Robot polishing method based on posture identification and iterative learning

Abstract

The invention relates to the field of intelligent mechanical equipment, in particular to a robot polishing method based on posture identification and iterative learning, and solves the problem of poor generalization ability of a mechanical arm in the prior art. The teaching track is reconstructed; identifying a posture; when the posture of the workpiece changes, a mechanical arm of the robot calls the changed track information, and machining is conducted according to the track information; and carrying out contact force iterative learning. According to the method, the path of the mechanical arm, the tail end posture and the point cloud data of the workpiece are bound together, the point cloud information before and after the posture of the workpiece changes is compared, the posture change of the workpiece is obtained through point cloud registration, and new track information is called for machining; and the force sensor at the tail end records interaction force data, and the path is iteratively updated according to the set iterative learning algorithm updating rate, so that the interaction force reaches an expected value, and an accurate polishing task is completed.

Description

Technical Field

[0001] The invention relates to the field of intelligent mechanical equipment, and in particular to a robot polishing method based on posture recognition and iterative learning. Background Art

[0002] In recent years, with the advancement of technologies such as electronic hardware, image processing, and artificial intelligence, robotics has gradually been applied to many fields such as home services, medical rehabilitation, firefighting and disaster reduction. The application scope of traditional robotic arms has also been significantly expanded, and has begun to involve more complex processing, production, and testing. Path planning is the core operation of robotic arm motion control, and different path planning problems require different solutions. In addition to traditional direct programming methods, current path planning solutions mainly include teaching programming, visual guidance, and brain-computer interfaces. These solutions handle different tasks and are suitable for different working environments.

[0003] There are two main ways to use teaching programming in industrial robotic arms:

[0004] Online teaching: In this method, the technician guides and controls the movement of the robot arm, records the program points during the operation, and inserts the necessary robot arm commands to complete the program writing.

[0005] Offline teaching: The operator does not directly teach the operating robot arm, but programs in an offline programming system or simulation environment, generates a teaching program, and indirectly controls the robot arm's control cabinet through a PC.

[0006] The limitation of online teaching is that it is difficult to achieve accurate tracking of a specific task path due to the limitations of human vision and control capabilities. When the operator drags the robot arm along the target path, any slight jitter may cause the robot arm controller to record the wrong path point, thus affecting the tracking effect of the end of the robot arm on the target path. In addition, during the drag teaching process, it is impossible to ensure that the end of the robot arm maintains an accurate distance from the target path, especially in high-precision tasks such as welding and laser cutting. Simple online teaching is difficult to meet the requirements of these tasks.

[0007] Offline teaching requires the operator to set the robot arm's working path and other parameters in advance, which has certain advantages when performing single and repetitive operations. However, in scenarios where work objectives change frequently, offline teaching is relatively inefficient because the robot arm's control program needs to be constantly updated as the work objectives change.

[0008] However, the above teaching method is not applicable to complex situations. For example, after the posture of the same workpiece changes, the previously taught trajectory cannot be used directly and needs to be re-taught. This low-generalization algorithm is not suitable for changing task scenarios.

[0009] There is an urgent need for a new type of robotic arm control method that can solve the above problems. Summary of the invention

[0010] The present invention proposes a robot polishing method based on posture recognition and iterative learning, which solves the problem of poor generalization ability of the robot arm in the prior art.

[0011] The technical solution of the present invention is implemented as follows: a robot polishing method based on posture recognition and iterative learning comprises the following steps:

[0012] 1. Teaching trajectory reconstruction: The human partner teaches the polishing path of the current workpiece multiple times, reconstructs the teaching trajectory through dynamic time warping and Gaussian mixture model, and the depth camera mounted on the end of the robot's mechanical arm captures the point cloud data of the workpiece and sends it to the information processing module of the microprocessor to generate processing trajectory information, and store the trajectory information and point cloud data in the storage module;

[0013] 2. Posture recognition: The point cloud data of the workpiece is captured by the depth camera and sent to the microprocessor, which is then saved in the storage module for storage. The point cloud is registered through the information processing module to determine whether the workpiece has undergone a posture change, and the trajectory information generated after the posture change is saved;

[0014] 3. When the posture of the workpiece does not change, the robot's mechanical arm calls the trajectory information of the storage module and performs processing according to the trajectory information;

[0015] 4. When the posture of the workpiece changes, the robot's mechanical arm calls the trajectory information after the posture change in step 2 and performs processing according to the trajectory information;

[0016] 5. Iterative learning of contact force: The force sensor installed at the end of the robot arm sends the contact force information during the processing process to the information processing module; the information processing module iteratively updates the processing path according to the set iterative learning update rate and sends it to the robot arm end, and the cycle continues until the contact force meets the expected contact force.

[0017] The iterative learning update rate is designed as follows:

[0018]

[0019] The initial trajectory X r,0 =W′, the transformed trajectory is the initial value; i represents the number of iterations, L represents the iteration update rate, η i is the variable weight factor, F d represents the desired contact force, F e,i represents the real interaction force between the end of the robot and the workpiece at the i-th iteration, K erepresents the ambient stiffness and can be calculated by Hooke's law in two iterations.

[0020] The step 2 is specifically:

[0021] First, calibrate the depth camera to obtain the internal parameters of the depth camera and perform hand-eye calibration on the camera and the robotic arm to obtain the transformation relationship between the robotic arm and the camera; obtain two sets of point cloud data: point cloud data before and after posture transformation; define a target optimization function as follows:

[0022]

[0023] where p s and p t They represent the point cloud before and after posture transformation, R and t are the rotation matrix R and translation matrix t required between two point cloud lines, N is the total number of point clouds, and F(t) is the target optimization function;

[0024] Through the above two matrices, the trajectory W′=RW+t under the current posture is calculated, where W is the teaching trajectory in step 1. After obtaining the trajectory, the information is sent to the robot side through the software.

[0025] The dynamic model of the closed-loop system of the robotic arm controller in the spatial domain is:

[0026]

[0027] Among them, M d , C d , K d are the desired moment of inertia, damping matrix and stiffness matrix respectively; is the speed of the robot in the direction of motion; represents the spatial differential operator; E(s) is the tracking error vector; F h (s) is the interaction force.

[0028] The mechanical arm is provided with a damping controller:

[0029]

[0030] Where X r is the reference path of the robot, F d is the desired contact force, K is the equivalent stiffness parameter;

[0031] According to the robot-environment contact force model, considering the i-th iteration of path update, we have:

[0032] F e,i′ =K e (X i′ -X0)

[0033] where F e,i′ is the contact force, X i′ is the position of the end effector, K e and X0 are the stiffness and surface position of the environment respectively; considering two consecutive iterations, the stiffness of the environment is obtained:

[0034]

[0035] The DTW-ILC method is used to design the reference path update rule:

[0036]

[0037] The initial reference path is set to X r,0 =X r,p ,z is the unit matrix representing the path update direction, K e is the estimated equivalent stiffness, is the contact force; Q represents a discrete-time low-pass filter.

[0038] The expression of Gaussian mixture model to reconstruct the teaching trajectory is:

[0039]

[0040] Where N(ξ j |u k ,∑k) is the probability function of the Gaussian mixture model, {α k ,∑k} are Gaussian mixture model parameters.

[0041] The present invention discloses a robot polishing method based on posture recognition and iterative learning. The method obtains the initial path and terminal posture of the workpiece to be polished based on teaching programming, obtains the point cloud data of the workpiece using a depth camera, and binds the robot arm path with the terminal posture and the point cloud data of the workpiece; after the posture of the workpiece changes, the point cloud information of the changed workpiece is obtained using a depth camera, and the posture change of the workpiece is obtained through point cloud registration; after the posture change is obtained, the previous teaching trajectory is retrieved for polishing, and the force sensor at the end records the interactive force data, and the path is iteratively updated according to the set iterative learning algorithm update rate, so that the interactive force reaches the expected value and completes the precise polishing task. It avoids the problem that the previous trajectory skill information cannot be reused due to the change of the workpiece posture, and introduces iterative learning in the polishing task, which enhances the control of the contact force and improves the accuracy of polishing. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1 : System module composition diagram;

[0044] Figure 2 : Control algorithm system flow chart;

[0045] Figure 3 : The teaching process uses dynamic time warping and Gaussian mixture model to reconstruct the before and after comparison images;

[0046] Figure 4 : Comparison of posture recognition and generalized trajectory information;

[0047] Figure 5 : Force tracking performance of the robot during iteration;

[0048] Figure 6 : Control framework for position-based impedance control method. DETAILED DESCRIPTION

[0049] The following will be combined with the 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 described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The present invention discloses a robot polishing method based on posture recognition and iterative learning, comprising the following steps:

[0051] 1. Teaching trajectory reconstruction: The human partner teaches the polishing path under the current workpiece multiple times, and reconstructs the teaching trajectory through dynamic time warping and Gaussian mixture model to reduce the instability of the human partner's teaching. The depth camera at the end of the robot's mechanical arm captures the point cloud data of the workpiece and sends it to the information processing module of the microprocessor to generate processing trajectory information, and store the trajectory information and point cloud data in the storage module;

[0052] 2. Posture recognition: The point cloud data of the workpiece is captured by the depth camera and sent to the microprocessor, which is then saved in the storage module for storage. The point cloud is registered through the information processing module to determine whether the workpiece has undergone a posture change, and the trajectory information generated after the posture change is saved;

[0053] 3. When the posture of the workpiece does not change, the robot's mechanical arm calls the trajectory information of the storage module and performs processing according to the trajectory information;

[0054] 4. When the posture of the workpiece changes, the previous teaching trajectory will not be applicable to the new posture. The robot's mechanical arm calls the trajectory information after the posture change in step 2 and performs processing according to the trajectory information;

[0055] 5. Iterative learning of contact force: The force sensor installed at the end of the robot arm sends the contact force information during the processing to the information processing module; the information processing module iteratively updates the processing path according to the set iterative learning update rate and sends it to the end of the robot arm, and the cycle continues until the contact force meets the expected contact force. Due to the instability of human partner teaching and the error of posture recognition, the interaction force between the end of the robot arm and the workpiece during the actual processing is not necessarily the expected contact force, and accurate contact force is the prerequisite for ensuring the polishing effect;

[0056] The iterative learning update rate is designed as follows:

[0057]

[0058] The initial trajectory X r,0 =W′, the transformed trajectory is the initial value; i represents the number of iterations, L represents the iteration update rate, η i is the variable weight factor, F d represents the desired contact force, F e,i represents the real interaction force between the end of the robot and the workpiece at the i-th iteration, K e represents the environmental stiffness, which can be calculated by Hu Ke's law during the two iterations; according to the above iteration update rate design, the trajectory X that satisfies the expected interaction force is obtained r,i+1 .

[0059] The step 2 is specifically:

[0060] First, calibrate the depth camera to obtain the internal parameters of the depth camera and perform hand-eye calibration on the camera and the robotic arm to obtain the transformation relationship between the robotic arm and the camera; obtain two sets of point cloud data: point cloud data before and after posture transformation; define a target optimization function as follows:

[0061]

[0062] where p s and p tThey represent the point cloud before and after the posture transformation, R and t are the rotation matrix R and translation matrix t required between two point clouds, N is the total number of point clouds, and F(t) is the target optimization function. Ideally, when F(t) = 0, the ideal rotation matrix and translation matrix are found. In practical applications, when F(t) is the global minimum, the optimal rotation matrix R and translation matrix t can be obtained.

[0063] Through the above two matrices, the trajectory W′=RW+t under the current posture is calculated, where W is the teaching trajectory in step 1. After obtaining the trajectory, the information is sent to the robot side through the software.

[0064] The dynamic model of the closed-loop system of the robotic arm controller in the spatial domain is:

[0065]

[0066] Among them, M d , C d , K d are the desired moment of inertia, damping matrix and stiffness matrix respectively; is the speed of the robot in the direction of motion; represents the spatial differential operator; E(s) is the tracking error vector; F h (s) is the interaction force.

[0067] The mechanical arm is provided with a damping controller:

[0068]

[0069] Where X r is the reference path of the robot, D d is the desired contact force, K is the equivalent stiffness parameter;

[0070] According to the robot-environment contact force model, considering the i-th iteration of path update, we have:

[0071] F e,i′ =K e (X i′ -X0)

[0072] where F e,i′ is the contact force, X i′ is the position of the end effector, K e and X0 are the stiffness and surface position of the environment respectively; considering two consecutive iterations, the stiffness of the environment is obtained:

[0073]

[0074] In order to improve the contact force tracking performance, the DTW-ILC method is used to design the reference path update law by combining the impedance control model and the environmental stiffness estimation:

[0075]

[0076] The initial reference path is set to X r,0 =X r,p ,z is the unit matrix representing the path update direction, K e is the estimated equivalent stiffness, is the contact force; Q represents a discrete time low-pass filter. Based on the proposed path learning method and GMM / GMR method, the robot's desired path and posture can be learned. Therefore, the designed ILC update law is used to update the reference path only in the axial direction of the end effector of the robot arm, thereby achieving compensation for the force tracking error.

[0077] The expression of Gaussian mixture model to reconstruct the teaching trajectory is:

[0078]

[0079] Where N(ξ j |u k ,∑k) is the probability function of the Gaussian mixture model, {α k ,∑k} are Gaussian mixture model parameters.

[0080] The robot arm used in the experiment is a Sawyer 7-axis collaborative robot arm, the depth camera is an interrealsense D435i camera, and the force sensor is a Robotiq FT300.

[0081] The human partner teaches the polishing path of the current workpiece multiple times. The software algorithm reconstructs the teaching trajectory through dynamic time warping and Gaussian mixture model to reduce the instability of the human partner's teaching. The depth camera mounted on the end of the robotic arm captures the point cloud data of the workpiece and sends it to the information processing module of the microprocessor to generate processing trajectory information, and stores the trajectory information and the point cloud data in the storage module; when the posture of the workpiece changes, the depth camera at the end of the robotic arm obtains the point cloud data of the workpiece in the current posture and sends it to the information processing module of the microprocessor to generate processing trajectory information, and stores the trajectory information and the point cloud data in the storage module; the optimal rotation matrix R and translation matrix t are obtained through point cloud registration, and the trajectory W′=R□+t in the current posture is calculated through the above two matrices, where W is the teaching trajectory of step one. After obtaining the trajectory, the information is sent to the robotic arm side through the software; after obtaining the trajectory information, the robotic arm side processes according to the trajectory information, and the force sensor at the end records the contact force information during the processing and sends it to the information processing module of the microprocessor. The machining path is iteratively updated according to the set iterative learning update rate and sent to the robot end, and the cycle is repeated until the contact force meets the expected contact force.

[0082] In step 1, the initial path of workpiece processing is obtained by multiple teachings of the human partner. First, the human partner drags the end of the robotic arm to teach according to the desired processing path three times. The robotic arm records the teaching trajectory and sends it to the information processing module of the microprocessor. The information processing module first aligns the experimental data of different experimental lengths using the dynamic time warping method. After the data is aligned, the Gaussian mixture model is used to reconstruct the three teaching data to reduce the instability of the human partner's teaching. The following formula is the expression of the Gaussian mixture model to reconstruct the teaching trajectory:

[0083]

[0084] Where N(ξ j |u k ,∑k) is the probability function of the Gaussian mixture model, {α k ,∑k} is the Gaussian mixture model parameter. Step 2 of the present invention is posture recognition and processing trajectory mapping based on point cloud data matching. After obtaining two sets of point cloud data ① and ②, a target optimization function is first defined as follows:

[0085]

[0086] where p s and p t Respectively represent point clouds ① and ②, R and t are the rotation matrix and translation matrix required between the two point clouds, N is the total number of point clouds, and F(t) is the target optimization function. Respectively represent the centroid of the original point cloud ① and the target point cloud ②, let Perform SVD decomposition on this 3×3 matrix and get H=U∑V T , then the optimal rotation matrix is ​​R = VU T , the optimal translation matrix is At this point, the optimal rotation matrix R and translation matrix t are obtained, and the new trajectory of the mapping is: W′=RW+t, where W is the teaching trajectory of step one.

[0087] Contact force adjustment, first of all, the underlying control of the present invention involves specifically:

[0088] First, the dynamic model of the system is introduced. The dynamic model of the robot arm in Cartesian space is as follows:

[0089]

[0090] in X(t) represents the acceleration, velocity and position of the end effector respectively; C x and G x They represent the Coriolis force matrix coefficient and the gravity term respectively; u(t) represents the joint torque of the robot arm, j represents the mapping relationship between the terminal velocity, angular velocity and joint angular velocity, that is, the Jacobian matrix; F h (t) represents the external force applied by the human partner and can be measured by a force sensor.

[0091] Let E = x(t)-x r (t), represents the expected trajectory x r The error between (t) and the true trajectory x(t) is Represent the velocity and acceleration errors respectively. Combining the dynamic model in Cartesian space, we can get:

[0092]

[0093] In the above formula, H e =H x , C e =C x , F(t)=J -T u(t).

[0094] The controller of the robotic arm is as follows:

[0095]

[0096] Where M d , C d , K d are the desired inertia matrix, damping matrix and stiffness matrix respectively.

[0097] Then we can get the expression of the closed-loop dynamic system in the time domain:

[0098]

[0099] So far, we have obtained the closed-loop system's interaction force F h The expressions of tracking error E(t) and its derivatives in time domain are given below. The specific control method will be given in combination with the actual robot operation environment.

[0100] The impedance control model, which is commonly used in robotic arm control, was proposed by Hogan. It approximately simulates the process of the robot's contact with the outside world into a second-order system of mass-spring-damper. In order to achieve ideal control performance, the stiffness information of the environment and the surface position information of the environment in the robot's base coordinate system need to be specified in advance before the actual operation of the robot. Then, the dynamic characteristics of the robot when in contact with the environment can be adjusted by adjusting the inertia, damping and stiffness parameters in the impedance model, thereby indirectly achieving control of the contact force.

[0101] like Figure 1 As shown, the entire system module composition of the present invention is specifically described. It mainly includes a visual system, namely a depth camera, a robotic arm workbench, a control algorithm system and a robotic arm body. The visual system is specifically a depth camera; the robotic arm workbench is specifically a human-computer interaction interface, a microprocessor or a server computer; the control algorithm system includes a coordinate generalization module, a motor control module, and an inverse kinematics module, namely, the dynamic model of the closed-loop system of the robotic arm controller in the spatial domain; the relationship between the four parts is explained in detail. Among them, the visual system is responsible for collecting point cloud data, the control algorithm system is responsible for processing point cloud information and controlling the movement of the robotic arm, the robotic arm workbench is responsible for displaying point cloud data and human-computer interaction, and the robotic arm body is responsible for polishing the workpiece.

[0102] like Figure 2 As shown in the figure, the specific process of the control algorithm system is described in detail, which is divided into two parts. The first part is responsible for teaching, GMM / GMR teaching trajectory reconstruction and initial point cloud acquisition. The second part is to calculate the initial path using point cloud information after the workpiece posture is transformed, and to calculate the final path using iterative learning, combined with the inverse kinematics of the robot arm and motor control to achieve the final grinding / polishing task.

[0103] like Figure 3 The comparison chart before and after the teaching process using dynamic time warping and Gaussian mixture model reconstruction is shown in the figure. The red curve represents the experimental results of GMM / GMR reconstruction through the teaching process, and the other three curves represent the three teaching trajectory curves. It can be seen from the two figures that after GMM / GMR reconstruction, the trajectory data is more stable and smooth, and the robot arm can better represent the teaching intention when executing.

[0104] like Figure 4As shown in the comparison chart of posture recognition and generalized trajectory information, the black curve represents the processing trajectory before the workpiece posture change and the teaching trajectory of step one, the red curve represents the generalized trajectory after posture recognition and the result of step two, and the blue curve represents the expected result after posture change. The error between the two is within 5 mm, which meets the polishing processing accuracy.

[0105] The polishing experiment was carried out under the same experimental conditions using the method proposed in the present invention, where the force tracking performance of the robot at the 0th, 2nd, 4th, and 7th iterations is as follows: Figure 3 As shown, despite the different polishing feed speeds (variable robot operating speed), using the proposed force tracking method, the robot can effectively track the required force as the iterations proceed, where the mean and standard deviation of the force tracking error are reduced to 0.18N and 0.12N, respectively, at the 7th iteration, meeting the required polishing force tracking requirements.

[0106] The present invention discloses a robot polishing method based on posture recognition and iterative learning. The method obtains the initial path and terminal posture of the workpiece to be polished based on teaching programming, obtains the point cloud data of the workpiece using a depth camera, and binds the robot arm path with the terminal posture and the point cloud data of the workpiece; after the posture of the workpiece changes, the point cloud information of the changed workpiece is obtained using a depth camera, and the posture change of the workpiece is obtained through point cloud registration; after the posture change is obtained, the previous teaching trajectory is retrieved for polishing, and the force sensor at the end records the interactive force data, and the path is iteratively updated according to the set iterative learning algorithm update rate, so that the interactive force reaches the expected value and completes the precise polishing task. It avoids the problem that the previous trajectory skill information cannot be reused due to the change of the workpiece posture, and introduces iterative learning in the polishing task, which enhances the control of the contact force and improves the accuracy of polishing.

[0107] Of course, without departing from the spirit and essence of the invention, technicians familiar with the field should be able to make various corresponding changes and modifications based on the invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the invention.

Claims

1. A robot polishing method based on posture recognition and iterative learning, characterized in that: The following steps are involved:

1. Teaching trajectory reconstruction: The human partner teaches the polishing path of the current workpiece multiple times, reconstructs the teaching trajectory through dynamic time warping and Gaussian mixture model, and the depth camera mounted on the end of the robot's mechanical arm captures the point cloud data of the workpiece and sends it to the information processing module of the microprocessor to generate processing trajectory information, and store the trajectory information and point cloud data in the storage module; 2. Posture recognition: The point cloud data of the workpiece is captured by the depth camera and sent to the microprocessor, which is then saved in the storage module for storage. The point cloud is registered through the information processing module to determine whether the workpiece has undergone a posture change, and the trajectory information generated after the posture change is saved; 3. When the posture of the workpiece does not change, the robot's mechanical arm calls the trajectory information of the storage module and performs processing according to the trajectory information; 4. When the posture of the workpiece changes, the previous teaching trajectory will not be applicable to the new posture. The robot's mechanical arm calls the trajectory information after the posture change in step 2 and performs processing according to the trajectory information; 5. Iterative learning of contact force: The force sensor installed at the end of the robot arm sends the contact force information during the processing process to the information processing module; the information processing module iteratively updates the processing path according to the set iterative learning update rate and sends it to the robot arm end, and the cycle continues until the contact force meets the expected contact force.

2. The robot polishing method based on posture recognition and iterative learning according to claim 1, characterized in that: The iterative learning update rate is designed as follows: The initial trajectory X r,0 =W′, the transformed trajectory is the initial value; i represents the number of iterations, L represents the iteration update rate, η u is the variable weight factor, F d represents the desired contact force, F e,i represents the real interaction force between the end of the robot and the workpiece at the i-th iteration, K e represents the ambient stiffness and can be calculated by Hooke's law in two iterations.

3. The robot polishing method based on posture recognition and iterative learning according to claim 2, characterized in that: The step 2 is specifically: First, calibrate the depth camera to obtain the internal parameters of the depth camera and perform hand-eye calibration on the camera and the robotic arm to obtain the transformation relationship between the robotic arm and the camera; obtain two sets of point cloud data: point cloud data before and after posture transformation; define a target optimization function as follows: where p s and p t They represent the point cloud before and after posture transformation, R and t are the rotation matrix R and translation matrix t required between two point cloud lines, N is the total number of point clouds, and F(t) is the target optimization function; Through the above two matrices, the trajectory W′=RW+t under the current posture is calculated, where W is the teaching trajectory in step 1. After obtaining the trajectory, the information is sent to the robot arm side through the software.

4. The robot polishing method based on posture recognition and iterative learning according to claim 3 is characterized in that: The dynamic model of the closed-loop system of the robotic arm controller in the spatial domain is: Among them, M d , C d , K d are the desired moment of inertia, damping matrix and stiffness matrix respectively; is the speed of the robot in the direction of motion; represents the spatial differential operator; E(s) is the tracking error vector; F h (s) is the interaction force.

5. The robot polishing method based on posture recognition and iterative learning according to claim 4, characterized in that: The mechanical arm is provided with a damping controller: Where X r is the reference path of the robot, F d is the desired contact force, K is the equivalent stiffness parameter; According to the robot-environment contact force model, considering the i-th iteration of path update, we have: F e,i′ =K e (X i′ -X0) where F e,i′ is the contact force, X i′ is the position of the end effector, K e and X0 are the stiffness and surface position of the environment respectively; considering two consecutive iterations, the stiffness of the environment is obtained: The DTW-ILC method is used to design the reference path update rule: The initial reference path is set to X r,0 =X r,p ,z is the unit matrix representing the path update direction, K e is the estimated equivalent stiffness, is the contact force; Q represents a discrete-time low-pass filter.

6. The robot polishing method based on posture recognition and iterative learning according to claim 5, characterized in that: The expression of Gaussian mixture model to reconstruct the teaching trajectory is: Where N(ξ j |u k ,∑k) is the probability function of the Gaussian mixture model, {α k ,∑k} are Gaussian mixture model parameters.

Images