Environment perception and autonomous decision system for embodied intelligent robot for logistics sorting

Abstract

This invention relates to the field of intelligent robot control technology, specifically to an embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting. It includes a multimodal perception module for establishing a coordinate system and mapping spatial coordinates from various coordinate systems to a global coordinate system of the sorting area; a target processing and feature optimization module for using a joint clustering algorithm including a similarity threshold to cluster multiple spatial coordinates in the global coordinate system of the sorting area, obtaining a geometric feature set; further splitting the geometric feature set into multiple planar coordinate subsets, and then using a linkage verification method to analyze the contour shape curves corresponding to the planar coordinate subsets, obtaining a target space set including virtual coordinates; reusing the joint clustering algorithm to analyze the virtual feature vectors corresponding to the virtual coordinates in the target space set; and an autonomous decision-making and motion control module for analyzing the coordinates of the intelligent robot's grasping points based on the spatial coordinates in the target space set and generating robot sorting control commands.

Description

Technical Field

[0001] This invention relates to the field of intelligent robot control technology, and more specifically, to an embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting. Background Technology

[0002] In modern intelligent logistics sorting operations, to meet the practical needs of high-speed sorting, multi-target mixing, and accurate identification and stable grasping of irregular packages, embodied intelligent robots equipped with 3D vision perception modules are typically used to complete core tasks such as target localization, spatial modeling, package segmentation, and autonomous grasping decisions. The overall working principle is as follows: Point cloud data and spatial coordinate information of the logistics sorting scene are collected in real time using 3D perception devices such as depth cameras and LiDAR to establish a 3D spatial model of the scene; the target package is segmented, extracted, and its features are analyzed; and then, based on spatial location, geometric shape, and physical attribute information, grasping point planning is completed. With autonomous decision-making and control, relying on the closed-loop interaction between environmental perception and decision-making commands, it fully meets the actual needs of rapid target identification, accurate positioning, and stable grasping in complex sorting scenarios. At the same time, the visual perception system can independently process the spatial coordinates and feature information of each frame in real time. Based on preset similarity thresholds and contour fitting rules, it completes point cloud data analysis and target attribution determination, accurately distinguishes target packages from background interference, and then carries out targeted grasping path planning, grasping posture adjustment, and closed-loop control of sorting actions. This effectively avoids the decline in sorting efficiency and potential equipment operation hazards caused by target missorting, omissions, unstable grasping, and package slippage.

[0003] However, due to objective differences in target packages in logistics sorting scenarios, such as varying sizes, irregular shapes, dense stacking, mutual occlusion, and diverse surface materials, even among targets to be sorted, their spatial point cloud distribution, geometric contour features, and surface physical properties will exhibit individual differences. It is impossible to completely maintain the uniformity of spatial distribution and the consistency of features. Therefore, in the process of accurately segmenting target packages and autonomously planning stable grasping points, conventional perception and decision-making logic often suffers from the deficiency of considering only one dimension. At the same time, since the same frame of point cloud data contains a large number of scattered, unevenly dense, partially missing, and noisy original spatial coordinates, and the virtual reconstructed coordinates lack directly corresponding physical feature information, relying solely on geometric position cannot achieve reliable grasping judgment. Existing technologies usually use independently set similarity thresholds and clustering parameters to perform target segmentation and feature segmentation respectively. Analysis reveals that the lack of a unified similarity judgment standard ignores contour fitting deviation, hierarchical regularization error, and dynamic threshold change patterns. It also lacks coordinated verification and integrated processing of spatial coordinates, geometric features, and physical attributes. This makes it impossible to simultaneously consider target regularization, accurate feature mapping, and iterative optimization convergence, leading to problems such as misclassification and omission of target packages, distorted contour fitting, unreliable virtual feature construction, and inconsistent similarity judgment. Consequently, it causes a decrease in environmental perception accuracy, target positioning deviation, and unreasonable grasping decisions. This not only makes it difficult to achieve stable and efficient sorting of complex packages but also increases the probability of robot grasping failures, repetitive actions, and collision interference, reducing the overall operating efficiency of the sorting line. Ultimately, this results in decreased stability of logistics sorting operations, reduced sorting accuracy, and increased equipment failure rates, failing to meet the high-speed, stable, and precise automated operation requirements of modern smart logistics. In view of this, we propose an embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problems of existing intelligent robots for logistics sorting, such as inconsistent spatial benchmarks of multi-source perception data, mixed interference of target point clouds, low clustering and segmentation accuracy, incomplete contour reconstruction, and poor reliability of grasping decisions.

[0005] To achieve the above objectives, the present invention provides an embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting, comprising: The multimodal perception module is used to establish coordinate systems, including the local coordinate system corresponding to the multi-source sensors, the robot base coordinate system corresponding to the intelligent robot, and the global coordinate system of the sorting area; and to uniformly map the spatial coordinates under each coordinate system to the global coordinate system of the sorting area. The target processing and feature optimization module employs a joint clustering algorithm, including a similarity threshold, to cluster multiple spatial coordinates in the global coordinate system of the sorting area, resulting in a geometric feature set. This geometric feature set is then further divided into multiple planar coordinate subsets. A linked verification method is then used to analyze the contour shape curves corresponding to these planar coordinate subsets, yielding a target space set including virtual coordinates. Finally, the joint clustering algorithm is reused to analyze the virtual feature vectors corresponding to the virtual coordinates in the target space set. The linkage verification method first constructs a contour shape model including calibration polynomial coefficients, and then conducts contour linkage verification based on cross-plane shape matching relationship. Under the premise of satisfying cross-plane shape adaptation and structure matching constraints, the calibration polynomial coefficients in the contour shape model are adjusted to modified polynomial coefficients for analysis; and the similarity threshold in the clustering process is adjusted according to the difference between the calibration polynomial coefficients and the modified polynomial coefficients. The autonomous decision-making and motion control module analyzes the coordinates of the intelligent robot's grasping point based on the spatial coordinates in the target space set and generates robot sorting control instructions.

[0006] As a further improvement to this technical solution, the multimodal sensing module includes a multi-coordinate system construction unit and a global coordinate transformation unit, wherein: The multi-modal fusion module's multi-coordinate system construction unit establishes a global coordinate system for the sorting field based on the plane of the conveyor belt at the sorting site, a robot base coordinate system corresponding to the intelligent robot, and local coordinates corresponding to the multi-source sensors. The global coordinate transformation unit establishes a coordinate transformation mapping relationship between the local coordinate system under the multi-source sensor and the global coordinate system of the sorting field, and maps and transforms each local coordinate system to the global coordinate system of the sorting field. Specifically, an independent local coordinate system is established with the detection center of each modal sensor. The rotation matrix and translation vector of each local coordinate system relative to the robot base coordinate system are solved through hand-eye calibration, and the local coordinates of each modality are transformed to the robot base coordinate system. The global rotation matrix and global translation vector of the robot base coordinate system relative to the global coordinate system of the sorting field are obtained again through hand-eye calibration. The coordinates of the base coordinate system are then further transformed to the global coordinate system of the sorting field.

[0007] As a further improvement to this technical solution, the target processing and feature optimization module includes a joint clustering segmentation unit, a contour fitting unit, and a physical feature completion unit, wherein: The joint clustering segmentation unit uses a joint clustering algorithm to cluster and group multiple similar spatial coordinates that are discretely distributed in the global coordinate system of the sorting field in the global coordinate transformation unit, forming several independent geometric feature sets, each geometric feature set corresponding to a target item to be sorted. The contour fitting unit fits and reconstructs multiple spatial coordinates in each geometric feature set to obtain the target space set of the corresponding items to be sorted. The physical feature completion unit defines the spatial coordinates in the target space set as virtual coordinates and measured coordinates; it matches the geometric feature set where the virtual coordinates are located with the physical feature vectors of the measured coordinates in the geometric feature set to construct the virtual feature vector corresponding to the virtual coordinates.

[0008] As a further improvement to this technical solution, the joint clustering algorithm in the joint clustering segmentation unit retrieves each spatial coordinate in the global coordinate system of the sorting field, as well as the multi-dimensional physical features obtained by scanning the spatial coordinates using different sensors. The spatial coordinates and multi-dimensional physical features are then fused in a fixed order to obtain the spatial feature vector of each spatial coordinate. Clustering of spatial coordinates based on the similarity between spatial feature vectors of each spatial coordinate yields multiple geometric feature sets composed of multiple spatial coordinates, specifically: Let the sorting yard be in the global coordinate system, the first... The spatial feature vectors corresponding to the spatial coordinates are ,in For spatial coordinates, Spatial coordinates The physical feature vector, The total number of dimensions of the physical features obtained from all sensor scans; Classification space feature vector To find a cluster, calculate the mean of the same feature dimension among all spatial feature vectors under the cluster, and use this mean as the cluster feature vector corresponding to the cluster. ; Select the spatial feature vectors corresponding to the spatial coordinates of the adjacent spatial nodes outside the cluster. Calculate cluster feature vectors Spatial feature vectors Overall similarity between ; Set similarity threshold Compare the overall similarity Does it meet the similarity threshold? The matching spatial coordinates are classified into clusters. Iteratively select spatial coordinates adjacent to the cluster, recalculate the comprehensive similarity, and determine whether clustering is possible until no new spatial coordinates meet the admission criteria and are added to the current cluster. Finally, a cluster containing multiple spatial coordinates is obtained, which is the geometric feature set. Then, select another spatial coordinate that is not in the geometric feature set for iterative calculation until all spatial coordinates in the global coordinate system of the sorting field are clustered, and finally obtain multiple geometric feature sets in the global coordinate system of the sorting field.

[0009] As a further improvement to this technical solution, the contour fitting unit first divides multiple spatial coordinates in the geometric feature set into a set of planar coordinates according to the planar layering method: setting a fixed coordinate axis, traversing all spatial coordinates in the geometric feature set, setting selection rules to filter the spatial coordinates on the fixed coordinate axis and group them into the same layered plane, and decomposing the spatial coordinates in the geometric feature set into several mutually parallel and non-overlapping planar coordinate subsets. The linkage verification method is used to analyze the contour shape curve corresponding to the planar coordinate subset. That is, the physical acquisition accuracy of the sensor is defined as the minimum spacing threshold, and the planar coordinate subset is fitted and reconstructed based on the minimum spacing threshold and the contour shape curve. The spatial coordinates within the fitted and reconstructed planar coordinate subset are minimized to obtain a target spatial set including multiple spatial coordinates.

[0010] As a further improvement to this technical solution, the linkage verification method in the contour fitting unit is as follows: Set calibration polynomial coefficients, construct a contour shape model, input the abscissa of each contour coordinate, the contour shape model combines the abscissa and calibration polynomial coefficients to output the corresponding fitting ordinate, and combine the fitting ordinate and the abscissa of the contour coordinate to obtain the fitting coordinate. In the process of constructing the contour shape curve, contour linkage verification is carried out based on cross-plane shape matching relationship. Under the premise of satisfying cross-plane shape adaptation and structural matching constraints, the calibration polynomial coefficients are finely adjusted to the modified polynomial coefficients, and then the fitted coordinates are updated according to the modified polynomial coefficients. Finally, multiple fitted coordinates are connected in sequence to obtain the contour shape curve.

[0011] As a further improvement to this technical solution, when setting the similarity threshold for the joint clustering segmentation unit, the optimal setting is achieved by minimizing the difference between the calibration polynomial coefficients and the modified polynomial coefficients in the contour fitting unit.

[0012] As a further improvement to this technical solution, after determining the contour shape curve corresponding to the subset of planar coordinates, the contour fitting unit specifically corresponds to the target space set as follows: The minimum spacing threshold is set based on the physical acquisition accuracy of the sensor. The entire contour curve is traversed segment by segment from the starting position. The minimum spacing threshold is strictly used as the fixed sampling interval. The entire contour curve is traversed uniformly and reference feature points are selected. It is ensured that the spatial distance between adjacent reference feature points is not less than the set minimum spacing threshold. At the same time, spatial coordinates in the planar coordinate subset that are not equal to the minimum spacing threshold are removed. The uniform reference feature points obtained by the selection are used as the contour shape skeleton and conform to the shape change law of the original contour curve. On this basis, the planar coordinate subset is regularized and reshaped by interpolation fitting and local shape correction. Multiple spatial coordinates are reconstructed under the premise of meeting the minimum spacing threshold. The point selection and fitting reconstruction of the planar coordinate subset based on the minimum spacing threshold and the contour curve are completed. Finally, the reconstructed spatial coordinates and the unreconstructed spatial coordinates form the target spatial set.

[0013] As a further improvement to this technical solution, the physical feature completion unit retrieves the physical feature vector of the corresponding geometric feature set in the joint clustering segmentation unit based on the measured coordinates, and then uses the joint clustering algorithm in the joint clustering segmentation unit to cluster and analyze the physical feature vector again. Based on the clustering convergence result, the overall geometric feature set is adaptively divided into several sub-geometric feature sets with the same geometric attributes and spatial distribution patterns. Retrieve the measured coordinates corresponding to each sub-geometric feature set and determine the corresponding spatial range. Specifically, construct an axis-aligned three-dimensional minimum bounding box with the extreme values ​​of each axis as the boundary, and directly define the spatial region of the bounding box as the exclusive spatial range corresponding to the sub-geometric feature set. Match the exclusive spatial range where the virtual coordinates are located, and construct the virtual feature vector corresponding to the virtual coordinates using the physical feature vectors of multiple measured coordinates within the exclusive spatial range.

[0014] As a further improvement to this technical solution, the autonomous decision-making and motion control module retrieves multiple spatial coordinates within the target space set in the contour fitting unit, as well as the physical feature vectors corresponding to the spatial coordinates, and analyzes the coordinates of the grasping point of the intelligent robot. Based on spatial coordinates, construct the three-dimensional axis-aligned minimum bounding box of the items to be sorted, calculate the center of gravity of the bounding box as the initial gripping reference point, traverse all spatial coordinates and extract spatial geometric features, then set filtering conditions to select the geometrically optimal point, and determine the spatial geometric features of gripping stability and posture adaptability. Set up multi-level filtering conditions in sequence to select stable grasping areas, match grasping postures, and ensure interference-free grasping; Based on multi-level screening conditions and spatial geometric features, each spatial coordinate is matched and verified one by one, and only spatial coordinates that simultaneously meet all screening conditions are retained and determined as the geometrically optimal points for the robot's stable grasping operation. In addition, an adaptive judgment is carried out by combining the physical feature vectors corresponding to each spatial coordinate in the geometrically optimized points. Among the remaining geometrically optimized points, the spatial coordinates that are closest to the centroid of the bounding box, have the best physical feature uniformity, the most concentrated spatial distribution, and are located within the rigid support area of ​​the target object are selected as the optimal gripping point coordinates.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: In this embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting, the joint clustering segmentation unit uses a joint clustering algorithm to efficiently collect similar spatial coordinates discretely distributed in the global coordinate system of the sorting field, accurately forming an independent geometric feature set corresponding to each item to be sorted, effectively solving the problems of mixed and misattributed multi-target point clouds; the contour fitting unit divides the geometric feature set into a subset of planar coordinates and accurately fits the contour shape curve by combining a linkage verification method, while simultaneously adaptively optimizing the similarity threshold of joint clustering, greatly improving the adaptability of clustering segmentation and the accuracy of contour reconstruction, avoiding over-segmentation or erroneous merging of targets; the contour fitting unit merges the subset of planar coordinates and fits and reconstructs spatial coordinates to generate a complete target space set containing virtual coordinates, effectively compensating for the defects of missing, sparse, redundant and outlier points in the original point cloud, restoring the true and complete shape of the items to be sorted. The physical feature completion unit reuses the joint clustering algorithm and similarity threshold of the joint clustering segmentation unit to generate virtual feature vectors for virtual coordinate matching, ensuring that the feature judgment standard is consistent and the calculation logic is coherent throughout the process, greatly reducing parameter calibration overhead and improving system operating efficiency. Finally, the autonomous decision-making and motion control module optimizes the gripping point coordinates and generates sorting control instructions based on accurate and reliable geometric and physical feature analysis, improving the robot's gripping stability and sorting accuracy, and enhancing the system's robustness and autonomous operation capability in complex logistics sorting scenarios.

[0016] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall module of the present invention.

[0018] The meanings of the labels in the diagram are as follows: 100. Multimodal perception module; 110. Multi-coordinate system construction unit; 120. Global coordinate transformation unit; 200. Target processing and feature optimization module; 210. Joint clustering and segmentation unit; 220. Contour fitting unit; 230. Physical feature completion unit; 300. Autonomous decision-making and motion control module. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] refer to Figure 1 As shown, the embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting includes: The multimodal sensing module is used to establish a multi-source coordinate system, including the global coordinate system of the sorting yard; and to uniformly map the multimodal sensing data to the global coordinate system of the sorting yard. The target processing and feature optimization module employs a joint clustering algorithm, including a similarity threshold, to cluster multiple spatial coordinates in the global coordinate system of the sorting area. After clustering, a geometric feature set is obtained, which is then further split into multiple planar coordinate subsets. Subsequently, a linked verification method is used to analyze the contour shape curves corresponding to the planar coordinate subsets, resulting in a target space set including virtual coordinates. Finally, the joint clustering algorithm is reused to analyze the virtual feature vectors corresponding to the virtual coordinates in the target space set. The linkage verification method first constructs a contour shape model including calibration polynomial coefficients, and then conducts contour linkage verification based on cross-plane shape matching relationship. Under the premise of satisfying cross-plane shape adaptation and structure matching constraints, the calibration polynomial coefficients in the contour shape model are adjusted to modified polynomial coefficients for analysis; and the similarity threshold in the clustering process is adjusted according to the difference between the calibration polynomial coefficients and the modified polynomial coefficients. The autonomous decision-making and motion control module analyzes the coordinates of the intelligent robot's grasping point based on the spatial coordinates in the target space set and generates robot sorting control instructions. In the implementation of the above embodiments, the multi-modal perception module maps the multi-source heterogeneous coordinate system to the same global coordinate system of the sorting field, establishing a globally consistent spatial reference. This effectively eliminates perception errors caused by differences in the perspectives of multiple sensors, installation deviations, and inconsistent coordinate references, providing a stable and reliable spatial reference for subsequent target recognition and decision-making. Then, the target processing and feature optimization module progressively processes the mapped spatial coordinates. First, it clusters the corresponding geometric feature sets of the items to be sorted, achieving accurate separation of multiple targets. Then, based on the surface distribution of the items, it divides the geometric feature sets into multiple planar coordinate subsets, based on each... The inherent morphological continuity and structural matching relationship between surfaces are subjected to cross-plane linkage fitting, contour deviation is corrected synchronously and clustering threshold is adaptively optimized, the contour reconstruction accuracy and target segmentation robustness are improved, and the defects such as missing, sparse and noise interference of the original point cloud are made up for. Finally, a complete and regular target space set that fits the real shape is obtained. The autonomous decision and motion control module relies on the precise geometric boundary and complete physical features provided by the target space set to efficiently complete the optimal gripping point positioning, stable posture matching and collision-free trajectory planning, which greatly improves the operation accuracy, gripping reliability and autonomous decision-making ability of the logistics sorting robot.

[0021] The working principles of the multimodal perception module, target processing and feature optimization module, and autonomous decision-making and motion control module are detailed below: The multimodal perception module includes a multi-coordinate system construction unit for establishing the global coordinate system of the sorting area, the robot base coordinate system, and the local coordinate system of the multi-source sensors, and for completing the distortion correction of the sensor raw data and the local coordinate normalization; and a global coordinate transformation unit for achieving two-level coordinate transformation from each local coordinate system to the robot base coordinate system and then to the global coordinate system of the sorting area through hand-eye calibration, and for completing the spatial reference normalization. The specific working principle is as follows: The multi-coordinate system construction unit establishes a global coordinate system for the sorting area based on the plane of the conveyor belt, a robot base coordinate system corresponding to the intelligent robot, and a local coordinate system corresponding to the multi-source sensors. Specifically: The global coordinate system of the sorting area in the multi-coordinate system construction unit is as follows: a fixed reference mark point on the conveyor belt plane (such as the calibration reference mark point on one side of the starting end of the conveyor belt) is selected as the origin of the global coordinate system of the sorting area, and the coordinate axes are defined according to the right-hand three-dimensional rectangular coordinate system rule; that is, the X-axis is laid out horizontally along the material conveying direction of the conveyor belt, the Y-axis is laid out horizontally in the conveyor belt plane perpendicular to the conveying direction, and the Z-axis is vertically upward perpendicular to the conveyor belt plane. The robot base coordinate system in the multi-coordinate system construction unit is as follows: the geometric center of the bottom surface of the robot base or the center of the base mounting flange is used as the origin of the coordinate system. The orientation is also based on the right-hand rectangular coordinate system rule: the Z-axis is vertically upward along the robot base, the X-axis points to the horizontal direction directly in front of the robot facing the conveyor belt operation, and the Y-axis is generated orthogonally to the X-axis and Z-axis through the right-hand rule. The multi-source sensors in the multi-coordinate system construction unit include vision, laser, spectral, and temperature sensing sensors. Their corresponding local coordinate systems are all established with their own hardware detection center / optical lens center as the origin, and each independently establishes its own local coordinate system. Specifically, according to the sensor's installation angle and detection orientation, a three-axis direction is defined, and an orthogonal three-dimensional rectangular coordinate system is constructed with the sensor's own detection plane as the reference. This system is only used to describe the spatial position and shape characteristics of the target point within the sensor's field of view relative to its own detection origin. Due to differences in installation position, pitch angle, and detection angle, the origin position and three-axis orientation of each sensor's local coordinate system are not consistent, and they can only realize the acquisition of their own original coordinates and feature recording. Furthermore, at the sorting site, multi-source sensors collect raw detection information such as the surface morphology, contour, and location of the items to be sorted on the conveyor belt in real time. Through coordinate mapping and transformation algorithms, the surface information of the items to be sorted collected by the sensors is uniformly projected and transformed to the local coordinate system of each sensor. Multiple sets of independent spatial coordinates of the items to be sorted under different sensor fields of view are calculated, specifically: Due to inherent errors such as lens distortion, probe plane tilt, physical deviation of the photosensitive array, sensor mounting angle, and origin offset, the raw detection information is only the relative original point position under the sensor detection array. The coordinate axes are not orthogonal, the origin is not standardized, and there are distortion and offset issues, making it unsuitable for direct use as standard local coordinates. In the coordinate mapping and transformation algorithm, a pre-calibrated intrinsic parameter correction matrix is ​​first used to correct distortion and normalize the scale of the original point position, resulting in a distortion-free ideal detection point position. Then, a preset local attitude normalization rotation matrix is ​​used to achieve orthogonal alignment of the coordinate axes, and a local origin translation vector is used to compensate for the offset of the reference origin. The specific expression is as follows: ; in: For sensors The output is the original detection information, i.e. the original 3D point location; For sensors The intrinsic parameter correction matrix is ​​used for distortion correction and scale normalization; For sensors Standard coordinates in its own local coordinate system; Let be the local attitude normalization rotation matrix of sensor a, used to correct the deflection angle between the original axis and the standard local coordinate system axis; For sensors The local origin translation vector is used to compensate for the offset between the original detection reference and the geometric origin of the sensor's local coordinate system; Meanwhile, since the multi-source sensors in the multi-coordinate system construction unit include different types such as vision, laser, infrared, and millimeter wave, and the detection and imaging mechanisms, signal acquisition principles, and perception dimensions of various sensors are different, in the process of synchronously collecting and calculating the spatial coordinates of each set of items to be sorted, based on the detection principle of each sensor itself, the multi-dimensional physical characteristics of the target item corresponding to each spatial coordinate point are obtained synchronously, such as shape and texture features, distance and depth features, material spectral features, contour size features, temperature and radiation features, etc. Furthermore, considering that in actual logistics sorting operations, although the multi-source sensors in the multi-coordinate system construction unit can independently collect multi-dimensional sensing data such as surface morphology, material spectrum, and temperature distribution of the items to be sorted, the origins and axial directions of the local coordinate systems of each sensor are inconsistent due to different installation positions and detection angles. This results in the spatial coordinates under different sensors not being able to be correlated and aligned under the same reference. Therefore, the global coordinate transformation unit establishes a coordinate transformation mapping relationship between the local coordinate systems of each sensor and the global coordinate system of the sorting area, uniformly mapping and transforming each local coordinate system to the global coordinate system of the sorting area. This unifies the spatial coordinates under each local coordinate system with the spatial reference of the corresponding multi-dimensional physical features, providing a consistent and fusionable spatial reference for subsequent target segmentation, contour reconstruction, and gripping point decision-making. Specifically: Establish an independent local coordinate system with the detection center of each sensor, and solve the rotation matrix and translation vector of each local coordinate system relative to the robot base coordinate system through hand-eye calibration, and transform each local coordinate system to the robot base coordinate system; The global rotation matrix and global translation vector of the robot base coordinate system relative to the global coordinate system of the sorting field are obtained again through hand-eye calibration, and the coordinates of the base coordinate system are further transformed to the global coordinate system of the sorting field. Finally, by constructing the overall mapping relationship of each sensor through the above two-level transformation cascade, the spatial coordinates collected by each sensor are rotated and aligned and translated and compensated point by point, so as to uniformly map and transform all local coordinates to the same global coordinate system of the sorting field, and realize the normalization of the spatial reference of multi-source sensing data. The aforementioned hand-eye calibration specifically involves establishing a rigid spatial transformation relationship between the sensor's local coordinate system and the reference coordinate system. This is achieved by having the robot's end effector or sensor observe the calibration marker points in multiple different postures, while simultaneously recording the relative position changes of the sensor to the calibration plate and the end effector position changes corresponding to the robot's own joint poses. Based on paired observation data from multiple postures, the spatial transformation parameters between the two are solved, ultimately yielding the rotation matrix and translation vector of the sensor coordinate system relative to the reference coordinate system. Similarly, by constructing a global coordinate system for the sorting area using fixed reference marker points on the conveyor belt plane and repeating the above calibration process, the global rotation matrix and global translation vector of the robot's base coordinate system relative to the global coordinate system of the sorting area can be obtained.

[0022] Furthermore, after the spatial coordinates under different coordinate systems are uniformly mapped to the global coordinate system of the sorting field through the global coordinate transformation unit, since there are multiple target items of different shapes and materials distributed on the conveyor belt at the same time, the discrete spatial coordinate points on the surface of each item will be mixed and intertwined in the global coordinate system of the sorting field. It is impossible to directly distinguish which coordinates belong to the same item. Therefore, the coordinate point clouds of multiple different items to be sorted will be mixed in the global coordinate system of the sorting field. In order to clarify the independent spatial range and characteristic attributes of each item, so that the subsequent autonomous decision-making and motion control module can accurately process individual targets; The target processing and feature optimization module includes a joint clustering segmentation unit for performing joint clustering on discrete spatial coordinates in the global coordinate system of the sorting field to generate independent geometric feature sets, a contour fitting unit for performing planar layering and cross-planar linkage verification fitting reconstruction on the geometric feature sets and adaptively optimizing the clustering similarity threshold, and a physical feature completion unit for generating physical feature vectors to complete all attribute information for matching virtual coordinates within the target space set.

[0023] The joint clustering segmentation unit employs a joint clustering algorithm to cluster and group multiple similar spatial coordinates discretely distributed in the global coordinate system of the sorting area within the global coordinate transformation unit, forming several independent geometric feature sets. Each geometric feature set corresponds to a target item to be sorted. The specific working principle of the joint clustering algorithm is as follows: Retrieve the spatial coordinates of each space in the global coordinate system of the sorting area, as well as the multi-dimensional physical features of the spatial coordinates obtained by scanning with different sensors. Merge the spatial coordinates and multi-dimensional physical features in a fixed order to obtain the spatial feature vector of each spatial coordinate. Clustering spatial coordinates based on the similarity between spatial feature vectors of each spatial coordinate yields multiple geometric feature sets composed of multiple spatial coordinates. The working principle is as follows: Let the sorting yard be in the global coordinate system, the first... ( , The spatial feature vectors corresponding to the total number of spatial coordinates in the global coordinate system of the sorting area are: ,in For spatial coordinates, Spatial coordinates The physical feature vector, The total number of dimensions of the physical features obtained from all sensor scans; Classification space feature vector To find a cluster, calculate the mean of the same feature dimension among all spatial feature vectors under the cluster, and use this mean as the cluster feature vector corresponding to the cluster. ; Select the spatial feature vectors corresponding to the spatial coordinates of the adjacent spatial nodes outside the cluster. Calculate cluster feature vectors Spatial feature vectors Overall similarity between Specifically, a weighted calculation method can be used, which involves assigning different similarity weights to spatial coordinates and dimensional physical features, and then summing the similarities of each part to obtain the comprehensive similarity. ; First, calculate the cluster feature vector. With spatial feature vectors Similarity in spatial coordinate dimension Specifically, the clustering feature vector The corresponding spatial coordinate components and spatial eigenvectors The similarity in the coordinate dimensions is calculated by subtracting the spatial coordinate components, then taking the L2 norm (Euclidean distance) of the difference vector, and finally adding 1 to the Euclidean distance and taking its reciprocal. A higher similarity in the coordinate dimensions indicates a stronger spatial feature vector. Clustering feature vectors The closer their spatial locations are; Secondly, for each physical feature dimension k, the clustering feature vector is calculated separately. With spatial feature vectors The similarity of the same physical feature dimension can be obtained by normalizing the difference. That is, first calculate the absolute value of the difference between the two feature values, then divide the absolute value by the maximum possible difference of the corresponding physical feature dimension, and subtract the normalized difference from 1 to get the similarity. Assign different similarity weights to spatial coordinates and dimensional physical features: Collect spatial coordinates and physical feature sample data of multiple sets of targets to be sorted in the logistics sorting scenario, calculate the discrete variance of the spatial coordinate dimension within the same type of target and the feature difference value between different targets, and calculate the ratio of inter-class distance to intra-class distance for each physical feature dimension. Use this ratio as the discrimination score of the corresponding feature. The larger the ratio, the stronger the feature's ability to distinguish between different targets to be sorted. Then, sum the discrimination scores of all feature dimensions, divide the discrimination score of a single feature by the sum of the discrimination scores of all features to obtain the initial weight of the feature, and then perform normalization processing on the initial weights of the spatial coordinate dimension and all physical feature dimensions to ensure that the sum of the spatial coordinate weight and the weights of each physical feature is 1. Finally, obtain the feature similarity weights adapted to the current scenario. Set similarity threshold If the similarity threshold is greater than or equal to the spatial coordinates, then the spatial coordinates are determined. They are highly consistent with clusters in terms of spatial location and characteristic representation, and are classified into the same cluster; In the process of continuously expanding and absorbing other spatial coordinates that meet the similarity conditions, the clustering feature vectors corresponding to the clusters are calculated synchronously and iteratively. The similarity judgment and classification of adjacent spatial coordinates and spatial coordinates that have been classified into the same cluster are completed in a loop until no new spatial coordinates can meet the admission conditions and be added to the current cluster. Finally, a cluster including multiple spatial coordinates is obtained, which is the geometric feature set. In the process of similarity judgment of spatial coordinates already classified into the same cluster, if the comprehensive similarity between the spatial coordinates already classified into the same cluster and the cluster feature vector after iterative clustering is less than the similarity threshold, it indicates that the geometric features of the spatial coordinates within the cluster are too different, the morphological attributes are not homogeneous enough, and there are problems such as misclassification of heterogeneous coordinates, mismatch of planar contour connection, and noise interference from measured data. The cluster as a whole does not have a unified geometric feature distribution pattern. Therefore, it is necessary to remove the heterogeneous coordinates in the current cluster, and then iteratively update the cluster feature vector of the corresponding remaining compliant coordinates and perform similarity verification again. Then, another spatial coordinate that is not in the geometric feature set is selected for iterative calculation until all spatial coordinates in the global coordinate system of the sorting field are clustered, and finally multiple geometric feature sets in the global coordinate system of the sorting field are obtained. In the process of using the joint clustering algorithm to cluster multiple spatial coordinates as geometric feature sets, the joint clustering segmentation unit is prone to misclassification and omission due to the proximity of spatial coordinates of different items, uneven point cloud density, or discrete sampling deviation. Thanks to the multi-dimensional physical features extracted by multi-source sensors, which can complement and fuse with spatial coordinate information, the representation dimensions of target items can be enriched, the deviation caused by single spatial coordinate clustering can be weakened, and when calculating the comprehensive similarity between samples, the similarity measurement basis can be constructed by relying on spatial location association and multi-dimensional physical feature attributes. This can effectively improve the distinguishability between different items to be sorted, enhance the accuracy and anti-interference ability of clustering, and avoid the classification confusion caused by spatial coordinate overlap or local sparsity. Furthermore, during the process of clustering and aggregating multiple similar spatial coordinates discretely distributed in the global coordinate system of the sorting field by the joint clustering algorithm, since the same item to be sorted occupies continuous physical space on the conveyor belt, all the spatial coordinate points detected on its surface are naturally located in the adjacent three-dimensional spatial region in the global coordinate system of the sorting field. There is no large-scale intersection with the detection points of other items. Moreover, the physical properties such as material, shape, spectrum, and temperature of the same item to be sorted are consistent, and the corresponding physical feature vectors collected by each sensor are highly similar. The spatial coordinates of different items to be sorted will be divided into different clusters due to spatial distance or feature differences. Therefore, each geometric feature set corresponds to one item to be sorted on the conveyor belt. Meanwhile, during the process of scanning items to be sorted using multi-source sensors in the multi-coordinate system construction unit, the complex surface morphology of the items to be sorted (such as unevenness, sharp edges and corners, irregular textures, etc., resulting in weak or blocked reflected signals in some areas during sensor detection), the differences in the viewing angles of the multi-source sensors (different sensor installation positions and detection pitch angles, resulting in deviations in the detection angle and distance of the same target point under different sensor fields of view, making it impossible to achieve complete overlap detection), the discrete distribution characteristics of the point cloud data itself (the points collected by the sensors can only characterize the local features of the target surface, and cannot cover the entire outline of the target, naturally resulting in gaps between points), and the susceptibility to various external interference factors (environmental noise from workshop equipment vibration and electromagnetic interference, causing false feedback in sensor detection signals; changes in illumination will change the reflectivity of the target surface, causing the outline points collected by the visual sensors to shift and feature distortion; conveyor belt vibration will cause slight displacement of the items to be sorted, resulting in deviations between the points collected by the sensors and the actual spatial position of the target). Although intrinsic parameter correction matrices, local attitude normalization rotation matrices, and local origin translation vectors have been used for correction and normalization in the multi-coordinate system construction unit, the complex topography, viewpoint differences, discrete characteristics, and various interference factors have randomness and complexity. Furthermore, a single correction and normalization method cannot completely offset the errors caused by the superposition of multiple factors (intrinsic parameter correction can only eliminate the optical distortion of the sensor itself, and cannot solve the deviations caused by external interference and viewpoint differences; attitude normalization and origin translation can only normalize the local coordinate system of the sensor, and cannot compensate for the positional deviations caused by inconsistent viewpoints of multiple sensors, target displacement, and noise). This still leads to various problems with the spatial coordinates in the global coordinate system of the sorting area. Outliers (false points unrelated to the actual target contour are collected by the sensor due to environmental noise and vibration interference, with spatial distance deviations from normal points exceeding reasonable range), missing data (the target surface is obscured or the reflected signal is weak, the sensor cannot effectively collect points, resulting in blank points in the target contour), local sparsity (flat areas on the target surface or blind spots of the sensor detection, the density of collected points is too low, and it cannot accurately represent the local shape of the target), and data redundancy (overlapping areas of multiple sensor views, repeated collection of the same target point, resulting in duplicate coordinate data and increased invalid calculations) are all problems that ultimately prevent the spatial distribution range corresponding to the geometric feature set from fully describing the sorted target items. To further avoid the above situation, and to accurately define the spatial distribution range of the geometric feature set corresponding to the items to be sorted, the contour fitting unit receives the geometric feature set containing multiple spatial coordinates from the joint clustering segmentation unit, and reconstructs multiple spatial coordinates in each geometric feature set to obtain the target spatial set corresponding to the items to be sorted. This target spatial set describes the complete spatial distribution range of the items to be sorted, effectively compensating for holes, outliers, and redundant points in the original point cloud. This ensures that the target spatial range is continuous, complete, and closely matches the real shape, providing a reliable spatial boundary basis for subsequent gripping point selection and trajectory planning. Specifically: Based on the planar layering method, multiple spatial coordinates in the geometric feature set are divided: taking any coordinate axis of X-axis, Y-axis or Z-axis in three-dimensional space as a fixed coordinate axis, all spatial coordinates in the geometric feature set are traversed, and selection rules are set to filter the spatial coordinates on the fixed coordinate axis and collect them into the same layer plane. The disordered discrete spatial coordinates in the geometric feature set are decomposed and divided into several mutually parallel and non-overlapping planar coordinate subsets. The selection rules in the planar layering method are as follows: Set a tolerance threshold, and select multiple spatial coordinates based on a fixed coordinate axis. The setting of the tolerance threshold needs to comprehensively consider the sensor measurement accuracy, point cloud noise level, and surface undulation characteristics of the target item. This ensures that the spatial coordinates within the same fixed coordinate axis are neither excessively dispersed nor fail to be selected due to an excessively small tolerance threshold. The inherent measurement error of the 3D vision sensor itself can be used as a benchmark, and error adaptation correction can be performed by combining the random discrete noise and coordinate offset deviation of the point cloud data collected on site. At the same time, the coordinate dispersion range caused by the natural undulations and local concave-convex deformation of the surface of the target item to be sorted should also be taken into account. In this way, a reasonable tolerance threshold can be comprehensively defined. The linked verification method is used to analyze the contour shape curves corresponding to the subset of plane coordinates; Furthermore, since the sensor's physical acquisition accuracy is an inherent parameter calibrated during the scanning process, even if there are outliers or acquisition anomalies in the local area, most of the spatial coordinates obtained by the scan can still stably conform to the minimum effective resolution distance represented by this accuracy. Therefore, in the fitting and reconstruction process, the sensor's physical acquisition accuracy is defined as the minimum spacing threshold, and the planar coordinate subset is fitted and reconstructed based on the minimum spacing threshold and the contour shape curve. Under the premise of preserving the real contour shape, redundant noise points with excessively dense spacing are removed, and local sparse missing areas are filled in. The invalid fluctuations and redundant stacking of spatial coordinates in the fitted and reconstructed planar coordinate subset are minimized, and finally, a target spatial set including multiple spatial coordinates is obtained, which further provides a continuous, complete and realistic spatial boundary description for the items to be sorted. The system uniformly traverses and samples reference feature points along the entire contour curve, ensuring that the spatial distance between adjacent reference feature points is not less than a set minimum spacing threshold. Since the spacing between adjacent spatial coordinates naturally fluctuates due to factors such as curvature variations and sensor detection angle differences on the real contour surface, a distance tolerance range can be set. Only when the deviation between the spatial distance between adjacent spatial coordinates and the minimum spacing threshold exceeds this tolerance range is the point considered invalid and discarded. This preserves the effective geometric information of the contour curvature variation area and avoids the loss of key morphological features due to forced spacing constraints. Specifically, based on the combined characteristics of sensor physical acquisition accuracy and contour curvature variation, its working principle is as follows: First, the minimum effective resolution distance calibrated by the sensor at the factory is used as the baseline spacing threshold. Then, a basic tolerance is set based on the sensor acquisition error and point cloud noise level. Typically, ±10% to ±20% of the baseline spacing threshold is taken as the basic fluctuation range to cover random measurement deviations under normal operating conditions. On this basis, an adaptive adjustment mechanism for contour curvature is introduced. A smaller tolerance is used for areas with gentle contour curvature changes to ensure uniform distribution of sampling points without redundant stacking. For corners and concave areas with drastic contour curvature changes, the tolerance range is automatically expanded to allow the point spacing to fluctuate within a larger range, thereby preserving key geometric information at curvature abrupt changes and avoiding shape distortion caused by forced spacing constraints. This tolerance range can be dynamically set through a preset parameter table or real-time curvature calculation. This eliminates abnormal noise points that deviate significantly from the normal distribution while ensuring that key contour feature points are not mistakenly removed, achieving a balance between point cloud sampling accuracy and contour shape fidelity. The working principle of the specific linkage verification method is as follows: Set calibration polynomial coefficients, construct a contour shape model, input the abscissa of each contour coordinate, the contour shape model combines the abscissa and calibration polynomial coefficients to output the corresponding fitting ordinate, and combine the fitting ordinate and the abscissa of the contour coordinate to obtain the fitting coordinate. Furthermore, although the point distribution and local morphology of the planar coordinate subsets corresponding to different planes differ, all planar coordinate subsets are located together and constitute a complete target object, naturally forming an inherent cross-planar shape matching relationship between adjacent planar coordinate subsets. Therefore, in the process of constructing the contour shape curve, contour linkage verification is carried out based on the cross-planar shape matching relationship. Under the premise of satisfying the constraints of cross-planar morphology adaptation and structural matching, the calibration polynomial coefficients are fine-tuned into modified polynomial coefficients, and then the fitted coordinates are updated based on the modified polynomial coefficients. Finally, multiple fitted coordinates are connected sequentially to obtain the contour shape curve. Specifically: Shape fitting constraint refers to the continuous transition and consistency of the target contours corresponding to adjacent layer planes in terms of shape, size, curvature trend and smoothness, without morphological distortions such as abrupt changes in cross-sectional area, jumps in contour curvature, and excessive differences in fitting residuals. Structural matching constraints refer to the alignment of adjacent planar contours in three-dimensional space, the consistency of topological association and normal orientation, the spatial distance deviation of corresponding key points and the angle deviation of contour normal vectors within the allowable range, and the absence of structural mismatch problems such as offset, misalignment, distortion, and flipping. The specific working principle of the linkage verification method for setting calibration polynomial coefficients is as follows: initial polynomial coefficients are randomly set, and the contour shape model is constructed by fitting the contour coordinates using the initial polynomial coefficients. The minimum residual between the fitted coordinates and the contour coordinates is defined as the objective function. Since the polynomial coefficients of each layer plane are optimized synchronously rather than solved independently, when the sum of the residuals between the fitted coordinates and the contour coordinates of all planes is minimized as the global objective function, the coefficient adjustment of each plane will be indirectly constrained by the contour data of the adjacent planes. In order to reduce the overall residual, the fitted curve of the current plane and the contour shape and position of the adjacent planes above and below should maintain a continuous and smooth transition. Otherwise, the fitting residual of the adjacent contours will increase as a whole, and the residual cannot be minimized. Therefore, when the global residual reaches the minimum, the contour shape of all adjacent planes will naturally satisfy the shape adaptability, and the spatial position and orientation will necessarily meet the structural matching. The initial polynomial coefficients are dynamically adjusted with the objective function as a constraint, and finally the calibration polynomial coefficients corresponding to the contour shape model are obtained. The specific outline shape model is as follows: ; in: To fit the ordinate; The x-coordinate of the contour coordinates; To calibrate the polynomial coefficients, specifically: The x-coordinates are respectively square, cube... The power is used to combine different coefficients in the calibration polynomial to create curves such as straight lines, oblique lines, and circular arcs, adapting to various geometric shapes. The highest order of the polynomial determines the complexity of how well the contour shape model can fit the contour; the highest order of the polynomial The larger the value, the better it can fit the edges of objects with many bends and complex shapes; Calibration of polynomial coefficients middle: This is a constant term that controls the overall vertical translation of the entire contour curve. The coefficient of the first-order term controls the slope of the curve and is responsible for fitting the straight sides of squares, triangles, and trapezoids; The coefficients of the quadratic term control the curvature and convexity of the curve, and are responsible for fitting circles and arcs. These are higher-order coefficients used to fit complex, winding, and irregular contours.

[0024] Furthermore, since the calibration polynomial coefficients in the aforementioned contour fitting unit represent the locally optimal fitting parameters in a single plane, and the correction polynomial coefficients represent the globally fitted parameters after correction under the cross-plane shape consistency constraint, the calibration polynomial coefficients are obtained only by minimizing the residual between the original coordinates and the fitted coordinates in a single plane, without considering the continuity or shape matching of the contour between adjacent planes. In contrast, the correction polynomial coefficients force the satisfaction of shape transition smoothness and structural consistency constraints in cross-plane linkage verification, causing the originally locally optimal parameters to shift in order to adapt to global geometric self-consistency. Therefore, the magnitude of the difference between the two is proportional to the degree of abandonment of local optima in order to meet the requirements of shape matching and structural continuity between adjacent planes, thereby directly quantifying the degree of fitting deviation of the contour shape model under this constraint. The similarity threshold used in the joint clustering segmentation unit to determine whether clustering is possible is the key discrimination boundary for determining whether different spatial coordinates can be classified into the same geometric feature set. If the similarity threshold is set too high, it is easy to merge coordinates belonging to different target items incorrectly; if it is set too low, it will over-segment the coordinates of the same target item. At this point, the smaller the difference between the calibration polynomial coefficients and the correction polynomial coefficients, the closer the local optimal fit in a single plane is to the global consistency constraint across planes, and the higher the accuracy of the contour model in describing the geometry of the real object. In this case, a stricter similarity threshold can be used to improve the purity of segmentation. Conversely, the larger the difference, the more serious the cross-plane shape conflict, and the larger the contour fitting error. If the similarity threshold is still used, the surface coordinates of the same object will be excessively fragmented. Therefore, the similarity threshold must be appropriately relaxed to ensure the integrity of the target. Based on the above inverse correlation, in order to achieve the clustering threshold to adaptively match the actual accuracy of contour fitting in the current scene; Therefore, when setting the similarity threshold, the optimal setting is to minimize the difference between the calibration polynomial coefficients and the modified polynomial coefficients. Furthermore, considering that the spatial coordinates and multidimensional physical features relied upon by cluster analysis, as well as the same set of original point cloud data relied upon by contour analysis, all originate from synchronous acquisition and unified coordinate mapping of multiple sensor sources, there is no necessary sequential dependency between the two in the calculation process. Therefore, the cluster segmentation unit for cluster analysis of geometric feature sets and the contour fitting unit for contour shape curve analysis can be performed simultaneously to avoid timing contradictions caused by sequential execution. This allows the difference between the calibration polynomial coefficients and the correction polynomial coefficients calculated during contour fitting to be used in real time or iteratively to guide the adaptive optimization of the similarity threshold. Its detailed working principle is as follows: After receiving the spatial coordinates in the global coordinate system of the sorting field in the same frame, the joint clustering segmentation unit and the contour fitting unit run in parallel. The target processing and feature optimization module first uses an initial similarity threshold, which can be set based on the sensor calibration accuracy or historical statistical values, to perform a joint clustering algorithm on the point cloud and generate a preliminary geometric feature set. At the same time, the contour fitting unit runs synchronously, performs rapid planar layering of the spatial coordinates of the same frame, and performs polynomial contour fitting independently on each layered plane based on the principle of proximity of vertical coordinates or conveyor belt plane. On this basis, cross-plane linkage verification is performed to calculate the difference between the calibration polynomial coefficients and the correction polynomial coefficients of the current frame. A preset mapping function is used to dynamically adjust the initial similarity threshold of the current frame based on the difference, resulting in a new similarity threshold. The joint clustering segmentation unit then uses the updated similarity threshold to re-determine the point cloud's affiliation, corrects the existing geometric feature set division, and provides the corrected geometric feature set to the contour fitting unit again. The contour fitting unit then recalculates the difference between the calibration polynomial coefficients and the corrected polynomial coefficients based on the corrected feature set. This process is repeated iteratively. To ensure stable convergence of the iteration process, two stopping conditions need to be set in advance: one is the convergence criterion, which is that when the relative rate of change of the coefficient difference between the current iteration step and the previous iteration step is less than a preset threshold (such as 0.1%), it is determined to be converged; the other is the maximum number of iterations (such as 100 times), which will forcibly terminate the iteration when the maximum number of iterations is reached. The convergence criterion uses the relative rate of change of coefficient differences rather than the absolute change to avoid the influence of the initial deviation. Simultaneously, by setting upper and lower limits for the similarity threshold, it prevents the threshold from being excessively widened or tightened during iteration, thus ensuring that the clustering and contour fitting results do not deviate from a reasonable range. Theoretically, this iterative process can be viewed as a convex optimization problem with the goal of minimizing the difference. Since the objective function is convex and the parameters are updated within the feasible region in each iteration, there exists a unique global optimum. As long as the iteration step size and threshold adjustment step size are controlled within a reasonable range, stable convergence of the iteration can be guaranteed, and divergence will not occur. When any stopping condition is met, the iteration terminates. At this point, the final output geometric feature set and contour fitting results are consistent, and the similarity threshold has been adaptively optimized to the value that minimizes the cross-plane fitting deviation.

[0025] Specifically, the difference between the calibration polynomial coefficients and the correction polynomial coefficients is the deviation between the initial calibration polynomial coefficient vector of the contour fitting and the polynomial coefficient vector after cross-plane linkage verification and correction on the same hierarchical plane. It can be quantified by the norm of the coefficient vector. Specifically, first, construct the calibration polynomial coefficient vector A and the correction polynomial coefficient vector B, where each dimension of the vector corresponds to the coefficients of different orders of the polynomial. Then, calculate the difference between the corresponding elements of the two vectors and construct the difference vector Δ = B − A. Subsequently, calculate the second norm of the difference vector Δ, which is the square root of the sum of the squares of the differences of each coefficient, as the quantified value of the overall deviation. Alternatively, the first norm, which is the sum of the absolute values ​​of the differences of each coefficient, can be used as the deviation index. The magnitude of this difference directly reflects the correction magnitude of the local optimal fitting result in a single plane under the constraints of cross-plane shape adaptation and structural matching. The smaller the difference, the higher the consistency between the local optimum and the global constraint, and the better the overall stability of the contour fitting. Conversely, the larger the difference, the greater the conflict between the local fitting and the global continuity requirement, and the greater the magnitude of the contour correction. The mapping function is specifically: to retrieve the L2 norm of the difference vector Δ between the calibration polynomial coefficient vector A and the correction polynomial coefficient vector B. As a deviation quantification value, a preset benchmark similarity threshold is introduced. Deviation adjustment coefficient Maximum allowable deviation from the scene Construct mapping function ,in This refers to the maximum permissible deviation value determined during historical statistics or calibration. To control the coefficient of the threshold decay rate, a mapping function is used to convert the magnitude of the coefficient deviation into the adjustment range of the similarity threshold. When the deviation is 0, the threshold takes the most stringent benchmark value. When the deviation increases, the threshold decreases linearly, thereby realizing the inverse correlation between the deviation and the threshold. The above benchmark similarity threshold The basic fixed similarity judgment threshold used when the polynomial coefficients have no deviation and the contour fitting effect reaches the ideal optimal state can be initially set based on the factory calibration accuracy parameters of the multi-source sensor. Then, it can be corrected and optimized by combining the experience data of geometric feature clustering judgment under normal error-free working conditions in historical sorting operations. At the same time, it can be comprehensively determined by referring to the lowest similarity critical value of successful matching of similar target shapes, and used as the benchmark reference value for all dynamically adjusted thresholds. Deviation adjustment coefficient The adjustment parameter used to control the rate at which the adaptive similarity threshold decreases as the polynomial coefficient deviation increases directly determines the range of change in the threshold leniency. It can be determined by experimental debugging and scenario adaptation. The value is iterated and adjusted multiple times within the standard test sample group. Combined with the two-way verification of clustering segmentation accuracy and target integrity recognition rate, if the fitting deviation is too large, it is easy to have fragmented segmentation scenarios, so a smaller value is selected. If the shape difference is large, it is easy to have mixed clustering scenarios, so a larger value is selected. Finally, a fixed coefficient value that adapts to the overall system operation effect is determined. Maximum deviation value In a sorting operation scenario, the maximum overall deviation limit that the polynomial calibration coefficients and correction coefficients can be accepted by the system represents the upper limit of the error that the contour fitting result can tolerate under the premise of satisfying the constraints of shape adaptation and structural matching. If the value is exceeded, the contour fitting is judged to be ineffective. The contour fitting test can be carried out based on a large number of actual sorting target samples on site, and the extreme distribution range of the L2 norm of the difference between the two sets of coefficients under different working conditions can be statistically determined. The value can be comprehensively determined by combining the equipment sensor calibration error, point cloud acquisition noise, and target shape deformation error. Alternatively, a fixed empirical value can be directly set according to the industry sorting and detection accuracy standards.

[0026] After determining the contour shape curve corresponding to the subset of planar coordinates, the contour fitting unit, specifically the target space set, is as follows: A minimum spacing threshold is set based on the sensor's physical acquisition accuracy. Starting from the initial position, the entire contour curve is traversed segment by segment, strictly using the minimum spacing threshold as a fixed sampling interval. Reference feature points are selected and sampled uniformly across the entire contour curve, ensuring that the spatial distance between adjacent reference feature points is not less than the set minimum spacing threshold. Simultaneously, spatial coordinates in the planar coordinate subset whose spatial distance between adjacent spatial coordinates is not equal to the minimum spacing threshold are removed. The selected uniform reference feature points serve as the contour morphology skeleton, conforming to the morphological variation rules of the original contour curve. Based on this, the planar coordinate subset is regularized and reshaped through interpolation fitting and local morphology correction. Multiple spatial coordinates are reconstructed while satisfying the minimum spacing threshold. This completes the point selection and fitting reconstruction of the planar coordinate subset based on the minimum spacing threshold and the contour curve. Specifically: The reference feature points obtained by uniformly sampling along the contour curve according to the minimum spacing threshold are used as the control skeleton nodes of the overall contour. The spatial coordinates and the arrangement interval are fixed to ensure that the spatial coordinates always meet the minimum spacing threshold and are evenly and orderly distributed. Based on the continuous morphological features of the original contour curve, the interpolation algorithm is used to densify and supplement the points between every two adjacent reference feature points. According to the actual curvature and concavity and convexity of the contour, the intermediate spatial coordinates with smooth transition are generated to fill the gaps in the sparse areas of the original spatial coordinates, so that the points of the entire contour are continuous and smooth and there are no excessively dense points less than the minimum spacing threshold throughout the entire process. Using the fitted contour curve as a reference, the spatial coordinates within the closed area enclosed by the contour curve are retrieved from the subset of planar coordinates. Reconstruction is then performed using the minimum spacing threshold: the effective range of the inner spatial coordinates is defined using the contour curve as the boundary; all effective coordinates within the contour curve are selected through coordinate comparison; then, using the minimum spacing threshold as the core requirement, the selected spatial coordinates within the contour curve are regularized, redundant points with spacing less than the minimum spacing threshold are removed, and effective coordinates are supplemented for sparse areas through interpolation. This ensures that the distribution of spatial coordinates within the contour curve meets the minimum spacing threshold requirement. Simultaneously, the morphological regularity of the contour curve serves as a constraint, and the spatial coordinates within the contour curve are fitted and corrected to both conform to the geometric shape of the inner contour and strictly adhere to the minimum spacing threshold, avoiding situations where points are too dense or too sparse. Finally, all the filtered and reconstructed spatial coordinates from multiple planar coordinate subsets are combined in a fixed order to obtain a target space set including multiple spatial coordinates. After the contour fitting unit constructs the target space set, although the multiple spatial coordinates within it are arranged in an orderly manner through planar layering and contour fitting, providing a regular spatial structure foundation for geometric feature extraction and contour morphology analysis, the subsequent autonomous decision-making and motion control modules, in analyzing the coordinates of the intelligent robot's grasping points, do not only rely on the geometric position information of the spatial coordinates to complete the grasping point determination and stability verification. They also need to combine the physical feature vectors corresponding to the spatial coordinates to perform material property discrimination, grasping reliability assessment, slippage and deformation risk screening, and rigid support area positioning. Since the virtual coordinates are generated by fitting reconstruction and interpolation normalization and do not have directly collected original physical features, they cannot directly participate in the accurate determination of physical properties. Therefore, the physical feature completion unit defines the spatial coordinates obtained from the target space set after fitting reconstruction and interpolation normalization according to the minimum spacing threshold as virtual coordinates. The unfitted and unconstrained spatial coordinates are used as measured coordinates. The virtual coordinates are matched with the geometric feature set where they are located, and the physical feature vectors of the measured coordinates within the geometric feature set are used to construct the virtual feature vectors corresponding to the virtual coordinates. This provides the normalized, continuous, and uniformly distributed virtual coordinates with complete and reliable virtual physical feature vectors. This ensures that the spatial coordinates at the geometric level are arranged regularly, without missing, uneven density, or noise interference, meeting the data structure requirements for planar layering, contour fitting, iterative optimization, and accurate similarity determination. It also ensures that all spatial coordinates at the physical level have usable physical attribute information such as material, strength, roughness, and temperature. Based on unified and complete geometric and physical features, this effectively avoids judgment bias and process interruptions caused by missing coordinates and feature anomalies, improving the accuracy, stability, and robustness of the grasping and positioning process. The specific working principle is as follows: The physical feature vectors of the geometric feature sets corresponding to the measured coordinates in the joint clustering segmentation unit are retrieved. The joint clustering algorithm in the joint clustering segmentation unit is used again to cluster and analyze the physical feature vectors. Then, based on the clustering convergence result, the overall geometric feature set is adaptively divided into several sub-geometric feature sets with the same geometric attributes and spatial distribution patterns. Retrieve the measured coordinates corresponding to each sub-geometric feature set and determine the corresponding spatial range. Specifically, construct an axis-aligned three-dimensional minimum bounding box with the extreme values ​​of each axis as the boundary, and directly define the spatial region of the bounding box as the exclusive spatial range corresponding to the sub-geometric feature set. To match the exclusive spatial range where the virtual coordinates are located, a virtual feature vector corresponding to the virtual coordinates is constructed using the physical feature vectors of multiple measured coordinates within the exclusive spatial range. Specifically: Calculate the spatial distance between multiple measured coordinates and virtual coordinates within the exclusive spatial range, and assign fusion weights for the weighted fusion of physical feature vectors to each measured coordinate based on the size of the spatial distance. That is, the closer the spatial distance, the larger the fusion weight, and the farther the spatial distance, the smaller the fusion weight. Then, the physical feature vectors of multiple measured coordinates are fused using fusion weights to obtain the virtual feature vectors corresponding to the virtual coordinates. Furthermore, during the clustering analysis process where the physical feature completion unit again employs the joint clustering algorithm from the joint clustering segmentation unit, the joint clustering segmentation unit uses similarity weights to calculate the comprehensive similarity between the clustering feature vector and the corresponding spatial coordinates. This reflects the inherent similarity level and distinguishing criteria of the spatial coordinates within the same target object in terms of geometric distribution, spatial continuity, and feature correlation. However, in the process of clustering multiple spatial coordinates within the geometric feature set, since all spatial coordinates to be clustered and the joint clustering segmentation unit process the same frame of source point cloud data, their spatial distribution patterns, noise levels, coordinate discreteness characteristics, and hierarchical fitting deviations are completely consistent. Moreover, the core objective of clustering is also to group spatially adjacent, geometrically homogeneous, and attribute-consistent coordinates into the same set, which aligns with the joint clustering segmentation. The similarity judgment target, feature differentiation logic, and physical meaning of the units are completely unified. Moreover, the similarity threshold is set based on the difference between the calibration polynomial coefficients and the correction polynomial coefficients, which can dynamically adapt to the actual fitting accuracy and spatial consistency of the point cloud in this frame. It has a unique, accurate, and adaptive differentiation scale. Therefore, the physical feature completion unit directly reuses the similarity weights and similarity thresholds in the joint clustering segmentation unit. The geometric feature set is divided into multiple sub-geometric feature sets through the similarity threshold, thereby ensuring that the feature division standard is unified throughout the process, the boundary judgment is consistent, and the calculation logic is coherent. This avoids the misalignment of set division, confusion of feature attribution, and distortion of virtual feature mapping caused by the independent setting of weights and thresholds. At the same time, it greatly reduces the overhead of parameter calibration and algorithm reconstruction, and improves the consistency, stability and processing efficiency of the system.

[0027] To further clarify the control commands for the intelligent robot, the autonomous decision-making and motion control module retrieves multiple spatial coordinates within the target space set in the contour fitting unit, as well as the physical feature vectors corresponding to the spatial coordinates, and analyzes the coordinates of the intelligent robot's grasping point. Based on spatial coordinates, a 3D axis-aligned minimum bounding box is constructed for the items to be sorted. The centroid of the bounding box is calculated and used as the initial gripping reference point. All spatial coordinates are traversed and spatial geometric features are extracted. Then, filtering conditions are set to select points whose robot gripping direction tends to be orthogonal, unobstructed, and without edge interference as geometrically optimal points. Specifically: Based on spatial coordinates, a three-dimensional axis-aligned minimum bounding box for the items to be sorted is constructed. By traversing all spatial coordinates within the target space set, the extreme values ​​of the X, Y, and Z axes in the global coordinate system of the sorting field are obtained respectively, and the bounding box boundary is defined accordingly. This forms a minimum cuboid closed area that is parallel and aligned with the three axes of the global coordinate system of the sorting field and can completely enclose the target items. The geometric centroid of the bounding box is then calculated by averaging the extreme values ​​of the three axes. This centroid is used as the initial gripping reference point that has the characteristics of being centered and having balanced force. Then, all spatial coordinates within the target space set are traversed, and spatial geometric features, including contour curvature and surface normal vectors, are extracted based on the differential operation of neighborhood points to characterize the surface morphology of the target object and determine the grasping stability and posture adaptability. Set up multi-level filtering conditions in sequence to select stable grasping areas, match grasping postures, and ensure interference-free grasping; Based on multi-level screening conditions and spatial geometric features, each spatial coordinate is matched and verified one by one, and only spatial coordinates that simultaneously meet all screening conditions are retained and determined as the geometrically optimal points for the robot's stable grasping operation. Simultaneously, by combining the physical feature vectors corresponding to each spatial coordinate in the geometrically optimized points, adaptive discrimination is carried out based on the attributes of material hardness, surface roughness, spectral reflectance characteristics, temperature distribution, and structural strength in the physical feature vectors. Geometrically optimized points that are fragile and easily deformed, have smooth surfaces that are easy to slip, have abnormal temperatures, or insufficient structural strength are eliminated. Finally, among the remaining geometrically optimized points, the spatial coordinates that are closest to the center of gravity of the bounding box, have the best physical feature uniformity, the most concentrated spatial distribution, and are located within the rigid support area of ​​the target object are selected as the optimal gripping point coordinates that combine stability and reliability.

[0028] Using the global coordinate system of the sorting area as a unified global reference, safe blank areas in the sorting scene that are not occupied by items to be sorted, conveyor belts, obstacles, and the robot body are marked as feasible paths. These feasible paths meet the constraints of robot joint range of motion, end effector working space, and safe obstacle avoidance distance, providing a safe passage space for robot movement without collisions or interference. Then, based on the real-time spatial coordinates of the robot base in the global coordinate system of the sorting area, the robot's kinematic parameters, and the coordinates of the optimal gripping point, the optimal gripping posture of the end effector is determined through the robot's forward and inverse kinematics calculations. This ensures that the gripping direction of the actuator is orthogonally matched with the surface normal vector of the optimal gripping point. At the same time, within the feasible path range, a path planning algorithm is used to generate the optimal motion trajectory that is collision-free, has the shortest path, and is stable. Finally, the optimal gripping point coordinates, optimal gripping posture, optimal motion trajectory, and actuator opening and closing control parameters are integrated to generate complete control commands that can be directly issued and drive the robot.

[0029] In summary: In this example, the multimodal perception module is used to construct the coordinates of objects in different coordinate systems and map them to a unified global coordinate system of the sorting field. Then, the clustering and segmentation unit is used to cluster multiple spatial coordinates, thereby automatically classifying the scattered, disordered, and mixed spatial coordinates according to spatial proximity and geometric homology. Finally, a set of geometric features representing the target items to be sorted is obtained. In addition to relying on spatial distance, the clustering process also incorporates the physical feature vectors of the objects during the scanning process of the multi-source sensors. This can improve the target segmentation accuracy, suppress noise interference, avoid misclassification caused by overlapping or local sparsity of spatial coordinates, and enhance the distinguishability between different items to be sorted and the stability of the clustering results. During the clustering process of the joint clustering segmentation unit, the contour fitting unit simultaneously performs planar layering, polynomial fitting, and cross-plane linkage verification on the geometric feature set to reconstruct a continuous and complete target space set that fits the real shape. At the same time, based on the difference between the calibration polynomial coefficients and the correction polynomial coefficients, the clustering similarity threshold is dynamically and iteratively optimized in reverse, so that the segmentation scale adaptively matches the current contour fitting accuracy. This avoids the target being over-segmented and prevents the mis-merging of dissimilar targets, achieving a two-way mutual coordination between clustering and fitting and a closed-loop improvement in accuracy. The physical feature completion unit further analyzes the physical feature vectors of each spatial coordinate in the target space set, providing complete and reliable feature support for the robot's gripping point selection, gripping posture planning and operation stability judgment in the subsequent autonomous decision-making and motion control modules.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting, characterized in that, include: The multimodal perception module (100) is used to establish a coordinate system, including the local coordinate system corresponding to the multi-source sensors, the robot base coordinate system corresponding to the intelligent robot, and the global coordinate system of the sorting field; And the spatial coordinates under each coordinate system are uniformly mapped to the global coordinate system of the sorting area; The target processing and feature optimization module (200) uses a joint clustering algorithm including a similarity threshold to cluster multiple spatial coordinates in the global coordinate system of the sorting field, and obtains a geometric feature set after clustering; the geometric feature set is further split into multiple planar coordinate subsets, and then the linkage verification method is used to analyze the contour shape curves corresponding to the planar coordinate subsets to obtain a target space set including virtual coordinates; the joint clustering algorithm is reused to analyze the virtual feature vectors corresponding to the virtual coordinates in the target space set; The linkage verification method first constructs a contour shape model including calibration polynomial coefficients, and then conducts contour linkage verification based on cross-plane shape matching relationship. Under the premise of satisfying cross-plane morphology adaptation and structural matching constraints, the calibration polynomial coefficients in the contour shape model are adjusted to modified polynomial coefficients for analysis. And adjust the similarity threshold in the clustering process based on the difference between the calibration polynomial coefficients and the correction polynomial coefficients; The autonomous decision-making and motion control module (300) analyzes the coordinates of the intelligent robot's grasping point based on the spatial coordinates in the target space set and generates robot sorting control instructions.

2. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 1, characterized in that: The multimodal sensing module (100) includes a multi-coordinate system construction unit (110) and a global coordinate transformation unit (120), wherein: The multi-modal fusion module multi-coordinate system construction unit (110) establishes a sorting field global coordinate system based on the sorting field conveyor belt plane, a robot base coordinate system corresponding to the intelligent robot, and local coordinates corresponding to the multi-source sensors. The global coordinate transformation unit (120) establishes the coordinate transformation mapping relationship between the local coordinate system under the multi-source sensor and the global coordinate system of the sorting field, and maps and transforms each local coordinate system to the global coordinate system of the sorting field. Specifically, it establishes an independent local coordinate system based on the detection center of each modal sensor, and solves the rotation matrix and translation vector of each local coordinate system relative to the robot base coordinate system through hand-eye calibration, and transforms each modal local coordinate to the robot base coordinate system. The global rotation matrix and global translation vector of the robot base coordinate system relative to the global coordinate system of the sorting field are obtained again through hand-eye calibration. The coordinates of the base coordinate system are then further transformed to the global coordinate system of the sorting field.

3. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 2, characterized in that: The target processing and feature optimization module (200) includes a joint clustering segmentation unit (210), a contour fitting unit (220), and a physical feature completion unit (230), wherein: The joint clustering segmentation unit (210) uses a joint clustering algorithm to cluster and collect multiple similar spatial coordinates that are discretely distributed in the global coordinate system of the sorting field in the global coordinate transformation unit (120) to form several independent geometric feature sets, each geometric feature set corresponding to a target item to be sorted. The contour fitting unit (220) fits and reconstructs multiple spatial coordinates in each geometric feature set to obtain the target space set of the corresponding items to be sorted. The physical feature completion unit (230) defines the spatial coordinates in the target space set as virtual coordinates and measured coordinates; it matches the geometric feature set where the virtual coordinates are located with the physical feature vectors of the measured coordinates in the geometric feature set to construct the virtual feature vector corresponding to the virtual coordinates.

4. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 3, characterized in that: The joint clustering algorithm in the joint clustering segmentation unit (210) retrieves each spatial coordinate in the global coordinate system of the sorting field, as well as the multi-dimensional physical features obtained by scanning the spatial coordinates using different sensors. The spatial coordinates and multi-dimensional physical features are fused in a fixed order to obtain the spatial feature vector of each spatial coordinate. Clustering of spatial coordinates based on the similarity between spatial feature vectors of each spatial coordinate yields multiple geometric feature sets composed of multiple spatial coordinates, specifically: Let the sorting yard be in the global coordinate system, the first... The spatial feature vectors corresponding to the spatial coordinates are ,in For spatial coordinates, Spatial coordinates The physical feature vector, The total number of dimensions of the physical features obtained from all sensor scans; Classification space feature vector To find a cluster, calculate the mean of the same feature dimension among all spatial feature vectors under the cluster, and use this mean as the cluster feature vector corresponding to the cluster. ; Select the spatial feature vectors corresponding to the spatial coordinates of the adjacent spatial nodes outside the cluster. Calculate cluster feature vectors Spatial feature vectors Overall similarity between ; Set similarity threshold Compare the overall similarity Does it meet the similarity threshold? The matching spatial coordinates are classified into clusters. Iteratively select spatial coordinates adjacent to the cluster, recalculate the comprehensive similarity, and determine whether clustering is possible until no new spatial coordinates meet the admission criteria and are added to the current cluster. Finally, a cluster containing multiple spatial coordinates is obtained, which is the geometric feature set. Then, select another spatial coordinate that is not in the geometric feature set for iterative calculation until all spatial coordinates in the global coordinate system of the sorting field are clustered, and finally obtain multiple geometric feature sets in the global coordinate system of the sorting field.

5. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 4, characterized in that: The contour fitting unit (220) first divides multiple spatial coordinates in the geometric feature set into a set of planar coordinates according to the planar layering method: set a fixed coordinate axis, traverse all spatial coordinates in the geometric feature set, set selection rules to filter the spatial coordinates on the fixed coordinate axis and group them into the same layer plane, and decompose the spatial coordinates in the geometric feature set into several parallel and non-overlapping planar coordinate subsets. The linkage verification method is used to analyze the contour shape curve corresponding to the planar coordinate subset. That is, the physical acquisition accuracy of the sensor is defined as the minimum spacing threshold, and the planar coordinate subset is fitted and reconstructed based on the minimum spacing threshold and the contour shape curve. The spatial coordinates within the fitted and reconstructed planar coordinate subset are minimized to obtain a target spatial set including multiple spatial coordinates.

6. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 5, characterized in that: The linkage verification method in the contour fitting unit (220) is as follows: Set calibration polynomial coefficients, construct a contour shape model, input the abscissa of each contour coordinate, the contour shape model combines the abscissa and calibration polynomial coefficients to output the corresponding fitting ordinate, and combine the fitting ordinate and the abscissa of the contour coordinate to obtain the fitting coordinate. In the process of constructing the contour shape curve, contour linkage verification is carried out based on cross-plane shape matching relationship. Under the premise of satisfying cross-plane shape adaptation and structural matching constraints, the calibration polynomial coefficients are finely adjusted to the modified polynomial coefficients, and then the fitted coordinates are updated according to the modified polynomial coefficients. Finally, multiple fitted coordinates are connected in sequence to obtain the contour shape curve.

7. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 6, characterized in that: When setting the similarity threshold, the joint clustering segmentation unit (210) uses the minimum difference between the calibration polynomial coefficients and the modified polynomial coefficients in the contour fitting unit (220) as the optimization objective for optimal setting.

8. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 7, characterized in that: After determining the contour shape curve corresponding to the subset of planar coordinates, the contour fitting unit (220) corresponds to the target space set as follows: The minimum spacing threshold is set based on the physical acquisition accuracy of the sensor. The entire contour curve is traversed segment by segment from the starting position. The minimum spacing threshold is strictly used as the fixed sampling interval. The entire contour curve is traversed uniformly and reference feature points are selected. It is ensured that the spatial distance between adjacent reference feature points is not less than the set minimum spacing threshold. At the same time, spatial coordinates in the planar coordinate subset that are not equal to the minimum spacing threshold are removed. The uniform reference feature points obtained by the selection are used as the contour shape skeleton and conform to the shape change law of the original contour curve. On this basis, the planar coordinate subset is regularized and reshaped by interpolation fitting and local shape correction. Multiple spatial coordinates are reconstructed under the premise of meeting the minimum spacing threshold. The point selection and fitting reconstruction of the planar coordinate subset based on the minimum spacing threshold and the contour curve are completed. Finally, the reconstructed spatial coordinates and the unreconstructed spatial coordinates form the target spatial set.

9. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 8, characterized in that: The physical feature completion unit (230) retrieves the physical feature vector of the geometric feature set corresponding to the measured coordinates in the joint clustering segmentation unit (210), and then uses the joint clustering algorithm in the joint clustering segmentation unit (210) to cluster and analyze the physical feature vector again. Based on the clustering convergence result, the overall geometric feature set is adaptively divided into several sub-geometric feature sets with the same geometric attributes and spatial distribution rules. Retrieve the measured coordinates corresponding to each sub-geometric feature set and determine the corresponding spatial range. Specifically, construct an axis-aligned three-dimensional minimum bounding box with the extreme values ​​of each axis as the boundary, and directly define the spatial region of the bounding box as the exclusive spatial range corresponding to the sub-geometric feature set. Match the exclusive spatial range where the virtual coordinates are located, and construct the virtual feature vector corresponding to the virtual coordinates using the physical feature vectors of multiple measured coordinates within the exclusive spatial range.

10. The embodied intelligent robot environmental perception and autonomous decision-making system for logistics sorting according to claim 9, characterized in that: The autonomous decision-making and motion control module (300) retrieves multiple spatial coordinates within the target space set in the contour fitting unit (220), as well as the physical feature vectors corresponding to the spatial coordinates, and analyzes the coordinates of the grasping point of the intelligent robot. Based on spatial coordinates, construct the three-dimensional axis-aligned minimum bounding box of the items to be sorted, calculate the center of gravity of the bounding box as the initial gripping reference point, traverse all spatial coordinates and extract spatial geometric features, then set filtering conditions to select the geometrically optimal point, and determine the spatial geometric features of gripping stability and posture adaptability. Set up multi-level filtering conditions in sequence to select stable grasping areas, match grasping postures, and ensure interference-free grasping; Based on multi-level screening conditions and spatial geometric features, each spatial coordinate is matched and verified one by one, and only spatial coordinates that simultaneously meet all screening conditions are retained and determined as the geometrically optimal points for the robot's stable grasping operation. In addition, an adaptive judgment is carried out by combining the physical feature vectors corresponding to each spatial coordinate in the geometrically optimized points. Among the remaining geometrically optimized points, the spatial coordinates that are closest to the centroid of the bounding box, have the best physical feature uniformity, the most concentrated spatial distribution, and are located within the rigid support area of ​​the target object are selected as the optimal gripping point coordinates.

Images