Adaptive terrain recognition and navigation method for explosion-proof environment four-legged robot

Abstract

The application provides a four-legged robot adaptive terrain recognition and navigation method for an explosion-proof environment, and relates to the field of high-definition positioning technology.The application quantifies endogenous disturbance and sensing uncertainty of an actuator by synchronously extracting electromagnetic torque pulsation entropy and point cloud shielding entropy rate;cross-modal feature space alignment is achieved by using bilinear interpolation, and matching is performed based on Mahalanobis distance and a high-precision semantic map;by using semantic space consistency score and a left multiplication disturbance model in a Lie group space, a pose correction vector is output to perform manifold projection correction, thereby eliminating nonlinear drift caused by skidding;high-definition navigation or a tactile stepping strategy is triggered according to positioning reliability, and driving gain is dynamically optimized in combination with energy evolution attenuation gradient.The application effectively solves the problems of sensing non-determinacy and positioning misalignment under explosion-proof working conditions, and improves the navigation robustness and sub-centimeter positioning resolution of the robot in an extreme environment.

Description

Technical Field

[0001] This invention relates to the field of high-definition positioning technology, specifically to an adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments. Background Technology

[0002] In high-risk working environments such as coal mines and chemical plants, quadruped robots are widely used in automated inspections due to their excellent terrain adaptability. However, explosion-proof environments are often filled with high concentrations of dust, water mist, and flammable gases, which severely interfere with the imaging accuracy of lidar and vision sensors, leading to missing geometric features or severe artifacts. Existing navigation technologies mostly rely on single-modal perception. In unstructured terrains such as slippery or soft surfaces, the nonlinear drift of the odometer caused by foot slippage is difficult to compensate for effectively, resulting in decreased positioning accuracy over long distances. Furthermore, existing systems lack adaptive strategy switching mechanisms for edge cases of perception failure, making them highly susceptible to collision risks due to misjudging obstacles under sensor-limited conditions. Therefore, a robust navigation solution that integrates endogenous mechanical feedback and multimodal spatiotemporal alignment is urgently needed.

[0003] In the prior art, publication number CN121275003A, entitled "A Terrain-Adaptive Navigation Method and System for Robots," relates to the field of robot terrain navigation technology. The method includes: acquiring the robot's movement speed within a past time window, collecting current road surface data, and identifying and obtaining a road surface topography map; randomly selecting a first movement position, acquiring the corresponding first movement speed and amplitude, and combining this with the first road surface topography information to obtain a first movement stability coefficient; configuring a movement error range, dividing and obtaining a first movement position range, and calculating a first fusion stability coefficient based on the road surface topography information within the first range; calculating a first navigation fitness, and optimizing the movement position to obtain the optimal movement position. This invention solves the problem in traditional terrain-adaptive navigation where it is difficult to balance safety and smoothness during robot movement, leading to high risk of tipping over or frequent interruptions in the movement rhythm, thus affecting navigation performance.

[0004] However, in practical applications, the uncertainty of perception in high-risk environments (a reliability bottleneck) arises: explosion-proof environments (such as chemical plants and underground coal mines) are often accompanied by refraction interference from high concentrations of dust, water mist, or flammable gases. This can lead to a large amount of noise in the laser point cloud or geometric distortion due to multipath effects. Existing technologies rely on a single point cloud to construct a terrain map and use random location selection for stability verification. In low visibility or edge cases where sensors are limited, the positioning resolution drops drastically, easily causing the robot to misjudge small obstacles (such as explosion-proof thresholds or thin pipes), leading to hardware collisions and safety risks.

[0005] Existing technologies configure the movement error range by using the "movement speed within a past time window." However, the terrain in explosion-proof environments often has unstructured features (such as slippery, soft, and sloping terrain). When the robot performs gait switching, foot slippage can cause non-accumulative drift in the odometer. Simply relying on historical speed and random position compensation cannot achieve high-resolution pose alignment in a global coordinate system, resulting in significant "positioning drift" in the navigation system during long-distance operations, making it difficult to achieve sub-centimeter-level positioning accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments, comprising the following steps: S1. Simultaneously acquire the foot dynamic pulse data stream and global environmental point cloud stream of the quadruped robot in the explosion-proof environment, calculate the electromagnetic torque pulsation entropy that characterizes the intrinsic micro-disturbance of the actuator through signal processing, and calculate the point cloud occlusion entropy rate used to quantify the uncertainty of the remote sensing environment. S2. Load the preset configuration parameters in the structured external data carrier, perform cross-modal spatiotemporal alignment processing on the electromagnetic torque pulsation entropy and the point cloud occlusion entropy rate according to the preset configuration parameters, and perform association matching logic operation on the aligned feature vector and the node mechanical texture features stored in the preset high-precision semantic map. S3. Based on the residual distribution generated by the correlation matching logic operation, generate a high-reliability positioning confidence signal to characterize the reliability of the current positioning state, and calculate the deviation of the absolute pose of the current quadruped robot based on the coherence result of the correlation matching logic operation to output the pose correction vector. S4. Based on the pose correction vector, synchronously calibrate the global positioning coordinates of the quadruped robot, and at the same time determine whether the high reliability positioning confidence signal meets the preset high-precision positioning threshold, so as to execute the corresponding navigation path: if it meets the threshold, maintain the high-definition navigation strategy for path planning; if it does not meet the threshold, trigger the tactile stepping safety strategy, reduce the step frequency and combine the feedback of the electromagnetic torque pulsation entropy to perform blind zone obstacle avoidance navigation, and synchronously adjust the actuator drive control gain parameters.

[0008] Furthermore, in S1, the calculation of the electromagnetic torque pulsation entropy characterizing the intrinsic micro-perturbation of the actuator includes: acquiring the phase current pulse sequence of the drive motor and extracting the high-frequency harmonic components reflecting the micro-interaction response between the foot and the ground; performing frequency domain feature transformation on the high-frequency harmonic components to construct a drive energy fingerprint spectrum characterizing the current foot-ground mechanical response intensity; and using the information entropy model to calculate the energy distribution disorder of the drive energy fingerprint spectrum within a preset sampling time window to obtain the electromagnetic torque pulsation entropy used to characterize the terrain physical characteristics.

[0009] Furthermore, in S1, the process of determining the sampling window includes: acquiring the point cloud occlusion entropy rate calculated synchronously; dynamically adjusting the width of the sampling window according to the magnitude of the point cloud occlusion entropy rate, wherein the width of the sampling window is negatively correlated with the point cloud occlusion entropy rate, so as to improve the capture bandwidth of high-frequency mechanical response features by shortening the sampling window width, thereby using high-resolution endogenous mechanical features to complementarily correct the uncertainty of external visual positioning in visual perception-limited environments.

[0010] Furthermore, the cross-modal spatiotemporal alignment processing in S2 specifically includes: The sampling timestamp sequence of the point cloud occlusion entropy rate is obtained and used as a reference clock; the electromagnetic torque pulsation entropy in the high-frequency feature stream is resampled using a bilinear interpolation algorithm so that each electromagnetic torque pulsation entropy sample is aligned with the corresponding point cloud occlusion entropy rate sample on the time axis, thereby constructing an augmented spatiotemporal feature vector. In the resampling process, the interpolation step size is dynamically scaled according to the instantaneous rate of change of the point cloud occlusion entropy rate to eliminate asynchronous observation bias between heterogeneous sensors.

[0011] Furthermore, the node mechanical texture features stored in the preset high-precision semantic map in S2 include: Semantic retrieval ring domain based on Gaussian mixture model; extract the mechanical texture features of all candidate nodes in the semantic retrieval ring domain, and calculate the Mahalanobis distance between the augmented spatiotemporal feature vector and each candidate node as the matching residual.

[0012] Furthermore, in S3, the dynamic component in the augmented spatiotemporal feature vector is extracted and normalized; the preset mechanical impedance spectrum feature corresponding to the coordinates is retrieved from the preset high-precision semantic map, and the temporal similarity between the dynamic component and the preset mechanical impedance spectrum feature is calculated using the Pearson correlation coefficient operator. The temporal similarity is exponentially weighted and corrected using the point cloud occlusion entropy rate to obtain a semantic space consistency score. The weight coefficient of the exponential weighting correction is positively correlated with the point cloud occlusion entropy rate, which is used to forcibly increase the decision weight of mechanical matching when visual information is missing.

[0013] Furthermore, the output pose correction vector in S3 specifically includes: When the semantic space consistency score is greater than a preset confidence threshold, the coordinates of the node with the highest matching degree in the preset high-precision semantic map are obtained as logical anchor points; the Euclidean distance residual between the current predicted coordinates of the quadruped robot and the logical anchor points is calculated, and the Euclidean distance residual is mapped to a six-degree-of-freedom pose increment using a left-multiplication perturbation model in Lie group space; the six-degree-of-freedom pose increment is used as the pose correction vector to perform manifold space projection correction on the global positioning coordinates of the quadruped robot to eliminate nonlinear odometry drift accumulated by foot slippage.

[0014] Furthermore, the multi-level policy switching in S4 specifically includes: A high-reliability location confidence signal is obtained; a gain attenuation model based on the sigmoid logistic distribution function is constructed, and the proportional gain scaling factor is calculated; wherein, the proportional gain scaling factor is determined by obtaining the difference between the high-reliability location confidence signal and a preset safety threshold; multiplying the difference by a sensitivity constant and taking the opposite as the exponent, the power of the natural constant is calculated; the constant 1 is divided by the sum of the constant 1 and the power to obtain the proportional gain scaling factor; The proportional gain of the position loop of the motion controller is adjusted in real time by using a proportional gain scaling factor to achieve a seamless and smooth transition of control stiffness from a high-gain tracking state to a low-gain compliant state when the position confidence decreases.

[0015] Furthermore, in S4, the blind spot obstacle avoidance navigation in the haptic step safety strategy specifically includes: The instantaneous rate of change of the electromagnetic torque pulsation entropy is monitored in real time; when the instantaneous rate of change exceeds the preset mechanical collision threshold, it is determined that the foot has interacted with an unmodeled obstacle, triggering the impedance space foot trajectory reshaping logic. The impedance space foot trajectory reshaping logic includes obtaining the polar coordinate phase of the current foot; superimposing a vertical displacement component positively correlated with the amplitude of the instantaneous rate of change on the current foot teaching trajectory; synchronously lowering the step frequency cycle to extend the static friction support time between the foot and the ground; and through the dynamic compensation of the vertical displacement component, automatically crossing small obstacles by relying on foot torque feedback in blind zone environments where visual navigation fails, while maintaining the logical continuity of global positioning coordinates.

[0016] Furthermore, in S4, adjusting the drive control gain parameter of the actuator specifically includes: The power spectral density distribution of the actuator before and after gain adjustment is obtained, and the energy evolution attenuation gradient of the actuator is calculated based on the energy dissipation criterion; when the amplitude fluctuation of the pose correction vector causes the energy evolution attenuation gradient of the actuator to be lower than the preset damping critical value, the compensation gain branch is activated. The compensation gain branch includes: calculating the second derivative of the pose correction vector, extracting the spurious acceleration component caused by positioning jitter, and injecting the spurious acceleration component in reverse into the damping gain matrix of the foot compliance controller to counteract the transient impact oscillation caused by strategy switching, and keeping the total drive power consumption of the quadruped robot within the explosion-proof safety power limit at the moment of switching between the high-definition navigation strategy and the tactile stepping safety strategy.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention synchronously acquires foot dynamics pulse data streams and global environmental point cloud streams, and calculates the electromagnetic torque pulsation entropy characterizing the intrinsic micro-perturbations of the actuator and the point cloud occlusion entropy rate used to quantify the uncertainty of the remote sensing environment. This constructs a multimodal feature space covering micro-mechanical response and macro-geometric perception at the physical level. The point cloud occlusion entropy rate is used to exponentially weight the temporal similarity, generating a semantic space consistency score that can quantify the reliability of the positioning state in real time. This drives a gain attenuation model based on an S-shaped logistic distribution function to calculate the proportional gain scaling factor, achieving a seamless and smooth transition from a high-definition navigation strategy to a tactile stepping safety strategy. Especially in extreme conditions where laser point clouds generate noise and geometric distortion due to dust or water mist interference, by triggering impedance space foot trajectory reshaping logic and superimposing a vertical displacement component positively correlated with the instantaneous rate of change amplitude, the robot can automatically perceive and cross small obstacles such as explosion-proof thresholds based on foot torque feedback, improving the robot's perception robustness and operational safety in low-visibility edge scenarios. This invention also employs a bilinear interpolation algorithm to resample the electromagnetic torque pulsation entropy and dynamically scales the interpolation step size according to the instantaneous change rate of the point cloud occlusion entropy rate to construct an augmented spatiotemporal feature vector, thereby eliminating asynchronous observation bias between heterogeneous sensors from the underlying algorithm level. By calculating the Mahalanobis distance between the augmented spatiotemporal feature vector and the preset high-precision semantic map candidate nodes, and introducing a left-multiplication perturbation model in Lie group space to map the Euclidean distance residual into a six-degree-of-freedom pose increment, manifold space projection correction is performed on the global positioning coordinates, fundamentally compensating for the nonlinear odometry drift caused by unstructured terrain. Simultaneously, by combining the actuator energy evolution decay gradient activation compensation gain branch, the extracted spurious acceleration components are injected inversely into the damping gain matrix of the foot compliance controller, which not only effectively offsets the transient impact oscillations caused by strategy switching, but also ensures that the robot maintains sub-centimeter-level pose calculation accuracy during long-distance complex navigation, solving the technical problem of positioning resolution degradation in dynamic environments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall structure of the method of the present invention; Figure 2 This is a schematic flowchart of the overall method of the present invention; Figure 3 This is a schematic diagram of the process framework for S3 and S4 in this invention. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Please see Figure 1 , Figure 2 and Figure 3 This invention provides a technical solution: an adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments, the specific steps of which include: S1. Simultaneously acquire the foot dynamic pulse data stream and global environmental point cloud stream of the quadruped robot in the explosion-proof environment, calculate the electromagnetic torque pulsation entropy that characterizes the intrinsic micro-disturbance of the actuator through signal processing, and calculate the point cloud occlusion entropy rate used to quantify the uncertainty of the remote sensing environment. S2. Load the preset configuration parameters in the structured external data carrier, perform cross-modal spatiotemporal alignment processing on the electromagnetic torque pulsation entropy and the point cloud occlusion entropy rate according to the preset configuration parameters, and perform association matching logic operation on the aligned feature vector and the node mechanical texture features stored in the preset high-precision semantic map. S3. Based on the residual distribution generated by the correlation matching logic operation, generate a high-reliability positioning confidence signal to characterize the reliability of the current positioning state, and calculate the deviation of the absolute pose of the current quadruped robot based on the coherence result of the correlation matching logic operation to output the pose correction vector. S4. Based on the pose correction vector, synchronously calibrate the global positioning coordinates of the quadruped robot, and at the same time determine whether the high reliability positioning confidence signal meets the preset high-precision positioning threshold, so as to execute the corresponding navigation path: if it meets the threshold, maintain the high-definition navigation strategy for path planning; if it does not meet the threshold, trigger the tactile stepping safety strategy, reduce the step frequency and combine the feedback of the electromagnetic torque pulsation entropy to perform blind zone obstacle avoidance navigation, and synchronously adjust the actuator drive control gain parameters.

[0022] In S1, the calculation of the electromagnetic torque pulsation entropy characterizing the intrinsic micro-perturbation of the actuator includes: acquiring the phase current pulse sequence of the drive motor and extracting the high-frequency harmonic components reflecting the micro-interaction response between the foot and the ground; performing frequency domain feature transformation on the high-frequency harmonic components to construct a drive energy fingerprint spectrum characterizing the current foot-ground mechanical response intensity; and using the information entropy model to calculate the energy distribution disorder of the drive energy fingerprint spectrum within a preset sampling time window to obtain the electromagnetic torque pulsation entropy used to characterize the terrain physical properties.

[0023] Furthermore, regarding the determination of electromagnetic torque pulsation entropy, this embodiment employs a semantic calculation model based on power spectral density distribution. Specifically, by acquiring the stator phase current pulse sequence of the drive motor and using filtering or wavelet transform operators to remove the low-frequency fundamental wave, high-frequency harmonic components reflecting the microscopic interaction response between the foot and the ground are extracted. Frequency domain feature transformation is performed on the high-frequency harmonic components to calculate the power spectral density distribution on the frequency axis, which is defined as the "driving energy fingerprint spectrum" characterizing the current foot-ground mechanical response intensity, used to map the specific high-frequency feedback generated by different terrains on the motor load. The driving energy fingerprint spectrum is normalized using an information entropy model, converting the energy at each frequency point into a proportion probability, and calculating the statistical entropy value of this probability distribution to obtain the energy distribution disorder characterizing the degree of energy dispersion. The energy distribution disorder is used as the electromagnetic torque pulsation entropy characterizing the physical properties of the terrain. By identifying the energy characteristics with low disorder under flat terrain and the energy distribution differences with high disorder under complex terrain, accurate quantification of the microscopic force state at the foot of the quadruped robot is achieved.

[0024] In S1, the process of determining the sampling window includes: acquiring the point cloud occlusion entropy rate obtained by synchronous calculation; dynamically adjusting the width of the sampling window according to the magnitude of the point cloud occlusion entropy rate, wherein the width of the sampling window is negatively correlated with the point cloud occlusion entropy rate, so as to improve the capture bandwidth of high-frequency mechanical response features by shortening the sampling window width, thereby using high-resolution endogenous mechanical features to complementarily correct the uncertainty of external visual positioning in visual perception-limited environments.

[0025] The operating environment of this embodiment needs to include a drive control interface capable of outputting three-phase stator current in real time, and a laser sensing interface capable of outputting three-dimensional point cloud flow at a frequency of not less than 10Hz. To achieve the adaptive terrain recognition and navigation, a data acquisition and feature extraction process needs to be performed, including: a drive motor stator phase current pulse sequence, hereinafter denoted as Iph, used to characterize the microscopic waveform generated by the motor when executing motion commands; electromagnetic torque pulsation entropy, hereinafter denoted as Ert, used to quantify the degree of disorder in the interaction response between the foot and the ground, with its value range limited to the real number interval [0, 1] by normalization; point cloud occlusion entropy rate, hereinafter denoted as Hoccl, used to quantify the amount of perceived information lost per unit time due to environmental interference, with its value range also limited to the real number interval [0, 1]; and dynamic sampling window width, hereinafter denoted as Wdyn, which dynamically adjusts the length of data participating in the information entropy calculation, with its value range between 10 milliseconds and 200 milliseconds.

[0026] Regarding the aforementioned method for determining the electromagnetic torque ripple entropy Ert, this embodiment employs a linguistic calculation model based on power spectral density distribution. Specific steps include: acquiring the collected stator phase current pulse sequence Iph of the drive motor; projecting the three-phase currents onto the quadrature-axis coordinate system using rotating coordinate transformation logic to extract the quadrature-axis current component directly related to the output torque; performing high-pass filtering on this quadrature-axis current component to remove quasi-static components with frequencies below 50 Hz, obtaining a high-frequency ripple signal reflecting microscopic slippage characteristics; performing a discrete Fourier transform on the high-frequency ripple signal to obtain the drive energy fingerprint spectrum; dividing the fingerprint spectrum into N equal-width frequency bands, calculating the percentage of energy value in each band relative to the total energy, obtaining a set of probability distribution sequences; calculating the product of each element in the probability distribution sequence with its value and the natural logarithm of that value, summing all products and taking their negative; dividing the summed result by the natural logarithm of N to complete the normalization process, obtaining the final electromagnetic torque ripple entropy Ert.

[0027] Regarding the determination of the point cloud occlusion entropy rate Hoccl, this embodiment adopts an evolutionary calculation logic based on spatial probabilistic voxel fields. The space around the quadruped robot is divided into several three-dimensional voxel grids. The probability of each voxel grid being occupied in the current frame point cloud is calculated and compared with the occupation probability at the previous moment. The uncertainty change of the voxel occupation state per unit time is calculated using the information gain principle. Specifically, the entropy value of the occupation probability distribution of each voxel grid at the current moment and the previous moment is calculated using the Shannon entropy formula. The entropy increment (i.e., information gain) reflecting the deterministic evolution of space is obtained by the difference between the two. Then, the ratio of this increment to the time step between the two frames is calculated to accurately quantify the evolution rate of the uncertainty of the voxel field per unit time, thereby obtaining the point cloud occlusion entropy rate.

[0028] The calculation logic of the linkage adjustment mechanism for the dynamic sampling window width Wdyn includes real-time monitoring of the point cloud occlusion entropy rate Hoccl; setting a benchmark sampling window, which is set to 100 milliseconds in this embodiment; calculating the difference between constant 1 and the point cloud occlusion entropy rate Hoccl to obtain the environmental transparency factor; and multiplying the benchmark sampling window by the environmental transparency factor to obtain the current dynamic sampling window width Wdyn. Furthermore, when the quadruped robot enters a smoke area, causing the point cloud occlusion entropy rate Hoccl to increase (e.g., reaching 0.8), the calculated dynamic sampling window width Wdyn will automatically shorten (to 20 milliseconds). This means that the mechanical fingerprint of the foot will be captured with a higher temporal resolution, ensuring that sub-centimeter-level absolute pose correction is maintained through more frequent endogenous mechanical feature matching when the navigation system loses external geometric references. This results in a synergistic gain effect that cannot be achieved by a single perception mode.

[0029] The output consists of a joint feature vector composed of the electromagnetic torque pulsation entropy Ert and the point cloud occlusion entropy rate Hoccl, which serves as the input to the subsequent spatiotemporal coherence mechanical semantic adjudication model to trigger different navigation strategies. If the positioning confidence decreases due to extremely harsh environments, the drive gain will be automatically reduced based on the current Ert value, switching to haptic stepping mode to ensure movement safety in explosion-proof environments. This solution improves the reliability of high-precision positioning without requiring additional expensive explosion-proof sensor hardware, making it highly valuable for engineering implementation.

[0030] Cross-modal spatiotemporal alignment processing in S2 specifically includes: The sampling timestamp sequence of the point cloud occlusion entropy rate is obtained and used as a reference clock; the electromagnetic torque pulsation entropy in the high-frequency feature stream is resampled using a bilinear interpolation algorithm so that each electromagnetic torque pulsation entropy sample is aligned with the corresponding point cloud occlusion entropy rate sample on the time axis, thereby constructing an augmented spatiotemporal feature vector. In the resampling process, the interpolation step size is dynamically scaled according to the instantaneous rate of change of the point cloud occlusion entropy rate to eliminate asynchronous observation bias between heterogeneous sensors.

[0031] The node mechanical texture features stored in the preset high-precision semantic map in S2 include: Semantic retrieval ring domain based on Gaussian mixture model; extract the mechanical texture features of all candidate nodes in the semantic retrieval ring domain, and calculate the Mahalanobis distance between the augmented spatiotemporal feature vector and each candidate node as the matching residual.

[0032] In S3, the dynamic component in the augmented spatiotemporal feature vector is extracted and normalized; the preset mechanical impedance spectrum feature corresponding to the coordinates is retrieved from the preset high-precision semantic map, and the temporal similarity between the dynamic component and the preset mechanical impedance spectrum feature is calculated using the Pearson correlation coefficient operator. The temporal similarity is exponentially weighted and corrected using the point cloud occlusion entropy rate to obtain a semantic space consistency score. The weight coefficient of the exponential weighting correction is positively correlated with the point cloud occlusion entropy rate, which is used to forcibly increase the decision weight of mechanical matching when visual information is missing.

[0033] Furthermore, the preset high-precision semantic map also pre-stores a tactile semantic layer associated with spatial coordinates. The tactile semantic layer consists of a sequence of foot interaction force data recorded by the robot during its pre-traversal of the environment, which is stored as mechanical impedance spectrum features of different coordinate nodes through discretization processing. The process of retrieving the corresponding coordinates includes determining the three-dimensional coordinates (X, Y, Z) of the foot-ground contact point in the map coordinate system based on the robot's current global pose given by the current positioning algorithm. Subsequently, the tactile semantic layer is retrieved in the local neighborhood of the coordinate using a hash table index space search operator. The preset mechanical impedance spectrum features refer to a set of mechanical response vector sequences that are pre-calibrated at the coordinate point and reflect the stiffness and damping characteristics of the ground, which evolve over time. These vectors contain the expected normal force fluctuation and tangential friction damping distribution patterns when the foot contacts this type of surface (including grass, gravel, or metal grid). Through this retrieval logic, an "ideal reference waveform" is obtained, providing benchmark comparison data for subsequent calculation of the temporal similarity between the currently measured dynamic components and the map's true value.

[0034] The output pose correction vector in S3 specifically includes: When the semantic space consistency score is greater than a preset confidence threshold, the coordinates of the node with the highest matching degree in the preset high-precision semantic map are obtained as logical anchor points; the Euclidean distance residual between the current predicted coordinates of the quadruped robot and the logical anchor points is calculated, and the Euclidean distance residual is mapped to a six-degree-of-freedom pose increment using a left-multiplication perturbation model in Lie group space; the six-degree-of-freedom pose increment is used as the pose correction vector to perform manifold space projection correction on the global positioning coordinates of the quadruped robot to eliminate nonlinear odometry drift accumulated by foot slippage.

[0035] For cross-modal spatiotemporal alignment processing, this embodiment synchronously acquires the sampling timestamp sequence of the point cloud occlusion entropy rate using the reference clock of the navigation controller. Since the high-frequency feature stream (kilohertz level) from motor feedback and the environmental perception feature stream (hertz level) differ by orders of magnitude in sampling frequency, this embodiment employs a bilinear interpolation resampling algorithm with adjustable step size. The specific calculation process is as follows: First, obtain the timestamp of the current point cloud observation; second, retrieve two neighboring sampling points before and after the timestamp on the time axis of the electromagnetic torque pulsation entropy; third, obtain the point cloud occlusion entropy rate at the current moment and calculate its absolute value of change relative to the previous moment; fourth, determine the dynamic interpolation scaling factor based on this absolute value of change. In this embodiment, the baseline value of this factor is 1.0, and it increases linearly with the increase of the change rate, with a maximum upper limit of 2.0; fifth, use the dynamic interpolation scaling factor to perform a weighted average calculation on the two neighboring sampling points to obtain the resampled torque feature that is completely aligned with the point cloud observation on the time axis, and combine it with the point cloud feature to construct an augmented spatiotemporal feature vector. The technical consideration behind setting a dynamic scaling step size is that, under conditions of drastic environmental changes, increasing the interpolation precision can effectively eliminate phase hysteresis bias between heterogeneous sensors.

[0036] After obtaining the augmented spatiotemporal feature vector, the process proceeds to the association matching stage based on a high-precision semantic map. The high-precision semantic map used in this embodiment not only includes geometric topological information but also pre-defined node mechanical texture features based on a Gaussian mixture model (GMM). Specifically, a semantic retrieval ring with a radius of 0.5 meters is defined, centered on the current coordinates predicted by the quadruped robot's odometry. Within this ring, the mechanical texture features of all candidate nodes are extracted, including the probability distribution mean vector and covariance matrix of the surface impedance. Subsequently, the calculation logic for the Mahalanobis distance is executed: the difference between the augmented spatiotemporal feature vector and the mean vector of the candidate node is calculated; the covariance matrix of the node is retrieved and its inverse matrix is ​​calculated; the difference vector, the inverse matrix, and the transpose of the difference vector are multiplied consecutively; the square root operation is performed on the result of the multiplication operation to obtain the matching residual. To determine the reliability of the current positioning status, this embodiment further calculates the semantic space consistency score. The specific determination logic includes extracting the dynamic component in the augmented spatiotemporal feature vector, normalizing it by subtracting the mean and dividing by the standard deviation; calculating the temporal similarity between the dynamic component and the map's preset mechanical impedance spectrum features using the Pearson correlation coefficient operator; and performing exponential weighting correction on the similarity using the point cloud occlusion entropy rate. The exponential weighting correction process includes using the natural constant as the base and the product of the point cloud occlusion entropy rate and the shape coefficient as the exponent for power operation, multiplying the result by the temporal similarity as a weight coefficient, and finally mapping the result to the closed interval [0,1] using the hyperbolic tangent operator. When visual information is completely missing due to smoke (i.e., the occlusion entropy rate approaches 1), the decision weight of mechanical matching is forcibly increased to ensure that the quadruped robot can achieve accurate semantic node positioning based solely on foot perception. The specific calculation logic for calculating time-domain similarity using the Pearson correlation coefficient operator includes: obtaining dynamic component sequences of consistent length and preset mechanical impedance spectrum feature sequences within the current sampling period; calculating the product of deviations of corresponding elements of the two sequences after subtracting their respective sequence means; and dividing the sum of the deviation products by the product of the standard deviations of the two sequences to obtain a scalar value reflecting the consistency of the waveform evolution trend of the two sequences. Furthermore, the shape coefficient is defined as a dimensionless adjustment constant for adjusting the sensitivity of weight allocation, and its value range is set between 1.0 and 5.0, preferably 2.5 in this embodiment; this coefficient is used to determine the transition slope of the system to mechanical perception mode when visual information degrades, and by adjusting the growth rate of the exponential operation, the system's response sensitivity to changes in environmental occlusion and its anti-interference robustness are balanced. Further semantic space consistency scores are calculated to determine the reliability of the current positioning status. Specifically, this includes extracting the dynamic components in the augmented spatiotemporal feature vector, normalizing them by subtracting the mean and dividing by the standard deviation; calculating the sum of the products of the deviations of the corresponding elements of the normalized dynamic component sequence and the map preset mechanical impedance spectrum feature sequence after subtracting their respective means, and dividing the sum of these products by the product of the standard deviations of the two sequences to obtain the temporal similarity reflecting the consistency of the waveform evolution trends of the two sequences. The similarity is corrected by exponential weighting using the point cloud occlusion entropy rate: the product of the point cloud occlusion entropy rate and the shape coefficient (within the range of 1.0 to 5.0, preferably 2.5) is used as the exponent for power operation, and the result is used as the weight coefficient to multiply the temporal similarity to obtain the matching strength value of the coupling environment credibility. The matching intensity value is nonlinearly mapped to a closed interval between zero and one using the hyperbolic tangent operator. The resulting mapping is defined as the semantic space consistency score. When visual information is completely lost due to smoke (i.e., the occlusion entropy rate approaches 1), this logic forces an increase in the decision weight of mechanical matching, ensuring that the quadruped robot can achieve accurate semantic node localization solely based on foot perception.

[0037] When the semantic space consistency score is greater than a confidence threshold of 0.85, the pose correction vector generation process is triggered. This embodiment obtains the coordinates of the node with the highest matching degree in the map as a logical anchor point and calculates the distance residual between the quadruped robot's predicted coordinates and the anchor point. The determination logic for the confidence threshold (preferably 0.85) includes using 0.85 as a critical judgment point for strong semantic association, statistically filtering out random fluctuations in mechanical features caused by unstructured terrain (scores generated by fluctuations are below 0.75). Simultaneously, it ensures that the generation frequency of the pose correction vector is within the damping convergence range of the control system, avoiding frequent coordinate jumps and energy loss caused by low threshold triggering, thus ensuring both positioning reliability and the electrical power safety of the drive system. To ensure that the rotation matrix always satisfies orthogonality constraints during the correction process, this embodiment employs a left-multiplication perturbation model in Lie group space. The specific calculation process is as follows: the coordinate residual is mapped to a six-degree-of-freedom vector in Lie algebra space; this vector is transformed into a perturbation matrix in Lie group space through exponential mapping; this perturbation matrix is ​​then left-multiplied by the quadruped robot's current global pose matrix, thereby completing the projection correction in manifold space. The technical consideration behind this logic is that it avoids non-physical pose distortion caused by directly adding or subtracting coordinates in Euclidean space, and can eliminate nonlinear odometry drift accumulated from foot slippage, thus providing sub-centimeter-level stable pose output in complex, explosion-proof terrain.

[0038] The corrected pose correction vector is applied to the global navigation path in real time. Through the coordinated control of cross-modal spatiotemporal alignment and manifold space correction, the quadruped robot can maintain high reliability in positioning even when faced with sudden visual loss or physical slippage. This technical solution achieves a synergistic gain in perception redundancy and execution accuracy without relying on additional explosion-proof base stations, providing solid technical support for autonomous inspection in explosion-proof environments.

[0039] In this embodiment, the semantic space consistency score (AA) is limited to the real number closed interval [0,1]. Its numerical distribution characteristics have a definite mapping relationship with the positioning reliability state: when the score approaches the minimum value 0, it indicates that the cross-modal feature vector and the preset semantic map are seriously mismatched, corresponding to extreme environments such as dense smoke obscuring or severe slippage. At this time, the pose correction output is automatically blocked by identifying the low confidence state to avoid positioning collapse caused by incorrect projection; when the score approaches the maximum value 1, it indicates that visual perception and mechanical perception have reached a high degree of physical consensus and entered a sub-centimeter-level high-precision locking state. The matching residual, serving as a core quantitative indicator for evaluating the similarity between the current ground interaction features and preset map nodes, is obtained by calculating the Mahalanobis distance between the augmented spatiotemporal feature vector and the candidate node features within the semantic retrieval ring domain. This logic introduces a covariance matrix reflecting the terrain noise distribution to normalize multi-source sensor noise, thereby effectively suppressing random mechanical disturbances caused by unstructured terrain. This establishes a monotonically positive correlation between the matching residual and the Mahalanobis distance, accurately characterizing the degree of deviation of the current perceived features from the prior mechanical texture. Based on this quantitative logic, when the point cloud occlusion entropy rate surges due to environmental interference, the system uses this residual distribution to forcibly increase the decision weight of mechanical features in nonlinear weight allocation, ensuring that even in edge cases where visual information is lacking, the system can still complete the absolute pose deviation calculation and positioning alignment through highly sensitive mechanical feature identification.

[0040] The Mahalanobis distance is monotonically positively correlated with the matching residual. By introducing a covariance matrix to eliminate terrain noise interference, the sensitivity to complex mechanical textures is improved. Simultaneously, the resampling interpolation step size dynamically scales with the increase of the occlusion entropy change rate, thereby compensating for asynchronous observation biases between heterogeneous sensors in the time domain. This logic design based on deep parameter coupling enables the robot to autonomously adjust the localization decision weights according to the dynamic characteristics of the environment, ensuring rapid convergence of the localization residual while achieving a smooth switch and stability gain from high-definition navigation to tactile survival mode.

[0041] To verify the effectiveness and robustness of the computational logic described in this invention under extreme environments, this embodiment constructs an explosion-proof operation simulation environment based on a high-fidelity physics engine. This environment, by establishing a dynamic mathematical model of a quadruped robot, reproduces complex working conditions including dust cover and unstructured slippery ground. The simulation experiment designed three typical working scenarios for different interference intensities, and recorded key technical indicators by averaging multiple samples. Specific experimental results are shown in Table 1.

[0042] Table 1: Simulation data of adaptive positioning and navigation performance of quadruped robots under different explosion-proof conditions

[0043] In the data analysis process of this embodiment, a derived index is introduced to comprehensively evaluate system performance, namely the semantic positioning stability evaluation index (hereinafter referred to as SL). This index is obtained by the navigation management module by taking the reciprocal of the standard deviation of the "semantic space consistency score" within a 100-millisecond sliding window, and then multiplying the result with the exponential decay term of the "pose correction vector magnitude." This index is used to quantitatively characterize the positioning system's resistance to fluctuations under external interference. A higher SL value indicates smaller transient jumps during pose correction and a more stable positioning output.

[0044] During the transition from Scene 1 (clear environment) to Scene 3 (extreme smoke and oil spill environment), the point cloud occlusion entropy rate jumped from 0.04 to 0.91. Under the extreme conditions of Scene 3, due to the significant loss of visual point cloud features and the lack of foot-end mechanical feedback compensation, the positioning drift typically accumulates rapidly to over 0.8 meters, even triggering control system collapse. In contrast, through confidence assessment and cross-modal feature alignment, the maximum positioning drift in Scene 3 was controlled to 0.21 meters. Using the existing technology's drift of 0.8 meters as a benchmark, this invention improves positioning accuracy by approximately 73.75% under extreme conditions (the calculation process is: subtract the comparison value of 0.21 from the benchmark value of 0.8, and then divide the difference by the benchmark value of 0.8).

[0045] The multi-level policy switching in S4 specifically includes: A high-reliability positioning confidence signal is acquired; a gain attenuation model based on the sigmoid logistic distribution function is constructed, and a proportional gain scaling factor (hereinafter referred to as Bz) is calculated; wherein, the proportional gain scaling factor is determined by acquiring the difference between the high-reliability positioning confidence signal and a preset safety threshold; multiplying the difference by a sensitivity constant and taking the opposite as the exponent, the power of the natural constant is calculated; the constant 1 is divided by the sum of the constant 1 and the power to obtain the proportional gain scaling factor Bz; the proportional gain scaling factor Bz is used to adjust the position loop proportional gain of the motion controller in real time, so as to achieve a seamless and smooth transition of control stiffness from a high-gain tracking state to a low-gain compliant state when the positioning confidence decreases.

[0046] In S4, the blind-zone obstacle avoidance navigation in the tactile stepping safety strategy specifically includes: real-time monitoring of the instantaneous change rate of the electromagnetic torque pulsation entropy; when the instantaneous change rate exceeds a preset mechanical collision threshold, determining that the foot has interacted with an unmodeled obstacle, triggering the impedance space foot trajectory reshaping logic, including obtaining the polar coordinate phase of the current foot; superimposing a vertical displacement component positively correlated with the amplitude of the instantaneous change rate on the current foot teaching trajectory; synchronously lowering the step frequency cycle to extend the static friction support time between the foot and the ground; through the dynamic compensation of the vertical displacement component, in the blind zone environment where visual navigation fails, automatically crossing small obstacles by relying on foot torque feedback, and maintaining the logical continuity of the global positioning coordinates.

[0047] In S4, adjusting the drive control gain parameter of the actuator also includes stability verification logic based on energy dissipation criteria; obtaining the power spectral density distribution of the actuator before and after gain adjustment, and calculating the energy evolution attenuation gradient of the actuator; When the amplitude fluctuation of the pose correction vector causes the actuator energy evolution decay gradient to fall below the preset damping threshold, the compensation gain branch is automatically activated, including calculating the second derivative of the absolute pose correction vector and extracting the spurious acceleration components caused by positioning jitter. The spurious acceleration component is injected in reverse into the damping gain matrix of the foot compliance controller to counteract the transient impact oscillations caused by strategy switching; through the compensation gain branch, it is ensured that the total power consumption of the drive system remains within the explosion-proof safety power limit when the quadruped robot switches between the "HD navigation" and "tactile stepping" modes.

[0048] Furthermore, regarding the specific implementation of multi-level strategy switching, this embodiment obtains a high-reliability positioning confidence signal (hereinafter referred to as Con, with a value range limited to the real number interval [0,1]) in real time through the navigation management module. In order to achieve seamless and smooth transition of control stiffness between the high-definition navigation strategy and the haptic stepping strategy, this embodiment constructs a gain attenuation model based on the S-shaped logistic distribution function to calculate the proportional gain scaling factor (hereinafter referred to as Bz). The specific calculation process includes obtaining the difference between Con and the preset safety threshold (set to 0.65 in this embodiment, representing the critical point where the standard deviation of the positioning error begins to affect gait stability); multiplying the difference by the sensitivity constant (set to 10.0 in this embodiment, used to control the steepness of the switching curve) and taking its opposite, as the exponent with the natural constant e as the base, and performing a power operation; dividing the constant 1 by "the sum of the constant 1 and the result of the power operation" to obtain the proportional gain scaling factor Bz, with a value range in the interval [0,1]. Furthermore, in this embodiment, the high-reliability positioning confidence signal Con is obtained by mapping the initial matching residual through negative exponential normalization, and is used to characterize the statistical consistency between real-time perception features and map priors. When calculating the scaling factor Bz, the preset safety threshold is determined based on the landing point stability envelope of the quadruped robot gait planner, and the preferred value range is 0.6-0.7, to ensure that the strategy intervention is completed before the positioning drift reaches the instability critical point. The sensitivity constant is dynamically tuned based on the control loop main frequency, and the preferred value range is 8.0-15.0, aiming to eliminate the switching step caused by discrete sampling. This parameter coupling design based on S-shaped logic distribution enables the robot to achieve continuous nonlinear migration of control stiffness between the high-definition navigation strategy (Bz tends to 1) and the tactile stepping strategy (Bz tends to 0) according to the change of Con, avoiding the instantaneous torque impact caused by hard switching. After triggering the tactile stepping safety strategy, the quadruped robot enters the blind zone obstacle avoidance navigation mode. At this time, it no longer relies on high-precision external visual data, but instead monitors in real time the instantaneous rate of change of electromagnetic torque pulsation entropy (hereinafter referred to as Ert) fed back by the foot actuator. When the instantaneous rate of change exceeds the preset mechanical collision threshold (set to 50 joules / second in this embodiment to distinguish between ground friction fluctuations and contact with substantial obstacles), it is determined that the foot has interacted with an unmodeled obstacle, and then the impedance space foot trajectory reshaping logic is triggered. The preset mechanical collision threshold is preferably in the range of 30J / s-120J / s. In this embodiment, 50J / s is selected to filter out random friction noise (usually below 35J / s) generated by complex terrain, while ensuring that the energy entropy change peak caused by obstacle collision is captured before the joint load reaches the critical point of dynamic instability, thereby obtaining sufficient obstacle avoidance response margin in the time domain.

[0049] The impedance space foot trajectory reshaping logic includes obtaining the polar coordinate phase of the current swinging leg through the foot encoder; superimposing a vertical displacement component positively correlated with the Ert amplitude on the current foot teaching trajectory. The calculation logic of the vertical displacement component is to multiply the rate of change by a displacement mapping coefficient (with a value of 0.002 meters per unit rate of change); and simultaneously lowering the step frequency cycle by increasing the duration of the single-step support phase to prolong the static friction support time between the foot and the ground. The purpose of this process is that even in blind spots where vision is completely absent, the quadruped robot can adaptively cross small obstacles by relying on "tactile feedback" to avoid further expansion of positioning drift caused by hard collisions.

[0050] To ensure the system dynamics stability during the aforementioned strategy switching process, this embodiment introduces a stability verification logic based on the energy dissipation criterion. By acquiring the power spectral density distribution of the actuator before and after gain adjustment in real time, the actuator energy evolution decay gradient (hereinafter denoted as Ven) is calculated. When the pose correction vector's manifold projection causes an instantaneous displacement of the coordinates, resulting in Ven falling below a preset damping threshold, an oscillation propagation risk is identified, and the compensation gain branch is automatically activated. The operation of this branch is as follows: the second derivative of the absolute pose correction vector is calculated to extract the spurious acceleration component introduced by the abrupt change in positioning coordinates; this spurious acceleration component is inverted and injected as a correction term into the damping gain matrix of the foot compliance controller; the actuator energy evolution decay gradient reflects the energy time-varying rate of change of the power spectral density distribution within the high-frequency disturbance bandwidth; the energy dissipation criterion determines whether to trigger the compensation gain by judging whether the decay gradient reaches the minimum damping threshold required to maintain dynamic convergence.

[0051] The purpose of setting up the compensation branch is that pose correction is essentially a jump in logical coordinates, which manifests as a "virtual impact" at the physical layer. By extracting its second derivative and performing reverse damping compensation, an equivalent virtual inertial force can be generated to counteract the transient impact oscillation of the actuator caused by coordinate correction. This mechanism ensures that the total power consumption of the drive system remains within the explosion-proof safety power limit at the moment the quadruped robot switches from "high-definition navigation" to "tactile walking" mode, or at the moment it returns from "drift state" to "logical anchor point," thus avoiding electrical safety hazards caused by overcurrent.

[0052] Furthermore, the logic for determining the explosion-proof safety power limit is as follows: based on the maximum allowable surface temperature and minimum ignition energy limited by the explosion-proof environment level, combined with the inductive energy storage characteristics of the actuator drive circuit and the transient temperature rise dynamic model, the maximum instantaneous peak power consumption limit that the actuator can withstand during dynamic compensation is calculated, and this is used as the safety envelope boundary to constrain the current output amplitude of the compensation gain branch in real time.

[0053] It should be noted that all calculation formulas in this invention employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected relevant parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas of this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization. The technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random-access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of this invention.

[0054] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for terrain recognition and navigation of a quadruped robot for explosion-proof environment, characterized in that, The specific steps include: S1. Simultaneously acquire the foot dynamic pulse data stream and global environmental point cloud stream of the quadruped robot in the explosion-proof environment, calculate the electromagnetic torque pulsation entropy that characterizes the intrinsic micro-disturbance of the actuator through signal processing, and calculate the point cloud occlusion entropy rate used to quantify the uncertainty of the remote sensing environment. S2. Load the preset configuration parameters in the structured external data carrier, perform cross-modal spatiotemporal alignment processing on the electromagnetic torque pulsation entropy and the point cloud occlusion entropy rate according to the preset configuration parameters, and perform association matching logic operation on the aligned feature vector and the node mechanical texture features stored in the preset high-precision semantic map. S3. Based on the residual distribution generated by the correlation matching logic operation, generate a high-reliability positioning confidence signal to characterize the reliability of the current positioning state, and calculate the deviation of the absolute pose of the current quadruped robot based on the coherence result of the correlation matching logic operation to output the pose correction vector. S4. Based on the pose correction vector, synchronously calibrate the global positioning coordinates of the quadruped robot, and at the same time determine whether the high reliability positioning confidence signal meets the preset high-precision positioning threshold, so as to execute the corresponding navigation path: if it meets the threshold, maintain the high-definition navigation strategy for path planning; if it does not meet the threshold, trigger the tactile stepping safety strategy, reduce the step frequency and combine the feedback of the electromagnetic torque pulsation entropy to perform blind zone obstacle avoidance navigation, and synchronously adjust the actuator drive control gain parameters.

2. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 1, characterized in that: In S1, the calculation of the electromagnetic torque pulsation entropy characterizing the intrinsic micro-perturbation of the actuator includes: acquiring the phase current pulse sequence of the drive motor and extracting the high-frequency harmonic components reflecting the micro-interaction response between the foot and the ground; performing frequency domain feature transformation on the high-frequency harmonic components to construct a drive energy fingerprint spectrum characterizing the current foot-ground mechanical response intensity; and using the information entropy model to calculate the energy distribution disorder of the drive energy fingerprint spectrum within a preset sampling time window to obtain the electromagnetic torque pulsation entropy used to characterize the terrain physical properties.

3. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 2, characterized in that: In S1, the process of determining the sampling window includes: acquiring the point cloud occlusion entropy rate obtained by synchronous calculation; dynamically adjusting the width of the sampling window according to the magnitude of the point cloud occlusion entropy rate, wherein the width of the sampling window is negatively correlated with the point cloud occlusion entropy rate, so as to improve the capture bandwidth of high-frequency mechanical response features by shortening the sampling window width, thereby using high-resolution endogenous mechanical features to complementarily correct the uncertainty of external visual positioning in visual perception-limited environments.

4. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 3, characterized in that: Cross-modal spatiotemporal alignment processing in S2 specifically includes: The sampling timestamp sequence of the point cloud occlusion entropy rate is obtained and used as a reference clock; the electromagnetic torque pulsation entropy in the high-frequency feature stream is resampled using a bilinear interpolation algorithm so that each electromagnetic torque pulsation entropy sample is aligned with the corresponding point cloud occlusion entropy rate sample on the time axis, thereby constructing an augmented spatiotemporal feature vector. In the resampling process, the interpolation step size is dynamically scaled according to the instantaneous rate of change of the point cloud occlusion entropy rate to eliminate asynchronous observation bias between heterogeneous sensors.

5. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 4, characterized in that: The node mechanical texture features stored in the preset high-precision semantic map in S2 include: A semantic retrieval ring domain based on a Gaussian mixture model is constructed; the mechanical texture features of all candidate nodes within the semantic retrieval ring domain are extracted, and the Mahalanobis distance between the augmented spatiotemporal feature vector and each candidate node is calculated as the matching residual.

6. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 5, characterized in that: In S3, the dynamic component in the augmented spatiotemporal feature vector is extracted and normalized; the preset mechanical impedance spectrum feature corresponding to the coordinates is retrieved from the preset high-precision semantic map, and the temporal similarity between the dynamic component and the preset mechanical impedance spectrum feature is calculated using the Pearson correlation coefficient operator. The temporal similarity is exponentially weighted and corrected using the point cloud occlusion entropy rate to obtain a semantic space consistency score. The weight coefficient of the exponential weighting correction is positively correlated with the point cloud occlusion entropy rate, which is used to forcibly increase the decision weight of mechanical matching when visual information is missing.

7. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 6, characterized in that: The output pose correction vector in S3 specifically includes: When the semantic space consistency score is greater than a preset confidence threshold, the coordinates of the node with the highest matching degree in the preset high-precision semantic map are obtained as logical anchor points; the Euclidean distance residual between the current predicted coordinates of the quadruped robot and the logical anchor points is calculated, and the Euclidean distance residual is mapped to a six-degree-of-freedom pose increment using a left-multiplication perturbation model in Lie group space; the six-degree-of-freedom pose increment is used as the pose correction vector to perform manifold space projection correction on the global positioning coordinates of the quadruped robot to eliminate nonlinear odometry drift accumulated by foot slippage.

8. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 7, characterized in that: Multi-level policy switching in S4 specifically includes: A high-reliability location confidence signal is obtained; a gain attenuation model based on the sigmoid logistic distribution function is constructed, and the proportional gain scaling factor is calculated; wherein, the proportional gain scaling factor is determined by obtaining the difference between the high-reliability location confidence signal and a preset safety threshold; multiplying the difference by a sensitivity constant and taking the opposite as the exponent, the power of the natural constant is calculated; the constant 1 is divided by the sum of the constant 1 and the power to obtain the proportional gain scaling factor; The proportional gain of the position loop of the motion controller is adjusted in real time by using a proportional gain scaling factor to achieve a seamless and smooth transition of control stiffness from a high-gain tracking state to a low-gain compliant state when the position confidence decreases.

9. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 8, characterized in that: In S4, the blind spot obstacle avoidance navigation in the haptic step safety strategy specifically includes: The instantaneous rate of change of the electromagnetic torque pulsation entropy is monitored in real time; when the instantaneous rate of change exceeds the preset mechanical collision threshold, it is determined that the foot has interacted with an unmodeled obstacle, triggering the impedance space foot trajectory reshaping logic. The impedance space foot trajectory reshaping logic includes obtaining the polar coordinate phase of the current foot; superimposing a vertical displacement component positively correlated with the amplitude of the instantaneous rate of change on the current foot teaching trajectory; synchronously lowering the step frequency cycle to extend the static friction support time between the foot and the ground; and through the dynamic compensation of the vertical displacement component, automatically crossing small obstacles by relying on foot torque feedback in blind zone environments where visual navigation fails, while maintaining the logical continuity of global positioning coordinates.

10. The adaptive terrain recognition and navigation method for quadruped robots in explosion-proof environments according to claim 9, characterized in that: In S4, adjusting the actuator's drive control gain parameters specifically includes: The power spectral density distribution of the actuator before and after gain adjustment is obtained, and the energy evolution attenuation gradient of the actuator is calculated based on the energy dissipation criterion; when the amplitude fluctuation of the pose correction vector causes the energy evolution attenuation gradient of the actuator to be lower than the preset damping critical value, the compensation gain branch is activated. The compensation gain branch includes: calculating the second derivative of the pose correction vector, extracting the spurious acceleration component caused by positioning jitter, and injecting the spurious acceleration component in reverse into the damping gain matrix of the foot compliance controller to counteract the transient impact oscillation caused by strategy switching, and keeping the total drive power consumption of the quadruped robot within the explosion-proof safety power limit at the moment of switching between the high-definition navigation strategy and the tactile stepping safety strategy.

Images