Deep learning method for hole drilling movement of humanoid robot

Abstract

The application relates to the technical field of computer systems, and discloses a deep learning method for hole drilling movement of a humanoid robot, which comprises the following steps: collecting a robot tactile pressure space-time tensor and a laser point cloud intensity sequence, and concatenating a geometric manifold model to extract a spatial local curvature gradient vector; calculating a weight entropy increase rate of a hidden layer neuron according to the curvature gradient vector; comparing the weight entropy increase rate with a safe convergence threshold value; when the weight entropy increase rate is lower than the threshold value, updating an online gradient; when the weight entropy increase rate reaches or exceeds the threshold value, activating a truncation rule and locking a connection weight matrix parameter; and using the modulated or locked connection weight matrix for forward reasoning, outputting a joint trajectory error compensation control tensor to constrain a servo mechanism, so that the application eliminates semantic drift in a representation space, maintains smooth activation of neurons when a perception mode degenerates, and solves a control failure problem.

Description

Technical Field

[0001] This invention belongs to the field of computer system technology, and in particular relates to a deep learning method for humanoid robots to move through holes. Background Technology

[0002] Current multi-layer residual neural network computer systems solve multimodal spatial perception input feature matrices by topologically cascading hierarchical neuron nodes, construct motion representation spaces, and provide joint servo control command outputs. Under standard operating conditions, such computational models implicitly rely on continuous high-resolution features provided by the perception system, assuming that the external channel environment has geometric features that can be completely reconstructed and stable and distinguishable feature density.

[0003] However, when humanoid robots move through unstructured, physically confined spaces such as narrow cavities, the lack of external light sources, the obscuring effect of suspended particles, and high-frequency random friction on the robot's surface cause the heterogeneous multimodal feature stream output by the perception system to exhibit a degenerate form characterized by spatial sparseness, low signal-to-noise ratio, and localized fragmentation of geometric topology. Due to the lack of a self-balancing constraint mechanism for the nonlinear loss in the feature space in the computational model, semantic drift occurs when the hidden layer neurons solve the above-mentioned unevenly degenerate features. The neuron activation vectors are prone to saturation deadlock or gradient abrupt divergence, leading to interruption of the motion control chain or joint torque overshoot. To address the above-mentioned computational degradation, conventional improvement strategies tend to increase the depth of the hidden layer of the network to enhance the fitting ability, or add multi-sensor hardware to supplement redundant features. However, unilaterally increasing the parameters of the hidden layer spontaneously exacerbates the risk of overfitting the model when the features are broken over a large area, and increases the chip's computational latency. Adding hardware is limited by the physical volume and power consumption of the robot's operation, resulting in a static configuration between the topology adaptation capability and the boundary of computational resources, making it impossible to stabilize the neuron activation state from within the network topology.

[0004] Therefore, how to construct a structural self-balancing mechanism for the hidden layer feature manifold and an adaptive dissipation constraint mechanism for the network weight space, so as to achieve continuous spatial topological alignment of high-dimensional heterogeneous perceptual features and dynamic discrete truncation of backpropagation parameters, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention aims to solve the problems of spatial drift of intermediate parameters in deep learning networks and control output failure caused by weight cascade divergence when multimodal sensing data undergoes uneven degradation.

[0006] In this technical solution, a deep learning method for humanoid robots to move through holes includes the following steps: Step S1: Collect the spatiotemporal feature tensor of tactile pressure distributed on the robot surface and the intensity sequence of sparsed spatial laser point cloud output by the detection link. Step S2: The spatiotemporal feature tensor of tactile pressure and the intensity sequence of sparse spatial laser point cloud are concatenated and input into the geometric manifold model to construct the feature embedding manifold space. In the feature embedding manifold space, the spatial local curvature gradient vector of tactile pressure spatiotemporal feature tensor is extracted, and the discrete physical touch is converted into a geometric feature quantity that evolves continuously on the manifold surface. Step S3: Calculate the weight entropy increase rate of neuron nodes within the connection weight matrix of the hidden layer of the deep neural network based on the spatial local curvature gradient vector. Step S4: Compare the weight entropy increase rate with the safe convergence threshold. When the weight entropy increase rate is lower than the safe convergence threshold, update the online gradient of the connection weight matrix. When the weight entropy increase rate reaches or exceeds the safe convergence threshold, activate the weight nonlinear discrete truncation rule, suspend the gradient backpropagation of the current sampling period, and lock the parameter value of the connection weight matrix to the weight sedimentation value of the previous convergence time sequence. Step S5: Using the modulated or locked connection weight matrix for forward inference, output the joint trajectory error compensation control tensor to constrain the underlying servo mechanism.

[0007] Preferably, the activation of the nonlinear discrete truncation rule for weights in step S4 includes the following sub-steps: Step S41, when the weight entropy increase rate reaches or exceeds the safe convergence threshold, the backpropagation of the current time series full network feature gradient is paused, and the gradient update of the connected weight matrix is ​​rigidly suspended within the subsequent consecutive preset number of sampling time series windows to maintain the frozen protection state of the weight parameters. The weight precipitated value stored in the previous stable time series is called to complete the forward inference.

[0008] Preferably, after step S1, the following steps are also included: Step S11, using the feature space self-similarity metric operator to extract the first derivative of the tactile pressure spatiotemporal feature tensor to determine the tactile manifold energy index; Step S12, simultaneously extracting the spectral envelope similarity of the sparse spatial laser point cloud intensity sequence to determine the echo topology density index; Step S13, calculating the perceptual spatial manifold distortion scalar based on the real-time ratio of the tactile manifold energy index to the echo topology density index.

[0009] Preferably, step S2 further includes the following steps: Step S21, when the scalar of the perceptual space manifold distortion exceeds the preset geometric space mutation critical safety threshold of 2.50, the weight multiplication and addition operation of the hidden layer is stopped, and the orthogonal projection operation of the hidden layer feature manifold is started; Step S22, the orthogonal transformation basis matrix preset in the memory is called to project the global multimodal hidden representation tensor onto the low-dimensional self-similar feature subspace defined by the null term of the echo topological density index.

[0010] Preferably, step S22 includes the following sub-steps: Step S221, injecting a binary topological mask matrix into the low-dimensional self-similar feature subspace to dynamically modulate the connection weight matrix; Step S222, when the perceptual space manifold distortion scalar reaches or exceeds the geometric space abrupt change critical safety threshold of 2.50, changing the control bit element in the binary topological mask matrix corresponding to the image feature flow path from value 1 to value 0, thereby deactivating the visual processing sub-network inside the deep neural network; Step S223, keeping the control bit element in the binary topological mask matrix corresponding to the tactile and joint state perception path at value 1, and increasing the corresponding activation gain coefficient.

[0011] Preferably, step S3 further includes the following sub-steps: step S32, calculating the local spatiotemporal variance of the tactile pressure spatiotemporal feature tensor within a preset sliding time window; step S33, monotonically adjusting the spatiotemporal attention allocation matrix of the hidden neurons according to the reciprocal of the local spatiotemporal variance, and superimposing the Frobenius norm of the spatiotemporal attention allocation matrix in parallel in the total loss function of the deep neural network as a spatiotemporal regularization damping penalty term.

[0012] Preferably, the spatiotemporal regularization damping penalty term in sub-step S33 is determined by multiplying the dynamic adaptive adjustment coefficient by the square of the Frobenius norm of the spatiotemporal attention allocation matrix to obtain the spatiotemporal regularization damping penalty term, wherein the value of the dynamic adaptive adjustment coefficient is configured to be between 0.01 and 0.05.

[0013] Preferably, step S5 includes the following pre-steps: Step S51, embedding the joint physical interference manifold boundary model in the output layer, retrieving the joint spatial motion error compensation control instruction matrix generated by the current inference to calculate the predicted joint output torque tensor; Step S52, constructing a three-dimensional nonlinear difference between the predicted joint output torque tensor and the preset physical safety torque boundary tensor, and extracting the magnitude of the three-dimensional nonlinear difference to generate an interference overflow scalar.

[0014] Preferably, sub-step S52 includes the following processing: step S521, adding an edge gradient penalty operator for the interference overflow scalar in parallel to the total loss function of the deep neural network; step S522, setting the edge gradient penalty operator to 0 when the interference overflow scalar is less than or equal to 0; when the interference overflow scalar is greater than 0 and monotonically increasing, generating a quadratic step-like amplification loss function penalty weight based on the displacement of the interference overflow scalar exceeding 0, and superimposing the loss function penalty weight into the total loss function.

[0015] Compared with existing technologies, the deep learning method of this invention for humanoid robots to move through holes has the following advantages: 1. In deep learning for humanoid robots to move through holes, the contact stress spatiotemporal tensor and discrete spatial obstacle echo sequence are aligned across domains within the feature embedding manifold space. The computational model transforms the heterogeneous multimodal perception flow into a topologically aligned high-dimensional cascaded topological matrix. By extracting the spatial local curvature gradient vector, the physical tactile sensation of fracture and the discontinuous distance echo are transformed into geometric features that evolve continuously on the manifold surface. This eliminates the semantic drift of the representation space within the hidden layer, enabling the network to maintain the smooth activation state of neurons even under the condition of perception modality degradation. This ensures the stability of information mapping at subsequent processing ports and resolves the neuron activation deadlock caused by the uneven dissipation of the multimodal feature space.

[0016] 2. By introducing the feature weight manifold evolution operator, the computation model calculates the real-time entropy increase rate of the hidden layer parameter state transformation in the deep network. When a sudden change in the external physical environment causes a nonlinear sudden distortion in the perceptual input features, the nonlinear discrete truncation rule of the connection weights generates a step response based on the threshold of the real-time entropy increase rate. This promptly pauses the gradient backpropagation of the current temporal network feature weights and forcibly locks the value of the neuron connection weight matrix to the fixed parameter precipitate value stored in the previous convergence time sequence. This blocks the cascading diffusion of abnormal difference features between hidden layers and ensures the smooth convergence of the forward inference control instruction matrix at the output port.

[0017] 3. By introducing a spatiotemporal attention allocation matrix into the hidden layer of the deep residual topological neural network, the system maps the local spatiotemporal variance output by the tactile sensor array to the dynamic elements of the attention allocation matrix. The Frobenius norm of the attention allocation matrix is ​​superimposed on the total loss function to construct a spatiotemporal regularization damping penalty term. This damping penalty term applies an asymmetric evolution constraint in the gradient backpropagation process, effectively constraining the random and drastic drift of the weights of hidden neurons caused by high-frequency contact noise. It maintains the topological continuity of the representation space when the body surface is subjected to unsteady friction, thereby improving the anti-disturbance stability and convergence performance of the control system under unknown and harsh working conditions. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the deep learning method for humanoid robot drilling and moving in the present invention; Figure 2 This is a topological diagram of the deep learning system structure for the humanoid robot's drilling and movement according to the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A deep learning method for humanoid robots to move through holes includes the following steps: Step S1: Collect the spatiotemporal feature tensor of tactile pressure distributed on the robot surface and the intensity sequence of sparsed spatial laser point cloud output by the detection link. Step S2: The spatiotemporal feature tensor of tactile pressure and the intensity sequence of sparse spatial laser point cloud are concatenated and input into the geometric manifold model to construct the feature embedding manifold space. In the feature embedding manifold space, the spatial local curvature gradient vector of tactile pressure spatiotemporal feature tensor is extracted, and the discrete physical touch is converted into a geometric feature quantity that evolves continuously on the manifold surface. Step S3: Calculate the weight entropy increase rate of neuron nodes within the connection weight matrix of the hidden layer of the deep neural network based on the spatial local curvature gradient vector. Step S4: Compare the weight entropy increase rate with the safe convergence threshold. When the weight entropy increase rate is lower than the safe convergence threshold, update the online gradient of the connection weight matrix. When the weight entropy increase rate reaches or exceeds the safe convergence threshold, activate the weight nonlinear discrete truncation rule, suspend the gradient backpropagation of the current sampling period, and lock the parameter value of the connection weight matrix to the weight sedimentation value of the previous convergence time sequence. Step S5: Using the modulated or locked connection weight matrix for forward inference, output the joint trajectory error compensation control tensor to constrain the underlying servo mechanism.

[0021] Preferably, the activation of the nonlinear discrete truncation rule for weights in step S4 includes the following sub-steps: Step S41, when the weight entropy increase rate reaches or exceeds the safe convergence threshold, the backpropagation of the current time series full network feature gradient is paused, and the gradient update of the connected weight matrix is ​​rigidly suspended within the subsequent consecutive preset number of sampling time series windows to maintain the frozen protection state of the weight parameters. The weight precipitated value stored in the previous stable time series is called to complete the forward inference.

[0022] Preferably, after step S1, the following steps are also included: Step S11, using the feature space self-similarity metric operator to extract the first derivative of the tactile pressure spatiotemporal feature tensor to determine the tactile manifold energy index; Step S12, simultaneously extracting the spectral envelope similarity of the sparse spatial laser point cloud intensity sequence to determine the echo topology density index; Step S13, calculating the perceptual spatial manifold distortion scalar based on the real-time ratio of the tactile manifold energy index to the echo topology density index.

[0023] Preferably, step S2 further includes the following steps: Step S21, when the scalar of the perceptual space manifold distortion exceeds the preset geometric space mutation critical safety threshold of 2.50, the weight multiplication and addition operation of the hidden layer is stopped, and the orthogonal projection operation of the hidden layer feature manifold is started; Step S22, the orthogonal transformation basis matrix preset in the memory is called to project the global multimodal hidden representation tensor onto the low-dimensional self-similar feature subspace defined by the null term of the echo topological density index.

[0024] Preferably, step S22 includes the following sub-steps: Step S221, injecting a binary topological mask matrix into the low-dimensional self-similar feature subspace to dynamically modulate the connection weight matrix; Step S222, when the perceptual space manifold distortion scalar reaches or exceeds the geometric space abrupt change critical safety threshold of 2.50, changing the control bit element in the binary topological mask matrix corresponding to the image feature flow path from value 1 to value 0, thereby deactivating the visual processing sub-network inside the deep neural network; Step S223, keeping the control bit element in the binary topological mask matrix corresponding to the tactile and joint state perception path at value 1, and increasing the corresponding activation gain coefficient.

[0025] Preferably, step S3 further includes the following sub-steps: step S32, calculating the local spatiotemporal variance of the tactile pressure spatiotemporal feature tensor within a preset sliding time window; step S33, monotonically adjusting the spatiotemporal attention allocation matrix of the hidden neurons according to the reciprocal of the local spatiotemporal variance, and superimposing the Frobenius norm of the spatiotemporal attention allocation matrix in parallel in the total loss function of the deep neural network as a spatiotemporal regularization damping penalty term.

[0026] Preferably, the spatiotemporal regularization damping penalty term in sub-step S33 is determined by multiplying the dynamic adaptive adjustment coefficient by the square of the Frobenius norm of the spatiotemporal attention allocation matrix to obtain the spatiotemporal regularization damping penalty term, wherein the value of the dynamic adaptive adjustment coefficient is configured to be between 0.01 and 0.05.

[0027] Preferably, step S5 includes the following pre-steps: Step S51, embedding the joint physical interference manifold boundary model in the output layer, retrieving the joint spatial motion error compensation control instruction matrix generated by the current inference to calculate the predicted joint output torque tensor; Step S52, constructing a three-dimensional nonlinear difference between the predicted joint output torque tensor and the preset physical safety torque boundary tensor, and extracting the magnitude of the three-dimensional nonlinear difference to generate an interference overflow scalar.

[0028] Preferably, sub-step S52 includes the following processing: step S521, adding an edge gradient penalty operator for the interference overflow scalar in parallel to the total loss function of the deep neural network; step S522, setting the edge gradient penalty operator to 0 when the interference overflow scalar is less than or equal to 0; when the interference overflow scalar is greater than 0 and monotonically increasing, generating a quadratic step-like amplification loss function penalty weight based on the displacement of the interference overflow scalar exceeding 0, and superimposing the loss function penalty weight into the total loss function.

[0029] Example 1: When a humanoid robot performs a drilling and movement task, it is constrained by the narrow physical cavity. The tactile sensor array on the surface of the robot collects the spatiotemporal feature tensor of tactile pressure at a frequency of 100Hz. At the same time, the forward laser detection link outputs a sparse spatial laser point cloud intensity sequence. The computer system concatenates the spatiotemporal feature tensor of tactile pressure with the sparse spatial laser point cloud intensity sequence through a geometric manifold model, and extracts the spatial local curvature gradient vector of the spatiotemporal feature tensor of tactile pressure within the feature embedding manifold space. Here, the unit configuration of the spatial local curvature gradient vector is N / m². Since the physical metric to geometric representation is aligned in the same dimension when the feature is embedded in the manifold space, the system uses the normal change rate of the manifold surface to directly map the change of the contact normal force per unit area on the solid surface of the sensor array. By introducing a unit stress transformation operator in the manifold base space, the geometric curvature gradient is monotonically and linearly scaled to the physical mechanical dimension, so that the gradient evolution of the geometric space can be directly unified in dimension and interact with the physical stress unit of the underlying servo mechanism.

[0030] Computer system based on Calculate the hidden layer connection weight matrix of a deep neural network Weight entropy increase rate of inner neuron nodes , For connecting the weight matrix The change in the statistical distribution divergence of the weights within the current sampling period, when Below the safe convergence threshold At that time, the system executes the connection weight matrix. Online gradient update; when Greater than or equal to the safe convergence threshold At this time, the system activates the nonlinear discrete truncation rule for weights, suspends the gradient backpropagation for the current sampling period, and connects the weight matrix. The weight parameter values ​​are locked to the weight saturation values ​​of the previous convergence time series. During the activation of the weighted nonlinear discrete truncation rule, the system call In the forward inference process, the weight snapshot value in this invention refers to a snapshot of the steady-state parameter matrix of the connection weight matrix in a historically stable convergence state during the online update or evolution of the topological parameters of the deep neural network. Specifically, in the continuous sampling sequence of the humanoid robot's drilling movement, as long as the computational unit measures that the weight entropy increase rate of the current sampling period is strictly lower than the safe convergence threshold, it means that no semantic drift has occurred in the current internal representation space. After completing the online update of the backpropagation gradient for the current period, the system will spontaneously use the updated connection weight matrix parameters as the latest steady-state benchmark. The data is synchronously dumped in real time to a dedicated cache area of ​​the edge computing unit in the form of numerical matrix snapshots. This cache area is equipped with a rolling overwrite mechanism based on a sliding time window to ensure that, under normal convergence conditions, the cache area always maintains the latest legal parameter values ​​that have passed the safety convergence verification. However, if a drastic change occurs in the external environment, causing the weight entropy increase rate of a certain sudden time series to exceed the limit, the nonlinear discrete truncation rule of the weight is instantly activated, and the computing system immediately issues a bus-level write protection command to the dedicated cache area, interrupting the rolling overwrite during that sampling period and subsequent suspended periods. The input behavior forces the data state in the cache to be frozen at the last convergence time sequence that passed the security check. At this time, the parameter matrix snapshot frozen and locked in the dedicated cache constitutes the weight sedimentation value. By directly calling up the weight sedimentation value to replace the current time sequence weight parameters that have been contaminated or diverged by the sudden degradation perception flow, forward inference can be performed. Under the extreme condition of cutting off the back propagation path, the parameter matrix with historical convergence stability can be used to block the cascade diffusion of abnormal difference features between hidden layer neurons, ensuring the smooth convergence of the joint trajectory error compensation control tensor. Furthermore, the process of calculating the weight entropy increase rate based on the spatial local curvature gradient vector is as follows: First, the L2 norm of the spatial local curvature gradient vector is calculated to characterize the intensity of topological mutation in the external physical environment; next, the probability distribution density function of the weights of each neuron node in the connection weight matrix within the current sampling period is retrieved, and the time-domain first derivative of this probability distribution is calculated to obtain the information entropy change rate of the parameter space; finally, the topological mutation intensity is used as a variable factor to perform a nonlinear mapping with the information entropy change rate to determine the specific value of the weight entropy increase rate. The computer system then uses the modulated or locked connection weight matrix... Perform forward inference and output the joint trajectory error compensation control tensor. To achieve input alignment of high-dimensional heterogeneous features and safety constraints at the output end, the geometric manifold model adopts a geometric alignment network with a three-layer fully connected feedforward topology. Its input end is connected to the high-dimensional feature vector after cascading the tactile and laser data streams. The hidden layer is configured with 64 neuron nodes, and the discrete features are transformed into continuous manifold quantities through nonlinear manifold mapping. The joint physical interference manifold boundary model, which is embedded in the position before the output layer, is composed of the preset joint mechanical limit geometric boundary equation. It is cascaded and embedded between the output mapping layer of the deep neural network and the bottom servo drive interface. It is used to directly read the action command and calculate the predicted joint output torque before the forward inference outputs the joint trajectory error compensation control tensor, thereby providing deterministic structural boundary data for the subsequent interference overflow verification.

[0031] Example 2: When the humanoid robot performs the drilling and movement task, the control system operates within a general-purpose edge computing unit. To verify the anti-interference capability and system stability of the humanoid robot control scheme in an unstructured, confined space, a standard humanoid robot frame was selected as the test platform. Gaussian white noise with a signal-to-noise ratio of 15dB was injected into the sensing link through software simulation to simulate the nonlinear degradation of sensing data caused by sand and dust media and sensor limitations. The channel's geometric diameter was physically constrained to 1.1 times the maximum cross-sectional width of the humanoid robot. Under this condition, the sensing system collects the spatiotemporal characteristic tensor of tactile pressure and the intensity sequence of sparse spatial laser point clouds. The sampling period is set to 10ms to meet the real-time requirements for responding to sudden collisions in a narrow physical space. In actual operation, the spatial local curvature gradient vector serves as the external sensing object. The uncertainty measure of the underlying layer is mapped in real time to the dissipative constraint index of the internal parameter space through the forward control loop. The system establishes a control gating mechanism for parameter space evolution using the weight entropy increase rate, and pre-verifies the parameter stability of the hidden layer before backpropagation is triggered. The curvature gradient mutation caused by the external environment will directly increase the weight entropy increase rate through forward calculation. When it reaches or exceeds the safe convergence threshold, the system does not rely on the traditional error loss function backpropagation, but forcibly suspends the online gradient update flow of the current sampling period through the underlying microcode instruction, thereby realizing the cross-scale hard constraint of the external sensed physical quantity on the optimization state of the internal algorithm. By constructing multiple test sample groups with gradient differences, the stability of this technical solution under different degrees of sense loss is verified: Test sample group: scalar of manifold distortion caused by sense mode degradation. When the value exceeds 2.50, the system automatically initiates the orthogonal projection procedure of the hidden layer feature manifold and calculates the value based on the weight entropy increase rate. The execution of nonlinear truncation logic, specifically the suspension of weight multiplication and addition operations in the hidden layer mentioned above, refers to pausing the regular forward multiplication and addition iteration calculations of a specific hidden layer kernel receiving the degenerate input stream of multimodal perception, in order to prevent the diffusion of degenerate features to subsequent stages. Simultaneously, the system switches to the projection calculation path in parallel, calling the pre-stored orthogonal transformation basis matrix in memory to perform orthogonal projection mapping on the previously stable time-buffered global multimodal hidden representation tensor. Without needing to perform regular multiplication and addition operations, the projected feature representation located in the low-dimensional self-similar feature subspace is directly calculated to replace the original hidden layer output, thus ensuring the continuity of the global multimodal hidden representation tensor chain. First control group: The system architecture remains unchanged, and the nonlinear discrete truncation rule for weights is disabled. The system maintains the gradient backpropagation path even when reaching 3.50 bit / s. Second control group: keeping the system architecture unchanged, the safe convergence threshold is adjusted. The sampling rate was set to 6.00 bit / s to examine the hysteresis effect of the model's response to weight fluctuations. During the experiment, the spatiotemporal characteristic tensor of the robot's tactile pressure was synchronously acquired with the laser point cloud intensity sequence. When the side of the robot came into contact with the protrusion of the hole wall in the 45th sampling period, the local stress value of the spatiotemporal characteristic tensor of the tactile pressure jumped instantaneously from 12.4 N to 148.5 N. The system calculated the following... Sudden changes in modulus lead to an increase in the weight entropy rate. It is 4.65 bits / s.

[0032] The data evolution trend is as follows: During the 46th to 52nd sampling period after the wall was touched, the experimental sample group activated the weighted nonlinear discrete truncation rule and connected the weight matrix. The gradient update is forcibly suspended. Specifically, the pre-set orthogonal transformation basis matrix is ​​the orthogonal feature vector subspace matrix extracted after singular value decomposition of the hidden layer feature representation matrix under standard interference-free conditions during the system initialization phase, and is pre-stored in the static memory of the general edge computing unit. The binary topological mask matrix is ​​a control matrix with the same dimension as the current hidden layer connection weight matrix. The configuration rule of its internal control bit elements is as follows: the control bit elements corresponding to the image feature flow path are physically mapped to the channel grid responsible for the forward convolution propagation of visual features. When the system detects a spatial abrupt change and changes the image path control bit from value 1 to value 0, the activation output of the visual subnetwork is forcibly cleared to zero through the matrix element-wise multiplication operation, thereby realizing the complete deactivation and path decoupling of the visual processing subnetwork. The joint trajectory error compensation control tensor The peak fluctuation amplitude was limited to within 1.2% of the initial mean, and no divergence of the model output layer action control commands occurred. In the first control group, due to continuous weight updates, the connection weight matrix was established 20ms after the wall contact. Nonlinear drift occurs, and the output joint trajectory error compensation control tensor A step overshoot occurred, with the joint drive torque instantaneously reaching the mechanical hard limit. The second control group showed that the safe convergence threshold was exceeded. The setting was too high, causing a lag in the system response. This resulted in three high-frequency reciprocating oscillations of the joint at the moment of contact with the wall. In the 53rd sampling period, the tactile stress dropped back to 18.2N, and the weighted entropy increase rate... The connection weight matrix was restored after the speed dropped to 2.10 bit / s. Online gradient update, joint trajectory error compensation control tensor The output smoothly passes through the convex region. The above data confirms that the weighted nonlinear discrete truncation rule, under sudden high-frequency noise interference, achieves smooth transition by connecting the weight matrix. By locking onto steady-state parameters, a balance is established between the uncertainty of perception and the convergence of control commands. At speeds exceeding 3.50 bit / s, the system eliminates semantic drift in the representation space by cutting off the backpropagation path, achieving deterministic output of the control quantity. Experimental results show that, under extreme physical constraints where over 60% of the perception modes are lost, this scheme ensures the joint trajectory error compensation control tensor. High-frequency smooth convergence.

[0033] Example 3: The current humanoid robot drilling motion control system is equipped with a sensor data fusion and compensation module. By constructing a perceptual feature space alignment model, it eliminates the semantic drift of the representation space when multimodal perceptual data undergoes non-uniform degradation. The computer system cyclically collects the tactile pressure spatiotemporal feature tensor and the laser point cloud intensity sequence at a 10ms cycle. To address the nonlinear degradation of features caused by uneven physical stress distribution in the perception link, the system establishes a perceptual feature space alignment model, concatenating the tactile pressure spatiotemporal feature tensor and the laser point cloud intensity sequence to construct a joint feature vector. ,in It is a joint mapping vector of tactile pressure and spatial geometric features, with dimension . The computer system makes logical judgments using the feature weight manifold evolution operator in the neural network. The computing unit first calculates the joint feature vector. Characteristic divergence , The Euclidean distance between the joint eigenvector of the current period and the joint eigenvector of the previous period, expressed as a dimensionless constant, is given by the eigendivergence. Greater than the preset convergence threshold When the system determines that the sensing mode has undergone non-uniform degradation, to ensure the determinism of the control command output, the system activates the weighted discrete truncation procedure. The specific judgment logic is as follows: When At this point, the system instructs the hidden layer node parameters of the neural network to be locked, forcibly stopping backpropagation gradient updates, and stabilizing the baseline matrix according to the pre-stored weights. The hidden layer weights are adjusted. Under the robot drilling and moving test conditions, the average weight matrix generated by statistical analysis after 500 stable moving cycles is used. In this process, the feature divergence and the weight entropy increase rate in the aforementioned embodiment together constitute the system's dual safety defense. The feature divergence, as a feedforward feature index at the input end, identifies the non-uniform degradation of the perception flow by calculating the step change in the spatial distance of heterogeneous data. When the feature divergence exceeds the threshold, the system will automatically link and trigger the parameter protection of the hidden layer. At this time, the weight entropy increase rate inside the network synchronously shows an out-of-bounds state, thereby enabling the feedforward warning and the nonlinear discrete truncation rule of the weight inside the network to achieve logical alignment and coordinated control, ensuring that the cascading divergence of abnormal features can be blocked at different levels.

[0034] The experimental group exhibits characteristic divergence due to sensory mode degradation. When the value exceeds 2.50, the system automatically triggers the weight discretization truncation procedure. The first control group disables this truncation procedure under the same conditions, causing divergence in the backpropagation gradient calculation and severe jitter in the joint trajectory error compensation control tensor of the output layer. The second control group sets the safe convergence threshold. The threshold was set to 6.00 bit / s. Due to the excessively large threshold, the system failed to lock the damaged weights in time, resulting in a three-cycle delay in the response of control commands. The engineering basis for setting the critical safety threshold for geometric spatial mutation to 2.50 is as follows: According to the statistics of collision dynamics simulation experiments of robots in physically confined channels, when the spatial manifold distortion scalar of the multimodal perception flow is below 2.00, the conventional residual correction mechanism can still maintain the convergence of control commands; however, once the scalar reaches or exceeds 2.50, the semantic representation of the hidden layer neurons will undergo irreversible saturation deadlock, causing the probability of collision risk to surge to over 85%. Therefore, configuring the threshold precisely to 2.50 can provide a computational protection margin before a large-scale loss of features occurs.

[0035] The experimental data are recorded as follows: Under the influence of power frequency interference and Gaussian white noise on the sensing link, the peak fluctuation amplitude of the joint trajectory error compensation control tensor of the sample group of this invention is limited to within 1.2% of the initial mean. The comparative experiment shows that the weighted nonlinear discrete truncation program, by suppressing the semantic drift of the representation space in real time, ensures the high-frequency smooth convergence of the control command under the extreme condition that more than 60% of the sensing mode is damaged. This invention solves the decision-making deadlock problem caused by the step degradation of the sensing flow by mapping the evolution process of the neural network connection weight space to quantifiable feature divergence control, and ensures the continuity and stability of the control logic in a narrow space.

[0036] Example 4: This example combines Figures 1 to 2 This section describes a deep learning method for humanoid robots to move through holes, such as... Figure 1 As shown, in step S1, the spatiotemporal feature tensor of tactile pressure distributed on the robot surface and the intensity sequence of sparse spatial laser point cloud output by the detection link are collected. Then, in step S2, the spatiotemporal feature tensor of tactile pressure and the intensity sequence of sparse spatial laser point cloud are concatenated and input into the geometric manifold model to construct the feature embedding manifold space, and the spatial local curvature gradient vector is extracted to convert the discrete physical tactile sensation into a geometric feature quantity that evolves continuously on the manifold surface. Next, in step S3, the weight entropy increase rate of the neuron nodes in the connection weight matrix of the hidden layer of the deep neural network is calculated based on the spatial local curvature gradient vector. Then, in step S4, the weight entropy increase rate is compared with the safe convergence threshold. If it is lower than the threshold, the online gradient of the connection weight matrix is ​​updated. If it reaches or exceeds the threshold, the nonlinear discrete truncation rule of the weight is activated, the gradient backpropagation is suspended and the parameter value is locked as the weight sedimentation value of the previous convergence time sequence. Finally, in step S5, the modulated or locked connection weight matrix is ​​used for forward inference to output the joint trajectory error compensation control tensor to constrain the underlying servo mechanism.

[0037] like Figure 2 As shown, the external sensing hardware group includes a tactile sensor array and a forward laser detection link. The tactile sensor array is used to collect the spatiotemporal feature tensor of tactile pressure distributed on the robot surface and output a high-frequency tactile pressure data stream. The forward laser detection link is used to output a sparse spatial laser point cloud intensity sequence and provide a spatial distance point cloud stream. The general edge computing unit integrates a memory, a geometric manifold model module, a deep neural network hidden layer kernel, a vision processing subnetwork, and a joint physical interference manifold boundary model. The memory has a pre-set orthogonal transformation basis matrix and stable temporal weight deposit values, which are input to the geometric manifold model module by calling the orthogonal transformation basis matrix. The geometric manifold model module performs the construction of feature embedding manifold space and extracts the spatial local curvature gradient vector and provides a continuous curvature gradient vector. The input is fed into the hidden layer kernel of the deep neural network. The hidden layer kernel performs dynamic activation of the weights based on the entropy increase rate calculated by the spatial local curvature gradient vector, the nonlinear discrete truncation rule, and the spatiotemporal attention mechanism. It also dynamically deactivates the control bit through the injection mask and inputs it into the visual processing sub-network. The visual processing sub-network is dynamically deactivated by the binary topological mask matrix and connected to the joint physical interference manifold boundary model. The joint physical interference manifold boundary model performs a pre-embedded output layer to construct a physical safety torque constraint penalty and outputs a joint trajectory error compensation control tensor. The joint trajectory error compensation control tensor is finally transmitted to the bottom servo mechanism of the humanoid robot frame. The bottom servo mechanism is used to execute the joint trajectory error compensation control tensor to drive the joint to complete the movement within the restricted work space.

[0038] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A deep learning method for humanoid robots to move through tunnels, running on a computer system, characterized in that, Includes the following steps: Step S1: Collect the spatiotemporal feature tensor of tactile pressure distributed on the robot surface and the intensity sequence of sparsed spatial laser point cloud output by the detection link. Step S2: The spatiotemporal feature tensor of tactile pressure and the intensity sequence of sparse spatial laser point cloud are concatenated and input into the geometric manifold model to construct the feature embedding manifold space. In the feature embedding manifold space, the spatial local curvature gradient vector of tactile pressure spatiotemporal feature tensor is extracted, and the discrete physical touch is converted into a geometric feature quantity that evolves continuously on the manifold surface. Step S3: Calculate the weight entropy increase rate of neuron nodes within the connection weight matrix of the hidden layer of the deep neural network based on the spatial local curvature gradient vector. Step S4: Compare the weight entropy increase rate with the safe convergence threshold. When the weight entropy increase rate is lower than the safe convergence threshold, update the online gradient of the connection weight matrix. When the weight entropy increase rate reaches or exceeds the safe convergence threshold, activate the weight nonlinear discrete truncation rule, suspend the gradient backpropagation of the current sampling period, and lock the parameter value of the connection weight matrix to a snapshot of the steady-state parameter matrix of the connection weight matrix in the previous convergence sequence when it was in a historical stable convergence state. Step S5: Using the modulated or locked connection weight matrix for forward inference, output the joint trajectory error compensation control tensor to constrain the underlying servo mechanism.

2. The deep learning method for humanoid robot drilling and moving according to claim 1, characterized in that, Step S4 activates the nonlinear discrete truncation rule for weights, including the following sub-steps: Step S41, when the weight entropy increase rate reaches or exceeds the safe convergence threshold, the backpropagation of the current time series full network feature gradient is paused, and the gradient update of the connected weight matrix is ​​rigidly suspended within the subsequent consecutive preset number of sampling time series windows to maintain the frozen protection state of the weight parameters. The forward inference is completed by calling the steady-state parameter matrix snapshot stored in the previous stable time series.

3. The deep learning method for humanoid robot drilling and moving according to claim 1, characterized in that, Following step S1, the following steps are also included: Step S11, using the feature space self-similarity metric operator to extract the first derivative of the tactile pressure spatiotemporal feature tensor to determine the tactile manifold energy index; Step S12, simultaneously extracting the spectral envelope similarity of the sparse spatial laser point cloud intensity sequence to determine the echo topology density index; Step S13, calculating the perceptual spatial manifold distortion scalar based on the real-time ratio of the tactile manifold energy index to the echo topology density index.

4. The deep learning method for humanoid robot drilling and moving according to claim 3, characterized in that, Step S2 further includes the following steps: Step S21, when the scalar of the perceptual space manifold distortion exceeds the preset geometric space mutation critical safety threshold of 2.50, the weight multiplication and addition operation of the hidden layer is stopped, and the orthogonal projection operation of the hidden layer feature manifold is started; Step S22, the orthogonal transformation basis matrix preset in the memory is called to project the global multimodal hidden representation tensor onto the low-dimensional self-similar feature subspace defined by the null term of the echo topological density index.

5. A deep learning method for humanoid robot drilling and moving according to claim 4, characterized in that, Step S22 includes the following sub-steps: Step S221, injecting a binary topological mask matrix into the low-dimensional self-similar feature subspace to dynamically modulate the connection weight matrix; Step S222, when the perceptual space manifold distortion scalar reaches or exceeds the geometric space abrupt change critical safety threshold of 2.50, changing the control bit element corresponding to the image feature flow path in the binary topological mask matrix from value 1 to value 0, and forcibly clearing the activation output of the visual processing sub-network inside the deep neural network to zero; Step S223, keeping the control bit element corresponding to the tactile and joint state perception path in the binary topological mask matrix at value 1, and increasing the corresponding activation gain coefficient.

6. The deep learning method for humanoid robot drilling and moving according to claim 2, characterized in that, Step S3 further includes the following sub-steps: Step S32, calculate the local spatiotemporal variance of the tactile pressure spatiotemporal feature tensor within a preset sliding time window; Step S33, monotonically adjust the spatiotemporal attention allocation matrix of the hidden neurons according to the reciprocal of the local spatiotemporal variance, and superimpose the Frobenius norm of the spatiotemporal attention allocation matrix in parallel in the total loss function of the deep neural network as a spatiotemporal regularization damping penalty term.

7. A deep learning method for humanoid robot drilling and moving according to claim 6, characterized in that, The spatiotemporal regularization damping penalty term in sub-step S33 is determined as follows: the spatiotemporal regularization damping penalty term is obtained by multiplying the dynamic adaptive adjustment coefficient by the square of the Frobenius norm of the spatiotemporal attention allocation matrix, wherein the value of the dynamic adaptive adjustment coefficient is configured to be between 0.01 and 0.

05.

8. The deep learning method for humanoid robot drilling and moving according to claim 1, characterized in that, Step S5 includes the following pre-step: Step S51, embed the joint physical interference manifold boundary model in the output layer, and retrieve the joint spatial motion error compensation control instruction matrix generated by the current inference to calculate the predicted joint output torque tensor. Step S52: Construct a three-dimensional nonlinear difference between the predicted joint output torque tensor and the preset physical safety torque boundary tensor, and extract the modulus of the three-dimensional nonlinear difference to generate an interference overflow scalar.

9. A deep learning method for humanoid robot drilling and moving according to claim 8, characterized in that, Sub-step S52 includes the following processing: Step S521, the edge gradient penalty operator of the interference overflow scalar is connected in parallel to the total loss function of the deep neural network; Step S522, when the interference overflow scalar is less than or equal to 0, the edge gradient penalty operator is set to 0; when the interference overflow scalar is greater than 0 and monotonically increasing, a loss function penalty weight with a quadratic step-like abrupt amplification is generated according to the displacement of the interference overflow scalar exceeding 0, and the loss function penalty weight is superimposed on the total loss function.

Images