A singular-avoiding mechanical arm anisotropic damping inverse kinematics control method

Abstract

The application discloses a kind of anisotropic damping inverse kinematics control method for singular-avoiding mechanical arm.The method of the present application constructs state observation vector by splicing position error, attitude error, Jacobian singular value characteristics and current joint angle;The state observation vector is input into the policy network of the trained reinforcement learning model to output the damping action vector;The damping action vector is mapped into the damping parameter vector, and a diagonal damping matrix is constructed;According to the condition number of Jacobian matrix and the diagonal damping matrix, the joint speed at the current time is obtained by gating inverse solution of cartesian expected speed, and the joint angle to be executed at the current time is obtained by limiting amplitude and filtering processing in turn.The technical problems that fixed damping cannot consider different singular degree, unified scalar cannot distinguish the difference between axes, and analytical parameter adjustment has poor adaptability and depends on artificial in the prior art are solved.

Description

Technical Field

[0001] This invention belongs to the field of robot motion control, and specifically relates to an anisotropic damped inverse kinematics control method for a robotic arm oriented towards singularity avoidance. Background Technology

[0002] In industrial robots, assembly robots, inspection robots, and service robots, end effectors often need to continuously track the target pose in Cartesian space. Such tasks typically rely on the Jacobian matrix to map Cartesian velocities to joint velocities. However, the Jacobian matrix can become ill-conditioned or even rank-deficient under certain poses, leading to inverse kinematics instability. In engineering practice, the following phenomena can be observed: when the end effector error is still within an acceptable range, individual joint velocities suddenly exhibit high-amplitude spikes; frequent oscillations in the control signal cause increased actuator temperature rise; and the trajectory jitters near the target, making stable convergence difficult.

[0003] To alleviate this problem, the traditional damped least squares method introduces a damping term in the inverse solution. Its basic idea is to trade numerical stability for a certain level of accuracy. This strategy is effective in some scenarios, but it still has three typical limitations. First, fixed damping cannot simultaneously cover the different needs of "far from the singularity" and "near the singularity": if the damping value is too large, it will over-suppress even when far from the singularity, causing slow response and decreased accuracy; if the damping value is too small, it cannot adequately suppress sudden velocity changes when near the singularity. Second, uniform scalar damping cannot express the risk differences between different task axes. In many robotic arm configurations, singular sensitivity is not uniform between the translation axis and the attitude axis; equal-strength damping will unnecessarily slow down directions that could otherwise move efficiently. Third, traditional analytical parameter tuning has limited adaptability to changes in operating conditions, and in actual deployment, repeated manual tuning of thresholds and damping parameters is often required. Summary of the Invention

[0004] To address the problems existing in the background technology, this invention provides an anisotropic damping inverse kinematics control method for a robotic arm oriented towards singularity avoidance. This method solves the technical problems in the prior art, such as fixed damping being unable to take into account different singularities, uniform scalars being unable to distinguish between axis differences, and poor adaptability of analytical parameter tuning and reliance on manual intervention.

[0005] The technical solution adopted in this invention is:

[0006] I. An anisotropic damped inverse kinematics control method for a robotic arm oriented towards singularity avoidance:

[0007] S1. Acquire the current joint angle, current end-effector pose, and target end-effector pose of the robotic arm.

[0008] S2. Calculate and obtain position error, attitude error, and Jacobian singular value features based on the current joint angle, current end pose, and target end pose. Then, concatenate the position error, attitude error, Jacobian singular value features, and current joint angle to construct the state observation vector.

[0009] S3. Input the state observation vector into the policy network of the trained reinforcement learning model to output the damping action vector.

[0010] S4. Map the damping action vector to a damping parameter vector and construct a diagonal damping matrix.

[0011] S5. Calculate the Cartesian desired velocity based on the position error and attitude error. Then, perform a gated inverse solution on the Cartesian desired velocity based on the Jacobian matrix condition number and the diagonal damping matrix to obtain the joint velocity at the current moment.

[0012] S6. The joint speed at the current moment is sequentially limited and filtered to obtain the joint angle to be executed at the current moment, and the robotic arm is driven to execute according to the joint angle.

[0013] S7. Repeat steps S1-S6 to achieve control of the robotic arm.

[0014] The Jacobian singular value features include the reciprocal condition number of the Jacobian matrix, the minimum singular value normalization, and the maneuverability normalization.

[0015] The Jacobian matrix under the current configuration is calculated based on the current joint angle and the kinematic model of the robotic arm. The singular value set is obtained by decomposing the Jacobian matrix into singular values, and then the minimum singular value and the maximum singular value are obtained. The minimum singular value is divided by the maximum singular value to obtain the reciprocal condition number of the Jacobian matrix.

[0016] The Yar minimum singular value normalization and the operability normalization are obtained by processing according to the following formula:

[0017] ;

[0018] in, Represents the minimum singular value normalization; This represents the normalized quantity of maneuverability; Represents the minimum singular value; Indicates an index; Represents the set of singular values. One singular value; This represents the cumulative product of all singular values ​​in the set of singular values.

[0019] The diagonal damping matrix is ​​set according to the following formula:

[0020] ;

[0021] in, The diagonal damping matrix; Represents a diagonal matrix; , , , , and These are the six parameters in the damping parameter vector; Indicates an index; For the damping parameter vector, the first One parameter; For the first One damping action vector; and These are the preset minimum and maximum values ​​of the range in which the damping action vector output by the policy network lies; and These are the preset minimum and maximum values ​​for the range of damping parameters, respectively.

[0022] The gated inverse solution is specifically as follows: when the condition number of the Jacobian matrix is ​​less than a preset threshold, the joint velocity calculated by the pseudo-inverse solution is used as the joint velocity of the gated inverse solution; when the condition number of the Jacobian matrix is ​​not less than the preset threshold, the joint velocity calculated by the damped least squares inverse solution is used as the joint velocity of the gated inverse solution.

[0023] The pseudo-inverse solution and the damped least squares inverse solution are set according to the following formulas:

[0024] ; ;

[0025] in, This represents the joint velocity obtained from the pseudo-inverse solution; This represents the joint velocity obtained from the inverse least-squares solution of the damping; It is a Jacobian matrix; For Jacobi in a generalized inverse; This is the transpose of the Jacobian matrix; Descartes' expected velocity; The diagonal damping matrix; , , , , and These are the six parameters in the damping parameter vector.

[0026] The process of sequentially limiting and filtering the joint velocity at the current moment to obtain the joint angle to be executed at the current moment is specifically as follows: the joint angle to be executed at the current moment is obtained based on the joint velocity obtained from the inverse gate solution at the current moment, the joint velocity actually executed at the previous moment, and the joint angle actually executed at the previous moment.

[0027] The joint angle to be executed at the current moment is obtained by processing it using the following formula:

[0028] ;

[0029] in, This refers to the joint angle that is actually to be executed at the current moment. This refers to the joint angle actually executed at the previous moment; The sampling period; The actual joint speed to be executed at the current moment; The joint velocity is obtained from the inverse gate solution at the current moment; This indicates the joint velocity that was actually executed at the previous moment.

[0030] The reward function of the reinforcement learning model is set according to the following formula:

[0031]

[0032]

[0033] in, To reinforce the reward function of the learning model; As a stability reward; To avoid rewards for unusual behavior; Rewards for maneuverability; To match the damping bonus; The actual joint speed to be executed at the current moment. The condition number is the inverse of the Jacobian matrix. It is the minimum singular value normalizer; This is a normalized measure of maneuverability. For indexing; The first in the damping parameter vector One parameter; For average damping; This refers to the risk factor. , , , , and All are weighting coefficients; , and All are small positive constants; and All of these are threshold parameters.

[0034] II. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0035] 3. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0036] The innovation of this invention lies in constructing an anisotropic diagonal damping matrix using six-dimensional anisotropic damping parameters. Based on the diagonal damping matrix, independent adjustment of the translation axis and attitude axis is achieved, which can improve damping to suppress sudden velocity changes in singular sensitive directions and maintain smaller damping in non-sensitive directions, thereby reducing unnecessary dynamic performance loss.

[0037] The innovation of this invention lies in the proposal of a dual-channel gated inverse solution mechanism, which enables the use of pseudo-inverse solution in the low-risk region to ensure response speed and tracking accuracy, and switches to damped least squares inverse solution in the high-risk region to improve numerical stability, thus avoiding the beneficial effects of increased steady-state error and slower response caused by "large damping throughout the process".

[0038] The beneficial effects of this invention are:

[0039] 1. The present invention addresses the problem that fixed scalar damping cannot distinguish directional risks. The present invention uses six-dimensional anisotropic damping parameters to construct an anisotropic diagonal damping matrix, which allows the translation axis and attitude axis to be adjusted independently. In singular sensitive directions, it can improve damping to suppress sudden velocity changes, and in non-sensitive directions, it can maintain a small damping and reduce unnecessary dynamic performance loss.

[0040] 2. The method of the present invention uses a dual-channel gated inverse solution mechanism for the condition number of the Jacobian matrix. In the low-risk region, a pseudo-inverse solution is used to ensure response speed and tracking accuracy. In the high-risk region, the method switches to a damped least squares inverse solution to improve numerical stability and avoid the increase in steady-state error and slow response caused by "large damping throughout the process".

[0041] 3. The method of this invention adopts the DDPG strategy generation method of offline training and online forward inference. In the online stage, no parameter updates are performed. Only the damping parameters are output to participate in the calculation of interpretable analytical inverse solution, which takes into account the adaptive capability of learning methods and the real-time performance and verifiability in engineering deployment. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention.

[0043] Figure 2The graph shows the changes of the six damping parameters over time as obtained in the example.

[0044] Figure 3 The graphs showing the change of joint angles over time are shown in the example.

[0045] Figure 4 This is a comparison curve of the attitude error between the method of this invention and the constant damping method. Detailed Implementation

[0046] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0047] Example:

[0048] like Figure 1 As shown, in a specific robotic arm end-effector pose tracking task, the anisotropic damped inverse kinematics control method of this embodiment is implemented according to the following steps:

[0049] S0. Establish the robotic arm control scenario and complete parameter initialization. Construct a kinematic model of a six-DOF articulated robotic arm, set joint limits, sampling period, and velocity threshold; define the target pose set and initial joint angles shared by training and online control.

[0050] In this embodiment, the control parameter is set to: sampling period Maximum joint velocity Conditional number gating threshold .

[0051] S1. Obtain the current joint angles of the six-DOF articulated robotic arm. According to the current joint angle The current end effector pose is calculated. Based on the current end pose With the target end pose The position errors were calculated separately. and attitude error .

[0052] Specifically, position error Attitude error is obtained by directly subtracting the end-effector position from the target position. It is obtained by relative transformation based on the rotation matrices of the end-effector attitude and the target attitude.

[0053] S2, based on the current joint angle Calculate the Jacobian matrix under the current configuration using the kinematic model of the robotic arm. By analyzing the Jacobian matrix The singular value decomposition yields the singular value set. Thus, the minimum singular value is obtained. and maximum singular value According to the minimum singular value and maximum singular value Obtain the condition number of the reciprocal of the Jacobian matrix .

[0054] S3, respectively based on the minimum singular value and singular value sets Obtain the minimum singularity normalization and the maneuverability normalization.

[0055] The minimum singularity normalization and the operability normalization are obtained by processing them according to the following formulas:

[0056]

[0057] in, Represents the minimum singular value normalization; This represents the normalized quantity of maneuverability; Represents the minimum singular value; Indicates an index; Represents the set of singular values The Middle One singular value; Represents the set of singular values Multiply all the singular values ​​in the product.

[0058] S4. Sequentially adjust the positional error (3D) Attitude Error (3D) Minimum Singular Value Normalization (1-dimensional) Jacobian matrix reciprocal condition number (1-dimensional) Normalized controllability (1D) and current joint angle (6-dimensional) splicing yields (15-dimensional) state observation vector.

[0059] Specifically, the state observation vector is defined by the following formula:

[0060]

[0061] in, For the three-dimensional end position error, This refers to the three-dimensional end-effector attitude error. It is the minimum singular value normalizer. The condition number is the inverse of the Jacobian matrix. This is a normalized measure of maneuverability. The current joint angle, , , Representing respectively , , The transpose of the matrix, This represents the transpose of the matrix.

[0062] S5. Input the state observation vector into the policy network of the trained reinforcement learning model for processing, and output the damping action vector. and the damping action vector Mapped to a positive damping parameter vector .

[0063] The damping parameter vector is obtained by mapping according to the following formula:

[0064]

[0065] in, Indicates an index; For the damping parameter vector, the first One parameter; For the first One damping action vector; and These are the preset minimum and maximum values ​​of the range in which the damping action vector output by the policy network lies; and These are the preset minimum and maximum values ​​for the range of damping parameters, respectively.

[0066] Furthermore, specifically:

[0067] Damped action vector and damping parameter vector Defined by the following formula:

[0068] ,

[0069] , ,

[0070] in, , , , , and These are the six parameters in the damping parameter vector; , , , , and Damped action vector The six parameters in it.

[0071] In this embodiment, take , , , .

[0072] The changes of the six parameters in the damping parameter vector obtained in this embodiment over time are as follows: Figure 2 As shown, the horizontal axis represents time, and the vertical axis represents the damping coefficient values ​​corresponding to the six parameters.

[0073] S5. A diagonal damping matrix is ​​constructed using multiple damping parameters. ,in This represents a diagonal matrix.

[0074] Specifically, the diagonal damping matrix The off-diagonal elements are zero. It is used to adjust the damping strength of each task axis separately.

[0075] The diagonal damping matrix The principle behind this setup is to utilize the anisotropic characteristics of reinforcement learning output to achieve independent dynamic adjustment of the six degrees of freedom in the task space.

[0076] Unlike the traditional damping least squares method, which uses only a single scalar damping coefficient The anisotropic diagonal damping matrix constructed in this invention is configured such that all task axes share the same damping strength. It has the following technical effects:

[0077] Direction-Sensitive Decoupling in Task Space: When a robotic arm approaches a singular pose, the degradation of the Jacobian matrix is ​​often directional. For example, when the robotic arm is fully extended, singularity may occur only in translational motion along the arm's span, while rotation around the end effector may remain flexible. Traditional scalar damping... This approach applies equal damping to all axes in a "one-size-fits-all" manner, causing unnecessarily sacrificed motion performance in insensitive directions. This invention addresses this by using a diagonal matrix... This makes the translation axis ( ) and attitude axis ( It can independently adjust the damping according to the singular risk in that direction, achieving the effect of "precisely suppressing sensitive directions and protecting non-sensitive directions".

[0078] Adaptive matching of physical dimensions: Since the end-effector translational velocity (unit: m / s) and rotational angular velocity (unit: rad / s) are completely different in numerical magnitude and physical units, traditional methods use a single scalar. It is difficult to simultaneously meet the stability requirements of both. This invention utilizes a diagonal matrix structure and reinforcement learning to automatically learn the damping ratio that adapts to different physical meanings of the axes, thus solving the limitation of "choosing one over the other" in scalar parameter tuning and significantly improving the trajectory tracking accuracy under composite motion.

[0079] Real-time reconstruction of anisotropic features: due to the diagonal damping matrix The parameters are dynamically generated from the current state observation vector, meaning that the anisotropic characteristics of the diagonal damping matrix are reconstructed in real time as the robot arm configuration changes. When moving away from the singularity region, Approaching the zero matrix to ensure accuracy; when entering the singular region, It can spontaneously evolve into an ellipsoidal damping distribution, guiding joint velocities to avoid ill-conditioned solutions, thus solving the problem that traditional methods are difficult to handle in complex working conditions.

[0080] S6. Based on position error and attitude error Calculate Descartes' expected velocity According to the minimum singular value and maximum singular value Obtain the condition number of the Jacobian matrix .

[0081] S7. Based on the condition number of the Jacobian matrix Descartes' expected velocity The joint velocity at the current moment is obtained by performing condition number-gated inverse kinematics. .

[0082] The gating inverse solution is as follows:

[0083] When the condition number of the Jacobian matrix less than the condition number gate threshold At all times, a pseudo-inverse solution is used (such as...) Figure 1 The joint velocity calculated using the pseudo-inverse channel in the middle As the joint velocity in the inverse gate solution .

[0084]

[0085] in, This represents the joint velocity obtained from the pseudo-inverse solution; For Jacobi in a generalized inverse; For Descartes' expected velocity.

[0086] When the condition number of the Jacobian matrix Not less than the preset condition number gate threshold When using the damped least squares inverse solution (e.g.) Figure 1The joint velocity was calculated using the damped least squares channel. Joint velocity as a gated inverse solution .

[0087] ;

[0088] in, This represents the joint velocity obtained from the inverse least-squares solution of the damping; The generalized inverse of the Jacobian matrix (for the Jacobian matrix) (obtained by finding the generalized inverse) It is a Jacobian matrix; This is the transpose of the Jacobian matrix; Descartes' expected velocity; The diagonal damping matrix; , , , , and These are the six parameters in the damping parameter vector.

[0089] Furthermore, the gating inverse solution yields the joint velocity at the current moment. Then, calculate the joint velocity. 2-norm :like If the joint velocity exceeds the maximum joint velocity threshold, it is scaled proportionally to the threshold boundary. The joint velocity after scaling to the threshold boundary is used as the new joint velocity obtained from the gating inverse solution at the current moment. If the maximum joint velocity threshold is not exceeded, no action is taken.

[0090] By employing a dual-channel gated inverse solution mechanism based on the condition number of the Jacobian matrix, a pseudo-inverse solution is used in the low-risk region to ensure response speed and tracking accuracy, while switching to a damped least squares inverse solution in the high-risk region to improve numerical stability and avoid the increase in steady-state error and slow response caused by "large damping throughout the process".

[0091] S8. Joint velocities obtained from the inverse gating solution at the current moment. The actual joint velocity executed at the previous moment The joint angle actually executed at the previous moment Obtain the joint angle to be executed at the current moment. .

[0092] The joint angle to be executed at the current moment Obtain it using the following formula:

[0093] ;

[0094] in, This refers to the joint angle that is actually to be executed at the current moment. This refers to the joint angle actually executed at the previous moment; The sampling period; The actual joint speed to be executed at the current moment; The joint velocity is obtained from the inverse gate solution at the current moment; This indicates the joint velocity that was actually executed at the previous moment.

[0095] Specifically, the joint angle actually executed at the previous moment. Theoretically, it is equal to the current joint angle obtained in step S1. .

[0096] S9. The robotic arm determines the joint angle to be executed at the current moment according to step S8. Update the joint angles.

[0097] S10. Repeat steps S1-S9 to achieve control of the robotic arm.

[0098] Figure 3 This embodiment demonstrates the joint angles acquired in real time. (in , , , , and Joint angles The dynamic response results of the six corresponding joint parameters over time show that the method can effectively suppress abrupt joint angle changes and oscillations when approaching the singular region. The joint trajectory continuity and coordination are good, demonstrating high control stability and engineering feasibility. The posture error results obtained by the method in this embodiment are as follows: Figure 4 The method of this invention (anisotropic damping + gated inverse solution) is shown in the figure.

[0099] Furthermore, when the position error and attitude error When the threshold is met (the target is reached), it is determined that the robotic arm has accurately reached the target end pose, and the task is successfully completed.

[0100] The current process ends when the maximum number of steps is reached, and the task fails.

[0101] Specifically, the current process can be terminated and the task fails when a limit is exceeded. Limit exceeding includes situations such as joint angles exceeding limits, joint speeds exceeding safety thresholds, and Jacobian condition numbers being too low to reliably inversely solve the problem. Setting a maximum number of steps is to prevent infinite loops.

[0102] Furthermore, in this embodiment, the reinforcement learning model specifically adopts the DDPG network model, which includes a policy network and a value network. The policy network adopts a multi-layer fully connected structure, with a 15-dimensional state observation vector as input and a 6-dimensional action vector as output. The output layer uses the Tanh activation function to constrain the actions. The value network employs a dual-channel structure with state branches (multi-layer fully connected) and action branches (multi-layer fully connected), extracting features from the input state observation vector and action input respectively. These features are then fused and passed through a fully connected layer to output a value scalar. The value is used to evaluate the value of actions under the current policy and guide the iterative updates of the policy network.

[0103] The reward function for a reinforcement learning model is set according to the following formula:

[0104]

[0105]

[0106] in, To reinforce the reward function of the learning model; As a stability reward; To avoid rewards for unusual behavior; Rewards for maneuverability; To match the damping bonus; The actual joint speed to be executed at the current moment. The condition number is the inverse of the Jacobian matrix. It is the minimum singular value normalizer; This is a normalized measure of maneuverability. For indexing; The first in the damping parameter vector One parameter; For the damping parameter vector The average damping of the six damping parameters; This refers to the risk factor. , , , , and All are weighting coefficients; , and All are small positive constants; and All are threshold parameters; It is the natural logarithm.

[0107] The reward function constructed in this invention achieves stable rewards through multi-dimensional design. (Penalty for joint velocity) effectively suppresses velocity spikes and trajectory jitter; singularity avoidance reward With controllability reward Guiding the robotic arm away from the Jacobi pathological region while maintaining flexible movement; damping matching reward By dynamically adjusting the damping parameters based on singularity risk, damping can be increased in the danger zone to suppress abrupt changes, while damping can be decreased in the safe zone to avoid slow response and reduced accuracy. This solves three major technical problems: fixed damping cannot accommodate different singularity levels, uniform scalars cannot distinguish between axis differences, and reliance on manual parameter tuning.

[0108] Before using the policy network of the trained reinforcement learning model, a large amount of data is constructed with state observation vectors as inputs and damping action vectors as outputs to train the reinforcement learning model offline until training is complete.

[0109] The method of this invention adopts the DDPG strategy generation method of offline training and online forward inference. In the online stage, no parameter updates are performed, and only the damping parameters are output to participate in the calculation of interpretable analytical inverse solution. This method takes into account the adaptive capability of learning methods and the real-time performance and verifiability in engineering deployment.

[0110] This invention utilizes reinforcement learning, which demonstrates superior function approximation capabilities in continuous control. It maps state observation vectors to damping action vectors and integrates reinforcement learning with inverse control laws, enhancing adjustment flexibility while preserving the classical inverse law's physical structure. The method employs a learning strategy to provide damping parameters online, with the classical inverse law handling physical interpretability in a hierarchical manner, balancing control performance, interpretability, and stability verification requirements.

[0111] Comparative example:

[0112] To further demonstrate the beneficial effects of the method of the present invention, this comparative example employs a constant damping method in the same robotic arm end-effector pose tracking task as in the embodiment. The experimental results are as follows:

[0113] like Figure 4 As shown, the attitude error results obtained by the method of the present invention (anisotropic damping + gated inverse solution) and the comparative method (constant damping) demonstrate that the method of the present invention has superior performance in terms of error convergence speed, peak suppression, and overall smoothness. The method of the present invention is the method used in the embodiment, and the comparative method is the method used in this comparative example.

[0114] In summary, the method of this invention addresses inverse kinematic stability control during end-effector pose tracking, focusing on balancing stability and accuracy in singular neighborhoods. Compared to constant damping methods, this invention significantly reduces velocity abrupt changes when approaching singular poses, maintains higher tracking efficiency and control smoothness in non-singular directions, and is suitable for real-time motion control of articulated robotic arms with six degrees of freedom or more.

[0115] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for anisotropic damped inverse kinematics control of a robotic arm for singularity avoidance, characterized in that, Includes the following steps: S1. Acquire the current joint angles, current end-effector pose, and target end-effector pose of the robotic arm; S2. Calculate and obtain position error, attitude error, and Jacobian singular value features based on the current joint angle, current end pose, and target end pose. Then, concatenate the position error, attitude error, Jacobian singular value features, and current joint angle to construct the state observation vector. The Jacobi singular value features include the reciprocal condition number of the Jacobi matrix, the minimum singular value normalization, and the maneuverability normalization. The Jacobi matrix under the current configuration is calculated based on the current joint angle and the kinematic model of the robotic arm. The singular value set is obtained by decomposing the Jacobi matrix, and then the minimum and maximum singular values ​​are obtained. The reciprocal condition number of the Jacobi matrix is ​​obtained by dividing the minimum singular value by the maximum singular value. S3. Input the state observation vector into the policy network of the trained reinforcement learning model to output the damping action vector; S4. Map the damping action vector to a damping parameter vector and construct a diagonal damping matrix; S5. Calculate the Cartesian desired velocity based on the position error and attitude error. Then, perform a gated inverse solution of the Cartesian desired velocity based on the Jacobian matrix condition number and the diagonal damping matrix to obtain the joint velocity at the current moment. The gated inverse solution is specifically as follows: when the condition number of the Jacobian matrix is ​​less than a preset threshold, the joint velocity calculated by the pseudo-inverse solution is used as the joint velocity of the gated inverse solution; when the condition number of the Jacobian matrix is ​​not less than the preset threshold, the joint velocity calculated by the damped least squares inverse solution is used as the joint velocity of the gated inverse solution. S6. Perform amplitude limiting and filtering on the joint speed at the current moment in sequence to obtain the joint angle to be executed at the current moment, and drive the robotic arm to execute according to the joint angle. S7. Repeat steps S1-S6 to achieve control of the robotic arm.

2. The anisotropic damped inverse kinematics control method for a robotic arm according to claim 1, characterized in that: The minimum singular value normalization and the operability normalization are obtained by processing according to the following formula: ； in, Represents the minimum singular value normalization; This represents the normalized quantity of maneuverability; Represents the minimum singular value; Indicates an index; Represents the set of singular values. One singular value; This represents the cumulative product of all singular values ​​in the set of singular values.

3. The anisotropic damped inverse kinematics control method for a robotic arm according to claim 1, characterized in that, The diagonal damping matrix is ​​set according to the following formula: ； in, The diagonal damping matrix; Represents a diagonal matrix; , , , , and These are the six parameters in the damping parameter vector; Indicates an index; For the damping parameter vector, the first One parameter; For the first One damping action vector; and These are the preset minimum and maximum values ​​of the range in which the damping action vector output by the policy network lies; and These are the preset minimum and maximum values ​​for the range of damping parameters, respectively.

4. The anisotropic damped inverse kinematics control method for a robotic arm according to claim 1, characterized in that: The pseudo-inverse solution and the damped least squares inverse solution are set according to the following formulas: ； ； in, This represents the joint velocity obtained from the pseudo-inverse solution; This represents the joint velocity obtained from the inverse least-squares solution of the damping; It is a Jacobian matrix; For Jacobi in a generalized inverse; This is the transpose of the Jacobian matrix; Descartes' expected velocity; The diagonal damping matrix; , , , , and These are the six parameters in the damping parameter vector.

5. The anisotropic damped inverse kinematics control method for a robotic arm according to claim 4, characterized in that, The process of sequentially limiting and filtering the joint velocity at the current moment to obtain the joint angle to be executed at the current moment is as follows: The joint angle to be executed at the current moment is obtained from the joint velocity obtained by the inverse gating solution at the current moment, the joint velocity actually executed at the previous moment, and the joint angle actually executed at the previous moment. The joint angle to be executed at the current moment is obtained by processing it using the following formula: ； in, This refers to the joint angle that is actually to be executed at the current moment. This refers to the joint angle actually executed at the previous moment; The sampling period; The actual joint speed to be executed at the current moment; The joint velocity is obtained from the inverse gate solution at the current moment; This indicates the joint velocity that was actually executed at the previous moment.

6. The anisotropic damped inverse kinematics control method for a robotic arm according to claim 1, characterized in that: The reward function of the reinforcement learning model is set according to the following formula: in, To reinforce the reward function of the learning model; As a stability reward; To avoid rewards for unusual behavior; Rewards for maneuverability; To match the damping reward; The actual joint speed to be executed at the current moment. The condition number is the inverse of the Jacobian matrix. It is the minimum singular value normalizer; This is a normalized measure of maneuverability. For indexing; The first in the damping parameter vector One parameter; For average damping; This refers to the risk factor. , , , , and All are weighting coefficients; , and All are constants; and All of these are threshold parameters.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images