Robot massage learning and control method and system based on dynamic motion primitives
By combining RGB-D image segmentation and multi-track learning methods with multi-layer convolutional neural networks, along with admittance control strategies, the limitations of single-track learning in robotic massage are overcome, enabling efficient, smooth, and natural interaction in the robotic massage process.
Patent Information
- Application Number
- CN202311452843.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-03
- Publication Date
- 2026-05-15
- Estimated Expiration
- 2043-11-03
AI Technical Summary
Existing dynamic motion primitive technology mainly targets a single trajectory in robot massage learning, resulting in poor learning performance in diverse tasks or changing environments, and the interaction process between the robot and the human body is not smooth and natural enough.
A robot massage learning method based on dynamic motion primitives is adopted, which combines RGB-D image segmentation, multi-layer convolutional neural network, multi-trajectory learning and admittance control strategy to adjust the robot massage trajectory through real-time feedback and achieve adaptive control.
This improves the learning efficiency and smoothness of robotic massage in interacting with the human body, ensuring the naturalness and safety of the massage process.
Smart Images
Figure CN117532601B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of robotic arm trajectory planning technology, specifically relating to a robot massage learning and control method and system based on dynamic motion primitives. Background Technology
[0002] In recent years, research on the physical interaction between robots and the human body has continued to attract attention from the scientific community, gradually permeating into daily life, such as in areas like household assistance, health care, and therapeutic aids. Among these areas, teaching robots to perform massage movements through instruction on massage trajectories is increasingly becoming a challenging and practical research direction.
[0003] Traditional dynamic motion primitives technology has been widely used in the field of robot trajectory learning. Its core idea is to decompose an action or trajectory into a series of basic motion units, thus providing robots with a simple and modular learning method. However, a significant limitation of dynamic motion primitives is that it primarily targets the learning of single trajectories, meaning that the learning effect largely depends on the quality of the taught trajectory. When faced with diverse tasks or changing environments, a single taught trajectory may not provide sufficient information, resulting in poor learning outcomes.
[0004] Robot-environment interaction technology has become a hot topic in robotics research in recent years. Through this technology, robots can perceive and adapt to various changes in the external environment in real time. The application of deep learning image segmentation technology provides robots with more refined visual perception capabilities, enabling them to accurately identify and locate the back area of the human body, ensuring the accuracy and efficiency of massage operations. Real-time force feedback information acquired by the robot during contact with the human body plays a crucial role in the robot's real-time control and adjustment, further improving the smoothness and naturalness of its interaction with the human body. Summary of the Invention
[0005] This invention aims to address the shortcomings of existing technologies by proposing a robotic massage learning and control method and system based on dynamic motion primitives. This enables the robot to learn and optimize massage movements more efficiently, and makes the interaction between the massage process and the human body smoother and more natural.
[0006] To achieve the above objectives, the present invention provides the following solution:
[0007] A robotic massage learning and control method based on dynamic motion primitives includes the following steps:
[0008] S1: Acquire an RGB-D image of the massage working environment, and segment and recognize the RGB-D image based on a trained neural network to obtain an image of the human skin region;
[0009] S2: Based on the multi-trajectory learning method, the optimal massage motion trajectory is obtained, and a new complete massage motion trajectory is obtained by generalizing the current environmental information and the improved dynamic motion primitive model; the trajectory is adjusted based on the new complete massage motion trajectory and the human skin area image.
[0010] S3: Based on the adjusted trajectory and admittance control strategy, provide real-time feedback control to the robot so that the robot can perform massage tasks;
[0011] S4: During the execution of the massage task, the robot's control parameters are adaptively adjusted to complete the robot massage operation skill learning and control based on dynamic motion primitives.
[0012] Preferably, in step S1, the neural network is a multi-layer convolutional neural network structure, which is divided into an encoder and a decoder. It adopts a U-shaped network architecture and introduces an attention mechanism and a long short-term memory network.
[0013] Preferably, in step S2, the method for obtaining the optimal massage motion trajectory is as follows:
[0014] Based on the instructor's repeated demonstrations of a massage action, several massage action trajectories are obtained;
[0015] A common starting point and ending point are set for the several massage motion trajectories, and the corresponding rotation matrix is obtained by calculating the several massage motion trajectories based on the Rodgers rotation formula.
[0016] An expansion coefficient is introduced, and based on the expansion coefficient and the rotation matrix, the rotation-expansion matrix corresponding to each massage action trajectory is obtained;
[0017] Multiply each of the massage motion trajectories by the corresponding rotation-expansion matrix to obtain several aligned trajectories;
[0018] The mean trajectory is obtained by averaging the values of several alignment trajectories.
[0019] Calculate the average of the differences between the mean trajectory and the aligned trajectory, and construct an objective function based on the average value;
[0020] Based on the objective function, the optimal weights are obtained;
[0021] Based on the optimal weights and the dynamic motion primitive model, the optimal massage motion trajectory is obtained.
[0022] Preferably, the formula for calculating the rotation matrix is:
[0023]
[0024]
[0025]
[0026] R=I+sin(θ)×[v]+(1-cos(θ))×[v] 2
[0027] In the formula, g and x0 are the starting and ending vectors of the original trajectory, g′ and x′0 are the set starting and ending vectors, I is the identity matrix, and g is a vector. The angle between the vectors is given by [v], where [v] is the skew-symmetric matrix of vector v; and R is the rotation matrix.
[0028] Preferably, the formula for calculating the optimal weight is:
[0029] ψ×ω=F
[0030] ψ i (s)=exp(-h i ·(sc i ) 2 )
[0031] U(t,:)=[ψ1(s(t)),ψ2(s(t)),…,ψ N (s(t))]
[0032] ω=F×U -1
[0033] In the formula, U is the design matrix, ψ is the basis function, ω is the optimal weight, and h i c is the width of the basis functions. i Let be the center of the basis functions, t be a given time point, s(t) be the corresponding normalized phase variable, N be the number of basis functions, and F be the objective function.
[0034] Preferably, in step S2, the expression of the improved dynamic motion primitive model is:
[0035]
[0036]
[0037] In the formula, τ is the time scaling factor, y, The position, velocity, and acceleration of the moving point; a y β y K is a constant, y0 is the initial state; R is the desired state, f is the nonlinear forcing term, and ψ is the initial state. i Let ω be a basis function. i denoted as the basis function weights, and N is the number of basis functions.
[0038] Preferably, in step S3, the equation for the admittance control strategy is:
[0039]
[0040] Where Fc is the contact force, M, B, and T are pre-set mass, damping, and stiffness coefficients, respectively, and X, B, and T are the contact forces. To determine the position, velocity, and acceleration based on the contact force.
[0041] The present invention also provides a robot massage learning and control system based on dynamic motion primitives. The system uses the method described above, which includes: a human skin image segmentation module, an operation skill learning module, an operation control module, and a parameter adjustment module.
[0042] The human skin image segmentation module is used to acquire RGB-D images of the massage working environment, and to segment and recognize the RGB-D images based on a trained neural network to obtain human skin region images;
[0043] The operational skill learning module is used to obtain the optimal massage action trajectory based on a multi-trajectory learning method, and to obtain a new complete massage action trajectory by combining the current environmental information and the improved dynamic motion primitive model; and to adjust the trajectory based on the new complete massage action trajectory and the human skin area image.
[0044] The operation control module is used to provide real-time feedback control to the robot based on the adjusted trajectory and admittance control strategy, so that the robot can perform massage tasks.
[0045] The parameter adjustment module is used to adaptively adjust the robot's control parameters during the execution of the massage task, thereby completing the robot's massage operation skills learning and control based on dynamic motion primitives.
[0046] Compared with existing technologies, the advantages of this invention are as follows: This invention combines image segmentation technology, deep neural network learning, and real-time feedback control to provide an efficient, safe, and adaptive robotic massage method. It enables the robot to learn and optimize massage movements more efficiently, and the interaction between the massage process and the human body becomes smoother and more natural. Attached Figure Description
[0047] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0048] Figure 1This is a schematic diagram of the structure of the robot massage learning and control system based on dynamic motion primitives according to an embodiment of the present invention;
[0049] Figure 2 This is a flowchart of a robot massage learning and control method based on dynamic motion primitives according to an embodiment of the present invention;
[0050] Figure 3 This is a flowchart illustrating the multi-track learning process according to an embodiment of the present invention.
[0051] Figure 4 This is a schematic diagram illustrating parameter adjustment according to an embodiment of the present invention;
[0052] Figure 5 This is a schematic diagram of the neural network structure according to an embodiment of the present invention;
[0053] Figure 6 This is a schematic diagram of the massage trajectory adjustment according to an embodiment of the present invention. Detailed Implementation
[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0056] Example 1
[0057] like Figure 2 As shown, a robotic massage learning and control method based on dynamic motion primitives includes the following steps:
[0058] S1: The RealSense D435 depth camera acquires RGB-D images of the massage working environment in real time, and the trained neural network is used to segment and recognize the RGB-D images to obtain images of the human skin area (target area).
[0059] A further implementation method is, such as Figure 5 As shown, in step S1, the neural network is a multi-layer convolutional neural network structure, which is divided into an encoder and a decoder. It adopts a U-shaped network architecture and introduces an attention mechanism and a long short-term memory network to extract features and accurately segment the input image.
[0060] S2: Based on the multi-trajectory learning method, the optimal massage action trajectory is obtained, and a new complete massage action trajectory is obtained by generalizing the current environmental information and the improved dynamic motion primitive model; the trajectory is adjusted based on the new complete massage action trajectory and the human skin area image.
[0061] A further implementation method is, such as Figure 3 As shown, in step S2, the method for obtaining the optimal massage motion trajectory is as follows:
[0062] Based on the instructor's repeated demonstrations of a massage action, several massage action trajectories are obtained;
[0063] A common starting point and ending point are set for the several massage motion trajectories, and the corresponding rotation matrix is obtained by calculating the several massage motion trajectories based on the Rodgers rotation formula.
[0064] An expansion coefficient is introduced, and based on the expansion coefficient and the rotation matrix, the rotation-expansion matrix corresponding to each massage action trajectory is obtained;
[0065] Multiply each of the massage motion trajectories by the corresponding rotation-expansion matrix to obtain several aligned trajectories;
[0066] The mean trajectory is obtained by averaging the values of several alignment trajectories.
[0067] Calculate the average of the differences between the mean trajectory and the aligned trajectory, and construct an objective function based on the average value;
[0068] Based on the objective function, the optimal weights are obtained; specifically, the basic parameters of the DMP dynamic motion primitives are set, the loss function is calculated, and then the optimal weights are solved.
[0069] Based on the optimal weights and the dynamic motion primitive model, the optimal massage motion trajectory is obtained.
[0070] Specifically, the formula for calculating the rotation matrix is:
[0071]
[0072]
[0073]
[0074] R=I+sin(θ)×[v]+(1-cos(θ))×[v] 2
[0075] In the formula, g and x0 are the starting and ending vectors of the original trajectory, g′ and x′0 are the set starting and ending vectors, I is the identity matrix, and g is a vector. The angle between the vectors is given by [v], where [v] is the skew-symmetric matrix of vector v; and R is the rotation matrix.
[0076] Specifically, in order to keep the lengths of multiple motion trajectories the same, an expansion coefficient λ is introduced:
[0077]
[0078] The rotation-expansion matrix S for each motion trajectory is obtained by combining the rotation matrix R and the expansion coefficient λ, and is specifically defined as follows:
[0079] S=λR
[0080] Specifically, after all trajectories are aligned, the mean trajectory is calculated. Then, the difference between the mean trajectory and the aligned trajectories is calculated, and the average of these differences is used to construct the objective function F.
[0081]
[0082] Where Y is the aligned input trajectory, X is the mean trajectory, m is the number of teaching trajectories, and i represents the i-th trajectory;
[0083] A further implementation method involves using the following formula to calculate the optimal weight:
[0084] ψ×ω=F
[0085] ψ i (s)=exp(-h i ·(sc i ) 2 )
[0086] U(t,:)=[ψ1(s(t)),ψ2(s(t)),…,ψ N (s(t))]
[0087] ω=F×U -1
[0088] In the formula, U is the design matrix, ψ is the basis function, ω is the optimal weight, and h i c is the width of the basis functions. i Let be the center of the basis functions, t be a given time point, s(t) be the corresponding normalized phase variable, N be the number of basis functions, and F be the objective function. The design matrix is a matrix composed of multiple basis functions.
[0089] A further implementation method is that, in step S2, the expression of the improved dynamic motion primitive model is:
[0090]
[0091]
[0092] In the formula, τ is the time scaling factor, y, The position, velocity, and acceleration of the moving point; a y β y K is a constant, y0 is the initial state; R is the desired state, f is the nonlinear forcing term, and ψ is the initial state. i Let ω be a basis function. i represents the weights of the basis functions, and N is the number of basis functions. The start and end positions of the optimal massage motion trajectory (new trajectory) are input into the model for generalization to obtain a new, complete massage motion trajectory.
[0093] S3: Based on the adjusted trajectory and admittance control strategy, real-time feedback control is provided to the robot, enabling it to perform massage tasks. After trajectory generalization, accuracy is assessed by combining the obtained images of the human skin region. Specifically, the position coordinates of the trajectory points are analyzed to determine whether all trajectory points are completely within the target area.
[0094] A further implementation method is that, in step S3, the equation for the admittance control strategy is:
[0095]
[0096] Where Fc is the contact force, M, B, and T are pre-set mass, damping, and stiffness coefficients, respectively, and X, B, and T are the contact forces. To determine the position, velocity, and acceleration based on the contact force.
[0097] When the trajectory points in the generalized new massage trajectory are not entirely located within the target area, such as... Figure 6 Two strategies will be considered. First, trajectory points outside the target area can be directly deleted to ensure the new trajectory is entirely within the target area. Second, new trajectory points can be added at necessary locations. After adding a new trajectory point, its coordinates are mapped onto the taught trajectory as an intermediate point connecting the start and end points. The trajectory portion from the original start point to the new trajectory point is generated independently using dynamic motion primitives. Similarly, the portion from the new trajectory point to the original end point is also generated. This ensures the entire trajectory is constrained within the target area.
[0098] If adding a single point is insufficient to ensure the completeness and accuracy of the massage trajectory, more points can be added following the strategy described above. This flexible strategy not only ensures that the massage trajectory is entirely within the target area but also allows for appropriate adjustments based on the actual situation, thereby guaranteeing both the effectiveness and safety of the massage.
[0099] S4: Adaptively adjust the robot's control parameters during the massage task to complete the learning and control of robot massage operation skills based on dynamic motion primitives.
[0100] like Figure 4 As shown, during the massage operation, the robot converts the difference between its real-time position and the predetermined desired position at each time step into a velocity vector, which consists of three components: V in the horizontal direction. x and V y and V in the vertical direction z These three velocity components represent the robot's velocities in the X, Y, and Z directions of its coordinate system, respectively.
[0101] Based on the stiffness coefficient in the admittance control strategy, the specific parameter adjustment equation is as follows:
[0102]
[0103] T = T0 + α*δ
[0104] Where δ represents the force plus the reciprocal of a constant; ε is a tiny coefficient to ensure the denominator is not zero; T0 and α are predefined constants; and f... c T represents the real-time contact force between the robot and the human body, and T is the stiffness coefficient of the admittance control strategy.
[0105] Because the unevenness of the back causes the contact force between the robot and the back to be inconsistent during the massage process, an adaptive admittance controller (speed / force controller) is used to input the contact force between the robot and the skin during the massage. This controller adjusts the vertical speed in real time to adapt to changes in the external environment, ensuring high-precision movement of the robot.
[0106] Example 2
[0107] like Figure 1 As shown, the present invention also provides a robot massage learning and control system based on dynamic motion primitives. The method of the system application includes: a human skin image segmentation module, an operation skill learning module, an operation control module, and a parameter adjustment module.
[0108] The human skin image segmentation module is used to acquire RGB-D images of the massage work environment and to segment and recognize the RGB-D images based on a trained neural network to obtain images of human skin regions.
[0109] The operation skill learning module is used to obtain the optimal massage action trajectory based on the multi-trajectory learning method, and to obtain a new complete massage action trajectory by combining the current environmental information and the improved dynamic motion primitive model; the trajectory is adjusted based on the new complete massage action trajectory and the human skin area image.
[0110] The operation control module is used to provide real-time feedback control to the robot based on the adjusted trajectory and admittance control strategy, so that the robot can perform massage tasks.
[0111] The parameter adjustment module is used to adaptively adjust the robot's control parameters during the massage task, and to complete the robot's massage operation skills learning and control based on dynamic motion primitives.
[0112] A further implementation method is that, in the operational skills learning module, the method for obtaining the optimal massage motion trajectory is as follows:
[0113] Based on the instructor's repeated demonstrations of a massage action, several massage action trajectories are obtained;
[0114] A common starting point and ending point are set for the several massage motion trajectories, and the corresponding rotation matrix is obtained by calculating the several massage motion trajectories based on the Rodgers rotation formula.
[0115] An expansion coefficient is introduced, and based on the expansion coefficient and the rotation matrix, the rotation-expansion matrix corresponding to each massage action trajectory is obtained;
[0116] Each massage motion trajectory is multiplied by the corresponding rotation-expansion matrix to obtain several aligned trajectories;
[0117] The mean trajectory is obtained by averaging the values of several alignment trajectories.
[0118] Calculate the average of the differences between the mean trajectory and the aligned trajectory, and construct an objective function based on the average value;
[0119] Based on the objective function, the optimal weights are obtained; specifically, the basic parameters of the DMP dynamic motion primitives are set, the loss function is calculated, and then the optimal weights are solved.
[0120] Based on the optimal weights and the dynamic motion primitive model, the optimal massage motion trajectory is obtained.
[0121] Specifically, the formula for calculating the rotation matrix is:
[0122]
[0123]
[0124]
[0125] R=I+sin(θ)×[v]+(1-cos(θ))×[v] 2
[0126] In the formula, g and x0 are the starting and ending vectors of the original trajectory, g′ and x′0 are the set starting and ending vectors, I is the identity matrix, and g is a vector. The angle between the vectors is given by [v], where [v] is the skew-symmetric matrix of vector v; and R is the rotation matrix.
[0127] Specifically, in order to keep the lengths of multiple motion trajectories the same, an expansion coefficient λ is introduced:
[0128]
[0129] The rotation-expansion matrix S for each motion trajectory is obtained by combining the rotation matrix R and the expansion coefficient λ, and is specifically defined as follows:
[0130] S=λR
[0131] Specifically, after all trajectories are aligned, the mean trajectory is calculated. Then, the difference between the mean trajectory and the aligned trajectories is calculated, and the average of these differences is used to construct the objective function F.
[0132]
[0133] Where Y is the aligned input trajectory, X is the mean trajectory, m is the number of teaching trajectories, and i represents the i-th trajectory;
[0134] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A robotic massage learning and control method based on dynamic motion primitives, characterized in that, Includes the following steps: S1: Acquire an RGB-D image of the massage working environment, and segment and recognize the RGB-D image based on a trained neural network to obtain an image of the human skin region; S2: Based on the multi-trajectory learning method, the optimal massage action trajectory is obtained, and a new complete massage action trajectory is obtained by generalizing the current environmental information and the improved dynamic motion primitive model; the trajectory is adjusted based on the new complete massage action trajectory and the human skin area image. S3: Based on the adjusted trajectory and admittance control strategy, provide real-time feedback control to the robot so that the robot can perform massage tasks; S4: During the execution of the massage task, the robot's control parameters are adaptively adjusted to complete the robot massage operation skill learning and control based on dynamic motion primitives; In step S2, the expression for the improved dynamic motion primitive model is: In the formula, This is the time scaling factor. This refers to the position, velocity, and acceleration of the moving point; It is a constant. This is the initial state; For the desired state, It is a nonlinear forcing term. As basis functions, denoted as the basis function weights, and N is the number of basis functions.
2. The robot massage learning and control method based on dynamic motion primitives according to claim 1, characterized in that, In step S1, the neural network is a multi-layer convolutional neural network structure, which is divided into an encoder and a decoder. It adopts a U-shaped network architecture and introduces an attention mechanism and a long short-term memory network.
3. The robot massage learning and control method based on dynamic motion primitives according to claim 1, characterized in that, In step S2, the method for obtaining the optimal massage motion trajectory is as follows: Based on the instructor's repeated demonstrations of a massage action, several massage action trajectories are obtained; A common starting point and ending point are set for the several massage motion trajectories, and the corresponding rotation matrix is obtained by calculating the several massage motion trajectories based on the Rodgers rotation formula. An expansion coefficient is introduced, and based on the expansion coefficient and the rotation matrix, the rotation-expansion matrix corresponding to each massage action trajectory is obtained; Each massage motion trajectory is multiplied by the corresponding rotation-expansion matrix to obtain several aligned trajectories; The mean trajectory is obtained by averaging the values of several alignment trajectories. Calculate the average of the differences between the mean trajectory and the aligned trajectory, and construct an objective function based on the average value; Based on the objective function, the optimal weights are obtained; Based on the optimal weights and the dynamic motion primitive model, the optimal massage motion trajectory is obtained.
4. The robot massage learning and control method based on dynamic motion primitives according to claim 3, characterized in that, The formula for calculating the rotation matrix is: In the formula, and Let the start and end vectors of the original trajectory be the starting and ending vectors. and Given the start and end vectors, It is the identity matrix. For vectors , The included angle, For vectors A skew-symmetric matrix; R is a rotation matrix.
5. The robot massage learning and control method based on dynamic motion primitives according to claim 3, characterized in that, The formula for calculating the optimal weight is: In the formula, To design the matrix, As basis functions, For optimal weights, The width of the basis functions. As the center of the basis functions, For a given time point, For the corresponding normalized phase variables, Let F be the number of basis functions, and F be the objective function.
6. The robot massage learning and control method based on dynamic motion primitives according to claim 1, characterized in that, In step S3, the equation for the admittance control strategy is: in For contact force, , , For the pre-set mass, damping, and stiffness coefficients, , , To determine the position, velocity, and acceleration based on the contact force.
7. A robotic massage learning and control system based on dynamic motion primitives, wherein the system applies the method described in any one of claims 1 to 6, characterized in that, include: The system includes a human skin image segmentation module, an operation skills learning module, an operation control module, and a parameter adjustment module. The human skin image segmentation module is used to acquire RGB-D images of the massage working environment, and to segment and recognize the RGB-D images based on a trained neural network to obtain human skin region images; The operational skill learning module is used to obtain the optimal massage action trajectory based on a multi-trajectory learning method, and to obtain a new complete massage action trajectory by combining the current environmental information and the improved dynamic motion primitive model; and to adjust the trajectory based on the new complete massage action trajectory and the human skin area image. The operation control module is used to provide real-time feedback control to the robot based on the adjusted trajectory and admittance control strategy, so that the robot can perform massage tasks. The parameter adjustment module is used to adaptively adjust the robot's control parameters during the execution of the massage task, thereby completing the robot's massage operation skills learning and control based on dynamic motion primitives.