Dynamic obstacle avoidance control method for orbiting robot

Abstract

The application discloses a dynamic obstacle avoidance control method for a track inspection robot and relates to the technical field of robot control; the method comprises the following steps: based on real-time acquired laser point cloud data, visual image data, thermal distribution data and a robot motion state, obstacle fusion feature sets are calculated; the obstacle fusion feature sets are input into a pre-trained target recognition model to obtain obstacle preliminary classification labels and confidence scores; based on the obstacle fusion feature sets and the obstacle preliminary classification labels, the robot is controlled to actively physically probe the obstacles to obtain physical probe feature sets; based on the obstacle fusion feature sets and the physical probe feature sets, the obstacles are classified in terms of physical states to obtain obstacle final state labels; a pre-trained hybrid planning large model is used to perform real-time path planning; and the application overcomes the problem that the prior art only relies on material semantic recognition, which leads to the problem that obstacle disposal decisions do not match real track surface working conditions.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and more specifically, to a dynamic obstacle avoidance control method for a track-following robot. Background Technology

[0002] In applications such as rail transit, power line inspection, and factory track maintenance, track inspection robots serve as crucial equipment for automated inspection of track facilities. They typically operate autonomously along predetermined tracks, continuously monitoring the track surface condition and surrounding environment. During actual operation, track inspection robots inevitably encounter various track surface obstacles. Their ability to identify and handle these obstacles directly affects the continuity, safety, and overall maintenance efficiency of the inspection task.

[0003] Existing obstacle handling technologies for track-guided robots primarily rely on material semantic recognition of obstacles. By fusing multimodal perception information such as vision and lidar, obstacles are classified as movable or immovable, and corresponding handling strategies are executed accordingly. However, actual track surface conditions are significantly affected by factors such as rain, snow, icing, oil contamination, and the condition of the track structure. Obstacles of the same material category may exhibit different physical states due to attachment, embedding, or jamming, thus altering their actual manageability.

[0004] Current technologies rely solely on material properties as the primary basis for obstacle handling decisions, lacking an effective ability to determine the actual state of obstacles. They fail to differentiate between normally movable obstacles and those that are stuck or attached, easily leading to mismatches between handling decisions and actual working conditions. On one hand, misclassifying attached or stuck obstacles as movable obstacles may cause abnormal stress on the actuator, wheel-rail slippage, or equipment damage. On the other hand, misclassifying normally manageable obstacles as unmanageable obstacles will frequently trigger inspection stops, reducing inspection efficiency and increasing maintenance costs. Therefore, existing obstacle handling decision-making technologies for track inspection robots in complex track surface conditions still suffer from insufficient ability to determine the actual state of obstacles, limiting the further promotion and application of track inspection robots in the field of track maintenance.

[0005] In view of this, the present invention proposes a dynamic obstacle avoidance control method for a track-following robot to solve the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: a dynamic obstacle avoidance control method for a track-following robot, comprising:

[0007] Based on real-time acquired laser point cloud data, visual image data, thermal distribution data, and robot motion state, an obstacle fusion feature set is calculated.

[0008] The obstacle fusion feature set is input into the pre-trained target recognition model to obtain the visual category probability vector. The visual category probability vector is jointly corrected based on laser point cloud data and thermal distribution data to obtain the preliminary obstacle classification label and confidence score.

[0009] Based on the obstacle fusion feature set and the preliminary obstacle classification label, the robot is controlled to actively physically explore the obstacles and obtain the physical exploration feature set.

[0010] Based on the obstacle fusion feature set and the physical exploration feature set, the obstacles are classified into physical states to obtain the final state labels of the obstacles;

[0011] Based on the final state labels of obstacles and pre-loaded task constraints, real-time path planning is performed through a pre-trained hybrid planning model, outputting local control trajectories and execution strategy instructions, and controlling the track-following robot to perform obstacle avoidance, obstacle clearing, or obstacle detour.

[0012] Furthermore, methods for obtaining preliminary obstacle classification labels and confidence scores include:

[0013] The dimension of the visual category probability vector is equal to the number of preset visual categories, and each component of the visual category probability vector corresponds to the probability value of a different visual category.

[0014] The visual category corresponding to the component with the highest probability value in the visual category probability vector is taken as the visual determination result.

[0015] Based on laser point cloud data, the entry depth is extracted through a preset geometric analysis method. When the entry depth is greater than or equal to a preset depth threshold, an embedded risk marker is generated.

[0016] Based on thermal distribution data, temperature difference features are extracted using a preset temperature difference analysis method. When the temperature difference features are lower than a preset temperature difference threshold, a low-temperature adhesion risk marker is generated.

[0017] According to the preset physical consistency verification rules, if there is a logical conflict between the visual judgment result and the embedded risk mark or the low temperature adhesion risk mark, the preset probability reweighting process is performed on the component in the visual category probability vector corresponding to the visual judgment result.

[0018] The visual category probability vector after probability reweighting is normalized, and the visual category with the highest probability value in the normalized visual category probability vector is used as the initial obstacle classification label, and the corresponding probability value is used as the confidence score.

[0019] Furthermore, methods for controlling a robot to actively physically explore obstacles include:

[0020] Based on the initial obstacle classification labels and confidence scores, thrust strategy constraints are generated to determine the initial thrust level and thrust increment rate.

[0021] Based on the robot's motion state, control the robot to move toward the obstacle at a preset limited speed;

[0022] When the robot arm comes into contact with an obstacle, the robot arm is controlled to gradually increase the thrust on the obstacle according to the initial thrust level and the thrust increment rate, and the contact feedback force and motor load current are collected.

[0023] When the contact feedback force reaches the preset safety limit or the motor load current reaches the preset current limit, the active physical probe is terminated.

[0024] Furthermore, the thrust strategy constraints include:

[0025] The physical attribute type of the preliminary obstacle classification label is determined by a preset method to obtain the corresponding physical attribute type; the physical attribute type includes the lightweight and loose category and the heavy and dense structure category.

[0026] If the physical attribute type of the obstacle's initial classification label corresponds to the lightweight and loose category, a preset first initial thrust level is selected as the initial thrust level, and a preset first thrust increment rate is selected as the benchmark thrust increment rate.

[0027] If the physical attribute type of the obstacle's initial classification label corresponds to the dense heavy structure category, a preset second initial thrust level is selected as the initial thrust level, and a preset second thrust increment rate is selected as the benchmark thrust increment rate.

[0028] Furthermore, the thrust strategy constraints also include a reduction of the baseline thrust increment rate based on the confidence score;

[0029] Based on the confidence score, a lookup operation is performed on the preset reduction coefficient mapping table to obtain the corresponding thrust increment rate reduction coefficient. The thrust increment rate reduction coefficient is then multiplied by the benchmark thrust increment rate to obtain the thrust increment rate.

[0030] Furthermore, methods for calculating the obstacle fusion feature set include:

[0031] Spatial constraint filtering of laser point cloud data is performed using a pre-loaded trajectory map from the robot's built-in system;

[0032] The filtered laser point cloud data is segmented to obtain object clumps, which are sets of three-dimensional spatial coordinate points.

[0033] Based on object clumps and laser point cloud data, extract the geometric contour information and volume information of the object clumps;

[0034] Based on object clumps and visual image data, extract the color and texture information of the object clumps;

[0035] Based on object clumps and thermal distribution data, extract temperature distribution information of the object clumps;

[0036] Using object clumps as basic units, geometric contour information, volume information, color information, texture information, and temperature distribution information are combined to form an obstacle fusion feature set.

[0037] Furthermore, methods for obtaining physical exploration feature sets include:

[0038] The centroid coordinates of the point set are obtained by taking the arithmetic mean of the three-dimensional spatial coordinate points of the same object at different sampling times.

[0039] The displacement of the obstacle is obtained by calculating the Euclidean distance between the centroid coordinates of the point set at different times.

[0040] The displacement at each moment and the contact feedback force are combined into ordered point pairs to form a thrust displacement curve sampling sequence.

[0041] The obstacle displacement is accumulated over time to obtain the cumulative obstacle displacement.

[0042] The maximum contact feedback force, maximum motor load current, cumulative displacement, and thrust displacement curve sampling sequence during the active physical probing process are encapsulated to obtain a physical probing feature set.

[0043] Furthermore, methods for classifying the physical states of obstacles include:

[0044] Obstacles are categorized as fixed and stuck, loosely movable, and too heavy to move.

[0045] The cumulative obstacle displacement is compared with a preset displacement threshold. If the cumulative obstacle displacement is greater than the displacement threshold, the obstacle is determined to be a movable candidate; otherwise, the obstacle is determined to be an immovable candidate.

[0046] Among the movable candidates, the maximum contact feedback force is compared with the preset safety limit. If the maximum contact feedback force is less than the safety limit, the obstacle is classified as loosely movable; otherwise, it is classified as fixed and stuck.

[0047] Among the immovable candidates, the maximum contact feedback force is compared with the safety limit. If the maximum contact feedback force is greater than or equal to the safety limit, the obstacle is classified as fixed and stuck. Otherwise, impedance characteristics are introduced for further subdivision to determine the obstacle category.

[0048] Furthermore, the impedance characteristic subdivision determination includes:

[0049] Based on the thrust displacement curve sampling sequence, within the range where the cumulative displacement does not exceed the preset low displacement threshold, the thrust change and displacement change are obtained, the ratio of thrust change to displacement change is calculated, and the equivalent thrust impedance index is obtained.

[0050] The equivalent driving impedance index is compared with the preset high impedance threshold. If the equivalent driving impedance is greater than the high impedance threshold, the obstacle is determined to be fixed and stuck; otherwise, the obstacle is determined to be too heavy to move.

[0051] Furthermore, methods for obtaining the final state label of obstacles include:

[0052] Based on temperature distribution information, the obstacle categories obtained by physical state classification are checked for consistency. When the low temperature adhesion risk mark is established and the obstacle category is loose and movable, the final state label of the obstacle is set to fixed and stuck. When the low temperature adhesion risk mark is established and the obstacle category is too heavy and immovable or fixed and stuck, the final state label of the obstacle is set to the corresponding obstacle category.

[0053] Compared with the prior art, the technical effects and advantages of the dynamic obstacle avoidance control method for the track-following robot of the present invention are as follows:

[0054] This solution constructs an obstacle fusion feature set by fusing laser point cloud data, visual image data, thermal distribution data, and robot motion state. This allows for the simultaneous characterization of the obstacle's geometric embedding relationship, appearance features, and temperature distribution characteristics, overcoming the limitations of relying solely on material properties for judgment. Furthermore, by jointly correcting the visual category probability vector, the initial obstacle classification labels reflect the actual spatial relationship and environmental conditions between the obstacle and the track, reducing misjudgments caused by factors such as rain, snow, ice, and oil. Further, this solution acquires a physical exploration feature set through active physical probing and combines it with the obstacle fusion feature set to classify the obstacle's physical state. It directly determines the obstacle's physical mobility based on its actual mechanical response under controlled thrust, effectively distinguishing between a normally movable state and an attached, embedded, or stuck state, avoiding inferences about obstacle handling capabilities based solely on appearance or material properties.

[0055] Therefore, this solution can accurately determine the actual state of obstacles under complex track surface conditions, generating local control trajectories and execution strategy commands that match the actual physical state of the obstacles. This avoids both misjudging attached or stuck obstacles as manageable, thus preventing safety risks, and misjudging manageable obstacles as unmanageable, thus reducing inspection efficiency. This effectively improves the safety, accuracy, and engineering applicability of obstacle handling control for the track-tracking robot. Attached Figure Description

[0056] Figure 1 This is a flowchart of the dynamic obstacle avoidance control method for a track-following robot according to an embodiment of the present invention;

[0057] Figure 2 This is a flowchart illustrating the joint correction process for visual category probability vectors according to an embodiment of the present invention.

[0058] Figure 3 This is a flowchart of the active physics probing process according to an embodiment of the present invention;

[0059] Figure 4 This is a flowchart illustrating the physical state classification process according to an embodiment of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only for better illustrating and explaining the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and substance of the present invention, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in the present invention, and these should all be considered within the scope of protection of the present invention.

[0061] Example 1:

[0062] Please see Figure 1 As shown, this embodiment provides a dynamic obstacle avoidance control method for a track-following robot, including:

[0063] Based on real-time acquired laser point cloud data, visual image data, thermal distribution data, and robot motion state, an obstacle fusion feature set is calculated.

[0064] The laser point cloud data is obtained by real-time scanning and acquisition by the lidar sensor installed on the robot. The laser point cloud data includes a set of three-dimensional spatial coordinate points of the surrounding environment and track surface at the current sampling time. The three-dimensional spatial coordinate point set includes at least the spatial coordinate information of each point and the echo intensity information corresponding to the spatial coordinates. It is used to characterize the geometric shape of obstacles and track structure and for subsequent point cloud segmentation and contour volume calculation.

[0065] The visual image data is acquired in real time by a depth camera mounted in front of the robot. The visual image data includes at least one frame of color image data at the current sampling time and depth map information aligned with the color image. It is used to characterize the color, texture and shape details of the environment and serves as input data for subsequent obstacle category recognition based on deep learning.

[0066] The thermal distribution data is acquired in real time by an infrared thermal imager installed in the robot's forward field of view. The thermal distribution data includes a temperature matrix or thermal radiation intensity matrix within the field of view corresponding to the current sampling time. The temperature matrix includes at least pixel-level temperature values ​​of the obstacle surface and the track contact interface area, which is used to characterize the temperature difference between the obstacle and the track and support subsequent auxiliary identification of low-temperature adhesion or abnormal heat source targets.

[0067] The robot's motion state is provided by the robot's positioning and attitude measurement unit, which is composed of a wheel odometer and an inertial measurement unit. The wheel odometer outputs the robot's mileage increment, linear velocity, or wheel speed information at the current sampling moment through a wheel encoder. The inertial measurement unit outputs the robot's three-axis angular velocity, three-axis acceleration, and pitch and roll angle attitude information calculated from them at the current sampling moment. The robot's motion state is used to provide a unified timestamp and pose reference during multi-sensor data fusion, and is used to correct and constrain the robot's own positioning, slippage state, and vehicle stability.

[0068] Based on the robot's motion state, spatiotemporal alignment is performed between laser point cloud data and visual image data. The timestamp in the robot's motion state serves as a unified time reference, used to select the set of 3D spatial coordinate points corresponding to the sampling time from the laser point cloud data, and to select the color image data corresponding to the sampling time from the visual image data. The position information, pitch angle, and roll angle attitude information in the robot's motion state are used to determine the robot's pose relationship in the spatial coordinate system and to construct the mapping relationship from the robot's coordinate system to the spatial coordinate system. Let the track map coordinate system be... The robot's body coordinate system is The lidar coordinate system is The camera coordinate system is The mounting pose parameters of the LiDAR sensor, depth camera, and infrared thermal imager relative to the robot's body coordinate system are acquired and stored during the offline calibration phase. The mounting pose parameters represent the rigid body pose relationship between the sensor coordinate system and the robot's body coordinate system. This rigid body pose relationship is composed of rotation and translation parameters. The rotation parameters characterize the differences in coordinate axis directions, and the translation parameters characterize the differences in the position of the coordinate origin. The preferred output format for the mounting pose parameters is a 4x4 homogeneous transformation matrix. This homogeneous transformation matrix is ​​obtained by concatenating a 3x3 rotation matrix and a 3x1 translation vector, and includes homogeneous terms. This ensures that the extrinsic parameters are unique and easily accessible during spatial mapping and registration. The mounting pose parameters obtained during the offline calibration phase include... and ,in Indicates from coordinate system arrive The homogeneous transformation matrix. During online operation, the body pose at any time t is obtained from the robot's motion state. For the homogeneous coordinates of any point in the laser point cloud. Position in the orbital map coordinate system The calculation is as follows:

[0069] ;

[0070] To reduce cross-modal binding errors, the fusion computation adopts a unified time reference and a unified pose reference. The unified time reference is used to avoid pixel offset when the point cloud is projected onto the color image data, and the unified pose reference is used to avoid spatial offset of the point cloud in the orbital map coordinate system and deviation in safety limit judgment.

[0071] During single-frame point cloud acquisition using a rotating scanning lidar, the different sampling times for different points and the robot's movement can cause scanning distortion in the point cloud. This distortion is exacerbated when the lidar is rotating and the single-frame point cloud spans a time interval. To suppress scanning distortion, a unified reference time was selected. And perform motion compensation for each point:

[0072] ;

[0073] in The pose is obtained by interpolating the robot's motion state. The robot position is obtained by linear interpolation, and the robot posture is obtained by quaternion spherical linear interpolation to ensure the continuity of angles. This allows motion compensation to be implemented under any combination of sampling frequencies. The unified reference time is preferably the median time of the time interval spanned by a single frame point cloud.

[0074] The imaging intrinsic parameters of the depth camera are acquired during the offline calibration phase. These parameters include focal length, principal point, and distortion parameters, used to establish a correspondence between pixel positions in the depth camera's color image data and spatial directions in the spatial coordinate system. The infrared thermal imager further acquires its imaging geometric mapping parameters during the offline calibration phase. These parameters include equivalent infrared thermal imaging parameters and the imager's mounting pose. The equivalent infrared thermal imaging parameters are used to establish a correspondence between thermal imaging pixel positions and line-of-sight directions in the spatial coordinate system. Together, these parameters are used to map the thermal distribution data to the orbital map coordinate system. If necessary, a cross-spectral registration relationship is established using the corresponding points of the same calibration target under visible light and infrared imaging, ensuring that the thermal distribution data can be stably mapped to the spatial coordinate system and is feasible. The establishment of cross-spectral registration is triggered when the pixel deviation between the thermal imaging projection and the visible light projection exceeds a preset pixel registration tolerance threshold during the offline calibration and verification phase. The pixel registration tolerance threshold is set based on the reprojection error sequence calculated from the visible light image data and thermal distribution data collected during the offline calibration and verification phase, as well as the upper bound of the calibration residuals of the infrared thermal imager imaging geometric mapping parameters, the upper bound of the calibration residuals of the installation pose relationship between the infrared thermal imager and the robot body coordinate system, and the upper bound of the calibration residuals of the depth camera imaging intrinsic parameters. During the offline calibration and verification phase, visible light image data and thermal distribution data are collected simultaneously using a calibration target, and the cross-spectral reprojection error sequence is calculated. The upper bound of the pixel scale error budget is output during the offline calibration phase. The pixel registration tolerance threshold is the larger of the upper bound of the pixel scale error budget and the high quantile statistic of the reprojection error sequence, and the proportion of the reprojection error sequence exceeding the threshold does not exceed the preset upper limit of the false trigger probability. The reprojection error sequence is derived from repeatedly collected data during the offline calibration and verification phase. The upper bound of the pixel scale error budget is derived from the offline calibration solution output. The upper limit of the preset false trigger probability is derived from the robot's built-in system safety policy configuration file.

[0075] During the online operation phase, timestamps in the robot's motion state are used as a unified time reference to perform time matching on laser point cloud data, visual image data, and thermal distribution data. Time matching is performed using the most recent sampling time matching method. When the time difference between the sampling time of the laser point cloud data and the sampling time of the visual image data exceeds a preset synchronization tolerance threshold, the corresponding data set is discarded to avoid cross-sampling time fusion errors. Similarly, when the time difference between the sampling time of the thermal distribution data and the sampling time of the laser point cloud data exceeds the synchronization tolerance threshold, the corresponding data set is discarded to avoid cross-sampling time fusion errors. The logic for setting the synchronization tolerance threshold is based on the maximum allowable spatial alignment error, the robot calibration speed, and the upper bound of the time delay: the maximum allowable spatial alignment error is derived from offline playback evaluation data; the robot calibration speed is derived from the offline calibration condition configuration file; and the upper bound of the time delay is derived from the high quantile statistics of the corresponding sensor drive log difference sequence. The sensing sensors used for time synchronization include LiDAR, depth camera, and infrared thermal imager. The synchronization tolerance threshold is the smaller of the ratio of the maximum allowable spatial alignment error to the robot calibration speed and the upper bound of the time delay. The offline playback evaluation data comes from the playback results of the offline verification subset, and the sensor-driven log comes from the robot's built-in system sensor timestamp recording module.

[0076] The most recent sampling time matching method and the synchronization tolerance threshold discarding strategy constitute a time consistency constraint. The time consistency constraint is used to ensure that the laser point cloud data, visual image data and thermal distribution data meet the maximum allowable spatial alignment error budget within the fusion window.

[0077] The laser point cloud data is spatially constrained using a pre-loaded track map from the robot's built-in system. The track map contains spatial descriptions of the track structure and safety clearances. Based on the track boundary and height ranges defined in the track map, each 3D spatial coordinate point in the laser point cloud data is evaluated. Points outside the track safety clearances are identified as background noise and removed. Points within the track structure and a pre-defined allowable neighborhood are retained for subsequent point cloud segmentation. The track map stores the track centerline, left and right boundary lines, key geometric areas of fasteners and switches, and the track surface height envelope in a structured manner, and records the spatial coordinate and mileage references used during map construction. During online operation, the robot's positioning results are first registered to this map reference before safety clearance checks are performed. The allowed neighborhood range is determined based on the lateral position covariance and height covariance. The lateral position covariance and height covariance are derived from the internal state estimation output of the localization fusion module. The allowed neighborhood range is the sum of the pre-set confidence radius corresponding to the lateral position covariance and the track structure geometric margin, and in the height direction, it is the sum of the pre-set confidence radius corresponding to the height covariance and the track surface height envelope safety margin. The track structure geometric margin and track surface height envelope safety margin are derived from measured geometric statistics during the track map construction phase, and the pre-set confidence level is derived from the robot's built-in system safety policy configuration file. When a persistent inconsistency occurs between the map library and the real-time point cloud in a local segment, exceeding the preset map disabling duration threshold, a degraded mode is entered. Map point filtering is temporarily suspended, and instead, the track surface is fitted based on the real-time point cloud, with background removal performed using the track surface neighborhood. This ensures the system remains operational. The map disabling duration threshold is set based on the allowed upper limit of accumulated local map error mileage: this upper limit is derived from track map construction verification records. The preset duration is the ratio of the allowed upper limit of accumulated local map error mileage to the preset calibration speed, with a minimum time lower bound set. This minimum time lower bound is derived from the high quantile of the timestamp sequence sampling interval jitter statistics. The timestamp sequence is obtained from the sensor timestamp recording module. The determination of persistent inconsistency is achieved using a track surface height consistency index, which is the statistical measure of the height difference between the track surface height obtained from real-time point cloud fitting and the track surface height envelope on the track map at the same mileage position. Persistent inconsistency is determined when the absolute value of the height difference statistic continuously exceeds a preset inconsistency threshold within a continuous sampling period. The inconsistency threshold is set based on the cumulative offset statistic of the height difference data: the height difference data originates from the height difference between the track surface height obtained from real-time point cloud fitting and the track surface height envelope on the track map, and the cumulative offset statistic is obtained from the cumulative offset statistic. Figure 1 The upper bound of the high quantile in the scene is superimposed with the point cloud ranging resolution conversion margin, and the inconsistency judgment threshold is taken as the cumulative offset statistic. Figure 1The scene data is obtained from offline data collection in sections with stable track structures and reliable maps, and the point cloud ranging resolution is obtained from the lidar ranging accuracy calibration report.

[0078] The safety boundary space description information of the track map is used to constrain the retention area of ​​the laser point cloud data. The safety boundary space constraint can reduce the entry of background points around the track into the point cloud segmentation process. The allowable neighborhood range is determined by the lateral position covariance and height covariance, superimposed with the track structure geometric margin and the track surface height envelope safety margin. The allowable neighborhood range changes with the internal state estimation output of the positioning fusion module, which can make the point screening strategy consistent with the positioning reliability, thereby improving the robustness of point cloud screening. The track surface height consistency index and the continuous inconsistency judgment threshold are used to identify the local mismatch between the track map and the real-time point cloud. The continuous existence of the local mismatch will cause systematic rejection errors or systematic retention errors in the track map point screening. Systematic errors will directly affect the object cluster generation and embedding risk judgment results. The degradation mode switches to fitting the track surface based on the real-time point cloud and completing the background rejection by the track surface neighborhood. This can maintain the feasibility of background rejection and point cloud segmentation processing under the condition of track map mismatch, keep the obstacle fusion feature set generation process running continuously and reduce the risk of process interruption caused by track map mismatch.

[0079] Point cloud segmentation is performed on the retained laser point cloud data. This segmentation is based on the spatial distance relationships between sets of 3D spatial coordinate points. A clustering algorithm is used to aggregate 3D spatial coordinate point sets that meet the spatial proximity condition into independent object clusters. The clustering algorithm employs a point set aggregation method based on a preset spatial distance threshold and a minimum point count threshold. The spatial distance threshold is set based on the lidar angular resolution and the upper bound of the lidar ranging noise, and adapts to distance: the lidar angular resolution is derived from the lidar equipment specifications, the upper bound of the lidar ranging noise is derived from the ranging accuracy calibration report, and the spatial distance threshold is the weighted sum of the theoretical distance between adjacent points and the upper bound of the lidar ranging noise. The weighting coefficients are derived from the playback optimization results of the offline labeled point cloud dataset. The offline labeled point cloud dataset is derived from field-collected and manually labeled data. The minimum point count threshold is determined by inverse calculation based on the minimum detectable obstacle quantity requirement and the point cloud density model. The minimum detectable obstacle quantity requirement is derived from the inspection task risk target definition document. The point cloud density model is derived from the lidar angular resolution, scanning frequency, and offline point cloud density statistics. The minimum point count threshold is a conservative coefficient determined by multiplying the expected number of points by the upper bound of the occlusion rate and the upper bound of the reflection loss rate, and then rounding down. The upper bounds of the occlusion rate and the reflection loss rate are derived from offline playback statistics and material test records. The clustering algorithm preferably adopts Euclidean clustering based on nearest neighbor search. Specifically, any point is used as a seed point, and the nearest points of the seed point within the spatial distance threshold are searched and added to the same cluster. The nearest neighbor search is then iteratively performed on the newly added points until no new points are added. When the number of points in a cluster is less than the minimum point count threshold, the cluster is removed.

[0080] Based on object clusters and laser point cloud data, geometric contour and volume information of the object clusters are extracted. The geometric contour information of the object clusters is determined by the outer boundary of the set of 3D spatial coordinate points contained within the object cluster, which is generated using a 3D bounding box method. The volume information is calculated from the volume of the 3D bounding box or the volume occupied by voxels. For each object cluster, geometric contour information is calculated based on the set of 3D spatial coordinate points contained within the object cluster, and this geometric contour information is determined by the outer boundary of the set of 3D spatial coordinate points within the object cluster. The volume information is used to characterize the volumetric features of the object cluster in space. The volume information can be directly calculated as the volume of the 3D bounding box.

[0081] Brightness enhancement processing is performed on the visual image data. This is achieved by adjusting the pixel brightness distribution in the color image data to improve the discernibility of the visual image data in low-light or high-contrast environments. Brightness enhancement is implemented using histogram equalization or gamma correction with limited local contrast. The robot's built-in system stores a candidate set of enhancement parameters. This candidate set is a pre-calculated and stored series of specific parameter groups, providing optimal parameter options that balance detail enhancement and noise control for different scenarios, ensuring fast and stable processing. The setting logic for enhancement parameters is based on image brightness statistics: a preset number of images are collected under typical low-light and strong backlight scenarios, and the mean and standard deviation of image gray-level noise are statistically analyzed. Parameters that ensure the mean gray-level noise falls within a preset target range and the increase in the standard deviation does not exceed a preset noise amplification limit are selected as default enhancement parameters and stored in the candidate set. This ensures that during image processing, dark details are effectively enhanced while image noise is not excessively amplified.

[0082] Based on object clusters and visual image data, color and texture information of the object clusters are extracted. According to the established spatial mapping relationship between laser point cloud data and visual image data, the color and texture information in the brightness-enhanced visual image data are mapped to the corresponding object clusters. This ensures that each object cluster retains its original 3D spatial coordinates, geometric contour information, and volume information while being associated with the color and texture information of the corresponding region. Based on the installation pose relationship obtained from offline calibration within the imaging context, the 3D spatial coordinates of the object clusters are projected onto the imaging coordinates corresponding to the color image data. Color and texture information are extracted from the pixel regions where the projection points fall, thus associating the color and texture information with the corresponding object clusters. To ensure the feasibility of projection and pixel association, pixels are restored to 3D points and projected based on the depth map and depth camera imaging intrinsic parameters. The pixel region is selected using a preset neighborhood window centered on the projection point or a closed region formed by a block projection contour. The neighborhood window size is determined based on the depth camera imaging intrinsic parameter calibration residual, the depth camera mounting pose relationship calibration residual, and the upper bound of the depth recovery error. The depth camera imaging intrinsic parameter calibration residual is derived from the offline calibration solution output of the depth camera; the depth camera mounting pose relationship calibration residual is derived from the offline calibration solution output of the mounting pose relationship between the depth camera and the robot's body coordinate system; the upper bound of the depth recovery error is derived from the depth camera depth accuracy calibration report; and the neighborhood window size is the pixel error upper bound obtained by propagating the above errors through the projection model and rounded up. The depth camera depth accuracy calibration report is derived from the equipment calibration record.

[0083] Based on object cluster and thermal distribution data, temperature distribution information of the object cluster is extracted. According to the installation pose relationship obtained from offline calibration and the pose information in the robot's motion state, the temperature matrix is ​​mapped to a spatial coordinate system. The pixel-level temperature values ​​in the temperature matrix are assigned to the corresponding object clusters based on the spatial overlap between the projected area of ​​the temperature matrix in the spatial coordinate system and the projected area of ​​the object cluster in the spatial coordinate system, thus obtaining the temperature distribution information of the object clusters. To reduce the impact of absolute error in thermal imaging temperature measurement on the judgment, relative temperature difference features are preferred over absolute temperature values ​​for the temperature distribution information. Before assignment, abnormal points in the temperature matrix exceeding a preset abnormality threshold are removed and smoothed. The abnormality threshold is set based on the temperature statistics of the track background area collected under conditions without abnormal heat sources: the temperature statistics of the track background area are derived from the track background area of ​​the infrared thermal imager temperature matrix. The abnormality removal threshold is the product of the median temperature of the track background area plus a preset multiple and the estimated temperature noise scale value of the track background area. Only isolated connected regions with an area smaller than the connected region area threshold are removed. The temperature statistics of the track background area are derived from offline data collected under conditions without abnormal heat sources. The connected region area threshold is derived from the conversion result between the spatial resolution of the infrared thermal imager and the minimum discernible scale of the track background area. Combining the correspondence between the thermal distribution data and the visual image data at the sensor installation location and field of view, the temperature matrix in the thermal distribution data is mapped to a spatial coordinate system, and the pixel-level temperature values ​​in the temperature matrix are assigned to the corresponding object clusters, so that the object clusters have temperature distribution information corresponding to the obstacle surface and the track contact interface.

[0084] Using object clumps obtained from point cloud segmentation as basic units, the geometric contour information, volume information, color information, texture information, and temperature distribution information of the object clumps are uniformly organized and encapsulated according to the corresponding timestamps to form a structured obstacle fusion feature set.

[0085] The obstacle fusion feature set is input into the pre-trained target recognition model to obtain the visual category probability vector. The visual category probability vector is jointly corrected based on laser point cloud data and thermal distribution data to obtain the preliminary obstacle classification label and confidence score.

[0086] The target recognition model is implemented using a deep neural network. The deep neural network establishes a mapping between visual information in the obstacle fusion feature set and the preliminary obstacle classification labels. The training objective is to ensure that, given the input of the obstacle fusion feature set, the obstacle category prediction result output by the deep neural network is consistent with the true obstacle category label. The deep neural network adopts a convolutional neural network structure, which consists of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layers extract local spatial features from the visual image data corresponding to the object clumps; the pooling layers downsample and scale the local spatial features; and the fully connected layers perform nonlinear mapping on the high-dimensional features and output the category prediction result. The last layer of the convolutional neural network structure is the classification output layer.

[0087] The input data for the deep neural network consists of visual image data corresponding to object clumps. The obstacle fusion feature set uses object clumps obtained from point cloud segmentation as basic units. To enable the convolutional neural network structure to directly learn convolutional features from visual information, the color and texture information from the obstacle fusion feature set are provided to the deep neural network in the form of visual image data corresponding to object clumps. The visual image data is input into the deep neural network in the form of a pixel matrix, where each pixel contains color and texture information, serving as the raw input for feature learning. The geometric contour information, volume information, and temperature distribution information in the obstacle fusion feature set correspond to the object clump visual image data under the same object clump identifier and the same timestamp, and are used for subsequent strategy linkage, thereby ensuring the consistency and traceability of the input data organization.

[0088] The output data of the deep neural network is a visual category probability vector. This vector is a normalized multi-dimensional probability vector, with its dimension equal to the number of visual categories in a predefined set. Each vector component corresponds to a visual category, and the value of each component represents the predicted probability that the object cluster visual image data belongs to the corresponding visual category. The sum of all vector components is one. The visual category set describes the visual classification of obstacles. This set is predefined during the target recognition model construction phase based on the morphological characteristics, material characteristics, and color and texture differences of obstacles that may appear in the working environment. Each visual category in the set corresponds to a type of obstacle with significant consistency in appearance, color distribution, or texture structure. The visual category set includes obstacles such as stones, snow, metallic objects, plastic objects, and unknown obstacles. The reason for using object cluster visual image data as the visual input data for the deep neural network is that the appearance, color distribution, and texture structure of obstacles are the main observable features that distinguish different visual categories. Stone-like obstacles typically exhibit irregular geometric shapes and rough textures, snow-like obstacles possess high brightness and uniform texture, metallic obstacles display significant reflective properties, and plastic obstacles exhibit stable characteristics in color distribution and surface texture. Visual image data of object clumps can directly represent these appearance shapes, color distributions, and texture structures, enabling deep neural networks to effectively distinguish different visual categories based on visual differences, even when objects have similar geometric sizes or outlines.

[0089] Training of the deep neural network is completed in the offline training phase. During offline training, the deep neural network training dataset consists of a large number of obstacle fusion feature set samples with labeled obstacle categories. Each training sample uses an object cluster as the basic unit and contains at least the visual image data of the object cluster and the corresponding real obstacle category label. The visual image data of the object cluster and the real obstacle category label correspond at the same sampling time. The deep neural network training dataset is divided into training and validation datasets in a 7:3 ratio. The training objective of the deep neural network is to minimize the classification error between the predicted visual category probability vector and the real obstacle category label, which is measured using the cross-entropy loss function. The parameters in the convolutional neural network structure are updated using the backpropagation algorithm, and the parameter update method is batch gradient descent. During training, the learning rate is set as a necessary numerical parameter. The learning rate is either fixed or decreases with each training epoch to control the magnitude of parameter updates and prevent oscillations or divergence in the target recognition model training process. After offline training is completed, the parameters of the deep neural network remain fixed. During the online operation phase, the deep neural network only performs the forward inference process, outputting a visual category probability vector based on the input visual image data of the object clump, without performing parameter update operations, thus meeting the real-time requirements.

[0090] Before entering the deep neural network, the visual image data of the object clusters is uniformly scaled to a preset resolution and color normalized. The normalization parameters are fixed and stored along with the model. In a preferred embodiment, the resolution is a square resolution with a length and width of 224 pixels. The scaling method is to maintain the aspect ratio and fill in the insufficient edges to ensure consistent input size and avoid introducing uncontrollable deformation. The color normalization process is achieved by scaling the pixel channel values ​​to the range of 0 to 1 by a fixed ratio and subtracting the pre-statistical channel mean. The channel mean is obtained from the offline training dataset and fixed and stored to ensure consistency in online inference preprocessing.

[0091] In a preferred embodiment, the convolutional neural network employs a four-level convolutional feature extraction structure. Each level includes convolution operations and nonlinear activation operations, with pooling downsampling between levels. The nonlinear activation operation preferably uses rectified linear activation. A dropout strategy is preferably used before the fully connected layer outputs to suppress overfitting. The training dataset sources include on-site data collected under sunny, nighttime, low-light tunnel, rain, snow, and low-temperature environments. The on-site data includes visual image data of object clumps obtained through brightness enhancement and object clump cropping. The annotation is based on the actual category of obstacles on the track surface, with the annotation rule prioritizing the actual material and shape category of the obstacles rather than just their color and appearance, thus ensuring consistency between training and implementation.

[0092] Geometric embedding analysis is performed on laser point cloud data based on obstacle fusion feature sets. Using object clusters obtained from point cloud segmentation as analysis units, for each object cluster, entry depth and entry point ratio are extracted based on the set of 3D spatial coordinates contained within the object cluster and the track structure description in the track map. Specifically: The bottom 3D spatial coordinates of the object cluster are extracted from the set of 3D spatial coordinates contained within the object cluster. This extraction method involves sorting the 3D spatial coordinates of the object cluster by vertical coordinates and selecting the preset quantile point set with the smallest vertical coordinate as the bottom 3D spatial coordinates of the object cluster. The quantile can be preset based on the size of common actual obstacles during track inspection, for example, a quantile of 10%. The track surface height threshold and track gap region are extracted based on the track structure description in the track map. This extraction method involves retrieving the structural partition to which each point in the bottom 3D spatial coordinates of the object cluster belongs along a continuous interval in the mileage direction, querying the corresponding fitted track surface height and the upper bound of the residual distribution of the structural partition (low quantile), and superimposing the ranging resolution margin to obtain the track. The surface height threshold is used, and the range below the track gap region is determined by defining the track gap region in a continuous interval format along the mileage direction, based on the geometric gap between the two rail surface boundaries on the track map and the recessed area of ​​the fastener neighborhood. The set of three-dimensional spatial coordinate points at the bottom of the object cluster is filtered based on the track surface height threshold and the range below the track gap region. The three-dimensional spatial coordinate points at the bottom of the object cluster whose vertical coordinates are less than the track surface height threshold and fall below the range of the track gap region are formed into an entry point set. The entry depth is extracted based on the entry point set. The entry depth extraction method is to calculate the track surface height threshold for each point in the entry point set, subtract the vertical coordinate to obtain the negative height difference, and take the representative value of the negative height difference as the entry depth. A threshold judgment is performed on the entry depth. When the entry depth is greater than or equal to a preset depth threshold, the risk assessment result for the corresponding object cluster is that there is an embedding risk; otherwise, the risk assessment result for the corresponding object cluster is that there is no embedding risk. The depth threshold is a preset multiple of the upper limit of the resolution of the lidar ranging calibration. The preset multiple can be determined according to the error statistics of the upper limit of the lidar ranging calibration and the actual entry depth measured during track inspection. For example, if the upper limit of the lidar ranging calibration resolution is 2 mm, and the negative fluctuation of the entry depth in the sample without embedding risk does not exceed 8 mm at the 99th percentile, then the preset multiple is taken as 4, so that the depth threshold is 8 mm. This controls the situation where the entry depth exceeds the depth threshold due to normal ranging fluctuations within an acceptable range.

[0093] Temperature difference analysis is performed based on thermal distribution data from obstacle fusion feature sets. The thermal distribution data includes temperature matrices corresponding to the obstacle surface and track contact interface regions. Based on object clumps obtained through point cloud segmentation, the region correspondences in the temperature matrix are determined. The bottom 3D spatial coordinates of the object clumps are mapped to the imaging region corresponding to the temperature matrix to obtain the track contact interface region. The non-bottom 3D spatial coordinates of the object clumps are mapped to the imaging region corresponding to the temperature matrix to obtain the obstacle surface region. Pixel-level temperature values ​​of the obstacle surface region and the track contact interface region are statistically analyzed to obtain representative temperature values ​​for the obstacle surface region and the track contact interface region. The absolute value of the difference between the representative temperature values ​​of the obstacle surface region and the track contact interface region is used as the representative temperature value. Temperature difference characteristics: The ambient temperature is obtained through temperature statistics of the track background area in the temperature matrix or by an ambient temperature sensor. The selection rule for the track background area is that it is located at the edge of the field of view and does not include the obstacle projection. Pixels with abnormally high temperatures are removed before statistics to avoid interference from external heat sources on the ambient temperature estimation. The logic for setting the low-temperature condition threshold is based on historical ambient temperature records and material freezing and adhesion test records: historical ambient temperature records are from meteorological station historical records or line environmental monitoring records, and material freezing and adhesion test records are from adhesion test records. The low-temperature condition threshold is the safe and unfavorable value of the adhesion rise inflection point temperature and the low quantile temperature of the historical ambient temperature record, and it is fixed and stored. The fixed storage location is the threshold table of the robot's built-in system. When the ambient temperature is low, if the ambient temperature is less than the preset low-temperature condition threshold and the temperature difference characteristic is less than the preset temperature difference threshold, it is determined that there is a risk of low-temperature adhesion between the obstacle and the track. The temperature difference threshold is set based on fitting offline controllable temperature test data: the offline controllable temperature test data comes from the infrared thermal imager temperature matrix and the actual adhesion state labels, the actual adhesion state labels come from the separability test records or adhesion measurement records, and the temperature difference threshold is the output value of the temperature difference boundary curve that satisfies the adhesion false positive rate not exceeding the preset upper limit, and is updated according to the ambient temperature. The preset upper limit comes from the robot's built-in system safety policy configuration file.

[0094] Please see Figure 2 As shown, the visual category corresponding to the component with the highest probability value in the visual category probability vector is taken as the visual judgment result.

[0095] Based on the embedding risk assessment results, corresponding embedding risk markers are generated to characterize whether there is a spatial embedding relationship between the obstacle and the track structure; at the same time, based on the low-temperature adhesion risk assessment results, corresponding low-temperature adhesion risk markers are generated to characterize whether there is a possibility of physical adhesion between the obstacle and the track under low-temperature environmental conditions.

[0096] Physical consistency verification is performed between the visual judgment result and the embedded risk marker and the low-temperature adhesion risk marker. Physical consistency verification is implemented through pre-fixed physical consistency verification rules stored in the robot's built-in system. These rules, indexed by visual category, specify the verification relationship between the visual category and the embedded risk marker, and between the visual category and the low-temperature adhesion risk marker. The verification relationship is either consistent or conflicting. Specifically, when the embedded risk marker is valid, the verification conclusion between the visual judgment result and the embedded risk marker is determined according to the pre-given verification relationship between the visual category and the embedded risk marker in the physical consistency verification rules. When the low-temperature adhesion risk marker is valid, the verification conclusion between the visual judgment result and the low-temperature adhesion risk marker is determined according to the pre-given verification relationship between the visual category and the low-temperature adhesion risk marker in the physical consistency verification rules. If any verification conclusion is conflicting, the visual judgment result is determined to conflict with either the embedded risk marker or the low-temperature adhesion risk marker; otherwise, it is determined to be consistent.

[0097] When a visual judgment result conflicts with an embedded risk marker or a low-temperature adhesion risk marker, a probability reweighting process is performed on the visual category probability vector. The probability reweighting process is implemented using a multiplicative attenuation method: an attenuation coefficient is applied to the component in the visual category probability vector that corresponds to the obstacle category of the visual judgment result. The attenuation coefficient is determined by the number of risk markers that constitute a conflict. When only the visual judgment result conflicts with the verification conclusion of the embedded risk marker or only the verification conclusion of the low-temperature adhesion risk marker, the attenuation coefficient is a preset first attenuation coefficient. When both the visual judgment result and the verification conclusion of the embedded risk marker conflict, the attenuation coefficient is a preset second attenuation coefficient, and the second attenuation coefficient is less than the first attenuation coefficient. The preset logic of the first attenuation coefficient is determined based on the false alarm rate and false negative rate of the embedded risk judgment module. The false alarm rate and false negative rate of the embedded risk judgment module are derived from the statistics of the offline labeled dataset with real embedded status labels. The first attenuation coefficient is the maximum attenuation coefficient that ensures the proportion of immovable obstacles mistakenly judged as movable does not exceed the preset safety target under the condition of only embedded risk marker conflict. The preset safety target is derived from the robot's built-in system safety policy configuration file. The preset logic for the second attenuation coefficient is based on the false alarm rate and false alarm rate of the embedded risk judgment module and the false alarm rate and false alarm rate of the low-temperature adhesion risk judgment module. The false alarm rate and false alarm rate of the low-temperature adhesion risk judgment module are derived from statistics of offline labeled datasets with actual adhesion status labels. The second attenuation coefficient is the maximum attenuation coefficient that ensures the proportion of immovable obstacles being mistakenly judged as movable does not exceed the preset safety target under the conditions of embedded risk label conflict and low-temperature adhesion risk label conflict, and satisfies that the second attenuation coefficient is less than the first attenuation coefficient. The preset safety target is derived from the robot's built-in system safety policy configuration file. The visual category probability vector after probability reweighting is normalized, and the obstacle category with the highest normalized probability value is used as the initial obstacle classification label. The probability value corresponding to the component with the highest normalized probability value is used as the confidence score.

[0098] The target recognition model uses only visual image data as input. While visual image data directly represents appearance, color distribution, and texture structure, it cannot directly provide evidence of the spatial embedding relationship between obstacles and track structures, or the physical adhesion probability between obstacles and track. Embedding risk assessment results are derived from laser point cloud geometric embedding analysis, reflecting the spatial relationship below the track gap area. Low-temperature adhesion risk assessment results are derived from thermal distribution temperature difference analysis, reflecting the physical adhesion probability under low-temperature conditions. Physical consistency verification rules use visual categories as indexes to provide verification relationships. When a conflict is found, probability reweighting is performed to suppress the probability of conflicting categories. The normalization result after probability reweighting ensures that the jointly corrected preliminary obstacle classification labels satisfy the constraints of geometric embedding relationships and low-temperature adhesion risk assessment results. The first and second attenuation coefficients are determined by the false positive and false negative rates of the embedding risk assessment module and the low-temperature adhesion risk assessment module, and meet preset safety targets. The attenuation coefficient calibration constraints enhance the auditability of the joint correction analysis and reduce the probability of high-risk false positives.

[0099] Based on the obstacle fusion feature set and the preliminary obstacle classification label, the robot is controlled to actively physically explore the obstacles and obtain the physical exploration feature set.

[0100] Please see Figure 3 As shown, when an obstacle is determined to exist based on the obstacle fusion feature set and the target recognition model, the robot is controlled to perform active physical probing operations on the obstacle. The initial obstacle classification label and confidence score are used to generate thrust strategy constraints to ensure that the active physical probing process is conducted within a safe range. The thrust strategy constraints are used to convert the initial obstacle classification label and confidence score into thrust control boundaries and thrust growth rules for the active physical probing process, ensuring that thrust changes are controlled and reducing overshoot risk. The thrust strategy constraints include two types of constraint parameters: initial thrust level and thrust increment rate. These constraint parameters are jointly determined by the initial obstacle classification label and confidence score and are fixed as fields in the robot's built-in system.

[0101] Based on the output resolution and safety margin in the robot arm's built-in parameters, a first initial thrust level, a second initial thrust level, a first thrust increment rate, and a second thrust increment rate are set. Specifically, this includes: reading the output resolution from the robot arm's built-in parameters to obtain the minimum thrust change unit; and reading the safety margin and preset safety upper limit from the robot arm's built-in parameters to obtain the thrust level after deducting the safety margin from the preset safety upper limit, which is the allowable upper limit of the initial thrust. A candidate thrust level sequence is generated using the minimum thrust change unit as the unit scale, increasing incrementally from the minimum thrust change unit, and selecting the smallest candidate thrust level that does not exceed the allowable upper limit of the initial thrust as the first initial thrust level. Based on the first initial thrust level, the candidate thrust level is increased upwards by at least one minimum thrust change unit, and the smallest candidate thrust level that satisfies both the requirement of being not lower than the first initial thrust level and not exceeding the allowable upper limit of the initial thrust is selected as the second initial thrust level. The control cycle parameters are read, and the minimum thrust increment per cycle is obtained from the output resolution as the minimum thrust change unit. Simultaneously, based on the safety margin and the preset safety limit, the upper bound of the allowable thrust increment per cycle is determined, ensuring that an increase in thrust per cycle does not trigger the preset safety limit. The single-cycle thrust increment is set to the maximum candidate thrust increment within the upper bound of the allowable thrust increment per cycle, and the first thrust increment rate is calculated using the control cycle. The single-cycle thrust increment is set to the minimum candidate thrust increment that is not greater than the single-cycle thrust increment corresponding to the first thrust increment rate and does not exceed the upper bound of the allowable thrust increment per cycle, and the second thrust increment rate is calculated using the control cycle.

[0102] The physical attribute type of the preliminary classification label of the obstacle is determined to obtain the corresponding physical attribute type; the physical attribute type includes the light and loose category and the heavy and dense structure category.

[0103] Determining the physical attribute type of the initial obstacle classification label includes:

[0104] When an obstacle is initially classified as either a snow obstacle or a plastic foreign object obstacle, it is determined that the obstacle belongs to the lightweight and loose category. Snow obstacles are loosely piled up in structure, have low mass per unit volume, and are prone to shape changes; plastic foreign object obstacles have low overall mass, limited structural strength, and are prone to displacement or deformation under external forces.

[0105] When an obstacle is initially classified as a rock obstacle or a metallic object obstacle, it is determined that the obstacle belongs to the category of heavy, dense structures. Rock obstacles have high density and strong overall structural stability, while metallic objects usually have high mass and rigidity and are not easily deformed under external forces.

[0106] When an obstacle is initially classified as an unknown obstacle, it is classified into the dense heavy structure category by default to ensure safety. This is to avoid risks caused by using too low an initial thrust or too fast a thrust increment rate when the obstacle's properties are unclear.

[0107] Based on the physical attribute type corresponding to the initial obstacle classification label, an initial thrust level and a thrust increment rate are set. When the physical attribute type corresponding to the initial obstacle classification label is light and loose, the initial thrust level is set to the first initial thrust level, and the baseline thrust increment rate is set to the first thrust increment rate. When the physical attribute type corresponding to the initial obstacle classification label is heavy and dense, the initial thrust level is set to the second initial thrust level, and the baseline thrust increment rate is set to the second thrust increment rate. The first initial thrust level is lower than the second initial thrust level, and the first thrust increment rate is higher than the second thrust increment rate, thus allowing different thrust growth rules for different categories.

[0108] The confidence score is used to reduce the baseline thrust increment rate. Essentially, the confidence score characterizes the reliability of the initial obstacle classification label. The lower the reliability, the more likely a more aggressive thrust increase during active physics testing will amplify the classification error into structural and overshoot risks. Therefore, the confidence score needs to be mapped to a thrust increment rate reduction factor.

[0109] The confidence score is divided into multiple confidence score levels, and a corresponding thrust increment reduction coefficient is assigned to each confidence score level. A reduction coefficient mapping table is preset through offline historical data analysis or offline control testing. Specifically, firstly, based on the output resolution in the robot arm's built-in parameters, the minimum thrust increment that can be stably achieved within a single control cycle is determined, and the minimum thrust increment is converted into the minimum achievable change range of the thrust increment rate. Then, based on the safety margin, the maximum allowable thrust increment within a single control cycle is determined through offline safety testing. The lower and upper bounds of the thrust increment rate are jointly determined by the maximum and minimum thrust increments, thus obtaining the adjustable range of the thrust increment rate.

[0110] During the offline phase, based on historical operational data or offline control tests, the number of confidence rating levels is gradually increased within the adjustable range of the thrust increment rate, and different target reduced thrust increment rates are assigned to adjacent confidence rating levels. When the difference between the target reduced thrust increment rates corresponding to adjacent confidence rating levels is less than the minimum achievable change range of the thrust increment rate, or when a stable and repeatable difference in system response cannot be observed in historical data and offline test results, the increase in the number of confidence rating levels is stopped, thereby determining the number of confidence rating levels.

[0111] After determining the number of confidence levels, within the adjustable range of the thrust increment rate, the target reduced thrust increment rate corresponding to each confidence level is determined in order of confidence level, so that the upper bound of the thrust increment rate corresponding to the highest confidence level is set, and the lower bound of the thrust increment rate corresponding to the lowest confidence level is set. The target reduced thrust increment rate corresponding to the remaining confidence levels is determined by historical data statistics or offline control testing. Then, using the benchmark thrust increment rate determined by the preliminary obstacle classification label as the normalization benchmark, the ratio of the target reduced thrust increment rate corresponding to each confidence level to the benchmark thrust increment rate is calculated to obtain the thrust increment rate reduction coefficient corresponding to each confidence level. The thrust increment rate reduction coefficient is then clipped to between 0 and 1 to form a reduction coefficient mapping table.

[0112] During the online operation phase, the corresponding thrust increment rate reduction coefficient is obtained by directly querying the reduction coefficient mapping table according to the confidence score level. The benchmark thrust increment rate is then reduced to obtain the thrust increment rate used for the active physics test process. The thrust increment rate reduction coefficient is set between 0 and 1 and decreases monotonically as the confidence score level decreases.

[0113] Based on the robot's motion state, the robot is controlled to move towards the obstacle at a preset limited speed to avoid impact loads upon contact. During the robot's approach to the obstacle, contact force sensors are deployed on the robot's end effector to read the contact feedback force in real time. These sensors are positioned along the force transmission path where the robot and obstacle make physical contact, detecting the normal contact force or resultant force component generated during the robot's application of a thrust to the obstacle. To reduce noise interference from the contact feedback force, a moving average filter or low-pass filter is applied. Contact determination is performed using a continuous over-threshold determination method. When the filtered contact feedback force exceeds the contact threshold for multiple consecutive sampling periods, physical contact between the robot and the obstacle is determined, and the sampling time that first meets the continuous over-threshold determination condition is recorded as the contact start time. The contact threshold is determined based on contactless, no-load data acquisition: Contact feedback force sequences are continuously acquired while the actuator is not in contact with any object and the robot remains stationary. The contact threshold is set as the maximum value of the contact feedback force sequence, and a minimum detection margin determined by the contact force sensor resolution and zero-drift compensation residual is added to ensure that contact feedback forces in the no-contact state do not trigger contact judgment, and that actual contact can stably exceed the contact threshold. The number of consecutive judgments corresponding to multiple consecutive sampling periods is determined based on no-contact threshold-crossing events: The contact feedback force sequence is examined using the contact threshold in the no-contact state, and the longest continuous duration exceeding the contact threshold is counted. The number of consecutive judgments is set as the smallest integer greater than the number of sampling periods corresponding to the longest continuous duration, to suppress misjudgments caused by occasional threshold-crossing events in the no-contact state. The limited travel speed is restricted by relevant robot safety parameters to avoid impact loads when the robot contacts obstacles.

[0114] After physical contact is confirmed, the robot arm gradually increases the thrust applied to the obstacle according to the thrust strategy constraints. The thrust is generated by the output of the walking-driven joint motors of the robot arm. During the thrust increase, the contact feedback force and motor load current are simultaneously read at a fixed sampling period. When the contact feedback force reaches a preset safety limit or the motor load current reaches a preset current limit, the thrust is stopped from increasing further, and the robot either stops approaching the obstacle or performs a retraction action to ensure robot safety. The preset safety limit is determined based on the robot's structural strength and the rated load capacity of the actuator: the smaller of the allowable maximum contact feedback force corresponding to the structural allowable load and the rated load capacity of the actuator is taken as the upper limit benchmark, and safety reductions are made according to structural design safety specifications and material fatigue requirements to obtain the preset safety limit; the preset current limit is determined based on the motor thermal limit and the manufacturer's rated parameters. The retraction action is used to release the physical contact.

[0115] The contact feedback force and motor load current are time-aligned according to the timestamps in the robot's motion state, so that the contact feedback force, motor load current, and robot motion state at the same sampling moment constitute a synchronous data sequence. This synchronous data sequence is used to construct the thrust-displacement curve and the physical exploration feature set. During the thrust application process, the obstacle displacement is calculated using laser point cloud data. The obstacle displacement is calculated by taking the centroid coordinates of the three-dimensional spatial coordinate points corresponding to the same object mass at different sampling moments. The centroid coordinates are defined as the arithmetic mean position of each spatial coordinate point in the three-dimensional spatial point set along the three-dimensional coordinate axis. Specifically, the three-dimensional coordinate values ​​of each spatial coordinate point in the point set are summed and normalized to obtain the centroid coordinates of the point set at the corresponding sampling moment. The spatial distance between the centroid coordinates of the point set corresponding to different moments is taken as the obstacle displacement. The spatial distance is calculated using Euclidean distance in a unified reference coordinate system. The centroid coordinates of the point set are aligned according to the robot's motion state, so that the obstacle displacement and the robot's motion state are consistent in the same spatial coordinate system. To ensure the feasible association of the same object clusters in cross-time calculations, a cluster identifier is maintained for each cluster during online runtime, and cross-time matching is performed. The cross-time matching satisfies that the matching distance does not exceed a preset cluster tracking threshold. When the matching is uncertain, the displacement of the corresponding time period is invalidated, and the thrust increase of the active physical probing process is paused or terminated. The cluster tracking threshold is determined based on the static scene consistency constraint: under the condition that the obstacle is stationary and the robot is stopped, the sequence of centroid position changes at consecutive time moments is obtained by calculating the centroid position changes in the same segmentation and centroid. The cluster tracking threshold is set as the maximum value of the centroid position change sequence and superimposed with a minimum association margin. The minimum association margin is determined by the point cloud ranging resolution and the segmentation jitter residual, so that the clusters can continuously match under static conditions and the actual displacement can stably cross the cluster tracking threshold.

[0116] A thrust-displacement curve sampling sequence is constructed based on contact feedback force and obstacle displacement. During active physical probing, contact feedback force and obstacle displacement are simultaneously collected and aligned using timestamps. Starting from the moment of first stable contact, zero-point correction is performed on force and displacement, discarding free travel data before contact. Displacement values ​​at each moment are paired with contact feedback force to form ordered point pairs, creating a thrust-displacement discrete sampling sequence. This sequence is filtered to obtain the thrust-displacement curve sampling sequence, ensuring that under controlled thrust increments, the overall thrust-displacement curve satisfies the weak monotonicity of force and displacement, thus obtaining a thrust-displacement curve that reflects the obstacle's force response characteristics. The thrust-displacement curve is used to characterize the relationship between thrust changes and obstacle displacement changes. When the contact feedback force continuously increases with thrust and the obstacle displacement remains close to zero, the obstacle is considered to exhibit high thrust-resistance characteristics; when the contact feedback force is within a controlled range and the obstacle displacement accumulates over time, the obstacle is considered to exhibit low thrust-resistance characteristics. During active physical probing, the amount of obstacle displacement generated during the contact force phase is accumulated over time to obtain the cumulative displacement.

[0117] During active physics exploration, the robot's pitch and roll angle attitude information is continuously used to constrain its stability. When the pitch or roll angle exceeds a preset stability threshold, the thrust is limited or the thrust application process is terminated to ensure the robot's attitude stability during active physics exploration. The stability threshold is determined based on the robot's anti-tipping safety angle: the stability threshold is set to an angle lower than the upper limit of the robot's anti-tipping safety angle while retaining an attitude control margin. The attitude control margin is determined based on the upper limit of the robot's attitude control error and the upper limit of the attitude disturbance caused by the maximum permissible unevenness of the ground, ensuring that the attitude constraints do not approach the overturning boundary under the condition of superimposed disturbances.

[0118] The maximum contact feedback force, maximum motor load current, cumulative displacement, and thrust-displacement curve recorded during active physical probing are uniformly organized and encapsulated to form a physical probing feature set. This physical probing feature set is defined as a field table within the robot's built-in system. The obstacle fusion feature set provides a set of three-dimensional spatial coordinate points, geometric contour information, volume information, color information, texture information, and temperature distribution information. The obstacle's final state label characterizes the physical mobility state under controlled thrust, which requires mechanical response evidence. Active physical probing operations constrain the initial thrust level and thrust increment rate through thrust strategy constraints, and establish a contact determination and thrust increment sampling mechanism under limited travel speed conditions. The contact feedback force, motor load current, and obstacle displacement constitute the thrust-displacement curve sampling sequence. The thrust-displacement curve sampling sequence makes the thrust impedance characteristics explicit, which are used to distinguish the sources of insufficient displacement response and improve the reliability of determinations for fixed jamming, excessive weight preventing movement, and loose mobility. The pitch and roll angle attitude information of the robot in motion is used for stability constraints. The stability constraints, together with the preset safety upper limit and current upper limit, limit the risk boundary of the probing process, so that the active physical probing operation remains feasible under the constraints of structural strength and anti-tipping safety angle.

[0119] Based on the obstacle fusion feature set and the physical exploration feature set, the physical state of the obstacles is classified to obtain the final state label of the obstacles.

[0120] Please see Figure 4 As shown, the cumulative obstacle displacement is determined. The cumulative obstacle displacement is compared with a preset displacement threshold. Instantaneous displacement or single-frame displacement is easily affected by clutter segmentation jitter and pose alignment errors, and cannot stably reflect whether the obstacle has undergone actual physical movement. The cumulative obstacle displacement, by accumulating the displacement during the active physical probing process over time, can effectively suppress random errors, allowing the determination result to reflect the overall response trend of the obstacle under controlled thrust. Obstacle categories include fixed and stuck, loosely movable, and excessively heavy and immovable.

[0121] When the cumulative displacement is greater than the displacement threshold, it is determined that the obstacle has produced a real and effective displacement response under the control of the thrust, and the obstacle is classified as a movable candidate. When the cumulative displacement is less than or equal to the displacement threshold, it is determined that no effective displacement response has been produced, and the obstacle is classified as an immovable candidate.

[0122] The displacement threshold is determined based on the uniformity constraint of displacement in a static scene: laser point cloud data or visual image data are continuously collected under the condition that the robot is stopped and the obstacle is not under force. The obstacle displacement sequence is calculated according to the same method as the active physical exploration process, which includes clumping segmentation, clumping identification maintenance, centroid calculation and pose alignment. The displacement threshold is set as the maximum value of the displacement sequence and a minimum displacement discrimination margin is superimposed. The minimum displacement discrimination margin is determined by the ranging resolution and the segmentation jitter residual, so that no effective displacement judgment is generated under static conditions, while the actual movement can stably exceed the displacement threshold.

[0123] Among movable candidates, the maximum contact feedback force is determined. The maximum contact feedback force is the peak value of the contact feedback force recorded during active physical probing. Simply determining whether displacement occurs is insufficient to determine whether the corresponding displacement is sustainable within the safe force range; therefore, the maximum contact feedback force is introduced for constraint determination. A preset safety upper limit is established based on the minimum of the upper limit of contact feedback force specified by the robot's structural strength and the upper limit of contact feedback force specified by the actuator's rated load capacity. The maximum contact feedback force is compared with the safety upper limit. If the maximum contact feedback force is less than the safety upper limit, the obstacle is determined to be able to move continuously within the safe force range, and the obstacle category is identified as loosely movable. If the maximum contact feedback force is greater than or equal to the safety upper limit, the obstacle is determined to have reached the safe force limit while undergoing displacement, and the obstacle category is identified as fixed and stuck. This indicates that although the obstacle has undergone displacement, it has reached the safe force limit, and the corresponding displacement is not engineering feasible.

[0124] Among the immovable candidates, the maximum contact feedback force is compared with the safety upper limit to distinguish different physical reasons why the obstacle has not moved. If the maximum contact feedback force is greater than or equal to the safety upper limit, it is determined that the obstacle has not produced effective displacement after reaching the safety force limit, and the obstacle category is determined to be fixed and stuck. If the maximum contact feedback force is less than the safety upper limit, the determination result is in the immovable gray zone. The obstacle category is obtained through impedance characteristic analysis, indicating that the current thrust has not yet reached the safety limit, but the obstacle has not yet produced effective displacement. At this time, the specific physical reason cannot be determined by displacement and force magnitude alone.

[0125] Impedance characteristic analysis includes:

[0126] The reason for introducing impedance characteristic analysis is that fixed jamming and excessive weight immobility may both manifest as no effective displacement in the low displacement stage, but their essential characteristics in force-displacement relationship are different.

[0127] The equivalent thrust impedance index is calculated based on the thrust-displacement curve sampling sequence to quantify the mechanical response characteristics of an obstacle to controlled thrust during the low-displacement phase. In the thrust-displacement curve sampling sequence, sampling points whose cumulative displacement does not exceed a preset low-displacement threshold from the moment of initial contact between the obstacle and the actuator are selected as low-displacement intervals. Within these intervals, the force change is obtained by comparing the thrust value at the interval's termination point with that at the interval's beginning point, and the displacement change is obtained by comparing the displacement value at the interval's termination point with that at the interval's beginning point. The equivalent thrust impedance index is calculated as the ratio of the force change to the displacement change, and is used to characterize the obstacle's equivalent thrust impedance to controlled thrust during the low-displacement phase.

[0128] If the equivalent driving impedance is greater than the preset high impedance threshold, the obstacle is determined to have rigid, constrained characteristics, and the obstacle category is identified as fixed and stuck. If the equivalent driving impedance is less than the preset high impedance threshold, the obstacle is determined to have load-constrained characteristics, and the obstacle category is identified as too heavy to move. The high impedance threshold setting logic includes: during the robot's active physical exploration test phase, for typical fixed and stuck scenarios and typical too heavy to move scenarios, thrust displacement tests are conducted, and thrust displacement curve sampling sequences are collected and equivalent driving impedance index sequences are calculated. The lower bound distribution of the equivalent driving impedance index under fixed and stuck conditions and the upper bound distribution of the equivalent driving impedance index under too heavy to move conditions are statistically analyzed. At the same time, the noise fluctuation range of the equivalent driving impedance index caused by clumping jitter and force sensing resolution in the low displacement range is statistically analyzed. The high impedance threshold is set as the margin corresponding to the lower bound of the equivalent driving impedance index minus the upper bound of the noise fluctuation of the equivalent driving impedance index. This ensures that when the equivalent driving impedance index is greater than the high impedance threshold, even under the most unfavorable measurement error and initial contact instability conditions, the obstacle can still be stably determined to have rigid and restricted characteristics, thus avoiding misjudging the fixed jamming as being too heavy to move.

[0129] Because ambient temperature can cause phenomena such as adhesion and freezing, thus altering the actual physical mobility of obstacles under current environmental and stress conditions, temperature distribution information from the obstacle fusion feature set is introduced to perform consistency verification on obstacle categories, obtaining the obstacle's final state label. When the low-temperature adhesion risk marker is established and the obstacle category is loosely movable, the obstacle's final state label is set to fixed and stuck. When the low-temperature adhesion risk marker is established and the obstacle category is either too heavy to move or fixed and stuck, the obstacle's final state label is set to the corresponding obstacle category. The obstacle's final state label is used to characterize the physical mobility state of the obstacle under current environmental and stress conditions and serves as the basis for subsequent path planning and execution strategy generation.

[0130] The cumulative obstacle displacement and displacement threshold are used to determine the existence of effective displacement response, while the maximum contact feedback force and preset safety upper limit are used to determine whether the safe force limit has been reached. These two constitute the main criterion link for the final state label of the obstacle. In cases where the cumulative obstacle displacement is less than or equal to the displacement threshold and the maximum contact feedback force is less than the safety upper limit, there exists a state space where displacement is insufficient but the safe force limit has not been reached. The impedance characteristic subdivision process calculates the equivalent driving impedance index on the thrust-displacement curve. The equivalent driving impedance index is used to distinguish between high-impedance jamming and not yet entering the effective driving range. Temperature distribution information consistency verification negates the loose and movable judgment and adjusts it to fixed jamming when the risk of low-temperature adhesion is established. Temperature distribution information consistency verification introduces evidence of low-temperature physical adhesion into the final decision link, reducing the risk of mis-passage under low-temperature environmental conditions.

[0131] Based on the final state labels of obstacles and pre-loaded task constraints, real-time path planning is performed through a pre-trained hybrid planning model, outputting local control trajectories and execution strategy instructions, and controlling the track-following robot to perform obstacle avoidance, obstacle clearing, or obstacle detour.

[0132] Task constraints are generated and stored by the robot's built-in system during the offline configuration phase. These constraints include the robot's start point, end point, time constraints, and speed limit constraints. The start point is determined by the track node identifier obtained by mapping the robot's current localization result to the global track topology. The end point is determined by the target track node identifier or inspection node sequence recorded in the robot's built-in inspection task table. The time constraint is determined by the robot's built-in inspection cycle requirements and task completion time limits. The speed limit constraint is jointly determined by the speed limit field in the global track topology and the maximum permissible speed in the robot's kinematic constraints. The speed limit constraint takes the smaller value between the speed limit field and the maximum permissible speed. When the deceleration conditions are met at the junction marker, maintenance section marker, or curvature field, the speed limit constraint is reduced according to the robot's built-in deceleration rules. The deceleration rules are determined by offline tracking stability tests and are then stored.

[0133] The hybrid planning model employs a combined structure of strategy generation and classical planning. Strategy generation outputs the set of planning strategy parameters for path planning. Classical planning generates paths that satisfy drivability constraints on the global track topology; this is implemented by an alternative path search module and a cost evaluation module. A smoothing algorithm transforms the path results into local control trajectories that conform to robot kinematic constraints. Instruction generation forms directly executable execution strategy instructions. The global track topology is built offline and stored in the robot's built-in system. It is generated based on track design drawings, measured track geometry data, or historical inspection point cloud data. The global track topology uses a graph structure to represent track nodes and track connections; track nodes represent key track positions, and track connections represent drivable track segments between nodes. Obstacle final state labels are used to dynamically update drivability markers for track connections.

[0134] The hybrid planning model comprises a track topology mapping module, a strategy generation module, an alternative path search module, a cost evaluation module, a smoothing algorithm module, and an instruction generation module. The track topology mapping module updates the passable and impassable markings of track connections in the global track topology based on the final state labels of obstacles. When an obstacle's final state label is fixed, stuck, or too heavy to move, the track topology mapping module marks the track connection corresponding to the obstacle's location as impassable. When an obstacle's final state label is loose and movable, the track topology mapping module maintains the track connection corresponding to the obstacle's location as passable.

[0135] The policy generation module is implemented using a deep neural network with a Transformer decoder structure. The Transformer decoder structure consists of multiple attention layers and feedforward layers. It performs joint encoding on the global track topology, obstacle final state labels, and task constraints to generate a set of planning policy parameters. The planning policy parameter set includes the upper limit of the search depth for the alternative path search module, the upper limit of the number of candidate paths for the alternative path search module, and the weight vector for the cost evaluation module. The upper limit of the search depth restricts the expansion range of the alternative path search module within the global track topology, the upper limit of the number of candidate paths restricts the size of the candidate paths retained by the alternative path search module, and the weight vector weights the cost terms of different paths to form a unified cost score.

[0136] After the planning strategy parameter set is output online, range pruning is performed. The upper and lower bounds of the range pruning are determined by offline parameter scanning and stored in the robot's built-in system. Offline parameter scanning performs discrete combination tests on the upper limit of search depth and the upper limit of the number of candidate paths. The offline parameter scanning selects the parameter interval where the detour success rate reaches the detour success rate threshold and the planning time meets the real-time constraints as the upper and lower bounds of the range pruning. The detour success rate threshold is determined based on the time constraint and task completion rate requirement in the task constraints, and the detour success rate threshold is taken as the minimum completion rate allowed by the task system.

[0137] Real-time constraints are jointly determined by the online control cycle and the processor's computing power limit. The real-time constraint is the smaller of the fixed proportion of the online control cycle and the maximum allowable planning time corresponding to the processor's computing power limit. The fixed proportion of the online control cycle is determined by the control closed-loop time budget, which is jointly determined by the time occupied by the robot control thread's periodic tasks and the allowed time occupied by the planning thread. The fixed proportion of the online control cycle is the ratio of the allowed time occupied by the planning thread to the online control cycle. The maximum allowable planning time corresponding to the processor's computing power limit is determined by the upper bound of the single planning run time under typical topology scale and typical task constraints. The upper bound of the single planning run time is the maximum time observed in offline stress testing plus the minimum time margin corresponding to the system's timing resolution.

[0138] The set of planning strategy parameters after range trimming is used by the alternative path search module to perform alternative path searches. If the alternative path search module fails to find a feasible path, a backtracking is triggered. The backtracking strategy uses conservative default parameters and searches again. The conservative default parameters are selected based on a combination of parameters with a small search depth and a small number of candidate paths within the range trimming interval. The selection of conservative default parameters is based on reducing the topology expansion scale and the candidate path scale to ensure that the planning time meets the real-time constraint. The criteria for determining if the alternative path search module has failed to find a feasible path include two cases: the candidate expansion traversal has ended and the candidate path set is empty; the planning time reaches the real-time constraint and the candidate path set is empty.

[0139] The Transformer decoder structure consists of multiple stacked units, each including at least self-attention, feedforward, residual connections, and layer normalization operations. The self-attention operation models the correlation between the track topology encoding sequence and the task constraint encoding sequence, while the feedforward operation performs a nonlinear transformation on the attention output. The track topology encoding method employs a serialization encoding rule: track nodes are expanded into a sequence according to mileage; track node attributes and track connection relationship attributes are concatenated into a node feature vector input to the attention layer; impassable markers are encoded as additional features of track nodes or track connection relationships, making the mapping rule from graph structure to sequence input clear.

[0140] The alternative path search module takes as input the global track topology, impassable markers, task constraints, and a set of planning strategy parameters. It performs an alternative path search within the global track topology, using impassable markers as hard constraints to eliminate impassable track connections. The planning strategy parameter set constrains the upper limit of the search depth and the upper limit of the number of candidate paths. The module outputs a set of candidate paths, which consists of track connection sequences that satisfy the impassable marker constraints.

[0141] The cost evaluation module takes as input a set of candidate paths, task constraints, and a weight vector. It calculates a cost score for each candidate path in the set. The cost score is obtained by weighted summing of the path length, number of times a junction is traversed, number of times a maintenance section is traversed, and number of times a backup route is used. The cost evaluation module selects the candidate path with the lowest cost score as the optimal route and outputs the optimal route and its cost score. Before weighted summation, the cost evaluation module performs normalization on each cost item. Path length normalization is achieved by dividing the path length by the maximum path length in the candidate path set. Count normalization is achieved by dividing the count by its upper bound. The upper bound of the count is determined by statistics from the offline task set, taking the maximum value of the corresponding count item in the offline task set and adding the minimum counting margin corresponding to the counting resolution. The counting resolution is determined by a counting rule: incrementing the count by one for each junction traversed, each maintenance section traversed, and each backup route used. The counting resolution is set to 1. The minimum count margin is set to 1 to cover single extra traversals that may occur online but are not covered by the offline task set, making the normalization upper bound clear and implementable.

[0142] The smoothing algorithm module takes the optimal route and robot kinematic constraints as input. It transforms the discrete trajectory sequence of the optimal route into a continuous, trackable local control trajectory that satisfies the curvature and velocity variation constraints within the robot's kinematic constraints. The smoothing algorithm module outputs the local control trajectory.

[0143] The instruction generation module takes into account the obstacle's final state label, local control trajectory, task constraints, and cost score. Based on the obstacle's final state label, the module generates execution strategy instructions. When the obstacle's final state label indicates it is loosely movable, the execution strategy instructions include obstacle-clearing trajectory execution constraints, thrust control constraints, and safety force constraints. When the obstacle's final state label indicates it is fixed, stuck, or too heavy to move, the execution strategy instructions include detour route execution constraints, impassable warning information, and deceleration constraints. The execution strategy instructions and local control trajectory together serve as the control input for the robot's real-time execution. The deceleration constraints include a deceleration threshold and a minimum safe speed. The deceleration threshold and minimum safe speed are determined based on the track curvature and positioning error: multiple sets of candidate speed tracking experiments are performed in typical curvature and straight sections, recording the impact of control tracking error and positioning error on the control tracking error. The maximum candidate speed that meets the control accuracy requirements under both control tracking error and positioning error disturbances is selected as the speed corresponding to the deceleration threshold, and the minimum candidate speed that meets the control accuracy requirements is selected as the minimum safe speed. The control accuracy requirements are defined by the allowable deviation of the robot control system and stored in the robot's built-in system. The allowable deviation of the robot control system includes at least the upper limit of the lateral deviation of the trajectory and the upper limit of the speed deviation. The upper limit of the lateral deviation of the trajectory and the upper limit of the speed deviation are determined based on the robot's mission safety requirements and the track safety boundary. The upper limit of the lateral deviation of the trajectory is taken as the maximum lateral deviation allowed by the track's drivable boundary margin, and the upper limit of the speed deviation is taken as the maximum speed deviation allowed under the constraints of braking distance and obstacle safety distance.

[0144] Training of the Transformer decoder architecture is completed offline. The training dataset consists of global track topology, obstacle final state labels, task constraints, and planning policy parameter set labels. The planning policy parameter set labels are obtained offline by performing parameter scanning on the upper limit of search depth, the upper limit of the number of candidate paths, and the weight vector of the cost evaluation module. Under multiple sets of candidate parameters, the offline solution calls the alternative path search module and the cost evaluation module to generate a set of candidate paths and calculate the cost score. The offline solution selects the set of candidate parameters that satisfies the task constraints and has the smallest cost score as the planning policy parameter set labels. The training objective is to make the planning policy parameter set output by the Transformer decoder architecture, under the input global track topology, obstacle final state labels, and task constraints, approximate the planning policy parameter set labels.

[0145] The training dataset was divided into training, validation, and test subsets, with each subset comprising 0.8, 0.1, and 0.1% of the dataset. Batch gradient descent was used to update the Transformer decoder structure parameters, and the AdamW optimizer was employed. The learning rate was set to 0.00001, the batch size to 16, and the number of training epochs to 3. A weight decay coefficient of 0.01 was used to suppress overfitting. A gradient clipping threshold of 1.0 was set to limit gradient explosion and ensure training stability.

[0146] The loss function is a weighted combination of cross-entropy loss and numerical regression error. The cross-entropy loss function is used to constrain the prediction accuracy of the upper limit of search depth and the upper limit of the number of candidate paths, while the numerical regression error is used to constrain the numerical consistency of the weight vector. The weight coefficients of the cross-entropy loss function and the numerical regression error are set to 0.5 and 0.5, respectively, to balance the learning intensity of discrete policy parameters and continuous weight vectors.

[0147] Early stopping is employed during training to avoid overfitting: the cost score stability metric of the validation subset terminates training and retains the optimal parameters when there is no improvement after two consecutive evaluation cycles. The validation subset is used to evaluate the predictive consistency of the planning strategy parameter set and the cost score stability of the optimal route output by the alternative path search module. The test subset is used to evaluate the detour success rate under impassable marker constraints and the proportion of local control trajectories that satisfy the robot's kinematic constraints.

[0148] The input to the Transformer decoder structure uses consistent structured field encoding. This structured field encoding includes node attributes and connection attributes, which at least include mileage location, speed limit, gradient, curvature, maintenance section markers, and branch line markers. The upper limit of the track node sequence length is determined by the maximum number of nodes in the global track topology. When the track node sequence length is less than the upper limit, zero-padding is used for alignment. When the track node sequence length exceeds the upper limit, a subsequence containing nodes relevant to the current task is truncated by a mileage window, ensuring that the input dimension of the Transformer decoder structure is fixed and online inference is feasible.

[0149] During the online operation phase, the Transformer decoder structure performs forward inference and outputs a set of planning strategy parameters. This replaces the path search module, cost evaluation module, smoothing algorithm module, and instruction generation module in executing a real-time path planning process according to the planning strategy parameter set, outputting the local control trajectory and execution strategy instructions. The final obstacle state label is used to update the passable and impassable markers of track connections in the global track topology. Updating the passable and impassable markers incorporates the physical mobility state of obstacles into the global path constraints. Fixed jams and excessive weight triggering impassable markers reduce invalid search expansions of impassable track connections and improve the detour success rate. The strategy generation module outputs the planning strategy parameter set and performs range pruning. The upper and lower bounds of the range pruning are determined by offline parameter scanning and stored in the robot's built-in system. Range pruning limits the upper limit of the search depth and the upper limit of the number of candidate paths within the real-time constraint allowable range and controls the size of the candidate path set. The real-time constraints meet the requirements of the online control cycle and the upper limit of processor computing power. When the alternative path search module fails to find a feasible path, it triggers a backtrack and searches again. The backtrack strategy reduces the topology expansion scale and the candidate path scale to ensure that the planning time meets the real-time constraints. The improved stability of the planning time improves the stability of the local control trajectory and the execution strategy command output.

[0150] This embodiment constructs an obstacle fusion feature set by fusing laser point clouds, visible light images, infrared thermal imaging, and robot motion states under a unified time and pose reference, achieving a comprehensive representation of the obstacle's geometric shape, appearance features, temperature characteristics, and spatial relationship with the track structure. Obstacle appearance category recognition is completed based on a deep learning model, and the recognition results are physically consistent and jointly corrected by point cloud geometric embedding analysis and low-temperature adhesion risk analysis to improve classification reliability. Furthermore, through active physical probing under controlled thrust, the physical mobility state of the obstacle is determined using contact force and displacement response, distinguishing between fixed jammed, excessively heavy and immobile, and loosely movable obstacles. Finally, the obstacle state is incorporated into the obstacle avoidance control decision-making and execution strategy generation process, achieving safe and reliable autonomous inspection and obstacle avoidance control.

[0151] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0152] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic obstacle avoidance control method for a track-following robot, characterized in that, include: Based on real-time acquired laser point cloud data, visual image data, thermal distribution data, and robot motion state, an obstacle fusion feature set is calculated. The obstacle fusion feature set is input into the pre-trained target recognition model to obtain the visual category probability vector. The visual category probability vector is jointly corrected based on laser point cloud data and thermal distribution data to obtain the preliminary obstacle classification label and confidence score. Based on the obstacle fusion feature set and the preliminary obstacle classification label, the robot is controlled to actively physically explore the obstacles and obtain the physical exploration feature set. Methods for controlling robots to actively physically explore obstacles include: Based on the initial obstacle classification labels and confidence scores, thrust strategy constraints are generated to determine the initial thrust level and thrust increment rate. Based on the robot's motion state, control the robot to move toward the obstacle at a preset limited speed; When the robot arm comes into contact with an obstacle, the robot arm is controlled to gradually increase the thrust on the obstacle according to the initial thrust level and the thrust increment rate, and the contact feedback force and motor load current are collected. When the contact feedback force reaches the preset safety limit or the motor load current reaches the preset current limit, the active physical probe is terminated. Based on the obstacle fusion feature set and the physical exploration feature set, the obstacles are classified into physical states to obtain the final state labels of the obstacles; Based on the final state labels of obstacles and pre-loaded task constraints, real-time path planning is performed through a pre-trained hybrid planning model, outputting local control trajectories and execution strategy instructions, and controlling the track-following robot to perform obstacle avoidance, obstacle clearing, or obstacle detour.

2. The dynamic obstacle avoidance control method for a track-following robot according to claim 1, characterized in that, Methods for obtaining preliminary obstacle classification labels and confidence scores include: The dimension of the visual category probability vector is equal to the number of preset visual categories, and each component of the visual category probability vector corresponds to the probability value of a different visual category. The visual category corresponding to the component with the highest probability value in the visual category probability vector is taken as the visual determination result. Based on laser point cloud data, the entry depth is extracted through a preset geometric analysis method. When the entry depth is greater than or equal to a preset depth threshold, an embedded risk marker is generated. Based on thermal distribution data, temperature difference features are extracted using a preset temperature difference analysis method. When the temperature difference features are lower than a preset temperature difference threshold, a low-temperature adhesion risk marker is generated. According to the preset physical consistency verification rules, if there is a logical conflict between the visual judgment result and the embedded risk mark or the low temperature adhesion risk mark, the preset probability reweighting process is performed on the component in the visual category probability vector corresponding to the visual judgment result. The visual category probability vector after probability reweighting is normalized, and the visual category with the highest probability value in the normalized visual category probability vector is used as the initial obstacle classification label, and the corresponding probability value is used as the confidence score.

3. The dynamic obstacle avoidance control method for a track-following robot according to claim 1, characterized in that, The thrust strategy constraints include: The physical attribute type of the preliminary obstacle classification label is determined by a preset method to obtain the corresponding physical attribute type; the physical attribute type includes the lightweight and loose category and the heavy and dense structure category. If the physical attribute type of the obstacle's initial classification label corresponds to the lightweight and loose category, a preset first initial thrust level is selected as the initial thrust level, and a preset first thrust increment rate is selected as the benchmark thrust increment rate. If the physical attribute type of the obstacle's initial classification label corresponds to the dense heavy structure category, a preset second initial thrust level is selected as the initial thrust level, and a preset second thrust increment rate is selected as the benchmark thrust increment rate.

4. The dynamic obstacle avoidance control method for a track-following robot according to claim 3, characterized in that, Thrust strategy constraints also include reducing the baseline thrust increment rate based on confidence scores; Based on the confidence score, a lookup operation is performed on the preset reduction coefficient mapping table to obtain the corresponding thrust increment rate reduction coefficient. The thrust increment rate reduction coefficient is then multiplied by the benchmark thrust increment rate to obtain the thrust increment rate.

5. The dynamic obstacle avoidance control method for a track-following robot according to claim 1, characterized in that, Methods for calculating obstacle fusion feature sets include: Spatial constraint filtering of laser point cloud data is performed using a pre-loaded trajectory map from the robot's built-in system; The filtered laser point cloud data is segmented to obtain object clumps, which are sets of three-dimensional spatial coordinate points. Based on object clumps and laser point cloud data, extract the geometric contour information and volume information of the object clumps; Based on object clumps and visual image data, extract the color and texture information of the object clumps; Based on object clumps and thermal distribution data, extract temperature distribution information of the object clumps; Using object clumps as basic units, geometric contour information, volume information, color information, texture information, and temperature distribution information are combined to form an obstacle fusion feature set.

6. The dynamic obstacle avoidance control method for a track-following robot according to claim 5, characterized in that, Methods for obtaining physical exploration feature sets include: The centroid coordinates of the point set are obtained by taking the arithmetic mean of the three-dimensional spatial coordinate points of the same object at different sampling times. The displacement of the obstacle is obtained by calculating the Euclidean distance between the centroid coordinates of the point set at different times. The displacement at each moment and the contact feedback force are combined into ordered point pairs to form a thrust displacement curve sampling sequence. The obstacle displacement is accumulated over time to obtain the cumulative obstacle displacement. The maximum contact feedback force, maximum motor load current, cumulative displacement, and thrust displacement curve sampling sequence during the active physical probing process are encapsulated to obtain a physical probing feature set.

7. The dynamic obstacle avoidance control method for a track-following robot according to claim 1, characterized in that, Methods for classifying the physical state of obstacles include: Obstacles are categorized as fixed and stuck, loosely movable, and too heavy to move. The cumulative obstacle displacement is compared with a preset displacement threshold. If the cumulative obstacle displacement is greater than the displacement threshold, the obstacle is determined to be a movable candidate; otherwise, the obstacle is determined to be an immovable candidate. Among the movable candidates, the maximum contact feedback force is compared with the preset safety limit. If the maximum contact feedback force is less than the safety limit, the obstacle is classified as loosely movable; otherwise, it is classified as fixed and stuck. Among the immovable candidates, the maximum contact feedback force is compared with the safety limit. If the maximum contact feedback force is greater than or equal to the safety limit, the obstacle is classified as fixed and stuck. Otherwise, impedance characteristics are introduced for further subdivision to determine the obstacle category.

8. The dynamic obstacle avoidance control method for a track-following robot according to claim 7, characterized in that, Impedance characteristic subdivision determination includes: Based on the thrust displacement curve sampling sequence, within the range where the cumulative displacement does not exceed the preset low displacement threshold, the thrust change and displacement change are obtained, the ratio of thrust change to displacement change is calculated, and the equivalent thrust impedance index is obtained. The equivalent driving impedance index is compared with the preset high impedance threshold. If the equivalent driving impedance is greater than the high impedance threshold, the obstacle is determined to be fixed and stuck; otherwise, the obstacle is determined to be too heavy to move.

9. The dynamic obstacle avoidance control method for a track-following robot according to claim 8, characterized in that, Methods for obtaining the final state label of obstacles include: Based on temperature distribution information, the obstacle categories obtained by physical state classification are checked for consistency. When the low temperature adhesion risk mark is established and the obstacle category is loose and movable, the final state label of the obstacle is set to fixed and stuck. When the low temperature adhesion risk mark is established and the obstacle category is too heavy and immovable or fixed and stuck, the final state label of the obstacle is set to the corresponding obstacle category.

Images