A multi-robot arm cooperative pose adjustment and positioning construction method and system suitable for hoisting construction of a composite bridge deck

Abstract

The application provides a multi-mechanical arm cooperative pose adjustment and positioning construction method and system suitable for hoisting construction of superimposed bridge deck panels, and is applied to the technical field of data processing. First, full-dimensional sensing data of superimposed panels, hoisting equipment and the scene are collected, and key parameter settings such as scheduling, calculation, and swing suppression are completed according to requirements such as large load hoisting, high-precision positioning, etc. Through integrated processing of multi-source data, a real-time cooperative link from sensing to positioning is established to obtain equipment and operation state information. Combined with complex working conditions, communication and computing power constraints at the construction site, a cooperative strategy is determined, visual reasoning, pose fusion and other rule configurations are completed, and a control scheme suitable for the scene is formed. At the same time, distributed computing resources are adaptively allocated to ensure safe interaction between nodes. Taking multi-module cooperation as the core, the mapping relationship between working conditions and parameters is learned and optimized, the variable scene environment is adapted, the multi-level safety disposal strategy is matched, and the safe, accurate and stable automatic hoisting and positioning of large-quality bridge deck panels are realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multi-robotic arm collaborative posture adjustment and positioning construction method and system suitable for the hoisting and installation of composite bridge decks. Background Technology

[0002] In industrial scenarios such as industrial vision positioning, robot grasping, and disordered sorting, 6D pose estimation technology is the core support for achieving accurate target object positioning, posture perception, and automated operation. Its performance directly determines the accuracy and level of automation. The training of high-precision 6D pose estimation models relies heavily on large-scale, high-precision, scene-rich, and reliably labeled multimodal datasets.

[0003] Current methods for constructing 6D pose datasets mainly fall into two categories: real-world data acquisition and simulation generation. However, both methods suffer from significant technical limitations and fail to meet the demands of real-world applications. In real industrial scenarios, target objects are diverse, with complex poses and variable lighting, and issues such as occlusion, reflection, and dust interference are common. Manual data acquisition requires setting up complex acquisition environments and repeatedly adjusting object poses, resulting in long acquisition cycles and extremely high costs. Furthermore, 6D pose ground truth annotation relies on specialized equipment and manual calibration, leading to low annotation efficiency, susceptibility to human error, and difficulty in covering all target poses and working conditions. The datasets also suffer from a severe lack of diversity, hindering the training of model generalization capabilities.

[0004] To address the pain point of real data, the industry generally uses simulation software to generate 6D pose data: by using 3D models and physics engines to generate depth maps, mask maps, RGB maps and accurate 6D pose ground truth values ​​in batches, it has the advantages of low cost, high efficiency and absolutely accurate pose labels. However, traditional simulation data suffers from core technical bottlenecks such as the Sim2Real gap, resulting in insufficient visual realism: simulated images are mostly pure object renderings without backgrounds, lacking details such as textures, lighting, reflections, and shadows in real industrial scenes, and the imaging style differs significantly from real-world images; incomplete depth information: traditional simulations only generate depth data for foreground objects, leaving no effective depth data for the background region, which cannot support the training of pose estimation models with dense depth constraints; risk of foreground distortion: existing generative enhancement schemes lack rigid constraints on the foreground, which can easily alter the texture, contour, and pose of objects during optimization, disrupting the correspondence between the image and the true pose value; uncontrollable data quality: the lack of a systematic data quality assessment and adaptive parameter tuning mechanism makes it difficult to consistently meet the standards for realism, consistency, and scene adaptability of generated data; and poor data reusability: existing 3D models and pose data cannot be reused efficiently, making large-scale expansion difficult and hindering the rapid expansion of the dataset.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of this application, a multi-robotic arm collaborative posture adjustment and placement construction method suitable for composite bridge deck hoisting construction is provided, comprising: acquiring full-dimensional perception data of the composite bridge deck, cranes, multiple robotic arms, and construction site; based on the requirements of heavy load hoisting, millimeter-level placement, high-altitude wind disturbance suppression, and real-time collaboration, setting master-slave scheduling cycle, pose calculation times, sway compensation step size, segmentation confidence level, fusion threshold, and operation triggering conditions; integrating visual, pose, wind disturbance, and force data processing to construct a real-time collaborative link of perception-calculation-control-placement, reading pose, end-effector status, crane parameters, environmental interference, and operation status information; determining a collaborative strategy based on complex site conditions, edge computing power limits, low-latency communication, and high-altitude safety regulations, and considering factors such as lighting, wind disturbance, etc. The system sets inference resolution, fusion mode, sway suppression period, action increment, and planning rules based on attitude, space, and communication status to generate a multi-manipulator collaborative posture adjustment and positioning control scheme adapted to complex construction site scenarios. It adaptively allocates distributed computing resources to form batches of vision, calculation, control, and safety management threads, enabling local synchronization within nodes and secure interaction with the industrial LAN via ROS2 across nodes. With crane-multi-manipulator collaboration, lightweight visual segmentation, multi-view posture fusion, adaptive sway suppression, and autonomous positioning planning as the core framework, it learns the mapping relationship between working conditions and parameters, errors and strategies, and wind disturbance and compensation. It optimizes parameters by weighting according to load, geometry, semantics, and kinematic features, adapts to illumination, wind disturbance, scale, and communication status, and matches pause, rollback, reset, alarm, and emergency stop safety strategies.

[0008] Another aspect of this application is a multi-robot collaborative posture adjustment and placement construction system suitable for the hoisting and installation of composite bridge decks, used to execute executable instructions to perform the above-described multi-robot collaborative posture adjustment and placement construction method suitable for the hoisting and installation of composite bridge decks.

[0009] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described multi-robotic arm collaborative posture adjustment and placement construction method suitable for composite bridge deck hoisting construction by executing the executable instructions.

[0010] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described multi-robotic arm collaborative posture adjustment and positioning construction method applicable to the hoisting construction of composite bridge decks.

[0011] The beneficial technical effects of this application are as follows: First, it generates accurate depth maps, mask maps, and simulated RGB and 6D pose raw data based on simulation software; then, it introduces a diffusion model with ControlNet to optimize the background and lighting with the mask as a constraint; subsequently, it completes the background depth to obtain a dense depth map by using the DepthAnything model combined with rigid foreground constraints; next, it quantifies and scores the data in three dimensions—realism, consistency, and applicability—and adaptively tunes the parameters to purify a high-quality standard dataset; finally, through format parsing, batch rendering, and incremental network fine-tuning, it achieves large-scale data expansion and style unification, effectively bridging the gap between simulation and reality, and is suitable for scenarios such as robot grasping and industrial vision positioning.

[0012] This application aims to integrate simulation accuracy with generative realism, addressing the issues of monotonous backgrounds and poor realism in traditional simulation data. It significantly narrows the gap between Sim2Real and traditional methods by introducing rigid foreground constraints to prevent distortions in object texture and pose during generation, ensuring a one-to-one correspondence between pose and image. This results in high data accuracy, complete background depth mapping, meets the training requirements of models with dense depth constraints, and improves model generalization ability. Furthermore, it constructs a 3D quantization evaluation and adaptive parameter tuning mechanism, ensuring stable and controllable data quality, high efficiency in scalability, and reduced data acquisition costs.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0014] Figure 1 This application provides a flowchart illustrating a multi-robotic arm collaborative posture adjustment and positioning construction method for the hoisting and installation of composite bridge decks, according to an embodiment of this application.

[0015] Figure 2 This is a hybrid collaborative control architecture for cranes and multiple robotic arms;

[0016] Figure 3 This is a visual segmentation effect diagram of a composite panel;

[0017] Figure 4 A diagram illustrating the pose estimation effect of a composite plate.

[0018] Figure 5 This is a result image of autonomous identification of landing sites;

[0019] Figure 6This illustration shows a structural schematic diagram of a multi-robotic arm collaborative posture adjustment and positioning construction system for the hoisting and installation of composite bridge decks, provided in an embodiment of this application. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 2 As shown, Figure 2 This invention presents a crane-multi-robotic arm hybrid collaborative control architecture, which is the core support for the precise lifting and placement of large-mass composite slabs. In the civil engineering industry, composite slabs often weigh over 10 tons. Automating installation with robotic arms would require controlling at least 10 robotic arms working together, which is extremely difficult. Current crane lifting methods suffer from insufficient installation accuracy and uncontrollable posture. To address this, this invention proposes a crane-load-bearing + multi-robotic arm fine-tuning hybrid collaborative control mode, using a configuration of 1 crane + 4 robotic arms to overcome the dual bottlenecks of heavy load lifting and high-precision posture adjustment. The system adopts a master-slave distributed hybrid control architecture, with robotic arm 0 as the global master node and the other three robotic arms as slave nodes. Each robotic arm is equipped with an independent edge computing chip, forming a distributed collaborative computing terminal. Each node establishes a low-latency, high-reliability communication link with the industrial LAN through the ROS2 system, enabling unified issuance, real-time feedback, and synchronous execution of crane movement commands and multi-robotic arm action commands, ensuring reasonable load distribution and consistent action coordination throughout the lifting process. The master node coordinates the overall scheduling, responsible for the timing coordination of the crane and multiple robotic arms, pose data fusion, force balance control, and global anomaly handling. Slave nodes perform local visual data acquisition, local pose calculation, end-effector control, and attitude fine-tuning, working in real-time with the crane to stably lift large-mass composite slabs. This architecture fully leverages the complementary advantages of the crane's heavy-duty load-bearing capacity and the robotic arms' high-precision attitude control capabilities, achieving optimal coordination with a minimum number of robotic arms. It meets the safe lifting requirements for large-mass composite bridge decks weighing over 10 tons, while controlling lifting positioning errors and attitude deviations to the millimeter level, significantly improving the placement accuracy of the composite slabs and the level of construction automation.

[0022] In one implementation, Figure 1 The diagram illustrates a process flow of a multi-robotic arm collaborative posture adjustment and positioning construction method for the hoisting and installation of composite bridge decks, according to an embodiment of this application.

[0023] S101 acquires comprehensive perception data of composite bridge decks, cranes, multiple robotic arms, and the construction site.

[0024] In one implementation, the state perception of the composite bridge deck relies on the vision acquisition units mounted on each robotic arm terminal to collect information. Using images and depth data actually captured at the construction site as input, it continuously collects information on the composite bridge deck's appearance contour, spatial posture, edge coordinates, relative distance, and placement angle. The acquisition process can adapt to different lighting conditions, dust, and partial occlusion. It outputs the panel's real-time three-dimensional pose, edge positioning coordinates, horizontal tilt angle, and vertical deflection angle, providing fundamental data support for the robotic arm's adaptive gripping and positioning, posture fine-tuning, and stable clamping.

[0025] Crane operation status perception is achieved through the crane's built-in sensing unit and position detection device, which collects real-time information on hook spatial position, lifting height, running speed, braking status, load weight, and motion trend. The collected data covers hook horizontal coordinates, vertical height, running direction, real-time load and speed change rate, and outputs real-time hook coordinates, load values, motion state quantities, and stability level, which are used for time-series coordinated scheduling, force balancing distribution, and motion synchronization control between the crane and multiple robotic arms.

[0026] The multi-robotic arm status perception system collects operational information from each of the four robotic arms, acquiring in real-time information on joint rotation angles, end effector spatial positions, movement speeds, force magnitudes, load ratios, and reachable workspaces for each arm. It outputs real-time poses, end effector states, force constraints, load limits, and motion interference boundaries for each robotic arm, providing accurate equipment constraint data for the master node to coordinate master-slave collaboration logic, command distribution, action synchronization, and path obstacle avoidance.

[0027] The construction site environmental perception system comprehensively collects data in the complex scenarios of bridge construction, acquiring real-time information on ambient light intensity, wind speed, wind direction, dust concentration, location of spatial obstacles, boundaries of installed components, work area range, and communication signal strength. This environmental information is then converted into quantifiable levels, outputting environmental interference levels, work space safety boundaries, communication link stability, and operational condition adaptability levels. This data is used to improve the robustness of visual perception, mitigate collision risks, and ensure reliable communication and operational safety.

[0028] The four types of sensing data mentioned above are input into the master-slave collaborative control architecture according to a unified time base. The master node performs global timing alignment to eliminate time deviations caused by different sensors and terminals. Visual data, pose data, force data, and environmental data are then fused at the feature level to form a full-dimensional sensing dataset encompassing component status, equipment status, and environmental status. This dataset serves as a unified input, directly supporting the entire process of lightweight visual segmentation, multi-view pose calculation, high-altitude wind disturbance suppression, and autonomous placement planning, ensuring the system can stably, accurately, and safely complete hoisting operations under complex construction site conditions.

[0029] S102, based on the requirements of heavy load hoisting, millimeter-level positioning, high-altitude wind disturbance suppression, and real-time collaboration, sets the master-slave scheduling cycle, pose calculation times, sway compensation step size, segmentation confidence, fusion threshold, and operation triggering conditions.

[0030] In one implementation, the hoisting construction accuracy, coordination, disturbance rejection, and process indicators are broken down dimensionally, and features such as collaborative control, posture perception, wind disturbance suppression, and segmented management are extracted to form a subset of core control parameters. The accuracy, coordination, disturbance rejection, and process management indicators during the hoisting construction process are further broken down layer by layer to form a clear indicator system. Accuracy indicators mainly cover millimeter-level positioning control and composite slab attitude adjustment control to ensure that the final installation accuracy meets construction specifications. Coordination indicators focus on the unified timing of actions between the crane and multiple robotic arms, synchronized execution of commands, and balanced force distribution to ensure overall coordinated operation. Disturbance rejection indicators address gusts and environmental interference at high-altitude construction sites, specifying standards for sway suppression and attitude stability.

[0031] The process control indicators follow a standardized execution logic of six stages: initialization, locking, capture, stabilization, descent, and placement, ensuring the orderly progress of the entire process. Based on the above four types of indicators, key features required for system operation are further extracted. Cooperative control features are extracted from the collaborative scheduling and instruction distribution stage; pose perception features are extracted from the visual recognition and pose calculation stage; wind disturbance suppression features are extracted from the wind disturbance compensation and attitude stabilization stage; and segmented control features are extracted from the stage switching and condition judgment stage. These four types of features are integrated to form a subset of core control parameters covering the entire process of perception, calculation, regulation, and execution.

[0032] A crane-multi-manipulator master-slave collaborative control model is constructed, with the master-slave scheduling cycle as the core parameter, and the collaborative scheduling parameters are optimized in combination with the requirements of heavy load lifting. A master-slave collaborative control model is constructed with the zero manipulator as the global master node and the other three manipulators as slave nodes, forming a hybrid collaborative control architecture that combines crane load-bearing capacity with multi-manipulator fine-tuning. This model has three logical layers from top to bottom: a global scheduling layer, a command distribution layer, and an action execution layer. The global scheduling layer directly interacts with the command distribution layer, and the command distribution layer directly transmits commands to the action execution layer. Real-time communication and status feedback are maintained between each layer.

[0033] The global scheduling layer, using the edge computing unit mounted on the zero-arm robotic arm as its carrier, is responsible for coordinating the overall hoisting task, receiving multi-dimensional perception data, and making global decisions. The instruction distribution layer, based on the control logic output by the global scheduling layer, parses the unified scheduling instructions into executable instructions for different devices and time periods, and distributes them to the crane controller and each slave node robotic arm controller. The action execution layer consists of three slave node robotic arms and the crane actuator, which perform physical actions such as attitude adjustment, smooth hoisting, coordinated clamping, and precise posture adjustment according to the received instructions, and transmits the operating status and force data back to the instruction distribution layer and the global scheduling layer in real time.

[0034] Using the master-slave scheduling cycle as the core control parameter, and combining input data such as the constraints of lifting heavy loads of 10 tons or more at the construction site, the rated load limit of each robotic arm, and the safe operating speed range of the crane, a time-series collaborative algorithm is used to optimize the configuration of collaborative scheduling parameters. The scheduling rules follow the principle of unified time synchronization by the master node and synchronous response by each slave node. For example, the master node outputs a global control command once every scheduling cycle, and the slave nodes execute the command immediately upon receiving it, ensuring that the crane's lifting and lateral movements are strictly synchronized with the robotic arm's posture adjustment and clamping actions.

[0035] After the timing-coordinated algorithm is used, the system outputs the global scheduling command of the master node, the execution timing of the actions of each slave node, and the balanced distribution of the crane load and the mechanical arm force. This ensures that the system maintains coordinated actions, consistent posture, and reasonable force distribution between the crane and multiple mechanical arms while meeting the requirements for safe lifting of large loads. Ultimately, this achieves the unification of large load lifting capacity and millimeter-level high-precision coordinated control.

[0036] A joint model for visual segmentation, pose fusion, and adaptive sway suppression was constructed, setting segmentation confidence, fusion threshold, pose calculation iterations, and sway suppression compensation step size to meet accuracy and disturbance rejection requirements. A three-level linkage joint model of visual segmentation, pose fusion, and adaptive sway suppression was built to achieve accurate identification, stable attitude control, and smooth hoisting of composite bridge deck panels. This model adopts a sequential serial structure, with the output of the visual segmentation module directly connected to the pose fusion module as input, and the output of the pose fusion module directly connected to the adaptive sway suppression module as input, enabling real-time data transfer and seamless integration between modules.

[0037] The visual segmentation module, based on a lightweight YOLO-Seg network, is deployed on the edge computing terminals of each robotic arm. The module includes segmentation confidence parameters to filter out low-confidence recognition results, retaining only target regions that meet the confidence requirements. The module takes as input RGB and depth images acquired in real-time from the construction site. After feature extraction and target segmentation, it outputs the complete outline of the composite slab, edge coordinates, center point position, and slab surface range information, providing accurate target localization data for subsequent pose calculations.

[0038] The pose fusion module adopts a multi-view adaptive fusion architecture, featuring both front-fusion and back-fusion modes that autonomously switch between each other with a 10Hz threshold. The module has parameters for setting the fusion threshold and the number of pose calculations to control the data fusion intensity and calculation accuracy. The module receives multi-view perception data collected from multiple robotic arms and target contour information output from the visual segmentation module. It then performs calculations using a two-layer neural network of estimation and optimization, outputting complete spatial pose data for the composite plate, including its 3D spatial coordinates, horizontal tilt angle, and deflection angle.

[0039] The adaptive sway suppression module employs an integrated algorithm for wind disturbance response and attitude compensation. The module incorporates a sway compensation step size parameter to control the amplitude and stability of each adjustment. Inputs include real-time attitude data from the pose fusion module, on-site wind disturbance intensity data, and composite plate sway data. After calculating the sway frequency and amplitude, the module outputs a fine-tuning command along the reverse sway direction, controlling the robotic arm to perform small-amplitude, high-frequency attitude corrections.

[0040] The three modules work in sequence and collaboratively. The visual segmentation module first completes target localization, the pose fusion module completes accurate pose calculation based on the localization results, and the adaptive sway stabilization module completes real-time anti-disturbance pose stabilization based on the pose data. Through this three-level linkage mechanism, the system maintains high-precision target localization, stable attitude control, and reliable operation under complex construction site conditions such as changes in lighting, dust interference, and gusts of wind, meeting the positioning accuracy and anti-interference requirements for hoisting large-mass composite slabs.

[0041] A segmented operation control mechanism is established, determining operation trigger conditions based on equipment status and safety constraints, forming a complete collaborative control parameter setting scheme. This mechanism comprises six stages: initialization, locking, grabbing, stabilization, descent, and placement. The entire hoisting operation is divided into continuous and independent standardized processes, achieving fully automated progress without manual intervention. Based on the real-time operating status of the robotic arm, the stability of the stacked plate's posture, on-site space safety constraints, communication link status, and load stress conditions, the system sets clear operation trigger conditions for each stage. The conditions for each stage are independent, progressive, and sequentially connected. Once all conditions for the current stage are met, the system automatically proceeds to the next stage; otherwise, it maintains the current state and issues a safety warning.

[0042] The initialization phase is triggered when all robotic arms reach their zero position or a specified initial position, the communication link is detected as normal, the vision model is loaded, the sensors are calibrated, and there are no fault alarms from any devices. Upon entering this phase, the system completes self-checks and preparations, establishing a baseline for subsequent operations. The locking phase is triggered when initialization is complete, there is no interference within the range of motion of each robotic arm, the end effector points towards the stacked plate, and the vision segmentation successfully identifies the target. Upon entering this phase, the robotic arms maintain servo tracking, waiting for the opportunity to grasp.

[0043] The grabbing phase is triggered by the following conditions: locking in place, stable posture of the laminated plate, sufficient distance between the robotic arm and the plate, and no risk of collision. Upon entering this phase, the robotic arm synchronously approaches and smoothly grips the plate. The stabilization phase is triggered by the following conditions: grabbing completed, gripping force sufficient, and no significant swaying of the plate. Upon entering this phase, the system activates sway control to adjust the laminated plate to a stable posture. The descent phase is triggered by the following conditions: stable plate posture, safe lifting path, and identified placement area. Upon entering this phase, the crane and robotic arm descend in coordination, approaching the installation position. The placement phase is triggered by the following conditions: the plate descends to a predetermined height, the robotic arm completes its posture transition, and the posture error meets millimeter-level requirements. Upon entering this phase, precise placement and positioning are completed.

[0044] The system integrates collaborative control parameters, master-slave scheduling parameters, vision and pose model setting parameters, operation triggering parameters for each stage, and safety constraint parameters into a unified collaborative control parameter setting scheme covering the entire hoisting process, providing a unified and complete parameter basis for the automated, standardized, and safe operation of the overall system.

[0045] The S103 integrates visual, pose, wind disturbance, and force data processing to construct a real-time collaborative link of perception-calculation-control-positioning, and reads pose, end-point status, crane parameters, environmental interference, and operational status information.

[0046] In one implementation, based on the real-time nature and master-slave collaboration rules of multi-source data (visual, pose, wind disturbance, and force), the master control node and edge computing terminals collaboratively determine the data fusion and link construction strategy. Distributed acquisition and global temporal alignment determine the data processing format and collaborative control dimensions. Based on the real-time transmission characteristics of visual, pose, wind disturbance, and force data, and the master-slave collaboration rules of the crane and multiple robotic arms, the master control node centrally coordinates global scheduling, while each edge computing terminal synchronously collaborates to determine the multi-source data fusion method and the real-time collaborative link construction scheme. The master control node, with robotic arm number zero as its core carrier, is responsible for receiving and integrating data uploaded from all terminals. Edge computing terminals are deployed on the other three robotic arms, responsible for local acquisition, local computation, and local preprocessing, forming a master-slave distributed data processing architecture.

[0047] The system acquires independent sensing information from each terminal through a distributed acquisition method. Each edge terminal independently acquires image information, pose information, force information, and environmental information from its corresponding viewpoint, operating synchronously without interference. To ensure effective fusion of multi-source data, the system employs a global temporal alignment method, uniformly calibrating the timestamps of all sensor and terminal data to eliminate temporal deviations caused by transmission delays and different acquisition frequencies. For example, data from visual acquisition, pose calculation, wind disturbance detection, and force monitoring are uniformly aligned to the same time reference, ensuring that the data used for each frame of control commands maintains temporal consistency.

[0048] Building upon this foundation, the system further clarifies data processing format standards, uniformly converting image data, coordinate data, angle data, and force control data into a standard format that can be directly computed. Simultaneously, it defines the coverage dimensions of collaborative control, including spatial location, attitude adjustment, force balance, disturbance resistance stability, and safety constraints. These steps provide a unified, standardized, and reliable basis for subsequent integrated processing of multi-source data, real-time collaborative control, and closed-loop control throughout the entire process.

[0049] A four-stage linkage mechanism—sensing, calculation, control, and placement—is employed to integrate multi-source field data. First, each edge terminal completes local data acquisition and calculation. Then, the master node performs global data fusion, linking millimeter-level placement, wind disturbance suppression, and stress balance requirements to generate standardized control data. The sensing stage is executed independently by the edge computing terminals mounted on each robotic arm, simultaneously acquiring local visual images, multi-view pose data, field wind disturbance parameters, and equipment stress status. Each terminal operates independently, collecting multi-dimensional information about the target component, operating equipment, and construction environment in parallel, providing a raw and complete field data source for subsequent processing.

[0050] The calculation phase is still performed independently by each edge terminal. Based on locally collected data, a lightweight visual segmentation model is run to obtain target contour information, and a pose estimation algorithm is used to complete local pose calculation. Real-time wind disturbance intensity is identified by combining wind disturbance data, and preliminary force equilibrium analysis is performed based on force information. All calculation operations are completed at the edge, reducing transmission pressure and improving response speed.

[0051] The fusion phase is executed uniformly by the main control node. The main node receives the local calculation results uploaded by all edge terminals and performs global normalization and feature-level fusion on multi-view pose data, multi-node force data, and global wind disturbance data. During the fusion process, millimeter-level positioning accuracy, effective suppression of high-altitude wind disturbance, and balanced distribution of overall force are the core constraints. Data deviations are automatically calibrated, and data formats and time series benchmarks are unified.

[0052] After four stages of collaborative processing, the system outputs standardized control data with a unified format, consistent timing, and high accuracy. This data can be directly used for subsequent attitude adjustment, collaborative scheduling, sway control, and landing planning, providing a reliable instruction basis for fully automated operations.

[0053] The system performs real-time status verification on standardized control data, initiates adaptive correction processing for data fluctuations caused by complex on-site conditions, triggers a closed-loop compensation mechanism for pose calculation errors, and executes an early warning judgment process for abnormal operating states, generating optimized control information that includes data fusion, compensation correction, and status judgment. For data fluctuations caused by complex on-site conditions such as sudden changes in lighting, dust interference, and gusts of wind, the system automatically initiates adaptive correction processing. By judging the amplitude and frequency of fluctuations in real time, the system smooths, calibrates, and fits trends for visual data, pose data, and wind disturbance data that exceed the stable range, ensuring continuous and stable data output and preventing abnormal movements of the robotic arm and crane due to data jumps. For example, when a sudden increase in lighting causes a shift in the visual segmentation result, the system corrects the current result based on historical stable data to maintain continuous and reliable target recognition.

[0054] To address deviations in pose calculation, the system triggers a closed-loop compensation mechanism. The calculation results are compared with the standard placement posture and millimeter-level accuracy requirements. When the position error exceeds 1 millimeter or the posture deviation exceeds the allowable range, a compensation amount is immediately generated to correct the pose data in real time. For example, when a shift occurs in the multi-view pose fusion result, the system adjusts the calculation parameters in reverse and compensates for the posture angle, bringing the composite plate pose back to the target range.

[0055] For abnormal operating states of the robotic arm, such as overload, motion interference, communication interruption, and abnormal force, the system executes a graded early warning judgment process. The abnormalities are categorized into three levels: alert, warning, and fault, with status information recorded and corresponding handling logic triggered for each level. For example, when the force on a robotic arm joint approaches its rated limit, the system determines it to be in a warning state, promptly adjusting the collaborative load distribution to prevent equipment damage.

[0056] Through the combined operation of three processes—data fusion optimization, fluctuation compensation correction, and state judgment verification—the system eliminates invalid data, removes error fluctuations, and identifies operational risks, ultimately generating optimized control information with high stability and reliability. This provides a safe and accurate command basis for subsequent coordinated actions, attitude adjustments, and smooth landing.

[0057] The optimized control information is executed in a closed loop, achieving full-process linkage through a real-time collaborative link. This involves synchronously reading the composite plate pose, robotic arm end-effector status, crane operating parameters, environmental interference intensity, and global operational status information to generate complete feedback data supporting precise positioning. The optimized control information is input into the real-time collaborative link built by the system, executing closed-loop control according to master-slave distributed control logic to achieve full-process linkage operation between the crane and multiple robotic arms. Control information is distributed to each execution unit via the ROS2 communication framework and industrial LAN, ensuring strict synchronization between master node commands and slave node actions, maintaining consistent timing, timely response, and stable operation throughout the entire process.

[0058] During closed-loop execution, the system synchronously reads multiple key status information, including the real-time spatial pose and tilt information of the composite bridge deck, the clamping status and running position information of each robotic arm end effector, the hook height and translation speed information of the crane, the environmental interference intensity information such as wind speed and sunlight at the construction site, and the current global operation stage status such as initialization, locking, grasping, stabilization, descent, and placement.

[0059] The system integrates and aligns the aforementioned multi-dimensional status information in a unified manner to form complete feedback data, which is then transmitted back to the main control node. This feedback data is directly used for adaptive parameter adjustment of the visual segmentation module, online calibration of the pose fusion solution model, real-time optimization of wind disturbance suppression strategies, and dynamic correction of the placement path planning. For example, when feedback data shows a large pose error, the system automatically increases the number of pose calculations and improves fusion accuracy; when increased wind disturbance is detected, the suppression cycle is immediately shortened and the compensation step size is increased; when the operation enters the placement phase, the robotic arm posture is switched in a timely manner and the placement planning path is adjusted. Relying on the above complete feedback and real-time adjustment mechanism, the accurate, stable, and safe operation of the entire hoisting process is effectively guaranteed.

[0060] S104 determines the collaborative strategy by combining the complex working conditions of the construction site, the upper limit of edge computing power, low latency communication and high-altitude safety regulations. Based on the lighting, wind disturbance, attitude, space and communication status, it sets the inference resolution, fusion mode, sway suppression period, action increment and planning rules to generate a multi-robotic arm collaborative posture adjustment and positioning control scheme adapted to the complex construction site scenario.

[0061] In one implementation, the collaborative constraints in complex construction site conditions, edge computing power limits, low-latency communication, and high-altitude safety regulations are adapted, analyzed, and transformed into scenarios to generate master-slave collaborative scheduling characteristics, edge computing power allocation characteristics, real-time communication guarantee characteristics, and high-altitude operation safety characteristics, forming the basic information for collaborative strategy formulation. Adaptation, analysis, and scenario transformation are performed on the four types of collaborative constraints—complex construction site conditions, edge computing power limits, low-latency communication, and high-altitude safety regulations—and combined with the actual operational logic of composite bridge deck hoisting operations, abstract constraints are transformed into executable system control characteristics, providing a basis for collaborative strategy formulation.

[0062] To address the need for collaborative operations between multiple robotic arms and cranes, master-slave collaborative scheduling features are extracted. Using robotic arm number 0 as the master node and the other three robotic arms as slave nodes, a master-slave distributed scheduling system is constructed. The master node is responsible for issuing global commands, coordinating timing, and handling anomalies, while the slave nodes are responsible for local sensing, computation, and action execution. For example, the master node issues collaborative commands according to a fixed scheduling cycle, and the slave nodes respond immediately upon receiving the commands, ensuring strict synchronization between crane operation and robotic arm posture adjustment.

[0063] By considering the hardware performance limitations of edge computing terminals, edge computing power allocation characteristics are extracted. Based on the terminal's processor, memory, and power consumption limits, computing power allocation ratios are divided for four types of tasks: visual perception, pose calculation, collaborative control, and security management. For example, priority is given to ensuring computing power supply for visual segmentation and pose calculation, with the remaining computing power allocated to collaborative control and security management, ensuring the real-time operation of core functions under limited computing power conditions.

[0064] Based on the communication requirements of industrial LANs and ROS2, real-time communication assurance features are extracted. A data transmission latency upper limit and synchronization response cycle are set to establish a highly reliable communication link between master and slave nodes. For example, the command transmission latency is controlled to the millisecond level, and the synchronization cycle is kept consistent with the master-slave scheduling cycle to avoid asynchronous actions due to communication lag.

[0065] In accordance with the safety management requirements for high-altitude hoisting operations, the safety characteristics of high-altitude operations are extracted. Safety boundaries such as the robotic arm's range of motion, hook lifting speed, load limit, and collision avoidance distance are clearly defined. At the same time, areas where personnel are prohibited from approaching and emergency escape spaces for equipment are demarcated. For example, the robotic arm automatically decelerates when approaching the work boundary, and stops operating when the load exceeds the limit, comprehensively ensuring the safety of high-altitude operations.

[0066] By integrating master-slave collaborative scheduling characteristics, edge computing power allocation characteristics, real-time communication guarantee characteristics, and high-altitude operation safety characteristics, a foundational information for collaborative strategy formulation covering scheduling, computing power, communication, and security is formed, providing a unified and reliable basis for the generation of subsequent collaborative control schemes.

[0067] The system performs logical rule transformation and threshold boundary definition on dynamic changes in on-site factors such as lighting, wind disturbance, attitude, space, and communication status to generate visual inference environment parameters, wind disturbance suppression constraint parameters, pose control limitation parameters, workspace avoidance parameters, and communication link stability parameters, forming on-site state adaptation constraint information. Visual inference environment parameters are generated based on lighting intensity and dust variation conditions. The system automatically switches image acquisition gain, exposure parameters, and recognition thresholds according to lighting intensity and dust concentration to adjust the visual inference working mode. For example, a high-precision recognition mode is used when lighting is sufficient, while an enhancement mode is automatically activated when lighting is insufficient or dust levels are high to ensure stable output of target contour and position information by the YOLO-Seg segmentation model.

[0068] Based on real-time wind speed and sway amplitude, wind disturbance sway suppression constraint parameters are generated. The response speed, adjustment step size, and compensation magnitude of the sway suppression action are limited according to the wind speed level and the real-time sway amplitude of the composite slab. For example, a small step size is used for smooth sway suppression when the wind speed is low, while the response frequency and compensation amount are increased when the wind speed is high to ensure rapid attitude stabilization under different wind disturbances.

[0069] Based on the attitude and tilt angle of the composite plate, position and posture control limit parameters are generated. Allowable adjustment ranges for the horizontal tilt angle and deflection angle are set as constraint boundaries for position and posture correction and control. For example, when the tilt angle of the composite plate exceeds the allowable range, the system immediately initiates attitude correction control to ensure safe and controllable attitude during hoisting and placement.

[0070] Based on the distribution of the work area and obstacles, work space avoidance parameters are generated. Based on the locations of installed components, obstacles, and steel beams at the construction site, the safe movement zone, minimum avoidance distance, and reachable space boundaries for the robotic arm and crane are defined. For example, when the robotic arm approaches the boundary or enters a danger zone, it automatically decelerates and restricts its direction of movement to avoid collisions.

[0071] Based on the communication signal strength and packet loss rate, communication link stability parameters are generated. Signal strength thresholds and packet loss rate upper limits are set as criteria for determining data transmission interruptions and anomalies. For example, when the communication packet loss rate exceeds the threshold or the signal strength is too low, the system determines the link is unstable and automatically switches to a local secure operation mode.

[0072] By integrating five categories of parameters—visual reasoning environment parameters, wind disturbance and sway suppression constraint parameters, pose control limitation parameters, workspace avoidance parameters, and communication link stability parameters—a unified set of field status adaptation constraint information covering environment, equipment, space, and communication is formed, providing a complete and reliable field constraint basis for subsequent collaborative control strategy generation and parameter calibration.

[0073] For the objective function of crane-multi-robotic arm collaborative control, parameter calibration and collaborative accuracy evaluation are performed in conjunction with on-site state constraints. This generates visual inference resolution parameters, pose fusion mode parameters, wind disturbance suppression period parameters, end effector motion increment parameters, and placement path planning parameters, forming collaborative control rule optimization information. Calibration is carried out based on visual inference environment parameters to generate visual inference resolution parameters. The inference resolution is dynamically matched by considering the upper limit of the edge computing terminal's computing power and on-site lighting and dust conditions, controlling the computational load while ensuring recognition accuracy. For example, high-resolution inference is used when there is sufficient light and computing power to improve segmentation accuracy; the resolution is automatically reduced when the light is weak or computing power is limited to ensure real-time performance and keep the lightweight YOLO-Seg model running stably.

[0074] Based on pose control constraints and communication status, the system calibrates and generates pose fusion mode parameters. Using 10Hz as a threshold, the system autonomously selects between pre-fusion and post-fusion modes. For example, when communication is good and computing power is sufficient, pre-fusion is used to improve pose calculation accuracy to within 1 mm; when communication bandwidth is limited or computing power is insufficient, post-fusion is switched to reduce transmission pressure and ensure real-time control, with pose error controllable to within 5 mm.

[0075] Based on wind disturbance sway suppression constraint parameters, wind disturbance sway suppression period parameters are calibrated and generated. The cyclic adjustment interval for sway suppression is determined according to real-time wind speed and the sway amplitude of the composite slab. For example, the sway suppression period is extended in light wind conditions to ensure stability; in strong wind conditions, the sway suppression period is shortened to improve response speed and quickly converge sway deviation.

[0076] Based on workspace and posture constraints, calibration generates incremental parameters for the end effector's motion. This limits the displacement amplitude and angle increment of a single adjustment by the robotic arm, preventing excessive movement that could lead to collisions or overshoot. For example, the motion increment is reduced near the work boundary to achieve fine-tuning; in open spaces, the increment can be appropriately increased to improve work efficiency.

[0077] Based on the work area and safety constraints, calibration generates placement path planning parameters. Combining the steel beam location, the boundaries of installed components, and the distribution of shear studs, an optimal running route is planned that is collision-free, has the shortest distance, and exhibits a smooth trajectory. For example, the trajectory is automatically triggered when descending to approximately 10 centimeters from the steel beam, ensuring that the placement process does not interfere with surrounding components.

[0078] By integrating five types of parameters—visual reasoning resolution parameters, pose fusion mode parameters, wind disturbance and sway suppression cycle parameters, end-effector action increment parameters, and landing path planning parameters—a complete set of collaborative control rule optimization information is formed, providing accurate and reliable rule basis for the subsequent generation of control schemes adapted to complex construction site scenarios.

[0079] By integrating basic information on collaborative strategies, on-site state constraints, and optimized control rules, a comprehensive collaborative design is implemented for visual perception reasoning, pose fusion calculation, wind disturbance suppression and control, and landing path planning. This generates a multi-robotic arm collaborative posture adjustment and landing control scheme adapted to complex construction site scenarios. Furthermore, by integrating basic information on collaborative strategy formulation, on-site state adaptation constraints, and optimized collaborative control rules, a unified collaborative design is conducted for the four core functions: visual perception reasoning, pose fusion calculation, wind disturbance suppression and control, and landing path planning. This constructs a closed-loop control system covering the entire process of perception, calculation, control, and landing, ensuring orderly linkage, data interoperability, and synchronized execution among modules.

[0080] The visual perception inference module operates stably based on edge computing power allocation parameters and on-site visual inference environment parameters, completing local inference of the lightweight YOLO-Seg model at the edge terminal. It automatically adjusts the inference resolution and recognition threshold according to light intensity, dust concentration, and computing power margin, continuously outputting the contour, position, and region segmentation results of the composite bridge panel, providing reliable target information for subsequent pose calculation. For example, in scenes with strong light, backlight, or significant dust, the system automatically activates an enhanced recognition mode to ensure continuous, stable, and error-free target segmentation.

[0081] The pose fusion and calculation module autonomously switches between pre-fusion and post-fusion modes based on communication status, pose adjustment constraints, and collaborative accuracy requirements. When the system calculation frequency is higher than 10Hz and communication is good, the pre-fusion mode is used, where multi-view raw data is uniformly input into the master node for calculation, keeping the position error within 1 mm. When computing power is limited or transmission bandwidth is restricted, the module switches to post-fusion mode, where each edge terminal calculates independently and uploads the results, while the master node completes the fusion, keeping the pose error within 5 mm, thus balancing accuracy and real-time performance.

[0082] The wind disturbance and sway control module performs adaptive sway control based on real-time wind disturbance level, sway amplitude, and wind disturbance and sway period parameters. It calculates the sway frequency and amplitude through a sliding time window, outputs compensation commands in the opposite sway direction, and achieves rapid attitude stabilization by combining robotic arm fine-tuning and crane speed compensation. For example, when increased gusts cause more severe swaying of the composite slab, the system automatically shortens the sway period and increases the compensation step size, bringing the swaying back to a safe range within seconds.

[0083] The placement path planning module generates a collision-free, high-precision running route based on the workspace boundaries, attitude constraints, and placement path planning parameters. Upon descent to approximately 10 centimeters from the steel beam, an automatic attitude transition is triggered. The robotic arm's attitude is first adjusted to avoid shear spikes and already installed components before completing the precise placement. For example, in areas with narrow spaces and numerous surrounding obstacles, the system automatically plans an avoidance path and limits the end effector increment to ensure a safe and smooth placement process.

[0084] The four functional modules work together in a unified time sequence and data interface. Visual results are input for pose calculation, pose results are input for sway control and placement control, sway control status is fed back to pose calibration, and placement planning synchronously constrains the movements of the robotic arm and crane. Through end-to-end data interoperability and collaborative scheduling, a multi-robotic arm collaborative posture adjustment and placement control scheme that can adapt to complex construction site scenarios such as changes in lighting, wind disturbances, space constraints, and communication fluctuations is ultimately generated, providing a complete, stable, and reliable execution basis for automated hoisting operations.

[0085] S105 adaptively allocates distributed computing resources to form batches of vision, computation, control, and security management threads, enabling local synchronization within nodes and secure interaction with industrial LANs via ROS2 across nodes.

[0086] In one implementation, distributed computing resources are adaptively partitioned, including visual processing computing power, pose calculation computing power, collaborative control computing power, and security management computing power, generating task type partitioning results, resource allocation correspondences, and thread scheduling matching rules. Dedicated visual processing computing power is allocated to the lightweight visual segmentation module to support image acquisition, preprocessing, feature extraction, and target segmentation calculations at the construction site. This computing power prioritizes the real-time inference needs of the YOLO-Seg model, ensuring stable output of the contour and position information of the composite bridge deck under different lighting, dust, and occlusion conditions. For example, when image acquisition and target segmentation are running simultaneously, the system automatically allocates higher computing power weight to the visual processing task to avoid recognition delays or missed detections. Figure 3As shown, the lightweight composite slab on-site visual segmentation module is a lightweight visual perception unit designed for complex construction site conditions. It provides high-precision visual support for real-time positioning of composite slabs and adaptive grasping by robotic arms. Each robotic arm terminal independently carries this module, employing an edge computing distributed inference architecture. The YOLO-Seg segmentation model is directly deployed on the terminal edge chip, achieving local real-time perception and avoiding cloud transmission delays and interference from on-site network fluctuations. Model training uses real-world construction site scenarios as the core, constructing a 2000-image labeled dataset through a combination of simulation generation and on-site photography. This dataset comprehensively covers the morphology of composite slabs under different sizes, poses, lighting conditions, dust, and partial occlusion, closely reflecting the actual working conditions of bridge hoisting sites. The training process uses a batch size of 16 and an initial learning rate of 0.001, combined with a cosine annealing learning rate decay strategy and a weight decay coefficient of 0.0005. After 100 epochs of iteration and optimization, the robustness and generalization ability of the model in harsh construction site environments are significantly improved. To adapt to the limited computing power of edge computing chips, this module implements a two-layer lightweight design: first, the network undergoes channel pruning at a rate of 20%–30%, eliminating redundant feature channels and reducing computational load; then, the network weights are compressed using 8-bit quantization, significantly reducing the model size. To compensate for the accuracy loss caused by quantization, quantization-aware training is introduced, enabling the model to maintain the high-precision segmentation required in construction sites while achieving lightweight, low-power, and fast inference. Ultimately, this achieves a comprehensive advantage of lightweight terminal deployment, real-time inference, strong anti-interference capabilities, and accurate positioning, providing reliable visual support for stable robotic arm grasping.

[0087] Independent pose calculation power is allocated to the multi-view pose fusion estimation module to support the calculation of the composite plate's spatial coordinates, attitude angle estimation, multi-view data fusion, and error calibration. This computing power supports the operation of the estimation and optimization two-layer neural network, dynamically adjusting the computing power usage based on the number of pose calculations and the fusion mode to ensure the accuracy and real-time performance of the pose data output. For example, in the pre-fusion mode, the pose calculation computing power is increased to meet the high-precision calculation requirements; in the post-fusion mode, the computing power usage is appropriately reduced to improve overall operating efficiency. Figure 4As shown, the multi-view adaptive pose fusion estimation module for the composite plate is the core technology unit for achieving precise attitude control and smooth placement of the composite plate. The module employs a two-layer neural network of estimation and optimization as the core algorithm for pose calculation. It features an adaptive pre / post fusion switching mechanism, which can autonomously select the optimal fusion strategy based on real-time pressure calculations at the construction site, with a threshold frequency of 10Hz, balancing positioning accuracy and system real-time performance. In pre-fusion mode, the module directly inputs raw data such as images and depth from each robotic arm's perspective into the master node model, uniformly completing feature extraction and pose calculation. This achieves deep coupling of multi-source information at the data layer, improving pose estimation accuracy in complex occlusion scenarios. In post-fusion mode, each edge node independently completes local pose calculation based on its own camera data, only uploading the pose results to the master node for fusion. This significantly reduces transmission bandwidth and master node computing power consumption, ensuring the real-time performance of robotic arm motion control. The pose estimation model was trained with a batch size of 8, a learning rate of 0.0001, a learning rate decay coefficient of 0.8, a confidence weight of 0.0015, a confidence weight decay coefficient of 0.8, and a learning rate decay threshold of 0.0135. The network iterations were optimized to 4, resulting in a total of 500 training rounds, with each round repeated 20 times. Random translation noise of 0.01m was added during training to enhance on-site anti-interference capabilities. Ultimately, the model can accurately calculate the spatial coordinates and tilt angle of the composite plate in real time, with the pre-fusion position error stably controlled within 1mm and the post-fusion pose error controlled within 5mm. This provides highly reliable and real-time pose data support for hoisting posture stabilization, precise placement, and closed-loop correction.

[0088] Collaborative control computing power is allocated to support the generation of master-slave scheduling commands, timing synchronization control, command issuance, motion feedback reception, and force balance analysis for the coordinated operation of cranes and multiple robotic arms. This computing power ensures the real-time interaction of commands between the master and slave nodes, guaranteeing synchronized execution of crane movements and robotic arm posture adjustments. For example, during the simultaneous lifting and posture adjustment phase, the collaborative control computing power operates at full capacity to ensure continuous and lag-free movements.

[0089] Safety management computing power is allocated to segmented operation control and anomaly monitoring to support the logical operations of six-stage process status determination, equipment operation monitoring, anomaly identification, fault level judgment, and safety handling. This computing power monitors the status of the robotic arm, crane, composite plate, and environment in real time, and responds quickly to overload, interference, communication anomalies, and other situations. For example, when the system detects that the pose exceeds the limit or the load is abnormal, the safety management computing power immediately triggers the warning or emergency stop logic to ensure operational safety.

[0090] Based on the computational load, real-time response priority, and runtime sequence of each module, including visual segmentation, pose calculation, collaborative control, and security management, task type classification results, resource allocation correspondence, and thread scheduling matching rules are generated. The computing power limit, running priority, and scheduling order of each type of task are clarified, providing a complete and reliable basis for the standardized scheduling and efficient execution of distributed computing resources in the future.

[0091] Based on a master-slave distributed collaboration strategy, task types, resource allocation, and thread matching relationships are standardized and configured. Resource allocation granularity is set in conjunction with edge computing terminal performance, and the interaction synchronization cycle is determined based on low-latency communication requirements. Using a master-slave distributed collaboration strategy for cranes and multiple robotic arms, the pre-defined task types, resource allocation methods, and thread scheduling matching relationships are uniformly and standardized, ensuring that all computing tasks and computing resources strictly adapt to the operational logic of master node coordination and slave node execution.

[0092] Based on the actual computing power, storage space limits, and power consumption constraints of the edge computing terminal, corresponding resource allocation granularities are set for the four core tasks: visual processing, pose calculation, collaborative control, and security management. Visual processing tasks are prioritized for resource allocation to ensure real-time inference of the YOLO-Seg lightweight model on the terminal. For example, 40% of the edge terminal's computing power is allocated to visual processing to ensure continuous and stable image acquisition and target segmentation. Pose calculation tasks dynamically adjust resource usage based on pre-fusion and post-fusion modes. For example, computing power usage is reduced in post-fusion mode and increased in pre-fusion mode to meet high-precision calculation requirements. Collaborative control tasks maintain a fixed resource quota to ensure uninterrupted master-slave instruction transmission and action synchronization. Security management tasks are allocated a minimum guaranteed computing power to ensure timely response for anomaly monitoring, status determination, and security handling.

[0093] Based on the transmission performance of industrial LANs and the low-latency communication requirements of the ROS2 framework, the data interaction cycle and command synchronization cycle between the master control node and each edge terminal are determined. All nodes use a unified clock reference to maintain consistent data transmission and reception timing. For example, the synchronization cycle is set to be the same as the master-slave scheduling cycle, ensuring that command issuance, data feedback, and action execution are aligned throughout the process without delay or misalignment.

[0094] By constraining the master-slave collaboration strategy and precisely matching the hardware parameters of the edge terminal, the system's computing power is allocated reasonably, thread scheduling is orderly, and timing execution is unified, fully meeting the operational requirements of real-time collaborative control of cranes and multiple robotic arms, millimeter-level precise attitude adjustment, and safe high-altitude operations.

[0095] Based on resource allocation granularity and interaction synchronization cycle, various computing tasks are grouped and integrated to form job thread batches that include vision threads, calculation threads, control threads, and safety management threads. According to the pre-set resource allocation granularity and interaction synchronization cycle, four types of computing tasks—vision processing, pose calculation, collaborative control, and safety management—are orderly grouped and integrated, enabling similar operations to be executed together and different tasks to be scheduled in parallel, improving the utilization efficiency of distributed computing resources. All operations related to target segmentation, such as visual image acquisition, image preprocessing, target feature extraction, and YOLO-Seg lightweight model inference at the construction site, are unified and integrated into a vision thread. This thread prioritizes edge computing resources to ensure continuous output of the contour, edge, and position information of the composite bridge deck even under complex conditions such as changes in lighting, dust interference, and partial occlusion, providing stable input for subsequent grasping and positioning. For example, during the initialization, locking, and grasping phases, the vision thread maintains high-frequency operation, updating the target position in real time to ensure accurate tracking by the robotic arm.

[0096] The computations related to pose calculation, such as multi-view data input, coordinate calculation, pose estimation, data fusion, and error calibration, are unified into a single calculation thread. This thread dynamically adjusts the computational load based on the pre-fusion or post-fusion mode, outputting spatial pose data with millimeter-level accuracy. For example, during the stabilization, descent, and landing phases, the calculation thread runs synchronously with the vision thread, correcting pose deviations in real time to ensure landing accuracy.

[0097] The calculations related to master-slave instruction generation, timing scheduling, action synchronization, force balancing, and coordinated adjustment between the crane and the robotic arm are all integrated into a single control thread. This thread runs throughout the entire lifting process, ensuring strict synchronization between crane operation and robotic arm movements, and reasonable load distribution. For example, during the grabbing and lowering phases, the control thread simultaneously drives the crane and multiple robotic arms to perform coordinated actions without lag or misalignment.

[0098] The safety management thread integrates all related calculations, including equipment status monitoring, operation phase determination, anomaly identification, fault level assessment, and safety handling logic. This thread operates with low power consumption and high real-time performance, responding promptly to risks such as overload, interference, and communication anomalies. For example, when the robotic arm approaches its movement boundary, the safety management thread immediately triggers deceleration and limit commands to ensure operational safety.

[0099] Based on the runtime sequence and real-time priority of hoisting operations, the vision thread, calculation thread, control thread, and safety management thread are combined into a batch of work threads. This allows the four types of threads to run in parallel, communicate data, and not interfere with each other within the edge terminal and the master node. By merging threads and scheduling them into batches, resource contention and computational conflicts are reduced, significantly improving the overall computational efficiency and response speed of the system. This provides stable computing power support for fully automated, precise, and safe operations.

[0100] Local synchronization within each node is performed on each job thread. Based on the ROS2 communication mechanism and industrial LAN transmission requirements, secure cross-node data interaction is achieved, enabling coordinated scheduling of distributed computing resources. Local synchronization is performed in batches on each job thread, ensuring that the vision thread, calculation thread, control thread, and security management thread within the same edge computing terminal run in a unified timing sequence, avoiding timing discrepancies, data conflicts, or resource contention between multiple threads. Local synchronization uses a global clock as a reference, aligning the timing of the entire process from thread startup, data reading, computation execution, to result output, ensuring consistent computational pace within the same terminal. For example, when image acquisition, pose calculation, and status monitoring are performed simultaneously on an edge terminal, the local synchronization mechanism ensures that all three use the same acquired data at the same time, avoiding calculation deviations caused by time differences.

[0101] Based on the ROS2 communication mechanism and industrial LAN transmission standards, a stable and reliable data interaction channel is established between the master node robotic arm (number 0) and the other three slave node robotic arms. The master node is responsible for issuing collaborative scheduling commands, global fusion results, and safety control commands, while the slave nodes are responsible for uploading local sensing data, local pose calculation results, and equipment operating status. Data transmission adheres to low latency and high reliability requirements; all interactive information is verified and encrypted to achieve secure transmission and real-time sharing of control commands, sensing data, and status information. For example, attitude fine-tuning commands issued by the master node can reach each slave node within milliseconds, ensuring synchronization of the crane and multiple robotic arms.

[0102] Through local synchronization within nodes and secure data interaction across nodes, the system achieves unified and coordinated scheduling of distributed computing resources. The master node coordinates global computing power allocation and task scheduling, while each edge terminal executes tasks in parallel according to its allocated computing power, achieving efficient collaborative operation across the entire process of perception, calculation, control, and placement. This scheduling method fully leverages the advantages of real-time local processing at edge terminals and global fusion optimization at the master node, maintaining stable operation, rapid response, and reliable accuracy even under complex construction site conditions, providing robust computing power support and communication guarantees for the automated hoisting and placement of composite bridge decks. Figure 5As shown, the composite slab placement identification module is a key support for achieving fully automated placement and orderly construction of composite bridge decks. It accurately identifies the steel beam axis using point matching and calculates the boundaries and gaps of the already installed precast slabs using the aforementioned pose estimation model, automatically generating the optimal lifting position and completely eliminating reliance on manual command and positioning. Before lifting begins, a robotic arm equipped with a camera comprehensively scans and models the surrounding environment, constructing a digital benchmark for the lifting site. Then, a segmentation model is used to identify the steel beam position, axis, orientation, elevation, and the edges, joints, and relative postures of the installed composite slabs. Based on the analysis results, combined with constraints such as the crane's operating radius and the robotic arm's working space, the optimal placement area is calculated, and the crane's lifting and translation to the lifting point are planned. This module integrates target position identification and autonomous crane movement, achieving an upgrade from "passive positioning" to "active planning," significantly improving the efficiency and automation of lifting operations.

[0103] The S106 takes crane-multi-robotic arm collaboration, lightweight visual segmentation, multi-view pose fusion, adaptive sway suppression, and autonomous positioning planning as its core framework. It learns the mapping relationship between working conditions and parameters, errors and strategies, and wind disturbance and compensation. It optimizes parameters by weighting load, geometry, semantics, and kinematic features, adapts to lighting, wind disturbance, scale, and communication status, and matches pause, rollback, reset, alarm, and emergency stop safety strategies.

[0104] In one implementation, the core framework of crane-multi-robotic arm collaboration, lightweight visual segmentation, multi-view pose fusion, adaptive sway suppression, and autonomous placement planning is adapted to specific scenarios and matched for precision. This generates features for collaborative scheduling, visual perception, pose calculation, wind disturbance suppression, and placement planning, forming core control mechanism information. Using crane-multi-robotic arm collaboration, lightweight visual segmentation, multi-view pose fusion, adaptive sway suppression, and autonomous placement planning as the overall core framework, scenario adaptation is performed for the complex working conditions of bridge construction sites. Simultaneously, precision matching is conducted in conjunction with millimeter-level placement accuracy indicators, enabling the entire framework to adapt to actual site conditions such as sudden changes in lighting, dust interference, gusts of wind, limited space, and communication fluctuations, meeting the high-precision, high-stability, and high-safety requirements for the hoisting of large-mass composite bridge decks.

[0105] For the master-slave distributed operation mode of cranes and multiple robotic arms, collaborative scheduling features are extracted and generated. These features use robotic arm number 0 as the master node and the other three robotic arms as slave nodes. The master node is responsible for issuing global commands, coordinating timing, balancing forces, and handling anomalies, while the slave nodes are responsible for local sensing, local pose calculation, end-effector homing, and attitude fine-tuning, forming a collaborative operation rule with unified commands, synchronized actions, and reasonable loads. For example, during the grasping and descent phases, the master node outputs collaborative commands according to a fixed scheduling cycle, and the slave nodes synchronously execute attitude adjustments to ensure that the crane's operation and the robotic arm movements are seamless and without misalignment.

[0106] For real-time inference of the lightweight YOLO-Seg model at edge terminals, visual perception features are extracted and generated. These features are designed for complex lighting, dust, and partial occlusion conditions at construction sites, adapting to the upper limits of edge computing power. They possess the ability to automatically adjust image acquisition parameters, segmentation confidence, and inference resolution to ensure continuous and stable target recognition. For example, in backlit, low-light, or dusty scenes, the visual perception features automatically switch to an enhanced recognition mode, maintaining accurate output of the composite slab outline and edge coordinates.

[0107] For multi-view pose fusion estimation, pose calculation features are extracted and generated. These features support adaptive switching between pre-fusion and post-fusion, autonomously selecting a fusion strategy with a 10Hz threshold. They can output millimeter-level spatial coordinates and attitude angles, meeting the requirements for pose closed-loop correction. For example, pre-fusion is used when communication is good, controlling the pose error within 1 mm; when communication is limited, post-fusion is switched, controlling the error within 5 mm, balancing accuracy and real-time performance.

[0108] To address the sway suppression of large-mass composite slabs under high-altitude wind disturbance, a wind disturbance suppression feature is extracted and generated. This feature calculates the sway amplitude and frequency based on a sliding time window, and dynamically outputs the sway suppression compensation amount in conjunction with the upper limit of the robotic arm's load, linking the robotic arm and crane to jointly achieve posture stabilization. For example, when the on-site gusts intensify, the wind disturbance suppression feature automatically increases the sway suppression response frequency and the compensation step size, quickly bringing the composite slab swaying back to a safe range.

[0109] For autonomous identification of the placement area and path planning, placement planning features are extracted and generated. These features identify the steel beam axis based on point matching, and automatically calculate the optimal placement area by combining the boundaries of installed components and workspace constraints. A collision-free running path is planned, and attitude transition is automatically triggered approximately 10 centimeters from the steel beam. For example, in a narrow work area, the placement planning features automatically generate a safe path to avoid shear studs and installed plates, ensuring that the placement process is non-interfering and collision-free.

[0110] By integrating five types of features—cooperative scheduling features, visual perception features, pose calculation features, wind disturbance suppression features, and landing planning features—a core control mechanism information covering the entire process of perception, calculation, regulation, and landing is formed, providing a unified, complete, and reliable underlying control basis for subsequent working condition learning, parameter optimization, and safety strategy matching.

[0111] The system learns and calibrates the mapping relationships between operating conditions and parameters, errors and strategies, and wind disturbance and compensation, generating operating condition adaptation coefficients, error correction strategies, and wind disturbance compensation rules to form the basic information for dynamic control. The system continuously learns and dynamically calibrates the three types of mapping relationships—operating conditions and parameters, errors and strategies, and wind disturbance and compensation—enabling the control logic to autonomously optimize itself in accordance with the real-time status of the construction site, thus improving the adaptive adjustment capability throughout the entire process.

[0112] In response to real-time changes in working conditions such as light intensity, dust concentration, wind speed and direction, and communication signal strength at the construction site, the system learns by comparing historical data with the current state to generate corresponding working condition adaptation coefficients. These coefficients are used to automatically adjust key operating parameters such as visual inference resolution, communication synchronization cycle, and computing power allocation ratio. For example, when there is insufficient light or high dust levels, the working condition adaptation coefficient automatically increases the visual enhancement level; when the communication signal is weak, the data transmission frequency is reduced accordingly and the system switches to post-fusion mode to ensure stable and continuous operation.

[0113] To address error types such as pose calculation deviation, asynchronous coordinated actions, and end-effector positioning deviation, the system establishes a mapping relationship between errors and strategies, generating corresponding error correction strategies. When a pose error exceeds the allowable range, a closed-loop compensation mechanism is immediately triggered, eliminating the deviation by increasing the number of pose calculations, adjusting fusion weights, and correcting the coordination timing. For example, when the pose calculation error is greater than 1 mm, the error correction strategy automatically initiates compensation calculations, adjusting the robot arm's posture and the crane's position in reverse, enabling the composite plate to quickly return to the target pose.

[0114] In response to real-time disturbances such as wind intensity, sway amplitude, and sway frequency, the system learns the correspondence between wind disturbance and compensation, generating wind disturbance compensation rules. These rules automatically determine the sway suppression compensation step size, adjustment frequency, and action direction based on the magnitude of the wind disturbance, driving the robotic arm and crane to work together to suppress the sway. For example, a small, gentle compensation is used in light wind conditions; in strong winds or gusts, the compensation frequency and adjustment amplitude are increased to quickly reduce the sway of the composite plate.

[0115] The generated operating condition adaptation coefficients, error correction strategies, and wind disturbance compensation rules are integrated to form a unified set of dynamic control foundation information covering environmental adaptation, error elimination, and wind disturbance suppression. This information provides a complete and reliable underlying basis for the system's real-time dynamic adjustment, high-precision collaborative control, and stable and safe operation.

[0116] By combining load, geometric, semantic, and kinematic features for weighted optimization, and performing real-time adaptive matching based on illumination, wind disturbance, scale, and communication status, the system generates multi-feature fusion control parameters and full-state adaptation coefficients to form control optimization information. Combining the load characteristics, geometric dimensional characteristics, semantic environment characteristics, and kinematic constraint characteristics of the composite bridge deck, the system performs weighted optimization processing on core control parameters such as collaborative control, visual perception, pose calculation, wind disturbance suppression, and placement planning, making the control output more closely match the actual stress and motion laws of hoisting large-mass components.

[0117] Based on the characteristics of composite slabs with loads exceeding ten tons, priority is given to allocating weights to parameters related to force balance and load safety to ensure that the lifting process does not involve overloading or instability. Based on geometric dimensions, the gripping point location and attitude adjustment range are dynamically adjusted according to the slab length, width, and thickness. Based on semantic environment features, and considering the spatial distribution of steel beams, shear studs, and installed components, the weights for obstacle avoidance and placement accuracy are increased. Based on kinematic constraints, the motion parameters are optimized according to the limits of the robotic arm joints, speed limits, and acceleration limits to ensure smooth and controllable movement.

[0118] Based on changes in light intensity, wind disturbance level, component size differences, and real-time communication link status at the construction site, the system adaptively matches control parameters in real time, enabling the control logic to dynamically adjust according to site conditions. When lighting is strong or weak, it automatically matches visual enhancement parameters and recognition confidence levels to maintain stable target segmentation. When wind disturbance levels increase, it automatically increases sway compensation weights to improve attitude stability. When component dimensions change, it synchronously adjusts pose calculation and placement planning parameters to ensure adaptability. When communication link quality deteriorates, it automatically switches to a low-bandwidth, high-reliability post-fusion mode to ensure continuous system operation.

[0119] The system integrates and calculates multiple features, including load, geometry, semantics, and kinematics, to output a multi-feature fusion control parameter. This parameter combines visual, pose, force, and environmental information and can be directly used for collaborative command generation and motion adjustment, making the control output more precise and better suited to the actual working conditions.

[0120] A comprehensive assessment of environmental interference, equipment operation, operational phase, and communication health is performed to generate a full-state adaptation coefficient. This coefficient is used to adjust the global scaling intensity, synchronization cycle, and response threshold, ensuring the entire system operates within its optimal range under any combination of operating conditions.

[0121] Multi-feature fusion control parameters and full-state adaptation coefficients together constitute control optimization information, which is used to dynamically correct the entire process of collaborative scheduling, visual reasoning, pose calculation, sway control and landing planning, significantly improving the system's operational stability and millimeter-level control accuracy under complex construction site conditions.

[0122] By integrating core control mechanism information, dynamic regulation basic information, and control optimization information, and matching pause, rollback, reset, alarm, and emergency stop safety strategies according to operating conditions and error states, the system achieves full-process collaborative control and safety handling. Furthermore, by integrating core control mechanism information, dynamic regulation basic information, and control optimization information, the main control node, based on real-time collected data on construction site conditions, equipment operating status, and error exceeding limits, classifies and executes segmented safety handling strategies, achieving closed-loop safety management from early warning to protection.

[0123] When the system detects a minor anomaly in the equipment that has not reached a dangerous threshold, it immediately executes a pause action, maintaining the current position and stopping all movement commands to prevent the anomaly from escalating further. For example, if the joint temperature of a robotic arm slightly increases or there is a momentary fluctuation in communication, the system pauses all actions and resumes operation only after the situation returns to normal, ensuring on-site stability.

[0124] When a deviation in the coordinated posture between the crane and the multiple robotic arms is detected, or when the pose of the composite plate deviates from the target range, a state rollback is executed to restore the equipment and components to the previous stable position, and pose calibration and coordinated synchronization are re-performed. For example, if a coordination deviation occurs during the grasping phase, the system reverts to the completed locking phase state, reconfirms the pose, and then performs the grasping operation again to prevent posture instability.

[0125] When the triggering conditions for the work process are not met, a stage transition is abnormal, or the process is stuck, a reset operation is performed. This returns all robotic arms to the zero / initial calibration position, raises the crane hook to a safe height, clears the abnormal state, and restarts the initialization process. For example, if the process is interrupted during the descent stage due to an obstacle recognition anomaly, the system automatically resets to the initial state, completes a rescan of the environment, and then continues the operation.

[0126] When equipment operating data exceeds safety limits but does not trigger a dangerous state, an audible and visual alarm is triggered, abnormal data is recorded and uploaded to the backend, alerting on-site personnel without interrupting controllable operations. For example, when local wind speed approaches the limit or the robotic arm load approaches the rated upper limit, the system issues an alarm while continuing to maintain sway control and force balance to ensure continuous operation.

[0127] In the event of dangerous conditions such as severe overload, risk of motion interference, communication interruption, or attitude instability, an emergency stop command is immediately executed, cutting off all power output and locking the brakes to forcibly terminate all actions, maximizing the safety of personnel, equipment, and components. For example, if the system detects an impending collision of the robotic arm or significant uncontrolled shaking of the composite plate, it immediately stops to prevent accidents.

[0128] Through the orderly implementation of the above five-level segmented safety strategy, the system completes the collaborative control and full-cycle safety handling of the entire hoisting process, maintaining stable operation, precise control and reliable protection under complex construction site conditions, ensuring the accurate, stable and safe implementation of the entire hoisting operation of large-mass composite bridge decks.

[0129] In one implementation, such as Figure 6 As shown, this application also provides a multi-robotic arm collaborative posture adjustment and positioning construction system suitable for the hoisting and installation of composite bridge decks, including:

[0130] The 601 on-site all-dimensional perception module is used to acquire information on the operating status of composite bridge decks, cranes, and multiple robotic arms, as well as the construction site environment, providing complete raw perception data support for parameter configuration, data processing, and collaborative control.

[0131] The core control parameter configuration module 602 is used to set the master-slave scheduling cycle, pose calculation times, sway compensation step size, segmentation confidence, fusion threshold and operation triggering conditions based on the requirements of heavy load hoisting, millimeter-level placement, high-altitude wind disturbance suppression and real-time collaborative operation, forming a standardized set of control parameters.

[0132] The multi-source data collaborative processing module 603 is used to perform integrated processing of visual, pose, wind disturbance and force data, build a real-time collaborative link of perception-calculation-control-positioning, and read and output pose, end state, crane parameters, environmental interference and global operation status information.

[0133] The collaborative strategy and rule generation module 604 is used to determine the collaborative control strategy by combining the complex working conditions of the construction site, the upper limit of edge computing power, low latency communication and high-altitude safety regulations. Based on the lighting, wind disturbance, attitude, space and communication status, it sets the inference resolution, fusion mode, sway suppression period, action increment and path planning rules to generate a multi-manipulator collaborative posture adjustment and positioning control scheme adapted to the complex construction site scenario.

[0134] The distributed computing resource scheduling module 605 is used to adaptively allocate edge computing resources, divide and form batches of vision, calculation, control and security management operation threads, and complete local data synchronization within the node and secure data interaction across nodes ROS2 and industrial LAN.

[0135] The Intelligent Optimization and Safety Management Module 606 is used to learn the mapping relationship between working conditions and parameters, errors and strategies, and wind disturbance and compensation, based on the core framework of crane-multi-manipulator collaboration, lightweight visual segmentation, multi-view pose fusion, adaptive sway suppression, and autonomous positioning planning. It optimizes control parameters by weighting load, geometry, semantics, and kinematic features, adapts to changes in lighting, wind disturbance, scale, and communication status, and matches multi-level safety handling strategies such as pause, rollback, reset, alarm, and emergency stop.

[0136] The computer-readable storage medium provided in the above embodiments of this application and the multi-robotic arm collaborative posture adjustment and positioning construction method for composite bridge deck hoisting construction provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0137] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the multi-robotic arm collaborative attitude adjustment and placement construction method, system, electronic equipment, and readable storage medium suitable for composite bridge deck hoisting construction are relatively simple to describe because they are substantially similar to the embodiments of the multi-robotic arm collaborative attitude adjustment and placement construction method suitable for composite bridge deck hoisting construction described above. Relevant parts can be referred to in the descriptions of the embodiments of the multi-robotic arm collaborative attitude adjustment and placement construction method suitable for composite bridge deck hoisting construction described above.

Claims

1. A multi-robot collaborative posture adjustment and positioning construction method suitable for the hoisting and installation of composite bridge decks, characterized in that, include: Acquire comprehensive perception data of composite bridge decks, cranes, multiple robotic arms, and the construction site; Based on the requirements of heavy load hoisting, millimeter-level positioning, high-altitude wind disturbance suppression, and real-time collaboration, the master-slave scheduling cycle, pose calculation times, sway compensation step size, segmentation confidence, fusion threshold, and operation triggering conditions are set. The system integrates visual, pose, wind disturbance, and force data processing to construct a real-time collaborative link for perception, calculation, control, and placement, and reads pose, end-point status, crane parameters, environmental interference, and operational status information. Based on the complex working conditions of the construction site, the upper limit of edge computing power, low-latency communication and high-altitude safety regulations, a collaborative strategy is determined. According to the lighting, wind disturbance, attitude, space and communication status, the inference resolution, fusion mode, sway suppression cycle, action increment and planning rules are set to generate a multi-robotic arm collaborative attitude adjustment and positioning control scheme adapted to the complex construction site scenario. Adaptively allocate distributed computing resources to form batches of vision, computation, control, and security management threads, enabling local synchronization within nodes and secure interaction with industrial LANs via ROS2 across nodes; With crane-multi-robotic arm collaboration, lightweight visual segmentation, multi-view pose fusion, adaptive sway suppression, and autonomous positioning planning as the core framework, it learns the mapping relationship between working conditions and parameters, errors and strategies, and wind disturbance and compensation. It optimizes parameters by weighting them according to load, geometry, semantics, and kinematic features, adapts to lighting, wind disturbance, scale, and communication status, and matches safety strategies such as pause, rollback, reset, alarm, and emergency stop.

2. The method as described in claim 1, characterized in that, Based on the requirements of heavy load lifting, millimeter-level positioning, high-altitude wind disturbance suppression, and real-time coordination, the master-slave scheduling cycle, pose calculation count, sway compensation step size, segmentation confidence level, fusion threshold, and operation triggering conditions are set, including: The hoisting construction accuracy, coordination, anti-interference, and process indicators are broken down into dimensions, and the features of collaborative control, posture perception, wind disturbance suppression, and segmented management are extracted to form a subset of core control parameters; Construct a crane-multi-manipulator master-slave collaborative control model, with the master-slave scheduling cycle as the core parameter, and optimize the collaborative scheduling parameters in combination with the large load lifting requirements; A joint model for visual segmentation, pose fusion, and adaptive sway suppression is constructed, and the segmentation confidence, fusion threshold, number of pose calculations, and sway compensation step size are set to meet the requirements of accuracy and anti-interference. Establish a segmented operation control mechanism, determine the operation triggering conditions based on equipment status and safety constraints, and form a complete collaborative control parameter setting scheme.

3. The method as described in claim 1, characterized in that, The system integrates visual, pose, wind disturbance, and force data processing to construct a real-time collaborative link for perception, calculation, control, and placement. It reads pose, end-effector status, crane parameters, environmental disturbances, and operational status information, including: Based on the real-time nature and master-slave collaboration rules of multi-source data such as vision, pose, wind disturbance, and force, the data fusion and link construction strategy is determined by the master control node and the edge computing terminal in collaboration. The data processing form and collaborative control dimension are determined by distributed acquisition and global time sequence alignment. A four-stage linkage mechanism of perception-computation-control-positioning is adopted to process multi-source field data in an integrated manner. First, each edge terminal completes local data collection and computation, and then the master node completes global data fusion, which is associated with core requirements such as millimeter-level positioning, wind disturbance suppression, and force balance to generate standardized control data. Real-time status verification of standardized control data is performed, adaptive correction processing is initiated for data fluctuations caused by complex on-site working conditions, a closed-loop compensation mechanism is triggered for pose calculation errors, an early warning judgment process is executed for abnormal operating states, and optimized control information including data fusion, compensation correction, and status judgment is generated. The optimized control information is executed in a closed loop, and the entire process is linked through real-time collaborative links. The position and posture of the composite plate, the end state of the robotic arm, the crane operating parameters, the intensity of environmental interference, and the global operation status information are read synchronously to form complete feedback data to support precise placement.

4. The method as described in claim 1, characterized in that, Based on the complex working conditions of construction sites, the upper limit of edge computing power, low-latency communication, and high-altitude safety regulations, a collaborative strategy is determined. According to lighting, wind disturbance, attitude, space, and communication status, inference resolution, fusion mode, sway suppression period, motion increment, and planning rules are set to generate a multi-robotic arm collaborative attitude adjustment and positioning control scheme adapted to complex construction site scenarios, including: The collaborative constraints in complex working conditions, edge computing power limits, low-latency communication, and high-altitude safety regulations are adapted, analyzed, and transformed to generate master-slave collaborative scheduling characteristics, edge computing power allocation characteristics, real-time communication guarantee characteristics, and high-altitude operation safety characteristics, forming the basic information for collaborative strategy formulation. The system transforms dynamic changes in lighting, wind disturbance, attitude, space, and communication status into logical rules and defines threshold boundaries to generate visual reasoning environment parameters, wind disturbance swaying constraint parameters, pose control limitation parameters, work space avoidance parameters, and communication link stability parameters, thus forming on-site status adaptation constraint information. For the objective function of crane-multi-manipulator collaborative control, parameter calibration and collaborative accuracy evaluation are performed in combination with on-site state constraints. Visual reasoning resolution parameters, pose fusion mode parameters, wind disturbance sway suppression cycle parameters, end motion increment parameters, and placement path planning parameters are generated to form collaborative control rule optimization information. By integrating basic information on collaborative strategies, on-site status constraints, and control rule optimization information, a full-process collaborative design is carried out for visual perception reasoning, pose fusion calculation, wind disturbance suppression and control, and landing path planning to generate a multi-robotic arm collaborative posture adjustment and landing control scheme adapted to complex construction site scenarios.

5. The method as described in claim 1, characterized in that, Adaptive allocation of distributed computing resources to form batches of vision, computation, control, and security management threads, enabling local synchronization within nodes and secure interaction across nodes via ROS2 with the industrial LAN, including: Distributed computing resources are adapted and partitioned, including visual processing computing power, pose calculation computing power, collaborative control computing power, and security management computing power, generating task type partitioning results, resource allocation correspondence, and thread scheduling matching rules; Based on the master-slave distributed collaboration strategy, the task type, resource allocation, and thread matching relationship are configured in a standardized manner. The resource allocation granularity is set in combination with the performance of the edge computing terminal, and the interaction synchronization period is determined according to the low-latency communication requirements. Based on the resource allocation granularity and interaction synchronization cycle, each computing task is grouped and integrated to form a batch of job threads that includes a vision thread, a calculation thread, a control thread, and a security management thread. Each job thread performs local synchronization processing within the node. Based on the ROS2 communication mechanism and industrial LAN transmission requirements, it achieves secure data interaction across nodes and completes the collaborative scheduling of distributed computing resources.

6. The method as described in claim 5, characterized in that, With crane-multi-robotic arm collaboration, lightweight visual segmentation, multi-view pose fusion, adaptive sway suppression, and autonomous positioning planning as its core framework, the system learns the mapping relationships between working conditions and parameters, errors and strategies, and wind disturbance and compensation. It optimizes parameters by weighting based on load, geometry, semantics, and kinematic features, adapting to lighting, wind disturbance, scale, and communication status, and matching safety strategies such as pause, rollback, reset, alarm, and emergency stop. The core framework of crane-multi-robotic arm collaboration, lightweight visual segmentation, multi-view pose fusion, adaptive sway suppression, and autonomous landing planning is adapted to the scene and matched with the accuracy. Features of collaborative scheduling, visual perception, pose calculation, wind disturbance suppression, and landing planning are generated to form the core control mechanism information. The mapping relationship between operating conditions and parameters, error and strategy, and wind disturbance and compensation is learned and parameters are calibrated to generate operating condition adaptation coefficients, error correction strategies and wind disturbance compensation rules, forming the basic information for dynamic control. Weighted optimization is performed by combining load, geometry, semantics, and kinematic features. Real-time adaptive matching is completed based on illumination, wind disturbance, scale, and communication status to generate multi-feature fusion control parameters and full-state adaptation coefficients, forming control optimization information. Integrating core control mechanism information, dynamic regulation basic information, and control optimization information, and matching pause, rollback, reset, alarm, and emergency stop safety strategies according to operating conditions and error status, the system completes full-process collaborative control and safety handling.

7. A multi-robotic arm collaborative posture adjustment and positioning construction system suitable for the hoisting and installation of composite bridge decks, characterized in that, The system is used to execute executable instructions to perform the multi-robotic arm collaborative posture adjustment and positioning construction method for the hoisting construction of composite bridge decks as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: First processor; And a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the multi-robotic arm collaborative posture adjustment and placement construction method for the hoisting construction of composite bridge decks according to any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the multi-robotic arm collaborative posture adjustment and positioning construction method for the hoisting construction of composite bridge decks as described in any one of claims 1 to 6.

Images