A human-like arm hybrid inverse kinematics control method fusing stereo projection SEW angle and differential constraint

Abstract

The application discloses a kind of human arm hybrid inverse kinematics control method and system fusing stereo projection SEW angle and differential constraint.It is first to obtain human arm movement data by wearing exoskeleton, and combined with stereo projection geometric transformation, construct the paired data set consisting of end full pose and human preferred SEW angle, by constraining algorithm singularity in physically inaccessible area, ensure the parameter continuity of full workspace.Subsequently, redundant parameter sine and cosine components under target pose are predicted using neural network, and scalar SEW angle is reconstructed to eliminate phase boundary jump.Finally, a hybrid inverse kinematics solver is constructed, which combines main task tracking with redundant preference guidance based on null space projection, uses finite difference mechanism to approximate redundant gradient online, and implicitly maps the kinetic energy efficiency optimization characteristics.Experimental results show that the application guarantees end millimeter-level tracking accuracy while significantly improving joint comfort and reducing energy consumption, effectively avoiding algorithm singularity, and the generated motion has smooth and natural human-like characteristics.

Description

Technical Field

[0001] This invention relates to the fields of robot kinematics control, redundancy resolution, deep learning intelligent control, and teleoperation technology, specifically to a hybrid inverse kinematics control method for a humanoid arm that integrates stereo projection SEW angles and differential constraints. Background Technology

[0002] Seven-degree-of-freedom (7-DoF) humanoid arms, due to their redundant degrees of freedom, can achieve self-motion by adjusting their internal joint configuration while performing spatial pose tasks. However, effectively utilizing this redundancy to enable the robotic arm to both accurately complete tasks and exhibit human-like natural collaborative characteristics remains a core challenge in the field of robot control. Traditional inverse kinematics methods often employ mathematical criteria for optimization, resulting in mechanically stiff joint postures. Furthermore, the traditional SEW angle definition contains bidirectional singular lines spanning the available workspace, easily leading to computational failures. Additionally, while purely data-driven methods can capture human motion preferences, phase boundary jumps in neural network angle predictions can cause control discontinuities. Therefore, designing a hybrid control scheme that ensures high-precision tracking, avoids algorithmic singularities, and achieves smooth humanoid posture generation is of great significance. Summary of the Invention

[0003] In view of the above problems, this invention proposes a hybrid inverse kinematics control method for a humanoid arm that integrates stereo projection SEW angle and differential constraints. High-fidelity data is collected by an exoskeleton with spherical motion characteristics. Combined with the singular external stereo projection geometric representation and the phase closed-loop deep learning prediction model, it can adaptively generate smooth and natural humanoid postures while ensuring millimeter-level tracking accuracy at the end effector.

[0004] This invention provides the following solution:

[0005] S1: Use a wearable exoskeleton with shoulder axis intersection features to collect coordinated human arm movement data and calculate the corresponding stereoscopic projection. Corners, construct a pairing dataset;

[0006] S2: Establish a redundant parameter prediction model, and introduce full pose feature perception logic and dual-component circular mapping mechanism for model training.

[0007] S3: Real-time acquisition of the target end pose, outputting the sine and cosine components of the target redundancy parameters through the prediction model, and reconstructing them into scalar angles. ;

[0008] S4: Calculate the velocity mapping command in the main task space and establish a dynamic damping adjustment mechanism based on operability indicators to ensure numerical stability.

[0009] S5: The gradient is approximated online using the finite difference mechanism, the redundancy resolution speed is synthesized through the null space projection operator, and a dynamic consistency energy consumption optimization criterion is introduced;

[0010] S6: Synthesize the total joint velocity command and perform saturation trimming based on hardware physical boundaries, iteratively updating the robot arm joint state. .

[0011] Preferably, in step S1, the logic regarding the anthropomorphic exoskeleton and its data acquisition is as follows:

[0012] The wearable exoskeleton employs a passive, passive mechanical structure, with its shoulder roll joint ( ) and pitch joint ( The exoskeleton's joints are perpendicular to each other, and their extensions strictly intersect at a single point. Through adjustment of the mechanical support, this intersection point is aligned with the anatomical rotation center of the operator's shoulder (i.e., the center of the humeral head), thereby simulating an ideal spherical kinematic pair at the hardware level and eliminating parasitic torques and kinematic mapping distortions generated during traditional exoskeleton movement. Each joint of the exoskeleton is equipped with a high-resolution Hall effect sensor (resolution better than 0.1°) to capture angular change sequences in real time at a frequency of 100Hz. Furthermore, the kinematic isomorphism mapping operator is used to transform it into the teaching state of a 7-DOF humanoid arm, ensuring the high fidelity of the training data source.

[0013] Preferably, in step S1, regarding stereoscopic projection... The angle establishment process is as follows:

[0014] (1) Calculate the shoulder-wrist unit vector: ;

[0015] (2) Calculate the projection mapping vector: ;

[0016] (3) Determine the projection reference axis: and ;

[0017] (4) Calculate the final horn: .

[0018] By selecting a reference vector pointing towards the robot's back This confines the singularity of the algorithm to a physically inaccessible region. Because the physical limitations of the robotic arm restrict the elbow from crossing the back area, it ensures that the singular half-line will never enter the effective workspace. This solves the problem of numerical jumps in inverse kinematics solutions within the workspace from a geometric and topological perspective, ensuring parameter continuity throughout the entire workspace.

[0019] Preferably, in step S2, the input to the neural network prediction model is a 7-dimensional full-pose feature vector. Coupled with three-dimensional position vectors With quaternion attitude vector This is used to eliminate configuration ambiguity in inverse kinematics analysis; the feature vectors undergo standardization before input.

[0020]

[0021] in, and The first The statistical mean and standard deviation of the dimensional features; the mean squared error loss function minimized by the neural network model during the training phase. The formula is:

[0022]

[0023] in, and The sub-labels represent the cosine and sine prediction components of the network output. and It is the true component, used to solve numerical jumps in angle at phase boundaries.

[0024] Preferably, in step S4, the main task tracks the joint velocity. The following damped least squares mapping formula is used for calculation:

[0025]

[0026] in, The final geometric Jacobian matrix is... The mission space velocity vector is generated by mapping the target pose deviation. The identity matrix; the damping coefficient Satisfying the following operability indicators Nonlinear adjustment relationship:

[0027]

[0028] in, The maximum damping coefficient, This is the sensitivity coefficient.

[0029] Preferably, in step S5, the redundancy resolution joint speed The generation follows the null space projection formula:

[0030]

[0031] in, The main task is Jacobi's pseudo-reversal. To redundancy adjust the gain, A stereoscopic projection of the current moment horn;

[0032] The The gradient component of the angle relative to the joint angle is estimated using the finite difference operator, and is expressed as:

[0033]

[0034] in For perturbations;

[0035] The humanoid arm is constrained by the system dynamics equations:

[0036]

[0037] And guide the joint configuration through a neural network to enter the following energy efficiency evaluation functional. This process achieves low-energy-loss operation in the minimum region by implicitly transforming complex real-time dynamic optimization into geometric tracking of redundant parameters favored by the human body. This allows the inverse kinematics solver to track predicted scalar angles without explicitly solving the dynamic equations. This indirectly achieves the optimal energy consumption distribution with dynamic consistency, and the energy consumption formula is:

[0038]

[0039] in, For joint torque, The inertia matrix, This is the energy consumption weighting factor.

[0040] Preferably, in step S6, the joint state variable The iterative updates follow the following saturation pruning evolution criterion:

[0041]

[0042] in, The sampling period is and These are the lower and upper limits of the joint angles defined by the robotic arm hardware, used to ensure that the calculated control sequence strictly meets the mechanical physical limit constraints.

[0043] The beneficial effects of this invention are:

[0044] (1) Data acquisition is performed using exoskeleton hardware with shoulder axis intersection features, which simulates the ideal biological spherical kinematic pair, eliminates the kinematic distortion caused by heterogeneous mapping, and makes the collected limb coordination data have extremely high bionic fidelity, effectively improving the joint comfort index by 22.5%.

[0045] (2) Introducing stereoscopic projection The angular representation and singular point externalization strategy solves the analytical continuity problem of the 7-DOF arm in the entire workspace. Combined with the dynamic damping adaptive adjustment mechanism, it completely avoids numerical divergence under singular configurations.

[0046] (3) By combining the two-component circular mapping prediction model with the guidance of dynamic energy efficiency functional, not only was the numerical jump at the phase boundary solved, but the torque output of the robotic arm was also optimized, reducing the overall energy consumption by 11.3% and significantly extending the service life of the actuator.

[0047] (4) This hybrid control framework combines data-driven flexible intent capture with millimeter-level precise constraints obtained through differential solving. The overall solution time is controlled within 5ms, meeting the requirements of high-frequency real-time closed-loop control above 100Hz. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 is an overall logic flowchart of a hybrid inverse kinematics control method provided in an embodiment of the present invention;

[0050] Figure 2 is a schematic diagram of a wearable exoskeleton data acquisition interface provided in an embodiment of the present invention;

[0051] Figure 3 is a topology diagram of the MotionMLP-BN prediction network provided in an embodiment of the present invention;

[0052] Figure 4 is a comparison diagram of robotic arm configurations under different redundancy resolution strategies provided in the embodiments of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0054] Referring to Figure 1, a hybrid inverse kinematics control method for an anthropomorphic arm that integrates stereo projection SEW angles and differential constraints includes the following steps:

[0055] Step S1: Acquisition of exoskeleton data based on anatomical alignment.

[0056] Referring to Figure 2, this embodiment employs a humanoid wearable exoskeleton. Its shoulder rotation axis... With pitch axis Precisely configured so that their extensions are perpendicular to each other and intersect at a point. ,point Aligned with the anatomical center of the human body. The operator performs spatial object repositioning tasks, with high-resolution sensors at each joint capturing human joint sequences in real time at a frequency of 100Hz. And map it to the 7-dimensional full pose vector of the end effector. ,in For 3D position, The pose is represented by a 4-dimensional unit quaternion. Simultaneously, redundant parameters are calculated using the stereo projection formula. Angle. By using the reference vector Pointing to the back of the robotic arm, achieving a non-singularity in the parameter description of the entire workspace.

[0057] While acquiring the pose, the corresponding stereo projection is simultaneously calculated. The specific process for establishing the angle is as follows:

[0058] (1) Calculate the shoulder-wrist unit vector: ;

[0059] (2) Calculate the projection mapping vector: ;

[0060] (3) Determine the projection reference axis: and ;

[0061] (4) Calculate the final horn: .

[0062] By selecting a reference vector pointing towards the robot's back By utilizing the physical limitation that prevents the robotic arm's elbow from crossing its back, the problem of numerical jumps in inverse kinematics calculations within the workspace is solved from a geometric topological perspective, ensuring that singular semi-linear lines never enter the workspace.

[0063] Step S2: Training the neural network model.

[0064] Referring to Figure 3, in this embodiment, the neural network prediction model adopts a deep fully connected architecture called MotionMLP_BN, and the feature vectors undergo the following normalization preprocessing:

[0065]

[0066] in, and These are the statistical mean and standard deviation for the corresponding dimensions.

[0067] In terms of specific network topology, the model consists of a 7-input 256-output input mapping layer, three sets of 256-dimensional hidden layer blocks, a set of 128-dimensional dimensionality reduction layer computation blocks, and a 128-input 2-output linear output layer. Each fully connected linear layer is followed by a batch normalization (BN) layer and a ReLU activation layer. Furthermore, an L2 norm normalization operator is introduced at the end of the model's forward propagation, which forces the predicted components to be normalized by dividing the output vector by its magnitude. and It is mapped onto the unit circular manifold.

[0068] During model training, the AdamW optimizer minimizes the following mean squared error loss function based on 2D components. :

[0069]

[0070] The logic behind employing dual-component joint supervision is to avoid the interference of mathematical discontinuities in the scalar angle at the phase boundary (±π) on gradient backpropagation, thus forcing the model to learn a smooth angular manifold during the training phase. Experimental results show that the model achieves an average prediction accuracy of 1.13° on the validation set.

[0071] Step S3: Real-time reconstruction of target redundancy parameters.

[0072] The system acquires the current target pose sequence in real time. The standardized pose vector is then input into the prediction network to obtain the output component values. and .

[0073] The specific implementation details are as follows: Because the neural network output layer uses L2 norm normalization, the output vector is forcibly constrained to the unit circle manifold. Compared to directly regressing scalar angles, this two-component prediction mechanism is geometrically continuous; when human movement crosses the phase boundary of the SEW angle (i.e., from π to −π), the coordinates of the points on the unit circle... Evolution is smooth and continuous, and there is no step discontinuity that scalar regression models must learn at the boundary.

[0074] It can be reconstructed into a scalar angle using the following formula. :

[0075]

[0076] use The function reconstructs the phase of the circular coordinates, which ensures that the angle reference value obtained within the 2π range of the entire circle has numerical continuity, thereby completely eliminating the instantaneous increase in control commands caused by phase jumps and ensuring the smoothness of the angular velocity of the robotic arm joints.

[0077] Step S4: Solve the hybrid kinematics.

[0078] A control layer based on differential kinematics is constructed. First, the task space velocity vector is generated by mapping the spatial deviation between the target pose and the current end effector pose. The main task is to track joint velocity. The following damped least squares (DLS) mapping formula is used for calculation:

[0079]

[0080] in, It is a 6×7 dimensional end-geometric Jacobian matrix. The identity matrix is ​​used; to balance the tracking accuracy and numerical stability of the robotic arm near singular configurations, the damping coefficient is... As operability Perform the following nonlinear dynamic adjustment:

[0081]

[0082] Step S5: Dynamic consistency projection.

[0083] Using a finite difference mechanism to approximate redundant gradients at the current joint angle Apply small perturbations on the basis (In this embodiment, we take) ):

[0084]

[0085] This method eliminates the need to obtain analytical partial derivative formulas for redundant parameters, greatly improving the algorithm's adaptability to different robotic arm kinematic structures.

[0086] Human-like preferences are injected into the control sequence via null space projection, as shown in the formula:

[0087]

[0088] The system guides the joint configuration into the minimum region of the energy efficiency evaluation functional, as shown in the formula:

[0089]

[0090] Referring to Figure 4, the proposed method effectively avoids the elbow collapse phenomenon that occurs in the Fixed SEW strategy (as shown by the yellow arrows at poses A and B). Furthermore, while maintaining high operability, the proposed method also improves joint comfort. ) increased by 22.5%, and reduced energy consumption ( The efficiency was reduced by 11.3%, proving that the geometric tracking process implicitly achieves the optimal distribution of dynamic energy efficiency.

[0091] Step S6: Status update and execution.

[0092] Synthetic overall joint velocity command To ensure the executability and safety of the generated motion sequences on physical hardware, joint state variables... The iterative updates follow the following saturation pruning evolution criterion:

[0093]

[0094] The The function is based on the lower limit of the joint angle preset by the robotic arm hardware. and upper limit The instructions are constrained in real time; if the calculated joint angle exceeds the physical limit, it is automatically truncated to the boundary value; the overall calculation time of this hybrid control framework is controlled within 5ms, and it supports high-frequency closed-loop control above 100Hz, ensuring the smoothness and safety of the robotic arm when performing continuous tasks.

[0095] The above embodiments merely illustrate one implementation of the present invention, and while the description is relatively specific and detailed, it should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A hybrid inverse kinematics control method for a humanoid arm that integrates stereo projection SEW angles and differential constraints, characterized in that, Includes the following steps: S1 Data Acquisition and Paired Dataset Construction: The joint state trajectory data of human arm movement is obtained using wearable exoskeleton, and the corresponding robotic arm end effector pose is calculated using forward kinematics algorithm; Based on the geometry of the robotic arm, the stereo projection SEW angle is calculated, and a paired dataset consisting of the end-effector full pose and the human preferred SEW angle is constructed. S2 Redundancy Parameter Prediction Model Training: Construct a neural network model with the terminal 7-dimensional full pose feature vector as input; A dual-component projection mechanism is adopted to transform the target SEW angle into orthogonal cosine and sine components for prediction. The training dataset is used to optimize the model parameters so that the model can output redundant parameter prediction values ​​with human kinematic characteristics. S3 Real-time Target Attitude Prediction and Reconstruction: The target end-effector pose is acquired in real time, and the cosine and sine components of the target SEW angle are predicted through the neural network model. The SEW angle parameters are then reconstructed into scalar parameters through trigonometric function operations. ; S4 Hybrid Inverse Kinematics Command Synthesis: Construct the differential inverse kinematics equations in the main task space, and use a pseudo-inverse algorithm with dynamic damping factors to solve the main task joint velocities; at the same time, use the finite difference mechanism to approximate the gradient vector of the SEW angle relative to the joint angle, and inject the predicted redundancy preference into the self-motion space through null space projection to generate a redundancy resolution velocity that satisfies the dynamic consistency constraint. S5 State Synthesis and Execution: Synthesizes the total joint velocity command and performs joint amplitude limiting processing, updating the joint state of the robotic arm in real time. .

2. The control method according to claim 1, characterized in that, The wearable exoskeleton described in step S1 has a shoulder rotation joint, and the rotation axes of the roll joint and the pitch joint in the shoulder rotation joint are perpendicular to each other and their extensions intersect at a point. The intersection point is aligned with the anatomical center of the operator's shoulder to simulate the characteristics of spherical motion and realize the kinematic mapping of human posture to the humanoid arm.

3. The control method according to claim 1, characterized in that, In step S1, the stereoscopic projection SEW angle is transformed by geometric mapping. By constraining the singularity of the algorithm to a unidirectional half-line and using the physical limitation of the robotic arm to make the unidirectional half-line located in an unreachable region, the continuity of SEW parameters in the entire workspace is achieved.

4. The control method according to claim 1, characterized in that, The input preprocessing and training loss function of the neural network model in step S2 are as follows: The terminal 7-dimensional full pose feature vector Coupled with three-dimensional Cartesian coordinate vectors With quaternion attitude vector The feature vectors are standardized before input. in For the first The mean of the dimensional features, For the first Standard deviation of dimensional features; The neural network model minimizes the mean squared error loss function during the training phase. To optimize the internal parameters, the loss function The formula is expressed as: in, The total number of samples, and These are the cosine and sine predicted values ​​output by the neural network, respectively. and These are the true target cosine and true target sine values ​​in the dataset, respectively; by performing joint regression on the orthogonal cosine and sine components, the phase jump of the angle scalar at the boundary is eliminated.

5. The control method according to claim 1, characterized in that, The main task solution and numerical stability guarantee mechanism described in step S4 are as follows: The mission space velocity Mapped to the joint velocities required to achieve pose tracking The calculation formula is: in, The Jacobian matrix of the end effector. It is the identity matrix. The damping coefficient; The dynamic damping factor Satisfying the following operability indicators Mapping relationship: in, The preset maximum damping coefficient, This is the sensitivity coefficient.

6. The control method according to claim 1, characterized in that, The redundancy resolution and dynamic guidance mechanism described in step S4 is as follows: The redundancy resolution rate is generated using the null space projection formula: Wherein, the gradient components are estimated as ; The redundancy resolution process aims to minimize the following potential function, which reflects human posture comfort. : in, For the first The median range of each joint; The humanoid arm follows the following dynamic equation during its movement: in The inertia matrix, The matrix of centrifugal force and Coriolis force. It is the gravity vector. Joint torque; The joint configuration is guided by predicted redundancy parameters into the following energy efficiency evaluation functional. The region of minimum values: in, It is an energy consumption weighting factor used to improve joint comfort while reducing overall energy consumption.

7. The control method according to claim 1, characterized in that, The joint state update formula mentioned in step S5 is as follows: in, and Mechanical limits defined for hardware.

8. A redundant robotic arm control system implementing the control method according to any one of claims 1-6, characterized in that, include: Data acquisition module: used to acquire human arm motion data using wearable exoskeleton and generate pose-redundancy parameter paired dataset; Parameter prediction module: It has a built-in neural network trained with a cosine and sine dual-component prediction mechanism to calculate the redundant SEW angular components corresponding to human motion preferences in real time. Hybrid-driven solver: Includes differential inverse kinematics operator and null projection operator, used for real-time synthesis of total joint commands that balance high-precision end-effector tracking and biomimetic natural posture. ; Motor actuator: Used to receive calculated joint commands and drive the movement of each joint of the robotic arm.

Images