0030] See also Figure 1 and Figure 2 The embodiment of the present invention provides a scalable method for learning a robot's humanoid precision grasping skill, which may include the following steps:

[0031] Step 100: Obtain the motion trajectory of the robot imitating the demonstrator's grasping.

[0032] In this step, in a given unstructured task environment, the robot is set to gravity compensation mode, and the demonstrator conducts physical human-machine interaction with the robot, performing multiple traction demonstrations and teaching of humanoid grasping skills, and ensuring that the grasping task can be completed every time. During the demonstration process, the robot demonstration data is collected and the collected information is preprocessed, including noise reduction, outlier removal, and alignment of the teaching trajectory based on the dynamic time warping algorithm, so as to obtain multiple demonstration teaching trajectories of the robot's position and posture in the task space, which are expressed as where x t,m represents the position at time t in the mth demonstration, q t,m (x q ,y q ,z q ,w) represents the posture at time t in the mth demonstration, where x q is the vector part y in the x-axis direction q is the vector part in the y-axis direction, z q is the vector part in the z-axis direction, w is the scalar part, which is related to the angle of rotation, T is the amount of data for each demonstration trajectory after alignment, and M is the total number of demonstrations.

[0033] Step 200: Establishing a robot grasping skill model based on the motion trajectory imitating the grasping of the demonstrator.

[0034] In this step, the posture q t,m (x q ,y q ,z q ,w) for better grasping skill modeling and learning:

[0035]

[0036]

[0037] Where R represents the rotation matrix, Ω(α, β, γ) represents the Euler angle form of the posture, where α, β, and γ represent the angles of rotation around the x-axis, y-axis, and z-axis respectively.

[0038] Taking the demonstration motion trajectory as a reference, a robot grasping skill modeling and learning method based on an efficient kernel mapping mechanism is constructed. For a given reference trajectory x n ,Ω n is the position information and information corresponding to the nth point. There is a functional relationship between position and posture. For a given input x * The corresponding output function value f(x * ) and the sample output Ω can be expressed as:

[0039]

[0040] in, k * =[k(x * ,x1)k(x * ,x2)…k(x * ,x N )], k(.,.) represents the kernel function, K is the matrix containing the kernel function, Ν(·) represents the normal distribution, σ 2 is the variance of the noise. According to the joint probability distribution in the above formula and the conditional probability distribution of multivariate Gaussian, we can obtain Ρ(f(x * )|Ω), Ρ(·) represents the probability distribution, and its mean and variance can be expressed as:

[0041] μ(x)=k*(K+σ 2 I) -1 y

[0042] ∑=k(x * ,x * )-k * (K+σ 2 I) -1 k *T

[0043] This allows for the acquisition of experiential knowledge of the demonstrator's human-like grasping skills.

[0044] Step 300: Learning and optimizing the parameters of the robot grasping skill model to obtain a grasping reference trajectory guided by the demonstrator's experience and knowledge.

[0045] In this step, the demonstrator’s experience knowledge is introduced, and a reference trajectory can be obtained for a given input. They represent the mean and variance corresponding to the nth point in the parameter trajectory, respectively, and the following parameterized model is used:

[0046] ξ(x)=Φ T (x)ω

[0047] Among them, ξ(x) is the parameter model to be sought, weight ω, obeys the normal distribution ω~N(μ ω ,∑ ω ), for:

[0048]

[0049] in, Represents the B-dimensional basis function vector, minimizing the KL divergence between the probability distribution of the generated trajectory and the reference trajectory, that is:

[0050]

[0051] in, Φ T (x n )Σ ω Φ(x n )), Decompose the objective function to obtain the mean and variance after parameter optimization:

[0052] Ε(ξ(x * ))=k * (K+λ1Σ) -1 μ

[0053]

[0054] Among them, λ1>0,λ2>0 are regularization coefficients,

[0055] Step 400: Convert the grasping reference trajectory into a linear dynamic system, and optimize the parameters of the linear dynamic system to achieve global convergence of the robot's grasping skills.

[0056] In this step, the nonlinear reference trajectory is converted into a linear dynamic system, expressed as:

[0057] Ω=f(x)=h(x)(Ax+b)

[0058] Where A = D Ω (D x ) -1 ,b=Ε Ω -AE x , E x ,E Ω Represent the mean values ​​of position and posture respectively, D x ,D ΩRepresent the variance corresponding to the position and attitude respectively. In order to meet the global convergence of the system, the system should satisfy:

[0059]

[0060] Where h(x) is the proportional coefficient, P is any symmetric positive definite matrix, and the symbol '<0' indicates that the matrix is ​​negative definite. The parameterized quadratic Lyapunov function Ω(x) = (xx * ) T P(xx * ) Verify the global convergence of the linear system, learn and optimize the unknown parameters A and b of the linear parameter system, so that the robot humanoid grasping skill model can achieve robot humanoid precise grasping in an unstructured environment.

[0061] Step 500: Reconstruct the robot grasping skill model using the intermediate points to achieve rapid expansion of the robot grasping skill.

[0062] In this step, in order to better solve the problem that the traditional robot grasping system relies on manual experience and has poor scalability, the present invention introduces the method of the middle point after modeling the robot grasping skill. According to the task characteristics of different objects, multiple tasks, and cross-scenes, the robot grasping skill model is reconstructed. Assuming that the middle point is (x0, q0), the posture change of the middle point is R*V(x0, q0)=V′(x v ,Ω v ), and then the transformed intermediate point is introduced into the original robot grasping skill model for reconstruction. The mean and variance of the new model parameters after optimization are:

[0063]

[0064]

[0065] Then, through global stability strategies and Lyapunov function parameter optimization, the global stability of the reconstructed grasping skill model is achieved, thereby enabling the rapid expansion of grasping skills for different objects, multiple tasks, and cross-scenario situations.

[0066] Through the above framework, the rapid expansion capability of the robot grasping system and the accuracy of grasping skills in unstructured environments are effectively improved, which has important scientific research significance and application value.

[0067] Based on the same inventive concept, an embodiment of the present invention further provides an expandable robot humanoid precision grasping skill learning system, comprising:

[0068] A data acquisition module, which is used to obtain the motion trajectory of the robot imitating the demonstrator's grasping;

[0069] A model building module, which is used to build a robot grasping skill model based on the motion trajectory of the imitation demonstrator's grasping;

[0070] An experience guidance module, which is used to learn and optimize the parameters of the robot's grasping skill model to obtain a grasping reference trajectory guided by the demonstrator's experience knowledge;

[0071] A global convergence module, which is used to convert the grasping reference trajectory into a linear dynamic system and optimize the parameters of the linear dynamic system to achieve global convergence of the robot's grasping skill;

[0072] A rapid expansion module is used to reconstruct the robot grasping skill model using intermediate points to achieve rapid expansion of the robot grasping skill.

[0073] In the data acquisition module, the robot's motion trajectory imitating the demonstrator's grasping is obtained, including:

[0074] In a given unstructured task environment, the robot is set to gravity compensation mode. The demonstrator conducts physical human-machine interaction with the robot and performs multiple traction demonstrations and teachings on humanoid grasping skills. During the demonstration, the robot demonstration data is collected and the collected information is preprocessed to obtain multiple demonstration and teaching trajectories of the robot's position and posture in the task space. The multiple demonstration and teaching trajectories are determined as the motion trajectory of the robot imitating the demonstrator's grasping.

[0075]

[0076] Where x t,m represents the position at time t in the mth demonstration, q t,m (x q ,y q ,z q ,w) represents the posture at time t in the mth demonstration, x q is the vector part y in the x-axis direction q is the vector part in the y-axis direction, z q is the vector part in the z-axis direction, w is the scalar part, which is related to the angle of rotation, T is the amount of data for each demonstration trajectory after alignment, and M is the total number of demonstrations.

[0077] In the model building module, a robot grasping skill model is established based on the motion trajectory of the grasping imitating demonstrator, specifically including:

[0078] Convert posture information;

[0079]

[0080]

[0081] Where R represents the rotation matrix, Ω(α, β, γ) represents the Euler angle form of the posture, where α, β, and γ represent the angles of rotation around the x-axis, y-axis, and z-axis respectively;

[0082] For a given reference trajectory x n ,Ω n is the position information and information corresponding to the nth point. There is a functional relationship between position and posture. For a given input x * The corresponding output function value f(x * ) and the sample output Ω is expressed as:

[0083]

[0084] Where, k * =[k(x * ,x1)k(x * ,x2)…k(x * ,x N )], k(.,.) represents the kernel function, K is the matrix containing the kernel function, Ν(·) represents the normal distribution, σ 2 is the variance of the noise;

[0085] According to the joint probability distribution in the above formula and the conditional probability distribution of multivariate Gaussian, we can obtain Ρ(f(x * )|Ω), Ρ(·) represents the probability distribution, and its mean and variance can be expressed as:

[0086] μ(x)=k*(K+σ 2 I) -1 y

[0087] ∑=k(x * ,x * )-k * (K+σ 2 I) -1 k *T .

[0088] In the experience guidance module, the parameters of the robot grasping skill model are learned and optimized to obtain a grasping reference trajectory guided by the demonstrator's experience knowledge, specifically including:

[0089] By introducing the demonstrator’s experience knowledge, a reference trajectory can be obtained for a given input. They represent the mean and variance corresponding to the nth point in the parameter trajectory, respectively, and the following parameterized model is used:

[0090] ξ(x)=Φ T (x)ω

[0091] Where ξ(x) is the parameter model to be sought, and the weight ω follows the normal distribution ω~N(μ ω ,∑ ω ), for:

[0092]

[0093] Where, Represents the B-dimensional basis function vector, minimizing the KL divergence between the probability distribution of the generated trajectory and the reference trajectory, that is:

[0094]

[0095] Where, Φ T (x n )Σ ω Φ(x n )),

[0096] Decompose the objective function to obtain the mean and variance after parameter optimization:

[0097] Ε(ξ(x * ))=k * (K+λ1Σ) -1 μ

[0098]

[0099] Where λ1>0,λ2>0 are regularization coefficients,

[0100] In the global convergence module, the grasping reference trajectory is converted into a linear dynamic system, and the parameters of the linear dynamic system are optimized to achieve the global convergence of the robot's grasping skill, specifically including:

[0101] The nonlinear grasping reference trajectory is converted into a linear dynamic system, which is expressed as:

[0102] Ω=f(x)=h(x)(Ax+b)

[0103] Where A = D xq (D x ) -1 ,b=Ε q -AE x , E x ,E Ω Represent the mean values ​​of position and posture respectively, D x ,D Ω Represent the variance corresponding to position and posture respectively,

[0104] In order to satisfy the global convergence of the system, the system should satisfy:

[0105]

[0106] Where h(x) is the proportional coefficient, P is an arbitrary symmetric positive definite matrix, the symbol '<0' indicates the negative definiteness of the matrix, and the parameterized quadratic Lyapunov function Ω(x) = (xx * ) T P(xx * ) Verify the global convergence of the linear system, learn and optimize the unknown parameters A and b of the linear parameter system, so that the robot humanoid grasping skill model can achieve robot humanoid precise grasping in an unstructured environment.

[0107] In the rapid expansion module, the robot grasping skill model is reconstructed using the intermediate points to achieve rapid expansion of the robot grasping skill, specifically including:

[0108] Set the middle point to (x0, q0), and change the posture of the middle point to R*V(x0, q0)=V′(x v ,Ω v ), and then the transformed intermediate point is introduced into the original robot grasping skill model for reconstruction. The mean and variance of the new model parameters after optimization are:

[0109]

[0110]

[0111] Then, through global stability strategy and Lyapunov function parameter optimization, the global stability of the reconstructed grasping skill model is achieved, thereby realizing the rapid expansion of grasping skills for different objects, multiple tasks, and cross-scenarios.

[0112] Since this system is a system corresponding to the scalable robot humanoid precision grasping skill learning method of an embodiment of the present invention, and the principle of solving the problem by this system is similar to that of this method, the implementation of this system can refer to the implementation process of the above-mentioned method embodiment, and the repeated parts will not be repeated.

[0113] In the description of this specification, the reference terms "one embodiment", "some embodiments", "examples", "specific examples" or "some examples" mean 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 expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0114] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made based on the essence of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A scalable method for learning humanoid precision grasping skills in robots, characterized in that: include: Obtain the robot's motion trajectory imitating the demonstrator's grasp; Establishing a robot grasping skill model based on the motion trajectory of the grasping imitating demonstrator; Learning and optimizing the parameters of the robot grasping skill model to obtain a grasping reference trajectory guided by the demonstrator's experience and knowledge; Converting the grasping reference trajectory into a linear dynamic system and optimizing the parameters of the linear dynamic system to achieve global convergence of the robot's grasping skill; Reconstructing the robot grasping skill model using the intermediate points to achieve rapid expansion of the robot grasping skill; The grasping reference trajectory is converted into a linear dynamic system, and the parameters of the linear dynamic system are optimized to achieve the global convergence of the robot's grasping skills, specifically including: The nonlinear grasping reference trajectory is converted into a linear dynamic system, which is expressed as: Ω=f(x)=h(x)(Ax+b) Where A = D xq (D x ) -1 ,b=E q -AE x , E x ,E Ω Represent the mean values ​​of position and posture respectively, D x ,D Ω Represent the variance corresponding to position and posture respectively, In order to satisfy the global convergence of the system, the system should satisfy: Where h(x) is the proportional coefficient, P is an arbitrary symmetric positive definite matrix, the symbol '<0' indicates the negative definiteness of the matrix, and the parameterized quadratic Lyapunov function Ω(x) = (xx * ) T P(xx * ) Verify the global convergence of the linear system and learn and optimize the unknown parameters A and b of the linear parameter system, so that the robot humanoid grasping skill model can achieve humanoid precise grasping in an unstructured environment; The grasping reference trajectory is converted into a linear dynamic system, and the parameters of the linear dynamic system are optimized to achieve global convergence of the robot's grasping skills, specifically including: The nonlinear grasping reference trajectory is converted into a linear dynamic system, which is expressed as: Ω=f(x)=h(x)(Ax+b) Where A = D xq (D x ) -1 ,b=E q -AE x , E x ,E Ω Represent the mean values ​​of position and posture respectively, D x ,D Ω Represent the variance corresponding to position and posture respectively, In order to satisfy the global convergence of the system, the system should satisfy: Where h(x) is the proportional coefficient, P is an arbitrary symmetric positive definite matrix, the symbol '<0' indicates the negative definiteness of the matrix, and the parameterized quadratic Lyapunov function Ω(x) = (xx * ) T P(xx * ) Verify the global convergence of the linear system, learn and optimize the unknown parameters A and b of the linear parameter system, so that the robot humanoid grasping skill model can achieve robot humanoid precise grasping in an unstructured environment.

2. The scalable robot humanoid precision grasping skill learning method according to claim 1 is characterized in that: Obtain the robot's motion trajectory imitating the demonstrator's grasp, including: In a given unstructured task environment, the robot is set to gravity compensation mode. The demonstrator conducts physical human-machine interaction with the robot and performs multiple traction demonstrations and teachings on humanoid grasping skills. During the demonstration, the robot demonstration data is collected and the collected information is preprocessed to obtain multiple demonstration and teaching trajectories of the robot's position and posture in the task space. The multiple demonstration and teaching trajectories are determined as the motion trajectory of the robot imitating the demonstrator's grasping. Where x t,m represents the position at time t in the mth demonstration, q t,m (x q ,y q ,z q ,w) represents the posture at time t in the mth demonstration, x q is the vector part in the x-axis direction, y q is the vector part in the y-axis direction, z q is the vector part in the z-axis direction, w is the scalar part, which is related to the angle of rotation, T is the amount of data for each demonstration trajectory after alignment, and M is the total number of demonstrations.

3. The scalable robot humanoid precision grasping skill learning method according to claim 1 is characterized in that: The robot grasping skill model is established according to the motion trajectory of the grasping imitating the demonstrator, specifically including: Convert posture information; Where R represents the rotation matrix, Ω(α, β, γ) represents the Euler angle form of the posture, where α, β, and γ represent the angles of rotation around the x-axis, y-axis, and z-axis respectively; For a given reference trajectory x n ,Ω n is the position information and posture information corresponding to the nth point. There is a functional relationship between position and posture. For a given input x * The corresponding output function value f(x * ) and the output posture Ω is expressed as: Where, k * =[k(x * ,x1)k(x * ,x2)…k(x * ,x N )], k(.,.) represents the kernel function, K is the matrix containing the kernel function, Ν(·) represents the normal distribution, σ 2 is the variance of the noise; According to the joint probability distribution in the above formula and the conditional probability distribution of multivariate Gaussian, we can obtain Ρ(f(x * )|Ω), Ρ(·) represents the probability distribution, and its mean and variance can be expressed as: μ(x)=k*(K+σ 2 I) -1 y ∑=k(x * ,x * )-k * (K+σ 2 I) -1 k *T 。 4. The scalable robot humanoid precision grasping skill learning method according to claim 1 is characterized in that: The parameters of the robot grasping skill model are learned and optimized to obtain a grasping reference trajectory guided by the demonstrator's experience and knowledge, specifically including: By introducing the demonstrator’s experience knowledge, a reference trajectory can be obtained for a given input. They represent the mean and variance corresponding to the nth point in the parameter trajectory, respectively, and the following parameterized model is used: ξ(x)=Φ T (x)ω In the formula, ξ(x) is the parameter model to be sought, and the weight ω obeys the normal distribution for: Where, Represents the B-dimensional basis function vector, minimizing the KL divergence between the probability distribution of the generated trajectory and the reference trajectory, that is: Where, Decompose the objective function to obtain the mean and variance after parameter optimization: E(ξ(x * ))=k * (K+λ1Σ) -1 m Where λ1>0,λ2>0 are regularization coefficients, 5. A scalable robot humanoid precision grasping skill learning system, characterized by: include: A data acquisition module, which is used to obtain the motion trajectory of the robot imitating the demonstrator's grasping; A model building module, which is used to build a robot grasping skill model based on the motion trajectory of the imitation demonstrator's grasping; An experience guidance module, which is used to learn and optimize the parameters of the robot's grasping skill model to obtain a grasping reference trajectory guided by the demonstrator's experience knowledge; A global convergence module, which is used to convert the grasping reference trajectory into a linear dynamic system and optimize the parameters of the linear dynamic system to achieve global convergence of the robot's grasping skill; A rapid expansion module, which is used to reconstruct the robot grasping skill model using intermediate points to achieve rapid expansion of the robot grasping skill; In the global convergence module, the grasping reference trajectory is converted into a linear dynamic system, and the parameters of the linear dynamic system are optimized to achieve the global convergence of the robot's grasping skills, specifically including: The nonlinear grasping reference trajectory is converted into a linear dynamic system, which is expressed as: Ω=f(x)=h(x)(Ax+b) Where A = D xq (D x ) -1 ,b=E q -AE x , E x ,E Ω Represent the mean values ​​of position and posture respectively, D x ,D Ω Represent the variance corresponding to position and posture respectively, In order to satisfy the global convergence of the system, the system should satisfy: Where h(x) is the proportional coefficient, P is an arbitrary symmetric positive definite matrix, the symbol '<0' indicates the negative definiteness of the matrix, and the parameterized quadratic Lyapunov function Ω(x) = (xx * ) T P(xx * ) Verify the global convergence of the linear system and learn and optimize the unknown parameters A and b of the linear parameter system, so that the robot humanoid grasping skill model can achieve humanoid precise grasping in an unstructured environment; In the rapid expansion module, the robot grasping skill model is reconstructed using the intermediate points to achieve rapid expansion of the robot grasping skill, specifically including: Set the middle point to (x0, q0), and change the posture of the middle point to R*V(x0, q0)=V′(x v ,Ω v ), and then the transformed intermediate point is introduced into the original robot grasping skill model for reconstruction. The mean and variance of the new model parameters after optimization are: Then, through global stability strategy and Lyapunov function parameter optimization, the global stability of the reconstructed grasping skill model is achieved, thereby realizing the rapid expansion of grasping skills for different objects, multiple tasks, and cross-scenarios.

Images