Isomeric building material adaptive grasping control method and system

Abstract

The application discloses a heterogeneous building material adaptive grabbing control method and system, belongs to the technical field of building construction automation and robot grabbing control, and aims at solving the problem that single grabbing mode cannot take into account the significant differences of building materials such as bricks, steel bars, pipes and boards on a construction site. The application establishes a building material state vector through multi-source perception, instance segmentation and 6D pose estimation, generates a multi-mode grabbing candidate, constructs a building material-grabbing tool-contact graph, combines main animal nature detection, risk interval scoring and constraint model prediction control to output grabbing execution parameters, and realizes the technical effects of improving the grabbing adaptability of heterogeneous building materials and reducing the risks of slipping, pressure loss and overload.

Description

Technical Field

[0001] This invention relates to the field of building construction automation and robot grasping control technology, and in particular to an adaptive grasping control method and system for heterogeneous building materials. Background Technology

[0002] Construction sites typically contain a variety of building materials, including bricks, steel bars, pipes, and boards, which vary significantly in shape, size, weight distribution, surface roughness, rigidity, and fragility. With the development of construction automation and mobile robotics, utilizing robotic arms, grippers, suction devices, magnetic attachments, or combined grippers to handle building materials has become an important way to improve material handling efficiency at construction sites.

[0003] Existing material handling equipment mostly relies on preset building material categories or fixed gripper structures for grasping, typically employing single clamping, lifting, adsorption, or magnetic attraction strategies. While this approach can handle building materials with regular shapes and known parameters, it struggles to promptly identify accessible areas, center of gravity shifts, friction boundaries, and vulnerability risks in construction site environments characterized by stacking, obstruction, random postures, and mixed materials. This can lead to problems such as unstable gripping point selection, excessive gripping force causing pressure damage, and insufficient gripping force causing slippage.

[0004] Meanwhile, traditional grasping control processes often separate visual recognition, candidate grasping generation, and execution control. They do not make sufficient use of the physical property corrections obtained from low-force contact, short-distance trial lifting, and tactile feedback before execution. It is difficult to calibrate stability risks and damage risks before grasping, and it is also difficult to switch modes, reduce speed, or re-grasp control based on risk ranges during execution.

[0005] Therefore, there is a need for an adaptive grasping control method and system for heterogeneous building materials that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose an adaptive gripping control method and system for heterogeneous building materials. Addressing the problems in existing technologies where different building materials exhibit significant variations in shape, size, weight, and material properties, making it difficult to accommodate all materials with a single gripping method, and where insufficient pre-grip calibration leads to difficulties in timely control of slippage, pressure damage, and overload risks, this invention proposes a technical solution that involves multi-source sensing of building material status, construction of a building material-gripper-contact diagram, active property detection to generate calibration property vectors, selection of target gripping candidates based on stable risk intervals and damage risk intervals, and closed-loop execution of control through constraint model prediction. This invention achieves the technical effects of improving the adaptability of heterogeneous building material gripping, reducing damage and slippage risks, and increasing equipment utilization.

[0007] This invention provides an adaptive grasping control method for heterogeneous building materials, comprising: S1, collecting visual data, contact feedback data, and grasper state data of the building materials to be transported; obtaining building material recognition results, shape primitives, pose, accessible area, center of gravity estimation, and initial values ​​of material properties from the visual data; aligning these with the contact feedback data and grasper state data to form a building material state vector; S2, generating a grasping candidate set containing grasping modes, candidate contact points, candidate support points, and grasper posture based on the building material state vector; and constructing a building material-grasp-contact diagram; S3, performing low-force contact and short-distance trial lifting on the grasping candidate set, and collecting data on normal force change, tangential micro-displacement, tactile contact area, and surface texture response. S4. Generate a calibration property vector by measuring the change in the required lifting torque and the amount of the test lifting torque; S5. Write the calibration property vector, pose, center of gravity estimate, and load uncertainty into the building material-gripper-contact diagram to generate a stable risk range and a damage risk range; S6. Calculate the grabbing candidate score based on the stable risk range, damage risk range, collision gap, center of gravity offset, and slip index, and determine the grabbing candidate with the highest score that meets the collision and load boundaries as the target grabbing candidate; S7. Input the target grabbing candidate into the constraint model predictive controller, and output the grabbing mode, target pose, grabbing force curve, velocity curve, and abnormal re-grabbing command. Correct the clamping force, contact angle, and support position based on the tactile signal and torque signal.

[0008] Optionally, S1 includes:

[0009] Image frames and point cloud frames in the visual data are registered according to the sampling timestamp. Instance segmentation is used to obtain building material masks and category labels. Pose estimation is used to obtain the six-degree-of-freedom pose of the building material coordinate system relative to the gripper coordinate system.

[0010] Extract at least one shape element from the building material mask, point cloud boundary and normal distribution, including strip element, block element, tubular element and flat element;

[0011] The accessible area is determined based on the outer dimensions of the shape primitives and the point cloud density, and the centroid estimate is determined based on the point cloud voxel occupancy distribution and the initial values ​​of material properties.

[0012] After the torque sensor output, tactile array output, gripper opening, adsorption pressure, electromagnetic attraction state, and gripper end pose are transformed according to the building material coordinate system, they are combined with the building material recognition results to form a building material state vector.

[0013] Optionally, S2 includes:

[0014] For the shape primitives and initial values ​​of material properties in the building material state vector, select the gripping mode that matches the gripper actuator from the clamping, lifting, embracing, adsorption and magnetic gripping modes;

[0015] Candidate contact points and candidate support points are generated along the boundary points, planar support points, and curved envelope points of the accessible area;

[0016] Building material nodes are used to record shape primitives, dimensions, pose, center of gravity estimates and initial material properties. Gripper nodes are used to record gripper opening range, adsorption pressure range, electromagnetic attraction range, end load range and attitude reach range. Contact nodes are used to record candidate contact points, candidate support points, contact normal, contact area and contact mode.

[0017] Establish contactable edges between building material nodes and contact nodes, and establish execution reachable edges between gripper nodes and contact nodes to obtain the building material-gripper-contact graph;

[0018] Furthermore, the constrained model predictive controller also switches modes and slows down the transport based on the stability risk range and the damage risk range.

[0019] When the upper boundary of the stable risk range exceeds the stable risk threshold and the upper boundary of the damage risk range does not exceed the damage risk threshold, the grabbing mode of the target grabbing candidate is switched to the lifting mode or the embracing mode with an increased number of support points.

[0020] When the upper boundary of the damage risk zone exceeds the damage risk threshold, the upper limit of the handling speed in the speed curve is reduced, and the peak value in the clamping force curve is limited to the upper limit of the damage force determined by the local stiffness and contact area.

[0021] When the upper boundary of the stable risk range and the upper boundary of the damage risk range both exceed their respective thresholds, an abnormal re-grab instruction is output and S2 is returned to regenerate the grab candidate set.

[0022] Optionally, S3 includes:

[0023] The candidate contact point, initial material type, gripper attitude, and low-force contact command are used as active detection inputs.

[0024] During the low-force contact phase, the contact force is limited to the upper limit of the detection force determined by the rated load of the gripper and the initial value of the material properties, and the normal force change, tangential micro-displacement, tactile contact area and surface texture response are collected.

[0025] During the short-distance trial lifting phase, the trial lifting distance is limited to the upper limit of the trial lifting displacement determined by the building material size and collision gap, and the change in trial lifting torque is collected.

[0026] The local stiffness is determined by the ratio of the change in normal force to the amount of contact displacement; the lower limit of the friction coefficient is determined by the ratio of the tangential micro-displacement to the change in normal force; the surface roughness is determined by the tactile contact area and the surface texture response; the weight estimate is corrected by the change in the lifting torque; and the above results are used to form a calibration property vector.

[0027] Furthermore, active detection inputs are also used to trigger risk calibration;

[0028] The risk calibration includes generating a secondary detection command when the lower limit of the friction coefficient is less than the lower limit value of the material table corresponding to the initial material category, or when the local stiffness falls within the vulnerable stiffness boundary corresponding to the initial value of the material properties.

[0029] The secondary detection command reselects a detection point along the accessible area adjacent to the initial candidate contact point, and collects the normal force change, tangential micro-displacement, and tactile contact area again within the upper limit of the detection force.

[0030] The lower limit of the friction coefficient obtained from the secondary detection is compared with the lower limit of the friction coefficient obtained from the primary detection, and the lower limit of the value is taken. The local stiffness obtained from the secondary detection and the local stiffness obtained from the primary detection are weighted by the contact area to obtain the updated value of the calibration property vector.

[0031] Optionally, S4 includes:

[0032] The local stiffness, lower limit of friction coefficient, surface roughness, weight correction and slip index in the calibration property vector are written into the contact node corresponding to the candidate contact point, and the center of gravity estimate and load uncertainty of the building material node are corrected by the weight correction.

[0033] The stable risk interval is calculated based on the lower limit of the friction coefficient, the slip exponent, the estimated value of the center of gravity, and the load uncertainty. The lower boundary of the stable risk interval is determined by the lower limit of the friction coefficient and the slip exponent, and the upper boundary of the stable risk interval is determined by the load uncertainty and the center of gravity offset.

[0034] The damage risk range is calculated based on local stiffness, surface roughness, contact area, and initial values ​​of material properties. The lower boundary of the damage risk range is determined by local stiffness and contact area, and the upper boundary of the damage risk range is determined by surface roughness and initial values ​​of material properties.

[0035] Optionally, S5 includes:

[0036] The stable risk range, damage risk range, collision gap, center of gravity offset, and slip index are normalized to obtain the stable risk component, damage risk component, collision margin component, center of gravity offset component, and slip risk component.

[0037] The weights of each component are read from the preset weight table based on the building material category label and the calibration property vector, and candidate scores are calculated and captured.

[0038] The upper boundaries of the intervals for the stability risk component and the damage risk component are used in the scoring calculation, and the collision margin component is determined by the distance between the candidate capture trajectory and the on-site obstacle point cloud.

[0039] Grab candidates with a collision margin component not less than the collision threshold and a predicted grab load not exceeding the end load range of the gripper are selected as executable candidates. Among the executable candidates, the candidate ranked first in the grab candidate score is selected as the target grab candidate.

[0040] Optionally, S6 includes:

[0041] The constrained model predictive controller takes the grasping mode, target contact node, target support point and gripper attitude in the target grasping candidates as the control targets, and the gripper end load range, collision threshold, upper boundary of the stability risk interval, upper boundary of the damage risk interval, gripping force range and speed range as the constraints.

[0042] In each control cycle, the increments of the gripper end pose, gripping force, support position, and velocity are calculated, and the gripping force curve and velocity curve are generated accordingly.

[0043] During execution, the clamping force is adjusted based on the change in tactile contact area and torque residual, the contact angle is adjusted based on the contact normal deviation, and the support position is adjusted based on the pressure distribution at the support point.

[0044] When the slip index exceeds the slip threshold determined by the lower limit of the friction coefficient and the load uncertainty, or when the torque residual exceeds the torque threshold determined by the change in the lifting torque, an abnormal re-grabbing command is output.

[0045] On the other hand, the present invention also provides an adaptive gripping control system for heterogeneous building materials, including:

[0046] The module for state acquisition and recognition executes S1 and outputs the building material state vector; the module for candidate generation and contact map construction executes S2 and outputs the grasping candidate set and the building material-grabber-contact map; the module for master animal behavior detection executes S3 and outputs the calibration property vector; the module for risk interval generation executes S4 and outputs the stable risk interval and the damage risk interval; the module for candidate scoring executes S5 and outputs the target grasping candidates; and the module for predictive control execution executes S6 and outputs the grasping mode, target pose, grasping force curve, velocity curve, and abnormal re-grabbing command.

[0047] The beneficial effects of this invention are:

[0048] 1. By collecting visual data, contact feedback data, and gripper status data, and forming a building material state vector through instance segmentation, pose estimation, shape primitive extraction, and state alignment, the gripping control process can simultaneously obtain the building material category, shape, posture, accessible area, center of gravity estimate, and initial value of material properties, thereby overcoming the shortcomings of traditional equipment that only relies on a single gripping method or a single recognition result for handling.

[0049] 2. By generating gripping candidates such as clamping, lifting, hugging, adsorption, and magnetic attraction, and constructing a building material-gripper-contact diagram that includes building material nodes, gripper nodes, and contact nodes, the building material shape, gripper capability, candidate contact points, candidate support points, contact normals, and execution reachability can be uniformly expressed. This makes it easier to select a gripping mode that is more suitable for the target building material based on collision gaps, load boundaries, center of gravity offset, and slippage risk.

[0050] 3. By generating calibration property vectors through low-force contact and short-distance trial lifting during the pre-grabbing stage, and further generating stable risk range and damage risk range, the lower limit of friction coefficient, local stiffness, surface roughness and weight correction amount can be calibrated before formal gripping. This enables the constraint model predictive controller to trigger secondary detection, mode switching, deceleration handling or abnormal re-grabbing, reducing the risks of slippage, pressure loss and overload. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 Flowchart of the adaptive grasping control method for heterogeneous building materials;

[0053] Figure 2 This is a flowchart illustrating step S4 of the present invention, which generates the stable risk range and the damage risk range. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0055] refer to Figures 1-2An adaptive grasping control method for heterogeneous building materials includes: S1, collecting visual data, contact feedback data, and gripper state data of the building materials to be transported; obtaining building material recognition results, shape primitives, pose, accessible area, center of gravity estimation, and initial values ​​of material properties from the visual data; aligning these with the contact feedback data and gripper state data to form a building material state vector; S2, generating a grasping candidate set containing grasping mode, candidate contact points, candidate support points, and gripper posture based on the building material state vector; and constructing a building material-grip-contact diagram; S3, performing low-force contact and short-distance trial lifting on the grasping candidate set, and collecting data on normal force change, tangential micro-displacement, tactile contact area, surface texture response, and... S4. Test the change in lifting torque to generate a calibration property vector; S5. Write the calibration property vector, pose, center of gravity estimate, and load uncertainty into the building material-gripper-contact diagram to generate a stable risk range and a damage risk range; S6. Calculate the gripping candidate score based on the stable risk range, damage risk range, collision gap, center of gravity offset, and slip index. Determine the gripping candidate with the highest score that meets the collision and load boundaries as the target gripping candidate; S7. Input the target gripping candidate into the constraint model predictive controller, which outputs the gripping mode, target pose, gripping force curve, velocity curve, and abnormal re-grip command. Correct the clamping force, contact angle, and support position based on the tactile signal and torque signal.

[0056] In this specific embodiment, S1 includes:

[0057] After the handling robot enters the gripping station, the status acquisition and recognition module reads the synchronously triggered RGB-D image frames, three-dimensional point cloud frames, six-axis torque sensor readings, tactile array readings, gripper opening, adsorption pressure, electromagnetic engagement status, and gripper end pose. The RGB-D camera, torque sensor, and tactile array all complete extrinsic parameter calibration in the gripper coordinate system. The calibration records are saved as a sensor calibration table, which includes sensor identification, extrinsic parameter matrix, sampling period, effective time, and calibration residual fields. In this specific embodiment, the image frame sampling period is set to 33ms, and the tactile array and torque sensor sampling period is set to 5ms. Records with extrinsic parameter calibration residuals exceeding 2mm are marked as invalid and stop entering the state vector construction of this period.

[0058] The status acquisition and recognition module registers visual data, contact feedback data, and gripper status data according to the sampling timestamp, and performs visual frame collection within the same handling cycle. tactile frame set and gripper state set use and The registration rules are as follows: , and These are the timestamps for the image frame, haptic frame, and gripper status frame, respectively. The synchronization threshold is set to 20ms in this specific implementation. The missed tactile frames are generated by linear interpolation of two adjacent frames and the interpolation mark is written to the state cache. The building material targets that are missed for two consecutive visual cycles are removed from the recognition results of this cycle.

[0059] The visual recognition unit crops the registered image frames and point cloud frames into a 1.8m x 1.2m workspace in front of the gripper. It then uses a recognition model composed of a Mask R-CNN instance segmentation network and a PointNet-type point cloud classification head to output a building material mask, category labels, and recognition confidence. The training samples for the recognition model come from labeled transport images and point clouds of bricks, pipes, plates, profiles, and fragile panels. The training records store category labels, mask intersection-over-union ratios, and point cloud category labels. In this specific embodiment, when the recognition confidence... When the value is less than 0.65, the system retains the geometric target obtained by point cloud clustering but sets the initial value of the material property to a state of pending calibration to prevent low confidence categories from directly controlling the subsequent grasping mode selection.

[0060] The pose estimation unit establishes a local coordinate system for the building material using the point cloud within the building material mask as input. It then obtains the six-degree-of-freedom pose of the building material coordinate system relative to the gripper coordinate system using point-to-surface ICP registration and principal direction constraints. The pose matrix is ​​then calculated using... Calculate, where Let represent the pose of the building material coordinate system relative to the robot base coordinate system. Let represent the pose of the gripper coordinate system relative to the robot base coordinate system. Write the state vector Field, when ICP residual When the value is greater than 8mm, the system uses the point cloud bounding box spindle pose as the downgraded pose and... Write it as low confidence;

[0061] The shape resolution unit extracts shape primitives based on the building material mask, point cloud boundaries, and normal distribution. The lengths of the three principal axes are first calculated from the voxelized mesh of the point cloud. , and Then according to To determine elongated primitives, we consider that two adjacent principal axes are approximately equal and the cross-section is closed. To determine tubular primitives, we consider that... The flat primitive is determined, and the remaining targets that meet the solid voxel occupancy threshold are determined as block primitives, where... For a 1mm length zero protection quantity, the primitive tag, circumscribed dimension, point cloud voxel occupancy, and normal histogram are written. Field;

[0062] The accessible region extraction unit constructs candidate surface fragments along the boundary points, circumscribed patches, and surface sampling points of the shape primitives. For each surface fragment, it records the fragment identifier, center point, outward normal, radius of curvature, point cloud density, visibility scale, and occlusion marker, and then... and The rule preserves the accessible area, where For the surface segment outward normal, For the approach direction of the gripper, To accommodate the contact angle tolerance, this specific embodiment sets it to 35 degrees. For the density of point clouds in fragments, According to the camera resolution calibration table, segments that do not meet the density requirements are only considered as candidates for support and not as candidates for clamping contact.

[0063] The material property initialization unit reads initial material property values ​​based on category labels, color texture, point cloud reflection intensity, and historical batch records. The lookup keys for the material property table include category labels, surface coating labels, and batch identifiers, while the value fields include initial density. Vulnerability level Material surface friction lower limit Vulnerable stiffness boundary The version number and material property table are established by statistically analyzing compression tests, friction tests, and historical handling logs of specimens from the same batch. When the search key is not matched, the system reads conservative material records by category label and... Write it as pending secondary detection;

[0064] The centroid estimation unit is based on the voxel center. Voxels occupy volume Density in the initial values ​​of material properties Calculate the estimated value of the center of gravity ,in Located in the building materials coordinate system and the unit is mm. The unit is mm^3. From the material property table and maintaining consistent units within the same material zone, void areas Setting it to zero, the result is... Write the difference between the center and the bounding box The field is also provided for S4 to correct load uncertainties;

[0065] The state fusion unit integrates the output of the torque sensor, the output of the tactile array, the gripper opening, the adsorption pressure, the electromagnetic engagement state, and the end effector pose through... Transform to the building material coordinate system to form the building material state vector. The include , , , , , , , , , , , and The field, field version number, and synchronization flag are written together into the state cache as input for S2 to generate the crawl candidate set and construct the building material-grip-contact diagram.

[0066] In this specific embodiment, S2 includes:

[0067] The candidate generation and contact graph construction module reads the building material state vector written by S1. First based on , , A candidate generation context is established with the gripper status field. The context records the building material node identifier, current pose, center of gravity estimate, initial material type, gripper actuator list, end-load range, and attitude reachable range. In this specific embodiment, the actuator list includes grippers, lifting forks, circumferential belts, vacuum suction cups, and electromagnetic attraction units. The actuator capability table stores the gripper opening range, adsorption pressure range, electromagnetic attraction range, end-load range, attitude reachable range, control interface, and version number.

[0068] The pattern selection unit filters the grasping patterns based on the combination of material type and shape element. For long, thin elements, it prioritizes retaining clamping, lifting, and encircling patterns; for flat elements, it retains adsorption, lifting, and clamping patterns; for tubular elements, it retains encircling and clamping patterns; and for magnetic profiles, it retains magnetic attraction patterns. The selection rules adopt... and The predicate in the formula For candidate crawling modes, By material-shape pattern mapping table according to shape primitives and initial values ​​of material properties Read, Indicates the current gripper configuration The actuator constraints that satisfy this mode are retained only when the mapping table is not matched, and the conservative candidate flag is written.

[0069] The candidate point generation unit samples candidate contact points and candidate support points along the boundary points, planar support points, and curved envelope points in the accessible region. Support points are generated for the planar region at a 40mm grid spacing, envelope points are generated along the principal curvature direction for the curved region, and paired clamping points are generated on both sides of the principal axis for elongated primitives. The candidate point record includes... 3D coordinates Contact normal Local curvature Available contact area The system removes entries with occlusion distance and contact patterns where the angle between the normal and the approach direction of the gripper exceeds a certain limit. or contact area is smaller Candidate points, Read from the execution agency capability table using the capture mode;

[0070] The candidate point generation unit simultaneously establishes a contact point index. and support point index The Using building material node identifier, shape element type, and contact mode as search keys, the value fields include candidate point coordinates, normal, area, curvature, occlusion distance, and source region. The system uses building material node identifier, support surface number and support mode as search keys. The value fields include support point coordinates, support normal, load-bearing area and distance to center of gravity projection. The index records are built by the same candidate generation cycle and are accompanied by a version number. When the number of candidates under the same search key exceeds 40, the system performs sorting and pruning according to contact area, occlusion distance and distance to center of gravity projection, and retains the candidates that are ranked first for attitude generation.

[0071] The gripper posture generation unit constructs the gripper posture on the reserved candidate contact points and candidate support points. For the clamping mode, the midpoint of the paired contact points is used as the end target point; for the lifting mode, the support triangle formed by the support point and the center of gravity projection is used as the target point set; for the adsorption mode, the normal and reverse directions of the flat plate element are used as the approach direction of the suction cup; for the embracing mode, the tubular or long strip element main axis is used to generate the embracing closed posture; and for the magnetic suction mode, the center of the magnetic surface segment is used as the contact point. The candidate postures satisfy... Furthermore, the gripper envelope does not intersect with the building material outer enclosure box, in the formula The gripper's attitude rotation matrix is ​​used to write records that satisfy the attitude reachability range into the gripping candidate set. ;

[0072] The building material-grip-contact diagram construction unit consists of building material nodes, gripper nodes, and contact nodes forming a heterogeneous diagram. The building material node records the shape primitive, circumscribed dimensions, pose, estimated center of gravity, and initial material properties. The gripper node records the gripper opening range, adsorption pressure range, electromagnetic attraction range, end-effector load range, and achievable attitude range. The contact node records candidate contact points, candidate support points, contact normals, contact areas, contact modes, and candidate attitudes. The node primary key is... , and The image is generated by splicing and the version number is written. Field;

[0073] The graph edge generation unit establishes reachable edges between material nodes and contact nodes, and establishes execution-reachable edges between gripper nodes and contact nodes. The predicate for generating reachable edges is that the candidate point is located in the reachable region and... Furthermore, the occlusion distance is not less than 5mm, and the predicate for generating reachable edges is that the candidate pose is within the reachable range of the gripper pose and the gripper mode is active. The execution mechanism is valid. The edge fields include edge type, source node, target node, contact mode, normal consistency, contact area, reachability residual and deletion reason. Edges that do not satisfy the predicate are written to the removal log and do not enter the S3 active probe input.

[0074] The candidate set encapsulation unit encapsulates the same building material node, one or more contact nodes, support nodes configured according to the gripping mode, and a gripper posture into a gripping candidate. Each include , , , , , and The field uses contact node deduplication and attitude distance deduplication rules during candidate generation. When two candidates have the same set of contact nodes and an attitude rotation difference of less than 5 degrees, only the candidate with the higher contact area is retained. The encapsulated field... and Save them together to the candidate cache;

[0075] Candidate load estimation unit in packaging Read the initial centroid estimate and initial material property values ​​of S1 at regular intervals, and then... The initial gravitational force is obtained from the formula. The initial density of the material, For the volume occupied by point cloud voxels, It is 9.81 m / s^2. Write in units of N If the set of candidate support points cannot form a support polygon that covers the centroid projection, then... Write down the reasons for insufficient support and retain the reason for deletion;

[0076] The reachability verification unit performs inverse kinematics calculation on the candidate gripper's posture. If the joint angles obtained from the calculation are... satisfy And end attitude residual If the joint angle is less than 3mm, the candidate is marked as reachable. If the joint angle exceeds the limit but there is a candidate with the same pattern in the adjacent contact point index, the system regenerates the attitude using the adjacent contact points. If there is still no reachable solution, the candidate is deleted, and the reason for deletion is written into the execution reachable edge. Field;

[0077] The initial collision check unit reads the on-site obstacle point cloud from the candidate buffer. and candidate close trajectory ,calculate In the formula For trajectory sampling points, For obstacle point cloud points, The unit is mm. In this specific embodiment, when When the value is less than 8mm, the candidate state is written as "to be decelerated for verification". If the gap is less than 3mm, the candidate is deleted, and the pre-collision gap field of the candidate is retained for S5 to calculate the collision margin component;

[0078] When S6 subsequently returns an abnormal re-capture instruction due to simultaneous risk exceeding limits, the candidate generation and contact graph construction modules read the upper boundary of the stable risk interval from the abnormal record. Upper boundary of the damage risk zone Stable risk threshold and damage risk threshold ,when and At that time The system forcibly retains either the lifting mode or the encircling mode that increases the number of support points and blocks clamping candidates that failed in the previous cycle. Write the deceleration and damage force upper limit requirements to the candidate fields when and Delete the previous target grab candidate contact node combination and regenerate along the adjacent contactable area. ,in and The higher the value, the higher the stability risk and the damage risk, respectively. and Read from the risk parameter table, mode switching, speed reduction, and re-capture are all triggered when the corresponding threshold is exceeded;

[0079] The re-fetch candidate update unit writes the abnormal re-fetch records as candidate generation constraints. The fields include the failure candidate identifier, the failure contact node, the risk exceeding the limit type, the disabled crawling mode, the number of support points to be added, and the reduction ratio. When regenerating candidates, the system prohibits the reuse of the same combination of failure contact nodes, and prioritizes the selection from... Select the items whose distance to the center of gravity projection is prioritized and whose load-bearing area meets the requirements. The support points are used to mark the source of the new candidate as re-generated, so that S5 can compare it with ordinary candidates using the same scoring rules;

[0080] The candidate generation and contact graph construction module will retrieve the candidate set. Building Materials - Grippers - Contact Diagram Candidate removal logs, pre-collision gap fields, and re-capture entry states are written to the current period partition of the graph database. The version number of the output object is consistent with the version number of the S1 state vector. S3 reads... The candidate contact points, initial material type, gripper attitude, and low-force contact commands are used as active detection inputs.

[0081] In this specific embodiment, S3 includes:

[0082] The main animal behavior detection module reads and captures the candidate set. Building materials - grippers - contact diagram For each crawling candidate Extract candidate contact points, initial material type, gripper attitude, and low-force contact commands; actively detect input records including... , Contact normal Candidate contact area Initial values ​​of material properties , rated load of gripper and close to the speed limit In this specific embodiment, the low-force contact speed is set to 6 mm / s, the tactile array sampling period is 5 ms, and the torque sensor sampling period is 5 ms.

[0083] During the low-force contact phase, the controller performs a closed-loop approach along the contact normal and limits the contact force to the upper limit of the detection force. Inside, in the formula The proportionality coefficient for the rated load of the gripper is taken as 0.08 in this specific embodiment. The unit is N. This represents the upper limit of the safe contact force corresponding to the initial value of the material properties, and the unit is N. The material conservatism coefficient is taken as 0.6 in this specific embodiment. Simultaneously, detection commands and security monitoring fields are written, and the real-time normal force exceeds [a certain value]. Immediately stop approaching and mark the candidate as having exceeded the detection limit;

[0084] The low-force contact data acquisition unit establishes a contact window after the contact force first exceeds 0.5N. ,exist Internal acquisition of normal force variation Contact displacement tangential micro-displacement tactile contact area and surface texture response ,in The value is obtained by subtracting the initial normal force from the normal force at the end of the window, and the unit is N. and The value is obtained jointly from the end effector encoder and the haptic array slip estimation, and the unit is mm. The result is obtained by accumulating the effective tactile unit surface area, with the unit being mm^2. It is obtained by normalizing the high-frequency response energy of the tactile array through the no-load response;

[0085] The physical property calculation unit calculates the local stiffness based on the ratio of the change in normal force to the amount of contact displacement. In the formula The zero protection amount is 0.05mm displacement and is consistent with Same unit, The unit is N / mm. The system simultaneously records the start and end times of the calculation window and the upper limit of the detection force. Less than and When the value is greater than 1N, the local stiffness is written as the upper limit of stiffness record instead of an infinite value. Write the calibration property vector Field;

[0086] The physical property calculation unit determines the lower limit of the friction coefficient based on the tangential micro-displacement and the change in normal force, starting with the tangential equivalent stiffness of the gripper. Will Converted to tangential force increment , then calculate In the formula The unit is N / mm and is read from the gripper no-load calibration table according to the gripper mode. This table stores the gripper mode, tangential equivalent stiffness, unit, calibration sample, version number, and recalibration trigger condition. The normal force is 0.2N with zero protection. The lower limit is dimensionless and clipped to the range of 0 to 1.5. If the haptic array detects continuous slip events, the system uses the last stable window before the slip. This serves as the conservative friction lower bound for the candidate.

[0087] The surface estimation unit determines the surface roughness index based on the tactile contact area and the surface texture response. In the formula This is the reference contact area read from the actuator capability table in the current grasping mode, and the unit is mm^2. and Assign weights to roughness, and the sum of the two is 1. The value is a dimensionless value between 0 and 1. The larger the value, the stronger the surface texture response or the less effective contact area. Candidates that have not formed a stable contact area are written into the low contact confidence flag.

[0088] During the short-distance trial lifting phase, a trial lifting command is executed in the opposite direction of gravity without leaving the safe working space. The upper limit of the trial lifting displacement is adopted. Determined, in the formula This refers to the smallest external dimension of the building material, expressed in mm. The S2 pre-collision gap is in mm. The unit is mm. In this specific embodiment, an absolute upper limit of 12 mm is also set. The change in lifting torque is collected during the trial lifting process. The trial lifting will be terminated if the collision gap in any direction is less than 3mm.

[0089] The weight correction unit corrects the weight estimate based on the change in the trial lifting torque. Calculate the weight correction amount, where The unit is N·mm. The lever arm length from the candidate contact point to the current estimated center of gravity, in mm. For a 1mm lever arm with zero protection, The unit is N. If a quality correction amount needs to be written, it is done through... Conversion, The force is 9.81 m / s², and the system retains the unit of force. It provides an update for the S4 load uncertainty and is only added to the weight and force terms that are both N in the subsequent load formula, and is not directly added to the kg mass term;

[0090] The slip estimation unit calculates the slip index based on the tangential micro-displacement and the material contact datum. In the formula The reference displacement for the material-gripper combination sliding is in mm, and is read from the sliding calibration table according to material type, gripping mode, and tactile array model. The value is a dimensionless value between 0 and 1, with a larger value indicating a higher risk of slippage at low force levels. This formula normalizes the displacement ratio in mm and does not include torque divided by lever arm conversion. The conversion of torque to N unit force is only done by the previous weight correction unit. Completed. The slip calibration table was established and its version number was saved by statistical analysis of micro-slip test quantiles of the same batch of material specimens under different normal forces.

[0091] The risk calibration unit will actively detect inputs to trigger secondary detections when... Less than the lower limit of the material table corresponding to the initial material category ,or Falling into the vulnerable stiffness boundary corresponding to the initial value of the material properties Within this timeframe, the system generates a secondary detection command. This secondary detection command reselects a detection point within the accessible area adjacent to the initial candidate contact point. The reselection rule is to maximize the distance from the initial contact point within the same accessible area while maintaining a contact normal angle of less than 20 degrees. The secondary detection is still subject to... Constrain and collect again , and ;

[0092] The secondary detection update unit updates the lower limit of the friction coefficient obtained from the secondary detection. Compared with the lower limit of the friction coefficient obtained from the initial detection Take after comparison and the local stiffness obtained from the secondary detection Compared with the local stiffness obtained from the initial detection Obtained by weighted average based on contact area In the formula With a zero protection level for an area of ​​1 mm², no secondary detection command was triggered. and ;

[0093] The main animal behavior detection module will , , , , , , Contact window quality markers, secondary detection markers, and corresponding markers Encapsulated as a calibration property vector The The detection results are written to the cache and used as input for S4 to write the building material-grip-contact diagram and generate the stable risk range and damage risk range.

[0094] In this specific embodiment, S4 includes:

[0095] The risk range generation module receives the calibration property vector calculated by S3 from real-time low-force contact and trial lifting data. Pose and center of gravity estimates and building material-grip-contact diagram ,according to Local stiffness Lower limit of friction coefficient Surface roughness Weight correction amount and slip index Write the corresponding contact node, where The force is expressed in N units, and subsequent load comparisons will remain in N units. If the mass (kg) field is required, then... Conversion and The value is 9.81 m / s^2. The contact node simultaneously stores the upper limit of the detection force, the upper limit of the trial lifting displacement, the secondary detection mark and the calibration version number. The contact node that has not completed active detection is marked as an unscoring node.

[0096] The building material node update unit corrects the center of gravity estimate and load uncertainty of the building material node based on the weight correction amount, first adjusting the initial weight estimate. and Synthesized ,in , and All values ​​are weight forces in N units; mass is only considered when the kg field is required. The conversion is then performed, and the center of gravity offset is updated according to the direction of the candidate contact lever arm. In the formula Zero protection for 1N weight. In this specific implementation, the upper limit ratio for center of gravity correction is set to 0.18. The unit is mm. The updated centroid estimate is the length correction amount in mm. Write to the building materials node, It is a unit vector pointing from the point of contact to the initial centroid;

[0097] The load uncertainty update unit calculates based on the trial lifting moment residual, the material's uncalibrated state, and the secondary detection markers. In the formula The initial load uncertainty is shown in the material property table. Here is the slip uncertainty factor. This is a quality penalty for detection, and is set to 1 when low contact confidence or secondary detection is triggered, otherwise it is set to 0. The values ​​are dimensionless values ​​ranging from 0 to 1, with larger values ​​indicating higher load uncertainty. Write to the building material node Field;

[0098] The stability risk calculation unit calculates the stability risk interval based on the lower limit of the friction coefficient, the slip index, the estimated value of the center of gravity, and the load uncertainty, first forming the lower boundary. Then form the upper boundary. In the formula This is a reference value for stable friction of the material-grip. As the center of the candidate supporting polygon, The maximum external dimension of the building material is in mm, and the weight is... , , and Read from the stable risk weight table and the sum does not exceed 1. and The larger the value, the higher the stability risk.

[0099] The damage risk calculation unit calculates the damage risk range based on local stiffness, surface roughness, contact area, and initial material properties. The upper limit of the damage force is obtained from the formula. The unit is mm^2. This is the calibration value for the material's vulnerable stiffness boundary. For zero protection quantity in the same unit, The unit is N, then calculate. as well as In the formula To maintain the minimum required clamping or adsorption capacity of this candidate, This is the normalized value for the material's vulnerability level. and The larger the value, the higher the risk of damage.

[0100] The risk parameter table is jointly established based on material specimen compression tests, gripper no-load calibration, friction slippage tests, and historical handling logs. Table fields include material type, gripping mode, etc. , , , Stable risk threshold Damage risk threshold The record in this specific embodiment uses the ceramic tile flat element and adsorption mode as the search key to save the friction reference value, vulnerable stiffness boundary, damage force upper limit and risk threshold. The record is updated after every 200 effective handlings or sensor recalibration.

[0101] When the risk parameter table does not match the material category and grabbing mode, the risk interval generation module reads the conservative record of the same shape primitive and... and After each is increased by 0.1 and then trimmed to 1, when there is a detection limit exceeding the limit or a low contact confidence mark at the contact node, the system writes the upper boundary of the corresponding risk interval as a conservative value of not less than 0.8. The branch results are also written into the graph edge attribute, which is used by S5 as a basis for punishment or elimination when scoring and screening executable candidates.

[0102] The risk range generation module will stabilize the risk range. Damage risk range Updated centroid estimate Load uncertainty Upper limit of damage force reference The risk parameter version is written into the building materials-grip-contact diagram. And save the updated graph object as The It serves as a common input for S5 to calculate candidate scores and for S6 to set control constraints.

[0103] In this specific embodiment, S5 includes:

[0104] Candidate scoring module reads risk map and fetch candidate set For each crawling candidate Extract the upper boundary of the stable risk interval Upper boundary of the damage risk zone Candidate capture trajectory Updated centroid estimate Supporting the center of the polygon Slip index And the end-load range of the gripper, scoring records are made. Use it as the primary key and store each normalized component, weight version, executable tag, and sort number;

[0105] The normalization unit converts the stable risk range and the damage risk range into risk components, respectively. and Both of these use the upper boundary of the interval in the scoring, with a larger value indicating higher risk. The system uses this boundary in the scoring formula. and As a safety benefit item, the score is reduced when the risk increases, and the original value of the risk component is still stored in the record for use by the S6 control constraints;

[0106] The collision margin calculation unit reads the obstacle point cloud at the scene. and candidate capture trajectory Calculate the minimum gap In the formula For trajectory sampling points, These are obstacle points, all in mm, and then normalized to... , The collision threshold, In this specific embodiment, the reference gap is unobstructed. Set to 10mm and Set to 80mm. The larger the value, the more ample the collision margin.

[0107] The center of gravity offset component is from Calculate, where The external dimensions of the building material are in mm. A larger value indicates a more significant deviation of the center of gravity from the support center. The slip risk component is calculated using the dimensionless slip exponent, which has already been obtained from the ratio of displacements per unit. Both are included in the scoring formula as scoring penalties when the candidate has no support point and adopts an adsorption or magnetic attraction mode. Find the geometric center of the contact surface and write the support center type into the scoring record;

[0108] The weight reading unit matches a preset weight table and obtains the data based on the building material category label, shape primitive, grasping mode, and the local stiffness and friction coefficient ranges calculated by real-time detection. , , , and The weight table fields include material category, shape primitive, gripping mode, local stiffness range, friction coefficient range, weight vector, version number, and calibration source. The weights are obtained by normalizing the logistic regression coefficients of successful and failed handling records in the quality control samples and satisfy the following conditions: If a match is missed, the conservative weight of the same material category is read and the stable risk weight is increased;

[0109] The scoring unit calculates and captures candidate scores. In the formula The scores are dimensionless, ranging from 0 to 1, with larger values ​​indicating more suitable candidates for target capture. Each component is a normalized quantity ranging from 0 to 1. The stability risk component, damage risk component, center of gravity offset component, and slip risk component are all expressed as a benefit term by subtracting the risk or offset from one. The collision margin component is directly expressed as a positive benefit term. The scoring results are written to... Field;

[0110] The load screening unit predicts the grasping load based on the updated weight estimate and the maximum acceleration estimate during the control phase, and adopts... Calculate, where The weight after S4 update is in N. For quality, Estimate the maximum acceleration for candidate velocity curves, in units of m / s². The dynamic additional load is read from the gripper motion calibration table and is in N. If the load exceeds the upper limit of the gripper end load range, the candidate will be written as a load that cannot be executed.

[0111] The executable candidate filtering unit will and Furthermore, the candidate contact node was not marked as an unscored crawling candidate by S4 and was thus considered an executable candidate. The collision margin threshold is set to 0.25 in this specific implementation. The upper limit of the load range at the end of the gripper is N. When the candidate set is empty, the system outputs an empty target record and returns a request to S2 to regenerate the candidate set. The request carries the rejected mode, the insufficient collision area and the reason for the load exceeding the limit.

[0112] The sorting unit is ordered by executable candidate. Sort by score from largest to smallest. If the difference between two candidate scores is less than 0.02, then the candidate with the collision margin is selected first. Candidates ranked first by numerical value are selected; if they are still the same, the damage risk component is chosen. If the candidates are still ranked first in ascending order, then the candidate with the largest attitude change is selected. The candidate ranked first that also satisfies the collision and load boundary conditions is determined as the target grabbing candidate. The sorting criteria, weight version, and component values ​​are written into the candidate scoring log;

[0113] The candidate scoring module outputs target capture candidates. This includes the grasping mode, target contact point, target support point, grasper attitude, target pose, and scoring. The stable risk range, damage risk range, collision margin, predicted grab load, and slip index are described below. Write to the control input buffer and pass it to the S6 constrained model predictive controller.

[0114] In this specific embodiment, S6 includes:

[0115] Predictive control execution module reads target and captures candidates ,Will The grasping mode, target contact node, target support point, and gripper attitude are used as control targets, and the gripper end-load range, collision threshold, and upper boundary of the stable risk zone are used as control targets. Upper boundary of the damage risk zone Clamping force range, adsorption pressure range, speed range, and upper limit of damage force reference. As a constraint, the control period is set to 20ms in this specific implementation, and the prediction time domain... Set to 15 control cycles;

[0116] Constructing state vectors for constrained model predictive controllers ,in For the end position of the gripper, For the contact angle, For gripping or adsorption equivalent grasping force, For the terminal velocity, The tactile contact area, For torque residual, The slip index is the control quantity. These represent the end-effector pose increment, clamping force increment, support position increment, and velocity increment, respectively. Both the status and control variables have unit fields and are stored in the control cache.

[0117] The controller uses a linearized discrete model. Predicting subsequent states, in the formula and Online updates of contact stiffness obtained from gripper kinematics and low-force detection. The perturbation input matrix is... Including the change in tactile contact area, torque residual and contact normal deviation, the model parameters are determined by the calibration property vector and gripper calibration table for this cycle. When the model parameter version is inconsistent with the S4 risk parameter version, the controller refuses to execute and outputs an abnormal re-grip command.

[0118] The controller solves the objective function in each control cycle. In the formula To predict the error vector between the state and the target contact node, the target support point, and the target pose, and To control the weight matrix, and As a risk penalty weight, The smaller the value, the better the trajectory, gripping force, and risk constraints. The solver uses quadratic programming with box constraints. When the solution timeout exceeds 8ms, the feasible control quantity of the previous control cycle is used and the speed limit is reduced.

[0119] The constraint set includes , , , , and In the formula and Effective ranges include the clamping force range or the adsorption pressure range. It is obtained from the minimum distance between the predicted trajectory and the obstacle point cloud. and To control the risk threshold during the control phase, when constraints are not feasible, the controller freezes the current gripper posture and generates an abnormal re-grip command;

[0120] The control output unit will obtain the solution Expand to include the grasping mode command, target pose, and grasping force curve. and velocity curve The grasping force curve has control cycle, target force, force rise rate and peak limit as fields, and the speed curve has control cycle, end velocity, acceleration and deceleration as fields. The curves are written into the execution queue and sent by the robot controller to the gripper, lifting fork, circumferential belt, suction cup or electromagnetic attraction unit according to the timestamp.

[0121] During execution, the feedback correction unit adjusts the amount of change in tactile contact area. and torque residual Correct the clamping force and adopt In the formula The unit is N / mm^2. The unit is 1 / mm. The target contact area is expressed in mm², and the corrected area is... The unit is N. If the tactile contact area decreases continuously for three control cycles and the torque residual increases, the upper limit of the speed in the next cycle of the speed curve will be reduced by 20%.

[0122] The contact angle correction unit is based on the contact normal deviation. Correct the contact angle and adopt Calculate the angle increment. The values ​​are read from the gripper attitude calibration table and are in rad. The support position correction unit is based on the pressure distribution at the support point. Calculate the support location increment In the formula To support the distribution of pressure for the target, The unit is mm / N, used to convert the support pressure difference into support position increments in mm units without involving torque-to-lever conversion. The angle increment and support position increment are limited to the reachable range of the attitude and the reachable range of the support point before being written into the next control cycle. ;

[0123] The slip and torque anomaly detection unit calculates the slip threshold. and torque threshold In the formula and From S4, and From the slip calibration table, This is the torque residual coefficient. The sensor noise threshold is expressed in N·mm, when the slip index is applied. Exceed or torque residual Exceed When an exception occurs, the controller outputs an abnormal re-capture command and records the trigger source;

[0124] when Exceeding the stable risk threshold and Not exceeding the damage risk threshold When the controller switches the target grabbing candidate's grabbing mode to a lifting mode or a wrapping mode with increased support points, it writes the mode switching record to the candidate regeneration entry. Exceed At this time, the controller reduces the upper limit of the conveying speed in the speed curve to 60% of the original upper limit, and limits the peak value in the gripping force curve to a value determined by local stiffness and contact area. ,when and When all exceed their respective thresholds, the controller outputs an abnormal re-grab command and returns to S2 to regenerate the grab candidate set;

[0125] The predictive control execution module writes the final grasping mode, target pose, grasping force curve, velocity curve, online correction record, risk mode switching record, and abnormal re-grasping instruction into the execution result cache. If there is no abnormal re-grasping instruction, the robot completes the grasping and handling of building materials according to the output curve. If there is an abnormal re-grasping instruction, it executes a safe release or maintains the posture and passes the failure candidate, trigger threshold, contact node, and risk range to the candidate regeneration entry of S2.

[0126] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0127] This invention connects multi-source sensing, building material state vector, building material-grabber-contact diagram, risk interval scoring, and constraint model predictive control into a continuous control chain, enabling building materials of different shapes and materials on the construction site to complete candidate generation, property calibration, risk assessment, and closed-loop execution under the same control framework. This provides a direct technical improvement to the problems of difficulty in coordinating different building material grasping methods, insufficient grasping stability, and insufficient equipment utilization.

[0128] This invention further incorporates active property detection and risk calibration in the pre-grabbing stage. A calibration property vector is obtained through low-force contact, short-distance trial lifting, and tactile feedback. This vector is then written into the building material-grabber-contact diagram to participate in the generation of stable risk and damage risk zones. This allows subsequent control to no longer rely solely on the initial identification results, but to perform secondary detection, mode switching, deceleration handling, or re-grabbing control based on the actual contact state. This better achieves the technical effect of reducing slippage, pressure damage, and overload risks.

Claims

1. A method for adaptive grasping and control of heterogeneous building materials, characterized in that, include: S1. Collect visual data, contact feedback data, and gripper status data of the building materials to be transported. Obtain the building material recognition result, shape primitives, pose, accessible area, center of gravity estimate, and initial material properties from the visual data. Align this data with the contact feedback data and gripper status data to form a building material state vector. S2. Generate a gripping candidate set containing gripping modes, candidate contact points, candidate support points, and gripper posture based on the building material state vector, constructing a building material-gripper-contact diagram. S3. Perform low-force contact and short-distance trial lifting on the gripping candidate set, collecting data on normal force change, tangential micro-displacement, tactile contact area, surface texture response, and trial lifting torque change, generating... S4. Calibrate the material property vector; S5. Write the calibrated material property vector, pose, center of gravity estimate, and load uncertainty into the building material-gripper-contact diagram to generate the stable risk range and damage risk range; S6. Calculate the gripping candidate score based on the stable risk range, damage risk range, collision gap, center of gravity offset, and slip index. Determine the gripping candidate with the highest score that meets the collision and load boundaries as the target gripping candidate; S7. Input the target gripping candidate into the constraint model predictive controller, which outputs the gripping mode, target pose, gripping force curve, velocity curve, and abnormal re-grip command. Correct the clamping force, contact angle, and support position based on the tactile signal and torque signal.

2. The adaptive grasping control method for heterogeneous building materials according to claim 1, characterized in that, S1 includes: Image frames and point cloud frames in the visual data are registered according to the sampling timestamp. Instance segmentation is used to obtain building material masks and category labels. Pose estimation is used to obtain the six-degree-of-freedom pose of the building material coordinate system relative to the gripper coordinate system. At least one shape primitive among strip primitives, block primitives, tubular primitives, and flat primitives is extracted based on the building material mask, point cloud boundary, and normal distribution. The accessible area is determined based on the circumscribed size of the shape primitives and the point cloud density. The centroid estimate is determined based on the point cloud voxel occupancy distribution and the initial value of the material properties. The torque sensor output, tactile array output, gripper opening, adsorption pressure, electromagnetic attraction state, and gripper end pose are transformed according to the building material coordinate system and combined with the building material recognition result to form the building material state vector.

3. The adaptive grasping control method for heterogeneous building materials according to claim 1, characterized in that, S2 include: For the shape primitives and initial material properties in the building material state vector, a gripping mode matching the gripper actuator is selected from the clamping, lifting, embracing, adsorption, and magnetic gripping modes. Candidate contact points and candidate support points are generated along the boundary points, planar support points, and curved envelope points of the accessible area. Building material nodes are used to record shape primitives, dimensions, pose, center of gravity estimates, and initial material properties. Gripper nodes are used to record the gripper opening range, adsorption pressure range, electromagnetic attraction range, end load range, and attitude reachability range. Contact nodes are used to record candidate contact points, candidate support points, contact normals, contact areas, and contact modes. Accessible edges are established between building material nodes and contact nodes, and execution reachable edges are established between gripper nodes and contact nodes, resulting in a building material-gripper-contact graph.

4. The adaptive grasping control method for heterogeneous building materials according to claim 1, characterized in that, S3 includes: using candidate contact points, initial material type, gripper attitude, and low-force contact commands as active detection inputs; limiting the contact force to an upper limit determined by the gripper's rated load and initial material properties during the low-force contact phase, and collecting the normal force change, tangential micro-displacement, tactile contact area, and surface texture response; limiting the test lifting distance to an upper limit determined by the building material size and collision gap during the short-distance test lifting phase, and collecting the test lifting torque change; determining the local stiffness based on the ratio of normal force change to contact displacement, determining the lower limit of the friction coefficient based on the ratio of tangential micro-displacement to normal force change, determining the surface roughness based on the tactile contact area and surface texture response, correcting the weight estimate based on the test lifting torque change, and using the above results to form a calibration property vector.

5. The adaptive grasping control method for heterogeneous building materials according to claim 1, characterized in that, S4 includes: writing the local stiffness, lower limit of friction coefficient, surface roughness, weight correction, and slip index from the calibration property vector into the contact node corresponding to the candidate contact point, and correcting the center of gravity estimate and load uncertainty of the building material node with the weight correction; calculating the stable risk interval based on the lower limit of friction coefficient, slip index, center of gravity estimate, and load uncertainty, wherein the lower boundary of the stable risk interval is determined by the lower limit of friction coefficient and slip index, and the upper boundary of the stable risk interval is determined by the load uncertainty and center of gravity offset; calculating the damage risk interval based on the local stiffness, surface roughness, contact area, and initial values ​​of material properties, wherein the lower boundary of the damage risk interval is determined by the local stiffness and contact area, and the upper boundary of the damage risk interval is determined by the surface roughness and initial values ​​of material properties.

6. The adaptive grasping control method for heterogeneous building materials according to claim 1, characterized in that, S5 includes: normalizing the stable risk interval, damage risk interval, collision gap, center of gravity offset, and slip index to obtain stable risk components, damage risk components, collision margin components, center of gravity offset components, and slip risk components; reading the weights of each component from a preset weight table based on the building material category label and calibration property vector, and calculating the grab candidate score; wherein the upper boundary of the interval of the stable risk component and the damage risk component participates in the score calculation, and the collision margin component is determined by the distance between the candidate grab trajectory and the obstacle point cloud on site; grab candidates with a collision margin component not less than the collision threshold and a predicted grab load not exceeding the load range of the grab's end effector are selected as executable candidates, and the candidate ranked first in the grab candidate score among the executable candidates is selected as the target grab candidate.

7. The adaptive grasping control method for heterogeneous building materials according to claim 6, characterized in that, S6 includes: a constraint model predictive controller that uses the grasping mode, target contact node, target support point, and gripper attitude from the target grasping candidates as control targets, and uses the gripper end load range, collision threshold, upper boundary of the stability risk interval, upper boundary of the damage risk interval, clamping force range, and velocity range as constraints; solves for the gripper end pose increment, clamping force increment, support position increment, and velocity increment in each control cycle, and generates grasping force curves and velocity curves accordingly; corrects the clamping force based on the change in tactile contact area and torque residual during execution, corrects the contact angle based on the contact normal deviation, and corrects the support position based on the support point pressure distribution; outputs an abnormal re-grab command when the slip index exceeds the slip threshold determined by the lower limit of the friction coefficient and the load uncertainty, or when the torque residual exceeds the torque threshold determined by the change in the trial lifting torque.

8. The adaptive grasping control method for heterogeneous building materials according to claim 4, characterized in that, The active detection input is also used to trigger risk calibration; the risk calibration includes: when the lower limit of the friction coefficient is less than the lower limit of the material table corresponding to the initial material category, or when the local stiffness falls within the vulnerable stiffness boundary corresponding to the initial value of the material property, a secondary detection command is generated; the secondary detection command reselects a detection point along the contactable area adjacent to the initial candidate contact point, and collects the normal force change, tangential micro-displacement, and tactile contact area again within the upper limit of the detection force; the lower limit of the friction coefficient obtained by the secondary detection is compared with the lower limit of the friction coefficient obtained by the initial detection, and the lower limit of the value is taken; the local stiffness obtained by the secondary detection and the local stiffness obtained by the initial detection are weighted by the contact area to obtain the updated value of the calibration material property vector.

9. The adaptive grasping control method for heterogeneous building materials according to claim 7, characterized in that, The constraint model predictive controller also switches modes and reduces transport speed based on the stable risk range and the damage risk range. When the upper boundary of the stable risk range exceeds the stable risk threshold and the upper boundary of the damage risk range does not exceed the damage risk threshold, the grasping mode of the target grasping candidate is switched to a lifting mode or a wrapping mode with an increased number of support points. When the upper boundary of the damage risk range exceeds the damage risk threshold, the upper limit of the transport speed in the speed curve is reduced, and the peak value in the clamping force curve is limited to the upper limit of the damage force determined by the local stiffness and contact area. When the upper boundary of the stable risk range and the upper boundary of the damage risk range both exceed their respective thresholds, an abnormal re-grab instruction is output and S2 is returned to regenerate the grab candidate set.

10. A heterogeneous building material adaptive gripping control system, used to execute the heterogeneous building material adaptive gripping control method according to any one of claims 1 to 9, characterized in that, include: The state acquisition and recognition module is used to execute S1 and output the building material state vector; the candidate generation and contact map construction module is used to execute S2 and output the grasping candidate set and the building material-grabber-contact map; the main animal property detection module is used to execute S3 and output the calibration property vector; the risk range generation module is used to execute S4 and output the stable risk range and the damage risk range. The candidate scoring module is used to execute S5 and output target grabbing candidates; the predictive control execution module is used to execute S6 and output grabbing mode, target pose, grabbing force curve, velocity curve and abnormal re-grab command.

Images