Intelligent grabbing control method and system based on tactile perception

Abstract

The invention provides an intelligent grabbing control method and system based on tactile perception, and the method comprises the steps: obtaining a tactile feedback signal through a sensor array of a humanoid robot, carrying out the frequency domain decomposition of the signal through a Fourier transform method, and carrying out the analysis of the pressure and deformation data of each contact point on the surface of a to-be-grabbed object, calculating to obtain frequency characteristic distribution of the tactile feedback signal; if the initial estimation value of the object rigidity gradient exceeds a threshold value, performing finite element analysis on the object rigidity gradient, and combining a time domain processing result of the tactile feedback signal to obtain object rigidity gradient distribution; carrying out probability modeling on the deformation field by adopting a Bayesian inference method according to the gradient distribution of the rigidity of the object, and carrying out iterative optimization on a modeling result to obtain dynamic update parameters mapped by the deformation field; and regenerating an action sequence of the end effector of the humanoid robot based on the optimized strategy according to a trigger condition of grabbing failure detection, and determining a final grabbing control parameter.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an intelligent grasping control method and system based on tactile perception. Background Art

[0002] Intelligent grasping control is a crucial research area in robotics, directly impacting the operational precision and safety of humanoid robots in complex environments. With the increasing demand for robot-human interaction, grasping control based on tactile perception has become key to improving robot autonomy and adaptability. However, existing grasping control methods often struggle to achieve precise force control and posture stability when dealing with objects with complex internal structures or uneven stiffness. Existing impedance control strategies are often based on preset stiffness parameters and lack the ability to perceive and adapt to the object's dynamic characteristics in real time, leading to problems such as uneven force distribution and object slippage during grasping. Especially when the object's internal stiffness distribution is uneven, how to use tactile feedback signals to adjust the stiffness parameters of the end-effector in real time to adapt to the object's deformation characteristics becomes an urgent technical challenge. The core challenge lies in processing the tactile feedback signals and dynamically mapping the stiffness parameters. The tactile signals must accurately reflect the relationship between the object's stiffness gradient and deformation field, but existing methods struggle to strike a balance between real-time performance and accuracy. Furthermore, how the grasping controller adaptively adjusts the impedance parameters of each joint based on these signals to account for local variations in the object's stiffness is another key difficulty. Unresolved technical issues like these make it difficult to maintain stable object posture during grasping, especially when grasping objects with complex internal structures, which can easily lead to torque imbalance or grasp failure. Therefore, key challenges in this research are how to achieve real-time adaptive adjustment of the end-effector stiffness parameters through tactile feedback signals and establish a precise mapping between the stiffness gradient and the deformation field to ensure stable object posture during grasping. Summary of the Invention

[0003] The present invention provides an intelligent grasping control method based on tactile perception, which mainly includes:

[0004] The tactile feedback signal is obtained through the sensor array of the humanoid robot. The signal is decomposed in the frequency domain using the Fourier transform method. Combined with the pressure and deformation data of each contact point on the surface of the object to be grasped, the frequency characteristic distribution of the tactile feedback signal is calculated.

[0005] Based on the frequency characteristic distribution of tactile feedback signals, the local and global frequency characteristics of tactile signals are captured by fusing convolution kernels with different receptive fields. Multi-scale feature vectors are extracted using a pre-trained image recognition model. The stiffness gradient of the grasped object is hierarchically classified to determine a preliminary estimate of the object's stiffness gradient.

[0006] If the preliminary estimated value of the object stiffness gradient exceeds the threshold, perform a finite element analysis on the object stiffness gradient, and combine the time-domain processing results of the tactile feedback signal to obtain the object stiffness gradient distribution;

[0007] According to the object stiffness gradient distribution, use the Bayesian inference method to probabilistically model the deformation field, and perform iterative optimization on the modeling results to obtain the dynamic update parameters of the deformation field mapping;

[0008] Using the dynamic update parameters of the deformation field mapping, with the dynamic update parameters of the deformation field mapping as the state input and the impedance parameters of each joint of the end effector of the humanoid robot as the action output, if the adjusted impedance parameters meet the preset torque balance condition, then optimize the grasping trajectory of the joint angle of the end effector of the humanoid robot to obtain the execution instruction of torque balance control;

[0009] Adopt a preset grasping control strategy, according to the execution instruction of torque balance control, correct the grasping action of the end effector of the humanoid robot, and judge the compliance of the grasping control accuracy according to the corrected grasping action. If the grasping control accuracy does not meet the standard, then adjust the weight factor of the stiffness parameter based on the analysis result of the tactile feedback signal to obtain the trigger condition for grasping failure detection;

[0010] According to the trigger condition for grasping failure detection, regenerate the action sequence of the end effector of the humanoid robot based on the optimized strategy to determine the final grasping control parameters.

[0011] Further, a tactile feedback signal is obtained through the sensor array of the humanoid robot, and the Fourier transform method is used to decompose the signal in the frequency domain. By combining the analysis of the pressure and deformation data of each contact point on the surface of the object to be grasped, the frequency feature distribution of the tactile feedback signal is calculated, including: collecting the pressure of the tactile points on the object surface through a high-sensitivity sensor array, obtaining the pressure time-series data according to the set sampling period, and identifying the position coordinates and pressure values of the tactile contact points from the pressure data. Gaussian filtering and noise reduction processing are performed on the obtained pressure time-series data, the median filtering method is used to eliminate outliers, and the filtered tactile signal is smoothed according to the preset signal-to-noise ratio threshold. The fast Fourier transform is performed on the noise-reduced tactile signal, the signal spectrum coefficients are calculated in the frequency domain space, the peak points are extracted according to the spectral waveform curve, and the main frequency component and amplitude are obtained from the frequency domain data. A spatial discrete grid is established for the position coordinates of the tactile contact points, the bicubic spline interpolation algorithm is used to calculate the pressure values between the grid nodes, and a continuous pressure distribution surface is constructed according to the node pressure values. The deformation amount of each tactile contact point is calculated using the pressure distribution surface, and Hooke's law is used to establish the relationship equation between pressure and deformation, and the deformation displacement field distribution is obtained by solving. The frequency domain characteristic parameters and the deformation displacement field data are put into one-to-one correspondence according to the tactile contact points, and the fusion data matrix is constructed by the feature vector splicing method, and the fusion features are extracted by the principal component analysis method. According to the fusion features, a radial basis support vector machine classifier is trained, with the frequency components of the tactile signal as the labels and the deformation displacement field data as the input, to obtain the frequency feature distribution of the tactile feedback signal.

[0012] Further, according to the frequency feature distribution of the tactile feedback signal, capture the local and global frequency features of the tactile signal by fusing the convolutional kernels of different receptive fields, use the pre-trained image recognition model to extract multi-scale feature vectors, and perform hierarchical classification on the stiffness gradient of the object to be grasped to determine the preliminary estimated value of the object stiffness gradient, including: according to the tactile feedback signal frequency distribution map, divide the tactile receptive fields according to the signal amplitude size, extract local tactile features using small-scale receptive fields, and obtain global tactile features using large-scale receptive fields to generate a two-scale tactile feature map. Extract the convolutional kernel parameters of the first three layers from the pre-trained image recognition network, perform multi-level convolutional operations on the two-scale tactile feature map to generate a multi-layer tactile feature map group containing different spatial resolutions. Construct a feature pyramid for the multi-layer tactile feature map group, use the upsampling algorithm to magnify the low-resolution feature map, and obtain a multi-scale fusion feature map through cross-layer feature weighted fusion. Perform a non-linear transformation on the fusion feature map, use max pooling to extract significant feature points, compress the feature dimension through a dimensionality reduction algorithm, and generate a fixed-dimensional feature vector. Establish a stiffness gradient mapping function based on the feature vector, use the piecewise linear interpolation method to model the stiffness change trend, and generate a stiffness gradient curve. Use the clustering algorithm to hierarchically classify the stiffness gradient curve, set the stiffness change threshold to divide different stiffness intervals, and obtain the object surface stiffness distribution map. Calculate the local area stiffness gradient value according to the stiffness distribution map, use the adaptive threshold segmentation method to classify and label the stiffness gradient, and output the preliminary estimated result of the object stiffness gradient.

[0013] Furthermore, if the preliminary estimate of the object's stiffness gradient exceeds a threshold, a finite element analysis is performed on the object's stiffness gradient. Combined with the time-domain processing results of the tactile feedback signal, the object's stiffness gradient distribution is obtained. This includes: based on the pressure data collected by the tactile sensor array on the object's surface, time-domain sampling of the tactile feedback signal is performed according to a preset sampling period, and noise is removed from the sampled data using a bandpass filter to obtain a filtered time-domain signal sequence. A pressure distribution map of the object's surface is calculated based on the filtered time-domain signal sequence. A three-dimensional reconstruction algorithm is used to construct the geometric contour of the object's surface, and object boundary point cloud data is obtained from the reconstructed contour. If the preliminary estimate of the object's stiffness gradient exceeds a preset stiffness determination threshold, a tetrahedral mesh is created based on the boundary point cloud data. The mesh is subdivided in areas with significant stress variations to obtain a high-precision mesh structure on the object's surface. An elastic constitutive equation, including Young's modulus and Poisson's ratio, is established based on the object's material property parameters. The stress-strain equations are solved using Newton-Raphson iteration to obtain displacement field data for the mesh nodes. The mesh node displacement field data is used to calculate the unit stress and strain distribution. The stiffness coefficients of the local regions on the object's surface are solved according to Hooke's law to construct a stiffness distribution matrix. A spatiotemporal feature fusion algorithm is used to fuse the filtered time-domain signal sequence with the stiffness distribution matrix. Bilinear interpolation is used to calculate the stiffness values of the mesh nodes and construct a continuous stiffness distribution field. The finite difference method is used to calculate the spatial gradient of the stiffness distribution field and obtain an accurate distribution map of the stiffness gradient on the object surface.

[0014] Furthermore, based on the object's stiffness gradient distribution, a Bayesian inference method is used to probabilistically model the deformation field. The modeling results are then iteratively optimized to obtain dynamically updated parameters for the deformation field mapping. This involves spatially discretizing the object surface using a tetrahedral mesh based on the object's stiffness gradient distribution, and then using cubic spline interpolation to continuousize the stiffness gradient values at the mesh nodes to obtain a stiffness gradient spatial distribution function. A priori probability model of the object's deformation field is constructed based on the stiffness gradient spatial distribution function, and a normal distribution is used to describe the probability distribution of mesh node displacements to obtain an initial deformation field probability density function. The deformation field probability density function is updated using Bayesian inference, and a Monte Carlo method is used to generate a sequence of sampling points. Markov chain sampling is then used to obtain the posterior probability distribution of node displacements. A deformation field mapping matrix is established based on the posterior probability distribution of node displacements. The matrix elements contain the mean and variance of the node displacements, and the mapping matrix parameters are calculated using the least squares method. The mapping matrix parameters are iteratively optimized using an alternating optimization algorithm, with the mean parameter updated using gradient descent and the variance corrected using the quasi-Newton method. The probability distribution of the deformation field is reconstructed based on the optimized mapping matrix. Maximum a posteriori estimation is used to extract the expected values of node displacements, and a deterministic mapping function for the deformation field is constructed. The dynamic update parameters of the deformation field mapping, including node displacement increments and deformation rates, are calculated using the deformation field mapping function. The evolution characteristics of the deformation field are then extracted from the parameter sequence.

[0015] Furthermore, by dynamically updating the parameters of the deformation field mapping, using the dynamically updated parameters of the deformation field mapping as the state input, and using the impedance parameters of each joint of the end effector of the humanoid robot as the action output, if the adjusted impedance parameters meet the preset torque balance condition, then the joint angles of the end effector of the humanoid robot are optimized for the grasping trajectory to obtain the execution instruction for torque balance control, including: constructing a state input vector according to the dynamically updated parameters of the deformation field mapping, filtering the state data using the recursive least squares method, smoothing the state sequence using the sliding window method to obtain stable deformation field state parameters. A state-action mapper is constructed using a multi-layer perceptron. The input layer receives the deformation field state parameters, the hidden layer uses a two-layer structure to calculate the feature mapping, and the output layer generates the joint impedance parameters, including the stiffness matrix, damping matrix, and inertia matrix. Based on the obtained joint impedance parameters, the actuator dynamics equation is constructed, and the Lagrangian equation is used to calculate the relationship between the joint forces and motion parameters, and the joint torque values are solved from the dynamics equation. A balance judgment is made on the obtained joint torque values, the torque synthesis algorithm is used to calculate the resultant joint force, and it is judged whether the torque balance degree is less than the preset deviation threshold. If the torque balance degree is less than the preset deviation threshold, then the cubic spline curve is used to interpolate the joint angles of the actuator, the trajectory optimization objective function is constructed based on the variational method, and the optimal grasping trajectory is solved. The joint motion sequence is generated according to the optimal grasping trajectory, the joint angular velocity is constrained using the velocity planning algorithm, and the joint control instruction is extracted from the planning result. For the case where the torque balance degree is greater than the preset deviation threshold, the gradient feedback method is used to adjust the impedance parameters, and the torque balance calculation is performed again until the balance condition is met.

[0016] Furthermore, a preset grasping control strategy is adopted to correct the grasping action of the end effector of the humanoid robot according to the execution instruction of torque balance control, and the compliance of the grasping control accuracy is judged according to the corrected grasping action. If the grasping control accuracy does not meet the standard, the weight factor of the stiffness parameter is adjusted based on the analysis result of the tactile feedback signal to obtain the trigger condition for grasping failure detection, including: constructing the grasping correction amount of the actuator according to the torque balance control instruction and the preset grasping control threshold, and using a proportional-integral controller to adjust the joint angle of the actuator in real time to obtain the corrected grasping pose sequence. Tactile sensor data is collected for the corrected grasping pose, and wavelet transform is used to decompose the tactile signal into four layers, and the tactile features are extracted from the high-frequency coefficients to obtain the grasping contact feature vector. The control accuracy index is calculated according to the grasping contact feature vector, and the grasping stability is evaluated by the root mean square error, and whether the grasping accuracy meets the preset control threshold is judged from the stability value. If the grasping accuracy is greater than the preset control threshold, it is marked that the current grasping action has been optimized, the optimized grasping parameter group is recorded, and the grasping optimization completion flag is output. If the grasping accuracy is less than the preset control threshold, a tactile predictor is constructed using a support vector regressor, the grasping deviation feature is extracted from the prediction result, and a failure detection feature set is constructed. The update amount of the stiffness parameter weight is calculated according to the failure detection feature set, and the weight factor is iteratively adjusted using an exponential decay learning rate to obtain a new weight coefficient matrix. The grasping control parameters are compensated using the updated weight coefficient matrix, the grasping correction amount is recalculated, and the trigger condition parameters for grasping failure detection are constructed.

[0017] Further, according to the triggering conditions of grasping failure detection, based on the optimized grasping strategy, regenerate the action sequence of the end effector of the humanoid robot, and determine the final grasping control parameters, including: constructing a robot state vector according to the triggering conditions of grasping failure detection, including joint angles, speeds, and force information; constructing a policy network using the Deep Deterministic Policy Gradient algorithm; optimizing the grasping strategy through an action value evaluation function. Use the optimized policy network to output joint action instructions, interpolate the joint motion trajectory using cubic spline curves to generate angular displacement, angular velocity, and angular acceleration sequences. Calculate the end pose of the actuator according to the angular displacement sequence, solve the position and attitude matrix of the actuator using the forward kinematic equation, and obtain the object grasping state description. Evaluate the stability of the object grasping state, filter the pose data using a Kalman filter, calculate the covariance eigenvalue of the pose matrix, and determine whether the attitude fluctuation is less than the preset stability threshold. If the attitude fluctuation is greater than the preset stability threshold, use the gradient descent method to adjust the joint control parameters, including the stiffness coefficient and damping coefficient, and update the control parameters through backpropagation. Use the updated control parameters to regenerate the grasping action sequence, process the action sequence using a trajectory smoother to eliminate trajectory jumps and jitters. Verify the attitude stability according to the smoothed action sequence, calculate the root mean square error of the pose deviation, and obtain the final grasping control parameter set.

[0018] The present invention provides an intelligent grasping control system based on tactile perception, mainly including:

[0019] A tactile signal processing module, which is used to obtain tactile feedback signals through the sensor array of the humanoid robot, perform frequency domain decomposition on the signals using the Fourier transform method, and calculate the frequency feature distribution of the tactile feedback signals by combining the pressure and deformation data analysis of each contact point on the surface of the object to be grasped;

[0020] A frequency feature extraction module, which is used to capture the local and global frequency features of tactile signals by fusing convolution kernels with different receptive fields according to the frequency feature distribution of tactile feedback signals, extract multi-scale feature vectors using a pre-trained image recognition model, and perform hierarchical classification on the stiffness gradient of the object to be grasped to determine the preliminary estimated value of the object stiffness gradient;

[0021] A stiffness gradient analysis module, which is used to perform finite element analysis on the object stiffness gradient if the preliminary estimated value of the object stiffness gradient exceeds the threshold, and combine the time domain processing results of the tactile feedback signals to obtain the object stiffness gradient distribution;

[0022] A deformation field modeling module, which is used to perform probabilistic modeling on the deformation field using the Bayesian inference method according to the object stiffness gradient distribution, and perform iterative optimization on the modeling results to obtain the dynamic update parameters of the deformation field mapping;

[0023] An impedance control module, which is used to dynamically update parameters through deformation field mapping, use the dynamically updated parameters of the deformation field mapping as state inputs, use the impedance parameters of each joint of the end effector of the humanoid robot as action outputs, and if the adjusted impedance parameters meet the preset torque balance condition, optimize the grasping trajectory of the joint angles of the end effector of the humanoid robot to obtain an execution instruction for torque balance control;

[0024] A grasping control module, which is used to adopt a preset grasping control strategy, correct the grasping action of the end effector of the humanoid robot according to the execution instruction of torque balance control, judge the compliance of the grasping control accuracy according to the corrected grasping action, and if the grasping control accuracy does not meet the standard, adjust the weight factor of the stiffness parameter based on the analysis result of the tactile feedback signal to obtain a trigger condition for grasping failure detection;

[0025] A grasping optimization module, which is used to regenerate the action sequence of the end effector of the humanoid robot based on the optimized strategy according to the trigger condition for grasping failure detection, and determine the final grasping control parameters.

[0026] An intelligent grasping control device based on tactile perception, the intelligent grasping control device includes: one or more processors; a storage device for storing one or more programs;

[0027] When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the above method for grasping control.

[0028] The technical solution provided by the embodiment of the present invention may include the following beneficial effects:

[0029] The present invention discloses an intelligent grasping control method based on tactile perception. By analyzing the frequency domain characteristics of tactile sensor signals and combining adaptive multi-scale convolutional neural networks and finite element analysis, the method realizes accurate estimation of the stiffness gradient of the object to be grasped. Further, Bayesian inference is used to probabilistically model the deformation field, dynamically update the deformation field mapping parameters, and optimize the impedance parameters and grasping trajectories of the robot end effector based on this. Finally, the present invention uses a reinforcement learning algorithm to online optimize the grasping control strategy, effectively improving the grasping accuracy and stability of the humanoid robot in complex environments. This method can adapt to objects with different stiffnesses and shapes, significantly enhancing the grasping ability and operation flexibility of the humanoid robot. Description of the Drawings

[0030] Figure 1 It is a flowchart of an intelligent grasping control method based on tactile perception of the present invention.

[0031] Figure 2 It is a structural schematic diagram of an intelligent grasping control system based on tactile perception of the present invention. Detailed implementation manners

[0032] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] The intelligent grasping control method based on tactile perception collects tactile feedback signals through the sensor array of a humanoid robot, combines frequency-domain analysis and deformation field calculation, and dynamically optimizes the grasping strategy. The tactile feedback signals include pressure time-series data and contact point coordinates, and are used to reflect the stiffness and deformation characteristics of the object surface. The goal of grasping control is to adaptively adjust the impedance parameters of the end effector according to the tactile signals to ensure the stability of the object posture during the grasping process.

[0034] Figure 1 The implementation process of the intelligent grasping control method based on tactile perception provided by the embodiments of the present application is shown and described in detail as follows:

[0035] S101. Obtain tactile feedback signals through the sensor array of the humanoid robot, perform frequency-domain decomposition on the tactile feedback signals by using the Fourier transform method, and calculate the frequency feature distribution of the tactile feedback signals in combination with the analysis of the pressure and deformation data of each contact point on the surface of the object to be grasped.

[0036] The acquisition and processing of tactile feedback signals are the basic steps for realizing intelligent grasping control. The sensor array realizes the real-time acquisition of pressure data through a high-sensitivity microelectrode matrix, and generates a tactile frequency feature distribution in combination with deformation analysis, providing a basis for stiffness estimation.

[0037] S1011. Collect the tactile feedback signals on the surface of the object to be grasped, extract the position coordinates and pressure values of the tactile contact points, and perform noise reduction processing.

[0038] The sensor array adopts an 8×8 microelectrode matrix layout, the electrode spacing is 2 mm, the sampling period is 0.5 ms, and the collected tactile feedback signals can accurately capture the pressure distribution on the object surface. The collected pressure time-series data contains the coordinates and pressure values of each tactile point, and a Gaussian filter is used for noise reduction, with a standard deviation of 0.8 and a window size of 5×5 pixels. In combination with a median filter with a 3×3 window to eliminate outliers, ensuring that the signal-to-noise ratio reaches more than 25 dB to obtain clear tactile signals.

[0039] S1012. Perform a fast Fourier transform on the noise-reduced tactile feedback signals, extract the frequency-domain features, and construct a pressure distribution surface according to the tactile contact point coordinates to calculate the deformation displacement field.

[0040] Perform a 1024-point fast Fourier transform on the filtered tactile signal for frequency-domain decomposition, extract the main frequency component and amplitude, and focus on analyzing the spectral characteristics in the range of 20 to 200 Hz, which contains the object surface texture information. For the tactile contact point coordinates, establish a spatial discrete grid with a spacing of 0.5 mm, use the bicubic spline interpolation algorithm to calculate the pressure values between grid nodes, and construct a pressure distribution surface. According to the surface data, use Hooke's law to calculate the deformation amount of each tactile point, set the material elastic modulus to 2000 MPa, the Poisson's ratio to 0.3, and the maximum deformation amount to about 0.2 mm, so as to generate a deformation displacement field.

[0041] S1013. Integrate the frequency-domain features and the deformation displacement field, and use a classification algorithm to generate a spatial distribution map of tactile frequency features to obtain the frequency feature distribution of the tactile feedback signal.

[0042] Correspond the frequency-domain feature parameters and the deformation displacement field one by one according to the tactile points, and construct a fusion data matrix through the feature vector splicing method. The frequency-domain features include 10 main frequency components and their amplitudes, and the deformation data includes the displacement values of 25 grid nodes, constituting a 45-dimensional fusion feature vector. Use a radial basis support vector machine classifier for training, set the Gaussian kernel function parameter to 1.5, the penalty factor to 100, the training sample size to 1000 groups, and the classification accuracy reaches 92%. The generated spatial distribution map of tactile frequency features can clearly reflect the tactile characteristics of the object surface. The low-frequency band of 20 to 50 Hz reflects the basic touch feeling, the middle-frequency band of 100 to 150 Hz contains texture information, and the high-frequency band above 200 Hz corresponds to the microstructural characteristics. The frequency feature distribution of the tactile feedback signal can be known through the generated spatial distribution map of tactile frequency features, providing a reliable basis for stiffness analysis.

[0043] Through multi-scale frequency-domain analysis and deformation field calculation, the spatial distribution map of tactile frequency features can accurately characterize the stiffness and texture characteristics of the object surface. Compared with traditional methods, this application introduces double filtering and a high-precision interpolation algorithm in signal processing, significantly improving the signal-to-noise ratio and spatial resolution of the tactile signal, laying a foundation for stiffness gradient estimation, and thus effectively improving the accuracy and stability of grasping control.

[0044] S102. According to the frequency feature distribution of the tactile feedback signal, capture the local and global frequency features of the tactile feedback signal by integrating convolution kernels with different receptive fields, perform hierarchical classification on the stiffness gradient of the object to be grasped, and determine the preliminary estimated value of the object stiffness gradient.

[0045] Based on the frequency feature distribution of the tactile feedback signal, through multi-scale feature extraction and fusion, generate a stiffness distribution map of the object surface, providing an accurate stiffness gradient estimation for grasping control.

[0046] S1021. Divide small-scale and large-scale receptive fields according to the amplitude characteristics of the tactile feedback signal, extract local and global tactile features respectively, and generate a dual-scale tactile feature map.

[0047] The division of the tactile receptive field refers to the human tactile perception mechanism, aiming to capture the multi-level characteristics of the object surface. The small-scale receptive field is set as a 2×2 mm area, suitable for extracting fine texture and local pressure change features; the large-scale receptive field covers a 10×10 mm area, used to perceive the overall shape and stiffness distribution of the object. Based on the amplitude of the tactile feedback signal, the signal is divided into high-amplitude and low-amplitude intervals, corresponding to local mutation and global smooth characteristics respectively. By extracting features from the pressure signal of 64×64 sampling points, a dual-scale tactile feature map containing frequency and amplitude information is generated. Each sampling point records the amplitude and the main frequency component of the signal, providing a high-resolution input for the convolution operation.

[0048] S1022. Use the convolution kernel parameters of the pre-trained image recognition network to perform multi-level convolution operations on the dual-scale tactile feature map, and generate a multi-scale fusion feature map.

[0049] The first three layers of convolution kernel parameters of the VGG16 network pre-trained on the ImageNet dataset are used for initialization. The first layer contains 64 3×3 convolution kernels, which are good at extracting edge and texture features; the second layer contains 128 3×3 convolution kernels, capturing more complex spatial patterns; the third layer contains 256 3×3 convolution kernels, extracting high-level abstract features. Perform multi-level convolution operations on the dual-scale tactile feature map to generate a multi-layer tactile feature map with resolutions of 64×64, 32×32, 16×16, and 8×8. Based on the multi-layer tactile feature map, a multi-scale fusion feature map is constructed. The original resolution is retained at the bottom layer, and downsampling is gradually performed upward to extract more abstract feature representations. Bilinear interpolation is used to perform 2-fold upsampling to align the low-resolution feature map to the high-resolution layer, and the weighted fusion coefficients are 0.4, 0.3, 0.2, and 0.1 respectively, highlighting the contribution of the fine features at the bottom layer to obtain a multi-scale fusion feature map and ensuring the integrity of the multi-scale features.

[0050] S1023. Perform non-linear processing and feature dimensionality reduction on the multi-scale fusion feature map, extract multi-scale feature vectors, construct a stiffness gradient mapping function, and generate a stiffness distribution map of the object surface.

[0051] Apply the ReLU non-linear activation function to the multi-scale fusion feature map to enhance the expression ability of the features. Subsequently, extract significant feature points through max pooling operation to generate multi-scale feature vectors. Use the principal component analysis method to compress the original 4096-dimensional feature space to 128 dimensions, retaining more than 95% of the information. The multi-scale feature vectors after dimensionality reduction can efficiently characterize the stiffness characteristics of the object surface. Based on the feature vectors, construct a stiffness gradient mapping function, and map the eigenvalues to the stiffness range of 0 to 100 through piecewise linear interpolation. The stiffness values of soft materials such as rubber are distributed in the range of 0 to 30, hard plastics are distributed in the range of 31 to 70, and metal materials are distributed in the range of 71 to 100. Use the improved K-means clustering algorithm to divide the stiffness values into five levels: ultra-soft, soft, medium, hard, and extremely hard. The number of cluster centers is 5, and the inter-class variance reaches more than 85%, ensuring the robustness of the classification results. The generated stiffness distribution map can clearly reflect the stiffness change trend of the object surface.

[0052] S1024. Calculate the stiffness gradient according to the stiffness distribution map, and use the adaptive threshold segmentation method to classify the stiffness gradient, and output the preliminary estimated value of the object stiffness gradient.

[0053] Calculate the stiffness gradient using a 4×4 sliding window, and use the stiffness difference between adjacent points within the window as the gradient value. Set an adaptive threshold to divide the stiffness gradient into three categories: gently changing area, moderately changing area, and severely changing area, corresponding to the areas where the gradient values are less than 10, between 10 and 30, and greater than 30 respectively. The gently changing area reflects the area with uniform material, such as the stiffness gradient value inside soft rubber is about 5 to 8; the moderately changing area corresponds to the material transition area, such as the gradient value at the junction of rubber and plastic is about 25; the severely changing area appears at the contact edge of hard materials, such as the gradient value at the junction of plastic and metal can reach 45. Through threshold segmentation, classify and mark the stiffness gradient to generate a preliminary stiffness gradient estimation result, providing an accurate input for grasping parameter optimization.

[0054] By combining the multi-scale convolutional neural network and the feature pyramid, complex features of tactile signals can be efficiently extracted, and the generated stiffness distribution map provides a key basis for grasping control. Compared with traditional methods, this application introduces a pre-trained model and adaptive threshold segmentation in feature extraction, improving the robustness and real-time performance of stiffness gradient estimation, and laying a foundation for moment balance and trajectory optimization.

[0055] S103. If the preliminary estimated value of the object stiffness gradient exceeds the threshold, then calculate the object stiffness gradient distribution by combining the time-domain processing result of the tactile feedback signal and the geometric contour of the surface of the object to be grasped.

[0056] When the initial estimate of the object's stiffness gradient exceeds a threshold, it indicates that significant stiffness variations may exist on the object's surface, necessitating further refined analysis to ensure the accuracy of grasping control. If the initial estimate of the stiffness gradient falls below the threshold, this may indicate that the object is relatively soft or easily deformed. In this case, the system selects a pre-set control strategy to perform the grasping operation. This step combines time-domain signal processing with finite element analysis to accurately calculate the stiffness distribution of the object's surface, providing a reliable basis for grasping parameter optimization and improving the accuracy and robustness of stiffness estimation.

[0057] S1031 , performing time-domain sampling on the tactile feedback signal of the tactile sensor array using a preset sampling period, obtaining a filtered time-domain signal sequence through band-pass filtering, and generating a pressure distribution map of the object surface.

[0058] The tactile sensor array utilizes an 8×8 matrix layout with a 2mm sensor pitch, a sensitivity of 0.1 Newton, and a sampling period of 1 millisecond, enabling real-time capture of surface pressure variations. Time-domain sampling processes the raw signal through a 50-200 Hz bandpass filter, effectively filtering out ambient noise and high-frequency interference, generating a filtered signal sequence with a signal-to-noise ratio (SNR) better than 20dB. Based on this filtered signal sequence, the sampled point data is processed using a cubic spline interpolation algorithm to generate a surface pressure distribution map with a surface resolution of 0.5mm. This clearly reflects the pressure variation characteristics of the surface and provides high-precision input for geometric modeling.

[0059] S1032. Process the pressure distribution map on the object surface using a three-dimensional reconstruction algorithm to obtain geometric contour data of the object surface.

[0060] Based on the surface pressure distribution map of the object, the Marching Cubes algorithm is used for 3D reconstruction to generate boundary point cloud data containing more than 1,000 spatial sampling points, with a reconstruction accuracy better than 0.5 mm. If the preliminary estimate of the stiffness gradient exceeds the threshold of 50 MPa per millimeter, high-precision meshing is triggered, and the object surface is modeled using a tetrahedral mesh. The initial mesh size is 2 mm, and the mesh size is refined to 0.2 mm in stress concentration areas through adaptive subdivision to ensure accurate capture of areas with drastic stress changes. The mesh structure contains approximately 5,000 units and can efficiently support subsequent finite element calculations.

[0061] S1033. Establish an elastic constitutive equation based on the geometric profile data of the object surface, use an iterative solution method to calculate the displacement field of the grid nodes, and generate a stiffness distribution matrix.

[0062] Based on the material properties of the object, a linear elastic constitutive equation is constructed. Taking aluminum alloy as an example, the Young's modulus is set to 70 GPa and the Poisson's ratio is set to 0.33. The Newton-Raphson iterative algorithm is used to solve the stress and strain equations, and the convergence accuracy is set to 0.001 to generate the displacement field data of grid nodes at the micron scale. Based on Hooke's law, the stress-strain ratio of each grid element is calculated to generate an 8×8 stiffness distribution matrix, and the stiffness coefficient of each element reflects the stiffness characteristics of the local area. The stiffness coefficient is calculated by dividing the element stress by the strain to ensure that the calculation results conform to the laws of mechanics of materials.

[0063] S1034. Integrate the filtered time-domain signal sequence and the stiffness distribution matrix through a spatio-temporal feature fusion algorithm, use the interpolation method to generate a continuous stiffness distribution field, and calculate the stiffness gradient distribution of the object.

[0064] A spatio-temporal feature fusion method with weighted average is adopted. The weight of the time-domain signal is set to 0.6, and the weight of the stiffness distribution matrix is 0.4. The fused data is extended to a high-resolution grid of 32×32 through bilinear interpolation, and the spatial resolution is increased to 0.5 mm. The finite difference method is used to calculate the spatial gradient of the stiffness distribution field to generate an accurate stiffness gradient distribution map and obtain the stiffness gradient distribution of the object. The test results show that the stiffness gradient in the soft rubber area is about 5 MPa / mm, 20 to 40 MPa / mm in the hard plastic area, and up to 80 MPa / mm at the metal edge. The gradient map clearly shows the stiffness mutation characteristics in the material transition area.

[0065] By combining the time-domain processing of tactile signals with finite element analysis, the stiffness change characteristics of the object surface can be efficiently captured. Compared with traditional methods, this application introduces high-precision grid subdivision and spatio-temporal feature fusion, significantly improving the accuracy of stiffness gradient estimation, with the error controlled within 10%, providing reliable support for the fine grasping operation of the robot.

[0066] S104. According to the stiffness gradient distribution of the object, use the Bayesian inference method to probabilistically model the deformation field and calculate the dynamic update parameters of the deformation field mapping.

[0067] Based on the accurate distribution of the stiffness gradient on the object surface, dynamically updated deformation field mapping parameters are generated through probabilistic modeling and iterative optimization to provide high-precision deformation prediction for grasping control. In this step, by combining physical constraints and statistical inference, the complexity and uncertainty of object deformation are fully considered to ensure the real-time performance and reliability of the deformation field mapping.

[0068] S1041. Generate a stiffness gradient spatial distribution function according to the stiffness gradient distribution of the object.

[0069] Based on the object's stiffness gradient distribution, the cubic spline interpolation algorithm is used to smooth the stiffness values between grid nodes, ensuring the continuity of the second derivative of the interpolation function, and generating the spatial distribution function of the stiffness gradient. This function uses grid nodes as control points to accurately describe the continuous change of stiffness in three-dimensional space, providing high-precision input for subsequent probability modeling. Compared with traditional linear interpolation, cubic spline interpolation can better capture the non-linear characteristics of the stiffness mutation region, significantly improving the smoothness and expressive ability of the distribution function.

[0070] S1042. Construct a prior probability model of the deformation field based on the spatial distribution function of the stiffness gradient, use the normal distribution to describe the node displacement probability, and generate the initial probability density function of the deformation field.

[0071] The prior probability model of the deformation field is constructed by combining physical constraints and statistical assumptions, assuming that the displacements of each grid node follow a normal distribution. The mean value of the node displacement is calculated through the static equilibrium equation, reflecting the expected deformation of the object under stiffness constraints; the standard deviation is set to 0.1 mm, characterizing the uncertainty of the displacement. The generated initial probability density function of the deformation field is in units of grid nodes, describing the displacement probability distribution of each node, and can effectively reflect the deformation characteristics of the object surface in different stiffness regions. For example, in the uniform material region, the probability density function shows a narrow distribution, indicating that the deformation prediction is relatively certain; in the material transition region, the distribution is wider, reflecting higher uncertainty.

[0072] S1043. Update the initial probability density function of the deformation field through the Bayesian inference method, use the Markov chain Monte Carlo method to generate the posterior probability distribution of the node displacement, and construct the deformation field mapping matrix.

[0073] Bayesian inference is used to integrate the stiffness gradient data and the tactile feedback signal to update the probability density function of the deformation field. The Markov chain Monte Carlo method is used for probability sampling, generating 1000 sampling points for each grid node, and the sampling chain length is set to 5000 steps, of which the first 500 steps are excluded as the warm-up period to ensure the stability of the sampling results. The posterior probability distribution of the node displacement is calculated through the Gibbs sampling algorithm, generating a 512×512-dimensional deformation field mapping matrix, and the matrix elements include the displacement mean and variance of each node. The mean reflects the expected deformation trend, and the variance quantifies the prediction uncertainty. The initial mapping parameters are fitted by the least squares method to ensure the high consistency of the matrix elements with the stiffness gradient data.

[0074] S1044. Use the alternating optimization algorithm to optimize the parameters of the deformation field mapping matrix and extract the dynamic update parameters of the deformation field mapping.

[0075] The alternating optimization algorithm iteratively updates the mapping matrix parameters in two directions: displacement mean and variance. The gradient descent method with an adaptive learning rate is used to optimize the mean, with the initial learning rate set to 0.01, and the learning rate is dynamically adjusted according to the change of the loss function to accelerate convergence; the variance is corrected by the quasi-Newton method, and the condition number threshold of the Hessian matrix is set to 1000 to ensure the stability of the optimization process. The optimized mapping matrix is used to reconstruct the probability distribution of the deformation field, and the expected value of the node displacement is extracted through the maximum a posteriori estimation to construct a deterministic mapping function of the deformation field. This function can accurately predict the dynamic deformation process of the object surface and output the dynamic update parameters of the deformation field mapping including displacement increment and deformation rate. The displacement increment describes the deformation difference between adjacent moments, and the deformation rate reflects the severity of deformation in the local area, providing a key basis for real-time grasping control.

[0076] Tests show that in soft rubber materials, the displacement error predicted by the deformation field mapping function is controlled within 0.1 mm, and the deviation of the deformation rate is less than 5%, demonstrating high-precision prediction ability. In the contact area between hard plastic and metal, due to the discontinuous deformation caused by the sudden change of material stiffness, the system significantly reduces the prediction variance by increasing the sampling density and the number of iterations, ensuring the prediction reliability in complex scenarios. This method improves the adaptability and robustness of deformation field prediction through probability modeling and dynamic optimization, providing strong support for precise robot operation.

[0077] S105. Use the dynamic update parameters of the deformation field mapping as the state input and the impedance parameters of each joint of the end effector of the humanoid robot as the action output. If the adjusted impedance parameters meet the preset torque balance condition, optimize the grasping trajectory of the joint angles of the end effector of the humanoid robot to obtain the execution instruction for torque balance control.

[0078] Based on the dynamic update parameters of the deformation field mapping, through the mapping of state input and action output, dynamically adjust the impedance parameters of the end effector to ensure torque balance and trajectory optimization during the grasping process. This step significantly improves the stability and accuracy of grasping through multi-layer perceptron and dynamics modeling, combined with closed-loop torque control, providing reliable support for fine operation in complex scenarios.

[0079] S1051. Generate a state vector according to the dynamic update parameters of the deformation field mapping.

[0080] The dynamic update parameters of the deformation field mapping include the displacement field, velocity field, and acceleration field, with a total of 16-dimensional features. Among them, the 8-dimensional displacement field describes the deformation distribution on the object surface, the 4-dimensional velocity field reflects the deformation propagation characteristics, and the 4-dimensional acceleration field characterizes the dynamic stress changes. The recursive least squares method with a sliding window length of 10 is used to filter the state data. The data within the window is weighted averaged to eliminate random fluctuations, ensuring the smoothness and stability of the state parameters. The filtered state vector can clearly reflect the spatio-temporal characteristics of object deformation. For example, when grasping a glass, the displacement field shows that the maximum deformation of the cup wall is about 0.2 millimeters, and the velocity field indicates that the deformation propagation rate is about 5 millimeters per second, providing a reliable basis for state-action mapping.

[0081] S1052. Use a multi-layer perceptron to construct a multi-layer perceptron state-action mapper to convert the dynamic update parameters of the deformation field mapping into joint impedance parameters, including stiffness, damping, and inertia matrix.

[0082] The multi-layer perceptron state-action mapper adopts a double-hidden-layer multi-layer perceptron structure. The input layer receives a 16-dimensional state vector. The first hidden layer contains 32 neurons and uses the rectified linear unit activation function to enhance the non-linear expression ability. The second hidden layer contains 16 neurons and uses the hyperbolic tangent function activation to balance feature convergence. The output layer generates 9-dimensional impedance parameters, corresponding to the stiffness, damping, and inertia coefficients of the 3 joints of the end effector. During the training process, the perceptron optimizes the weights through the backpropagation algorithm. The learning rate is set to 0.001, the batch size is 32, and the output is stable after 1000 rounds of training. Among the generated impedance parameters, the stiffness coefficient ranges from 50 to 200 Newton meters per radian, the damping coefficient is from 0.5 to 2 Newton meters per radian per second, and the inertia coefficient is about 0.1 kilogram square meters, which can adapt to the grasping requirements of objects with different materials.

[0083] S1053. Based on the joint impedance parameters, construct the actuator dynamics equation, solve the joint torque value, and judge whether to trigger trajectory optimization or parameter adjustment through torque balance.

[0084] The Lagrangian method is used to construct the dynamics equation of a 6-degree-of-freedom end effector, comprehensively considering the coupling effect between joints, the influence of gravity, and external contact forces. The dynamics equation is solved by the improved Euler method, with a calculation step of 1 millisecond and the iteration accuracy controlled within 0.1 Newton to generate the torque values of each joint. The torque balance judgment uses the torque synthesis algorithm to calculate the resultant force of the joint forces and compare it with the preset threshold of 2 Newton meters. If the torque deviation is less than the threshold, it indicates that the grasping state is stable and triggers trajectory optimization. If the deviation exceeds the threshold, the impedance parameters are adjusted through the gradient feedback method, with a learning rate set to 0.5, and usually converge within the threshold range after 3 to 4 iterations. The dynamic modeling and closed-loop adjustment mechanism ensure the balance of torque distribution during the grasping process.

[0085] S1054. For the joint torque that satisfies the torque balance condition, use a cubic spline curve to optimize the joint angle trajectory and generate an execution instruction for torque balance control.

[0086] The trajectory optimization uses a cubic spline curve to interpolate the joint angles, with an interpolation point interval of 0.1 radian, ensuring the continuity of the second derivative of the trajectory to achieve smooth transition. Based on the Euler-Lagrange variational principle, an optimization objective function is constructed, comprehensively considering the trajectory smoothness and execution time, and the conjugate gradient method is used to solve the optimal grasping trajectory. The generated joint motion sequence is constrained by a velocity planning algorithm to control the angular velocity within 2 radians per second, and the acceleration fluctuation is less than 5 radians per square second. The optimized trajectory can achieve smooth force transmission when grasping fragile objects such as glass cups, avoid stress concentration, and ensure the fluency and safety of the grasping action.

[0087] Tests show that when grasping a hard plastic and metal composite object, by dynamically adjusting the impedance parameters, the torque imbalance in the stiffness mutation area can be effectively dealt with, and the torque balance converges to within 2 Newton meters within 3 iterations, and the grasping success rate is increased to more than 95%. This method provides efficient support for the robot's fine grasping through the combination of state mapping and trajectory optimization, especially showing good adaptability in complex material scenarios.

[0088] S106. Adopt a preset grasping control strategy, according to the execution instruction of torque balance control, correct the grasping action of the end effector of the humanoid robot, and judge whether the grasping control accuracy meets the standard according to the corrected grasping action. If the grasping control accuracy does not meet the standard, trigger conditions for grasping failure detection are constructed.

[0089] Based on the torque balance control instruction, combined with the preset grasping control strategy, optimize the grasping performance of the end effector through real-time correction of the grasping action and tactile feedback analysis. If the grasping control accuracy meets the standard, grasp according to the corrected grasping method of the end effector of the humanoid robot. This step solves the problem of insufficient accuracy caused by material changes or unstable contact during the grasping process through closed-loop control and dynamic parameter adjustment, and provides technical support for high-precision grasping in complex scenarios.

[0090] S1061. According to the execution instruction of torque balance control and the preset grasping control threshold, use a proportional-integral controller to adjust the joint angle of the actuator in real time and generate a corrected grasping pose sequence.

[0091] In the embodiment of this application, the grasping correction amount is calculated by comparing the execution instruction of torque balance control with a preset grasping control threshold, and the threshold is set according to the material characteristics of the object. For example, the position accuracy threshold for metal products is 0.1 mm, and the torque accuracy threshold is 0.5 N·m; the position accuracy for rubber products is relaxed to 0.5 mm, and the torque accuracy is 0.2 N·m. A proportional-integral controller is used to adjust the actuator joint angle in real time. The control period is 2 ms, the initial proportional coefficient is set to 5, and the integral coefficient is 0.1. It is dynamically optimized through an adaptive adjustment mechanism. The proportional coefficient varies within the range of 2 to 10, and the integral coefficient is adjusted between 0.05 and 0.2. In the scenario of grasping a glass, the initial position error is 0.3 mm. Through 200 ms of dynamic adjustment, the proportional coefficient increases to 7.5, and the integral coefficient is adjusted to 0.15, reducing the error to 0.08 mm, generating a corrected grasping pose sequence, which provides a reliable input for accuracy evaluation.

[0092] S1062. Collect tactile sensor data for the corrected grasping pose sequence, extract the grasping contact feature vector using wavelet transform, and evaluate the grasping stability through the root mean square error.

[0093] According to the corrected grasping pose sequence, the tactile sensor collects data at a frequency of 500 times per second, generating a high-resolution tactile signal sequence. The db4 wavelet basis function is used to perform four-layer wavelet decomposition on the signal. The first layer extracts the high-frequency contact features of 50 to 100 Hz, reflecting the instantaneous contact changes; the fourth layer extracts the low-frequency pressure features of 3 to 6 Hz, characterizing the overall pressure distribution. By reconstructing the decomposition coefficients, a 64-dimensional grasping contact feature vector is generated, covering the pressure amplitude and frequency characteristics of the contact points. The grasping stability calculates the root mean square error using the sliding window method. The window length is 100 ms, and the position error and torque error are evaluated respectively. A comprehensive score greater than 0.9 is considered to meet the accuracy standard, and less than 0.7 triggers the grasping failure detection. Tests show that under the stable grasping state, the amplitude of the high-frequency component remains within 10% of the sensor range, while when the grasping slips, the high-frequency component suddenly increases to 30%, and the low-frequency component shows periodic fluctuations. The feature vector effectively captures these abnormal characteristics.

[0094] S1063. Judge the control accuracy according to the grasping stability. If the accuracy does not meet the standard, use a support vector regressor to construct a tactile predictor, extract the grasping deviation features, and adjust the stiffness parameter weights.

[0095] If the grasping stability is less than 0.7, it indicates that the grasping accuracy does not meet the standard, and it is necessary to further analyze the tactile signals to optimize the control parameters. A support vector regressor is used to construct a tactile predictor. The radial basis kernel function is selected, the kernel parameter is set to 0.8, the penalty factor is 100, and a 64-dimensional tactile feature vector is input to predict the tactile response in the next 100 milliseconds. If the prediction error exceeds 30%, it is determined that the grasping is abnormal. The deviation features are extracted to construct a failure detection feature set, the weight update amount of the stiffness parameter is calculated, and iterative optimization is carried out using an exponential decay learning rate. The initial learning rate is set to 0.1, the decay coefficient is 0.95, and it decreases by 5% each iteration. The weight coefficient matrix is 3×3, corresponding to the stiffness parameters of the three joints of the end effector respectively. The updated weights are used to compensate the grasping control parameters, recalculate the grasping correction amount and generate the trigger condition for grasping failure detection. During the test, when grasping a glass and slipping occurs, the prediction model warns of the failure risk 50 milliseconds in advance. After the weight adjustment, the position error is reduced from 0.2 mm to 0.09 mm, significantly improving the grasping stability.

[0096] If the grasping stability is greater than the preset grasping control threshold of 0.9, it indicates that the grasping action has been optimized. Record the optimized grasping parameter set, including joint angles, torque values, and stiffness weights, and output the grasping optimization completion flag. This method ensures the adaptability of the grasping action to objects of different materials through the fine analysis of tactile signals and dynamic parameter adjustment. Especially when grasping fragile or flexible objects, it avoids stress concentration and slipping phenomena, providing efficient support for the fine operation of the robot.

[0097] S107. Optimize the grasping control strategy according to the trigger condition of grasping failure detection, regenerate the action sequence of the end effector of the humanoid robot based on the optimized grasping control strategy, and determine the final grasping control parameters of each joint of the end effector of the humanoid robot.

[0098] Based on the trigger condition of grasping failure detection, the grasping strategy is dynamically optimized through deep reinforcement learning to generate a stable action sequence to ensure the stability of the object posture, improving the self-adaptability and robustness of the grasping control, and providing efficient support for high-precision grasping in complex scenarios.

[0099] S1071. Construct a robot state vector according to the grasping state parameters, optimize the grasping strategy using the deep deterministic policy gradient algorithm, and generate joint action instructions.

[0100] The robot state vector contains 16-dimensional features, where 6 dimensions represent joint angles, 6 dimensions represent angular velocities, and 4 dimensions represent joint forces, comprehensively describing the motion and mechanical states of the end effector. The deep deterministic policy gradient algorithm is used to construct the policy network, which contains two hidden layers with 128 neurons in each layer. The ReLU activation function is used to enhance the non-linear expression ability. The action value evaluation adopts a dual value network architecture, and by comparing the outputs of the main value network and the target value network, the overestimation problem of the action value is reduced. During the training process, the experience replay buffer stores 10,000 state-action pairs, the batch size is set to 64, the learning rate is set to 0.001, and after 1000 rounds of training, the policy network can stably output joint action commands, including the target angle and execution time, providing accurate input for trajectory generation.

[0101] S1072. Interpolate the joint action commands using cubic spline curves to generate an angular displacement sequence, and calculate the pose matrix of the end effector through forward kinematics.

[0102] The joint action commands generate a smooth angular displacement sequence through cubic spline curve interpolation. The interpolation point interval is 20 milliseconds, and the maximum angular acceleration is constrained within 50 radians per second squared to ensure the smoothness and dynamic controllability of the trajectory. Based on the angular displacement sequence, the forward kinematics equation is used to calculate the pose matrix of the end effector. The position and orientation are described by homogeneous transformation matrices, considering the coupling effect between joints. The improved iterative method is used in the solution process, the convergence threshold is set to 0.01 mm, the pose calculation accuracy reaches 0.1 mm, and the attitude accuracy is better than 0.1 degree. The generated pose matrix of the end effector can accurately reflect the spatial state of the object during the grasping process, providing high-precision data for stability evaluation.

[0103] S1073. Use a Kalman filter to suppress the noise of the pose matrix of the end effector, evaluate the attitude stability through covariance eigenvalues, and adjust the control parameters according to the stability results.

[0104] The Kalman filter is used to filter the noise in the pose matrix. The observation noise variance is set to 0.01, and the process noise variance is set to 0.001. The filtered data can effectively suppress environmental interference and sensor noise. The stability assessment is carried out by calculating the covariance eigenvalues of the pose matrix. A maximum eigenvalue less than 0.1 indicates a stable pose, and a value greater than 0.3 triggers the adjustment of control parameters. The adjustment process uses the gradient descent method to optimize the joint stiffness coefficient and damping coefficient. The stiffness coefficient ranges from 50 to 200 Newton - meters per radian, and the damping coefficient ranges from 0.5 to 2 Newton - meters per radian per second. The initial learning rate is 0.1, which decays exponentially at a rate of 0.95 with the number of iterations. In the glass - grasping experiment, the initial pose fluctuation caused the covariance eigenvalue to reach 0.45. By adjusting the finger joint stiffness to 150 Newton - meters per radian and the damping coefficient to 1.5 Newton - meters per radian per second, the eigenvalue dropped to 0.08, significantly improving the pose stability.

[0105] S1074. Use a trajectory smoother to process the adjusted action sequence, generate a smooth grasping action, and verify the final grasping control parameters of each joint of the humanoid robot's end - effector through pose deviation.

[0106] The trajectory smoother combines five - point averaging filtering and a second - order Butterworth low - pass filter to process the action sequence. The filter cut - off frequency is set to 10 Hz. The maximum acceleration of the filtered trajectory is reduced to 30 radians per second squared, eliminating trajectory jumps and jitters. The smoothed action sequence is used to recalculate the pose matrix. The pose deviation is evaluated through the root - mean - square error, and the deviation is controlled within 0.05 mm, and the amplitude of pose fluctuation is less than 0.1 degree, meeting the requirements for stable grasping. The final grasping control parameters of each joint of the humanoid robot's end - effector include the optimized joint angles, stiffness coefficients, and damping coefficients, which are recorded as a reference for the grasping task. Tests show that for rigid objects, the policy network tends to adopt a high - stiffness and low - damping strategy; for flexible objects, a low - stiffness and high - damping combination is selected, demonstrating the ability of adaptive adjustment.

[0107] The present invention provides an intelligent grasping control system based on tactile perception, and the system includes:

[0108] A tactile signal processing module, which is used to obtain tactile feedback signals through the sensor array of the humanoid robot, perform frequency - domain decomposition on the tactile feedback signals using the Fourier transform method, and calculate the frequency - feature distribution of the tactile feedback signals by combining the pressure and deformation data analysis of each contact point on the surface of the object to be grasped.

[0109] A frequency - feature extraction module, which is used to capture the local and global frequency features of the tactile feedback signals by fusing convolution kernels with different receptive fields according to the frequency - feature distribution of the tactile feedback signals, perform hierarchical classification on the stiffness gradient of the object to be grasped, and determine a preliminary estimated value of the object stiffness gradient.

[0110] A stiffness gradient analysis module, which is configured to, if the preliminary estimated value of the object stiffness gradient exceeds a threshold, calculate the object stiffness gradient distribution by combining the time-domain processing result of the tactile feedback signal and the surface geometric profile of the object to be grasped;

[0111] A deformation field modeling module, which is configured to probabilistically model the deformation field according to the object stiffness gradient distribution by using the Bayesian inference method, and perform iterative optimization on the modeling result to obtain the dynamic update parameters of the deformation field mapping;

[0112] An impedance control module, which is configured to, through the dynamic update parameters of the deformation field mapping, use the dynamic update parameters of the deformation field mapping as the state input and the impedance parameters of each joint of the end effector of the humanoid robot as the action output. If the adjusted impedance parameters meet the preset torque balance condition, optimize the grasping trajectory of the joint angles of the end effector of the humanoid robot to obtain the execution instruction for torque balance control;

[0113] A grasping control module, which is configured to adopt a preset grasping control strategy, correct the grasping action of the end effector of the humanoid robot according to the execution instruction of the torque balance control, judge the compliance of the grasping control accuracy according to the corrected grasping action. If the grasping control accuracy does not meet the standard, construct the triggering condition for grasping failure detection;

[0114] A grasping optimization module, which is configured to optimize the grasping control strategy according to the triggering condition for grasping failure detection, regenerate the action sequence of the end effector of the humanoid robot based on the optimized grasping control strategy, and determine the final grasping control parameters of each joint of the end effector of the humanoid robot.

[0115] An intelligent grasping control device based on tactile perception, the intelligent grasping control device includes: one or more processors; a storage device for storing one or more programs;

[0116] When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the above method for grasping control.

[0117] In addition, it should be noted that, among the various specific technical features described in the above specific embodiments, they can be combined in any suitable manner without conflict. To avoid unnecessary repetition, the present invention does not separately describe various possible combination methods. In addition, any combination can be made between various different embodiments of the present invention as long as it does not violate the idea of the present invention, and it should also be regarded as the content disclosed by the present invention.

Claims

1. An intelligent grasping control method based on tactile perception, characterized in that, The method includes: Obtaining a tactile feedback signal through a sensor array of a humanoid robot, performing frequency-domain decomposition on the tactile feedback signal using the Fourier transform method, and calculating the frequency feature distribution of the tactile feedback signal by combining the pressure and deformation data analysis of each contact point on the surface of the object to be grasped; According to the frequency feature distribution of the tactile feedback signal, capturing the local and global frequency features of the tactile feedback signal by fusing convolution kernels with different receptive fields, hierarchically classifying the stiffness gradient of the object to be grasped, and determining a preliminary estimated value of the object stiffness gradient; If the preliminary estimated value of the object stiffness gradient exceeds the threshold, then calculate the object stiffness gradient distribution by combining the time-domain processing result of the tactile feedback signal and the geometric contour of the surface of the object to be grasped; According to the object stiffness gradient distribution, perform probabilistic modeling on the deformation field using the Bayesian inference method, and calculate the dynamic update parameters of the deformation field mapping; Taking the dynamic update parameters of the deformation field mapping as the state input and the impedance parameters of each joint of the end effector of the humanoid robot as the action output, if the adjusted impedance parameters meet the preset torque balance condition, then optimize the grasping trajectory of the joint angles of the end effector of the humanoid robot to obtain an execution instruction for torque balance control; Adopting a preset grasping control strategy, according to the execution instruction of the torque balance control, correcting the grasping action of the end effector of the humanoid robot, judging the compliance of the grasping control accuracy according to the corrected grasping action, and if the grasping control accuracy does not meet the standard, then constructing a trigger condition for grasping failure detection; Optimizing the grasping control strategy according to the trigger condition of the grasping failure detection, regenerating the action sequence of the end effector of the humanoid robot based on the optimized grasping control strategy, and determining the final grasping control parameters of each joint of the end effector of the humanoid robot.

2. The method according to claim 1, wherein The obtaining a tactile feedback signal through a sensor array of a humanoid robot, performing frequency-domain decomposition on the tactile feedback signal using the Fourier transform method, and calculating the frequency feature distribution of the tactile feedback signal by combining the pressure and deformation data analysis of each contact point on the surface of the object to be grasped includes: Collecting the pressure of the tactile points on the surface of the object to be grasped to obtain the time-series data of the tactile point pressure, and the time-series data of the tactile point pressure includes the position coordinates and pressure values of the tactile contact points; Performing filtering processing according to the time-series data of the tactile point pressure, and performing a fast Fourier transform on the filtered tactile signal to obtain frequency-domain features, and the frequency-domain features include the main frequency component and the amplitude; Constructing a spatial discrete grid for the position coordinates of the tactile contact points, calculating the pressure values of the grid nodes using the bicubic spline interpolation algorithm to obtain a pressure distribution surface, and solving the deformation displacement field according to the pressure distribution surface; Fusing the frequency-domain features and the deformation displacement field using the eigenvector splicing method, and calculating the frequency feature distribution of the tactile feedback signal.

3. The method according to claim 1, characterized in that, The according to the frequency feature distribution of the tactile feedback signal, capturing the local and global frequency features of the tactile feedback signal by fusing convolution kernels with different receptive fields, hierarchically classifying the stiffness gradient of the object to be grasped, and determining a preliminary estimated value of the object stiffness gradient includes: Divide the small-scale receptive field and the large-scale receptive field according to the amplitude of the tactile feedback signal, and use the small-scale receptive field and the large-scale receptive field to extract features from the tactile feedback signal to obtain a dual-scale tactile feature map; Obtain the parameters of the first three convolutional kernels of the pre-trained image recognition model for the dual-scale tactile feature map, and perform multi-level convolutional operations on the dual-scale tactile feature map through the convolutional kernel parameters to obtain a multi-layer tactile feature map group; If there is a low-resolution feature map in the feature map group, use the upsampling algorithm to magnify the low-resolution feature map, and obtain a multi-scale fusion feature map through cross-layer feature weighted fusion; Perform non-linear transformation and max-pooling operations on the multi-scale fusion feature map, divide different stiffness intervals through the stiffness change threshold, and output a preliminary estimate of the object stiffness gradient.

4. The method according to claim 1, wherein If the preliminary estimate of the object stiffness gradient exceeds the threshold, then calculate the object stiffness gradient distribution by combining the time-domain processing result of the tactile feedback signal and the geometric contour of the object to be grasped, including: Perform time-domain sampling on the tactile feedback signal sent by the tactile sensor array, and perform filtering processing on the time-domain sampled data to obtain a filtered time-domain signal sequence; Calculate the object surface pressure distribution map according to the filtered time-domain signal sequence, and perform three-dimensional reconstruction on the pressure distribution map to obtain the object surface geometric contour data; Establish an elastic constitutive equation according to the object surface geometric contour data, and solve the elastic constitutive equation to obtain the grid node displacement field data; Combine the filtered time-domain signal sequence with the grid node displacement field data to calculate the object stiffness gradient distribution.

5. The method according to claim 4, characterized in that, According to the object stiffness gradient distribution, use the Bayesian inference method to perform probability modeling on the deformation field, and calculate the dynamic update parameters of the deformation field mapping, including: Establish an initial deformation field probability density function according to the object stiffness gradient distribution; Use the Bayesian inference method to update the deformation field probability density function, and obtain the posterior probability distribution of the node displacement through Markov chain sampling; Establish a deformation field mapping matrix according to the posterior probability distribution of the node displacement, perform iterative optimization on the mapping matrix parameters, and calculate the dynamic update parameters of the deformation field mapping.

6. The method according to claim 1, wherein Use the dynamic update parameters of the deformation field mapping as the state input, and the impedance parameters of each joint of the end effector of the humanoid robot as the action output. If the adjusted impedance parameters meet the preset torque balance condition, then optimize the grasping trajectory of the joint angle of the end effector of the humanoid robot to obtain an execution instruction for torque balance control, including: Construct a state input vector according to the dynamic update parameters of the deformation field mapping, construct a multi-layer perceptron state-action mapper, and obtain the impedance parameters of each joint of the end effector of the humanoid robot from the state-action mapper; Construct an actuator dynamics equation according to the joint impedance parameters, and solve the joint torque value from the actuator dynamics equation; If the joint torque value satisfies the condition that the torque balance degree is less than the preset deviation threshold, then perform interpolation calculation on the actuator joint angle, obtain the optimal grasping trajectory from the interpolation calculation result, and generate an execution instruction for torque balance control.

7. The method according to claim 1, characterized in that, Adopt a preset grasping control strategy, correct the grasping action of the end effector of the humanoid robot according to the execution instruction of torque balance control, and judge the compliance of the grasping control accuracy according to the corrected grasping action. If the grasping control accuracy does not meet the standard, trigger conditions for grasping failure detection are constructed, including: Construct an actuator grasping correction amount according to the torque balance control instruction and the preset grasping control threshold, and use a proportional-integral controller to adjust the actuator joint angle in real time according to the actuator grasping correction amount to obtain a corrected grasping pose sequence; Collect tactile sensor data for the corrected grasping pose, and decompose the tactile feedback signal through wavelet transform to obtain a grasping contact feature vector; Calculate a control accuracy index according to the grasping contact feature vector, and use the root mean square error to evaluate the control accuracy index to obtain the grasping stability; If the grasping stability is less than the preset grasping control threshold, construct a tactile predictor based on the grasping contact feature vector, and compare the deviation between the predicted result of touch and the corrected grasping pose; Determine the stiffness parameters to be adjusted according to the deviation situation, compensate and adjust the grasping control parameters based on the stiffness parameters to be adjusted, recalculate the grasping correction amount, and construct the trigger conditions for grasping failure detection.

8. The method according to claim 1, wherein Optimize the grasping control strategy according to the trigger conditions for grasping failure detection, regenerate the action sequence of the end effector of the humanoid robot based on the optimized grasping control strategy, and determine the final grasping control parameters of each joint of the end effector of the humanoid robot, including: Construct a robot state vector according to the trigger conditions for grasping failure detection, and use the deep deterministic policy gradient algorithm to optimize the policy of the robot state vector to obtain a policy network; Output joint action instructions according to the policy network, perform trajectory interpolation on the joint action instructions using a cubic spline curve to obtain an angular displacement sequence, and calculate the pose matrix of the end effector through the angular displacement sequence; Use a Kalman filter to filter the data of the pose matrix of the end effector, and judge whether the attitude fluctuation is greater than the preset stability threshold by calculating the covariance eigenvalue of the pose matrix of the end effector; If the attitude fluctuation is greater than the preset stability threshold, adjust the stiffness coefficient and damping coefficient of each joint of the end effector of the humanoid robot to generate the final grasping control parameters of each joint of the end effector of the humanoid robot.

9. An intelligent grasping control system based on tactile perception, characterized in that, The system includes: A tactile signal processing module, which is used to obtain tactile feedback signals through the sensor array of the humanoid robot, perform frequency domain decomposition on the tactile feedback signals using the Fourier transform method, and calculate the frequency feature distribution of the tactile feedback signals by combining the pressure and deformation data analysis of each contact point on the surface of the object to be grasped; A frequency feature extraction module, which is used to capture the local and global frequency features of the tactile feedback signal by fusing convolution kernels with different receptive fields according to the frequency feature distribution of the tactile feedback signal, classify the stiffness gradient of the object to be grasped in layers, and determine a preliminary estimated value of the object stiffness gradient; A stiffness gradient analysis module, configured to, if a preliminary estimated value of the stiffness gradient of an object exceeds a threshold, calculate the stiffness gradient distribution of the object by combining the time-domain processing result of the tactile feedback signal and the surface geometric profile of the object to be grasped; A deformation field modeling module, configured to perform probabilistic modeling on the deformation field according to the stiffness gradient distribution of the object by using a Bayesian inference method, and perform iterative optimization on the modeling result to obtain dynamic update parameters of the deformation field mapping; An impedance control module, configured to, by using the dynamic update parameters of the deformation field mapping, use the dynamic update parameters of the deformation field mapping as state inputs and use the impedance parameters of each joint of the end effector of the humanoid robot as action outputs. If the adjusted impedance parameters meet the preset torque balance condition, optimize the grasping trajectory of the joint angles of the end effector of the humanoid robot to obtain an execution instruction for torque balance control; A grasping control module, configured to adopt a preset grasping control strategy, correct the grasping action of the end effector of the humanoid robot according to the execution instruction of the torque balance control, judge the compliance of the grasping control accuracy according to the corrected grasping action, and if the grasping control accuracy does not meet the standard, construct a trigger condition for grasping failure detection; A grasping optimization module, configured to optimize the grasping control strategy according to the trigger condition for grasping failure detection, regenerate an action sequence of the end effector of the humanoid robot based on the optimized grasping control strategy, and determine the final grasping control parameters of each joint of the end effector of the humanoid robot.

10. An intelligent grasping control device based on tactile perception, characterized in that, The intelligent grasping control device includes: one or more processors; a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the method according to any one of claims 1-8 for grasping control.

Images