Urban update intelligent construction platform and intelligent scheduling method

Abstract

The invention discloses a city updating intelligent construction platform and an intelligent scheduling method, and relates to the technical field of city updating, and the system comprises a data fusion and construction unit, an intelligent capsule warehouse cluster, a self-adaptive scheduling unit and an augmented reality man-machine cooperation unit. The method comprises the steps of constructing a digital twinborn model, updating a dynamic layer in real time, extracting state features, constructing a space-time joint potential field, generating a scheduling instruction, executing the scheduling instruction and enhancing reality man-machine collaboration. According to the invention, the real-time performance and the performability of the scheduling decision can be improved, and the construction process can be finely controlled, so that the construction efficiency of urban update construction is improved.

Description

Technical Field

[0001] This invention relates to the technical field of urban renewal, specifically to an intelligent construction platform and intelligent scheduling method for urban renewal. Background Technology

[0002] As my country's urbanization process enters the stage of improving the quality of existing stock, urban renewal has become the main battlefield of urban construction. Urban renewal refers to the necessary and planned reconstruction activities of areas in urban built-up areas that are not adapted to modern urban life. These include projects such as the renovation of old residential areas, the restoration of historical blocks, the functional upgrading of existing buildings, and the upgrading of infrastructure, involving various engineering types such as pipeline renovation, facade renovation, elevator installation, and road repair.

[0003] Compared to new construction projects, urban renewal projects are typically located in densely populated, space-constrained built-up areas, characterized by complex construction environments, narrow sites, numerous surrounding disturbances, and sensitivity to residents' lives. These projects often require construction within the normal living environment of residents, placing higher demands on construction organization, schedule control, and environmental protection. How to efficiently organize multi-disciplinary construction within limited spaces, minimize disruption to residents, and ensure construction safety and quality has become a critical technical issue that urgently needs to be addressed in the field of urban renewal.

[0004] In the early stages of urban renewal projects, traditional manual construction methods were primarily relied upon. Construction organization employed a phased, sequential approach, meaning work proceeded in order of process. For example, in exterior wall renovation projects, scaffolding had to be erected first, followed by manual base treatment and painting. Finally, the scaffolding was dismantled before the pipeline renovation team could enter. On-site management relied mainly on manual inspections and paper records. Safety officers patrolled the construction site to check compliance with construction standards, and quality inspectors manually inspected key processes. There was a lack of effective coordination between construction machinery and workers.

[0005] To address these issues, several intelligent construction management technologies have emerged in recent years. One type of technology involves digital management platforms for the construction process. These platforms deploy on-site cameras and sensors to collect images and environmental data on construction progress, which are then visualized using building information models. Managers can remotely view the site and track construction progress through the platform. Some platforms also have basic data statistics and report generation functions. While this type of technology achieves the digital collection and visualization of construction information, it still relies heavily on manual decision-making, with the platform only providing information support.

[0006] Another type of technology involves the intelligent operation of individual automated construction equipment. For example, intelligent spraying robots can autonomously complete wall spraying operations and have basic obstacle avoidance functions; intelligent inspection robots can scan structures along preset paths; and intelligent material delivery vehicles can transport materials according to on-site or preset requirements. These types of equipment automate specific processes, improving the efficiency of individual steps, but they lack unified scheduling and collaborative operation capabilities between different devices.

[0007] Despite the progress made in the above aspects, existing technologies still have the following major drawbacks: Existing scheduling systems mostly employ centralized static planning or simple time window allocation, failing to respond in real-time to dynamic changes at the construction site. This is especially true in scenarios such as narrow alleyways in old residential areas where space for equipment to avoid each other is extremely limited. When unexpected situations occur, such as equipment failure, personnel movement, or temporary material storage, the system struggles to adjust its scheduling strategy promptly, leading to equipment waiting, path conflicts, and even work interruptions. Furthermore, existing technologies use a unified static constraint model, failing to provide real-time feedback on progress, quality, and safety data collected at the construction site to the scheduling decision-making process. Urban renewal scenarios involve numerous dynamically changing constraints, such as the need to suspend material transportation during peak commuting hours, restrict machinery operation directions around cultural relic protection areas, and regulate traffic flow in narrow passages. This makes it difficult to continuously improve the scheduling scheme as the project progresses, affecting the continuity of construction.

[0008] Regarding the aforementioned technical solutions, existing construction organization methods are inefficient, lack effective coordination and scheduling mechanisms, and are difficult to manage in a refined manner during the construction process, thereby reducing the construction efficiency of urban renewal projects. Summary of the Invention

[0009] The purpose of this invention is to provide an intelligent construction platform and intelligent scheduling method for urban renewal, in order to solve the problems mentioned in the background. The core improvements of this invention revolve around three dimensions: multi-source data fusion, intelligent decision-making, and real-time response, addressing the core pain points in traditional urban renewal construction supervision through technological innovation.

[0010] Firstly, the present invention provides an intelligent construction platform for urban renewal, which achieves its purpose through the following technical solutions:

[0011] The urban renewal intelligent construction platform includes the following modules:

[0012] Data fusion and construction unit: used to acquire multi-source data of urban renewal areas and construct a four-layer digital twin model of construction scenarios, including geometric, physical, semantic and dynamic layers. The dynamic layer maps the location data of smart capsule warehouses, personnel location data, environmental monitoring data and dynamic obstacle data in real time.

[0013] Intelligent capsule cluster: consists of multiple functionally heterogeneous intelligent capsules, each of which integrates an environmental perception and communication module and a pluggable task module;

[0014] Environmental perception and communication module: used to perceive the surrounding environment in real time, build a local occupancy grid map, and transmit data;

[0015] Pluggable task modules: These modules can be replaced and installed on the smart capsule via standardized interfaces to perform specified construction procedures;

[0016] Adaptive scheduling unit: used to generate scheduling instructions based on deep reinforcement learning, including a state space construction module and a decision scheduling module;

[0017] State space construction module: used to extract the system state features at the current moment from the digital twin model to form a high-dimensional state vector;

[0018] Decision scheduling module: Deployed with a pre-trained deep reinforcement learning model, which is used to output a collaborative scheduling strategy based on a high-dimensional state vector, generate scheduling instructions and send them to the corresponding smart capsules.

[0019] By adopting the above technical solution, a four-layer digital twin model comprising geometric, physical, semantic, and dynamic layers is constructed using data fusion and construction units. This achieves multi-dimensional digital mapping of urban renewal construction scenarios. The dynamic layer maps real-time location data of smart capsules, personnel location data, environmental monitoring data, and dynamic obstacle data, providing a real-time and accurate state awareness foundation for subsequent scheduling decisions. This solves the problems of delayed information acquisition and opaque on-site status in traditional construction management. By setting up a cluster of smart capsules, with each capsule integrating an environmental perception and communication module and a pluggable task module, modularity and functional reconfigurability of the construction execution terminal are achieved. The environmental perception and communication module constructs a local occupancy grid map in real time and transmits data, providing the platform with micro-environmental information surrounding the smart capsules. The pluggable task module allows for rapid replacement through standardized interfaces, enabling the same smart capsule to be flexibly configured with different functions according to construction process requirements, improving equipment utilization and... Scene adaptability: By setting up an adaptive scheduling unit, the state space construction module extracts the current system state features from the digital twin model to form a high-dimensional state vector, transforming the complex construction site state into a feature representation that can be processed by a deep reinforcement learning model. The decision scheduling module deploys a pre-trained deep reinforcement learning model, outputs a collaborative scheduling strategy based on the high-dimensional state vector, and generates scheduling instructions to be issued to the intelligent capsule, realizing intelligent decision-making based on global state perception. The modules are tightly coupled through data flow and control flow, forming a complete technical closed loop from environmental perception, state representation, intelligent decision-making to instruction execution. Compared with the existing architecture where the scheduling system and execution equipment are independent, this technical solution effectively solves the technical problem of balancing real-time performance and global optimality in multi-machine collaborative scheduling under complex scenarios through a deeply coupled modular design. It achieves a synergistic improvement in construction organization efficiency and safety, enables refined management of the construction process, and thus improves the construction efficiency of urban renewal construction.

[0020] Optionally, in the data fusion and construction unit, the semantic layer is used to store semantic attribute information in the urban renewal area, including regional functional attributes, time constraint attributes, and security attributes.

[0021] The adaptive scheduling unit also includes a conflict prediction and potential field construction module: based on semantic attribute information and real-time mapped intelligent capsule location data, personnel location data, environmental monitoring data and dynamic obstacle data in the dynamic layer, it constructs a spatiotemporal joint potential field to describe the risk distribution of the work space, and uses the gradient or potential field value of the spatiotemporal joint potential field as the input feature of the high-dimensional state vector. The deep reinforcement learning model is trained to learn and output a collaborative scheduling strategy under the potential field constraint.

[0022] By adopting the above technical solution, and by defining the functional attributes, time constraints, and safety attributes of the semantic layer storage area, the construction rules, time windows, and safety requirements unique to urban renewal scenarios are integrated into the digital twin model in a structured form. The functional attributes define the functions of different areas, such as construction areas, material areas, and personnel passages, providing a basis for determining area types in subsequent conflict prediction. The time constraint attributes define the time windows during which each area is allowed to operate, such as avoiding residents' rest periods, enabling scheduling decisions to automatically comply with disturbance control requirements. The safety attributes define safety rules such as safety distance thresholds and no-entry boundaries, providing a rule basis for the generation of dynamic safety fences. By adding a conflict prediction and potential field construction module, and limiting it to construct a spatiotemporal joint potential field based on semantic attribute information and real-time data from the dynamic layer, a conflict prediction mechanism based on the potential field method is introduced. The spatiotemporal joint potential field integrates static rules with dynamic information, forming a continuous description of the risk distribution in the work space. This allows potential conflicts to no longer rely on discrete threshold judgments, but rather on the continuous quantification of risk levels through potential field values, improving the sensitivity and accuracy of conflict prediction. The gradient or potential value of the combined potential field is used as an input feature of the high-dimensional state vector, realizing the deep integration of potential field information and deep reinforcement learning model. In traditional methods, the potential field method usually outputs obstacle avoidance instructions as an independent module, independent of the decision module of this invention. This technical solution incorporates the potential field features as part of the state space of the reinforcement learning model, enabling the model to autonomously learn the optimal strategy under potential field constraints during training, rather than simply following preset potential field rules. This fusion architecture of potential field guidance and reinforcement learning optimization retains the intuitive expressive ability of the potential field method for spatial constraints, while leveraging the decision-making advantages of reinforcement learning in multi-objective optimization. It solves the technical problems of difficulty in handling complex collaborative tasks by relying solely on the potential field method and the difficulty in training reinforcement learning in sparse reward environments. The deep reinforcement learning model is trained to learn and output collaborative scheduling strategies under potential field constraints, making the generated scheduling instructions conform to spatial safety constraints, reducing the computational cost of posterior conflict detection and correction, improving the real-time performance and executability of scheduling decisions, enabling refined control of the construction process, and thus improving the construction efficiency of urban renewal.

[0023] Optionally, in the conflict prediction and potential field construction module, the spatiotemporal joint potential field includes a static potential field and a dynamic potential field. The static potential field is constructed based on the regional functional attributes in the semantic layer, and its potential field strength has a direction dependence. The dynamic potential field is constructed and updated in real time based on the location data of the smart capsule pods, and is used to characterize the mutual avoidance requirements between smart capsule pods. The repulsion strength function of the dynamic potential field is: ,in, Let q be the Euclidean distance between the current smart capsule and the obstacle smart capsule, and let q represent the pose vector of the current smart capsule in the configuration space, which includes position coordinates and orientation angle information. The radius of influence of the pre-set repulsive force. The relative speed between the current smart capsule and the obstacle smart capsule. The maximum relative velocity is preset, η is the first gain coefficient, and α is the second gain coefficient.

[0024] By adopting the above technical solution, and by defining the spatiotemporal joint potential field as including a static potential field and a dynamic potential field, the potential field construction task is decomposed into two components with clear physical meanings. The static potential field is constructed based on the regional functional attributes in the semantic layer and is used to characterize the risk distribution generated by fixed obstacles and rule constraints. This potential field does not change with time or only changes with the construction stage and can be pre-calculated and stored, reducing the burden of real-time computing. The dynamic potential field is constructed and updated in real time based on the location data of the smart capsule pods and is used to characterize the mutual avoidance needs between mobile smart capsule pods and between smart capsule pods and mobile personnel. This potential field changes in real time with the movement of smart capsule pods and requires high-frequency updates. This decomposed architecture optimizes the allocation of computing resources while ensuring the accuracy of the potential field description, and solves the technical problem of excessive computing power overhead caused by the need for full real-time calculation of a single potential field. By defining the potential field strength of the static potential field as directionally dependent, the ability to express the special constraints of urban renewal scenarios is introduced. In the traditional potential field method, the repulsive potential field generated by obstacles is usually isotropic, that is, the repulsive strength from the center of the obstacle to all directions. Similarly, in the narrow spaces of urban renewal, the prohibition requirements in certain directions, such as those facing building walls, are much stronger than in other directions, such as along alleyways. The direction-dependent potential field can accurately express this asymmetric constraint, allowing the smart capsules to move more naturally along feasible directions when planning their paths, avoiding dead zones or oscillations caused by isotropic potential fields. By defining the repulsion strength function of the dynamic potential field, the relative velocity is considered. The distance term in the formula represents the basic repulsion strength that increases sharply as the distance decreases, meeting the basic requirements for safe avoidance. The velocity modulation term ensures that when two smart capsules move towards each other with a large relative velocity, the repulsion strength increases accordingly, thereby triggering avoidance behavior in advance and solving the problem of the traditional static potential field method's delayed response in dynamic environments. The gain coefficients η and α in the above formula are adjustable parameters that can be optimized through simulation experiments or online learning, enabling the potential field model to adapt to the avoidance characteristics requirements of different construction scenarios. This improves the scenario adaptability and generalization ability of the technical solution, allowing for refined control of the construction process and thus improving the construction efficiency of urban renewal.

[0025] Optionally, in the data fusion and construction unit, the semantic attribute information also includes marked anisotropic constraint regions, which include personnel passage direction, directional prohibition of entry into cultural relic protection areas, and passage direction of narrow passages;

[0026] In the conflict prediction and potential field construction module, a dynamic potential field is also constructed based on the anisotropic constraint regions marked by the semantic layer to reflect the anisotropic constraints in the urban renewal scenario. The repulsion strength varies with the angle between the smart capsule and the constraint direction. The repulsion strength function of the dynamic potential field is further updated as follows: Where θ is the current intelligent capsule's direction of motion relative to the obstacle intelligent capsule. β represents the main constraint direction marked by the semantic layer, and β is the third gain coefficient.

[0027] By adopting the above technical solution, and by limiting the semantic attribute information to include marked anisotropic constraint regions, and by integrating urban renewal-specific constraints such as pedestrian traffic direction, directional prohibition of entry into cultural relic protection areas, and traffic direction in narrow passages into the digital twin model in a structured form, the pedestrian traffic direction constraint indicates that in densely populated areas, smart capsule pods should maintain consistency with the direction of pedestrian flow as much as possible or minimize reverse passage; directional prohibition of entry into cultural relic protection areas indicates that certain directions, such as approaching the cultural relic itself, are strictly prohibited, while movement parallel to the boundary of the protection area may be permitted; and the traffic direction in narrow passages indicates that in narrow alleys, a first-in-first-out or specific directional traffic rule should be followed to avoid two-way congestion. These directional constraints are the essential characteristics that distinguish urban renewal scenarios from open industrial environments, and explicitly modeling them is the key technology for solving the orderly passage in narrow spaces. By defining a dynamic potential field based on anisotropic constraint regions and making its repulsion strength change with the angle between the smart capsule pod and the constraint direction, a mathematical characterization of directional constraints is achieved. Traditional isotropic potential fields cannot distinguish the degree of risk in different directions, which may lead to the smart capsule pod being subjected to... In constrained areas, rigid or non-compliant behavior can be addressed by introducing a direction modulation factor. This allows the potential field to sense the deviation between the movement direction and the constraint direction and adjust the repulsion strength accordingly, guiding the intelligent capsule pods to move along the compliant direction. By expanding the repulsion strength function of the dynamic potential field, the repulsion strength increases when the movement direction is consistent with the main constraint direction, encouraging the pods to maintain that direction; when the movement direction is opposite to the main constraint direction, the repulsion strength decreases, suppressing reverse movement; and when the movement direction is perpendicular to the main constraint direction, a neutral effect is produced. This direction-selective potential field modulation enables the intelligent capsule pods to autonomously form an orderly flow within the constrained area, avoiding congestion and conflicts caused by disorderly competition. The aforementioned direction modulation mechanism, along with the potential field features as state vector input and online fine-tuning of reinforcement learning, works synergistically. In the early stages of training, the direction-modulated potential field provides prior guidance for the reinforcement learning model, accelerating convergence. In the later stages of training, the reinforcement learning model can optimize the β coefficient through online fine-tuning, making the direction modulation strength adaptable to the measured traffic efficiency data on-site. This achieves refined control of the construction process, thereby improving the construction efficiency of urban renewal projects.

[0028] Optionally, in the decision scheduling module, a multi-agent deep deterministic policy gradient algorithm is adopted, treating each intelligent capsule as an agent to learn a distributed collaborative strategy. The training process of the deep reinforcement learning model includes: constructing a high-fidelity simulation environment based on a digital twin model, and designing a reward function R= ×Progress Rewards+ ×Efficiency Rewards- ×Conflict Punishment- ×Energy Penalty- × Disturbance punishment, among which As the first weighting coefficient, This is the second weighting coefficient. The third weighting coefficient, It is the fourth weighting coefficient. The fifth weight coefficient is used for offline training in a simulation environment, and the model parameters are fine-tuned online based on real feedback data.

[0029] The adaptive scheduling unit also includes a feedback learning module: used to collect the actual execution results of the intelligent capsule after executing the scheduling instructions, form feedback samples, and periodically use the feedback samples to incrementally train the deep reinforcement learning model to achieve adaptive optimization of the scheduling strategy.

[0030] By adopting the above technical solution and limiting the use of a multi-agent deep deterministic policy gradient algorithm, each intelligent capsule is treated as an agent, learning a distributed cooperative policy. An advanced reinforcement learning framework suitable for multi-agent continuous control scenarios is introduced, employing a centralized training and distributed execution architecture. During the training phase, global information guides the optimization of each agent's policy; during the execution phase, each agent relies only on local observations for decision-making. This architecture ensures both the global optimality of multi-agent cooperation and meets the low-latency decision-making requirements of real-time scheduling, solving the problems of excessive computational latency and fragmentation inherent in traditional centralized scheduling in large-scale cluster scenarios. Distributed scheduling faces the technical challenge of ensuring global coordination. This paper addresses this issue by defining the training process of the deep reinforcement learning model, including constructing a high-fidelity simulation environment based on a digital twin model. This virtual training-real execution technical path, where the simulation environment, built on the digital twin model, can accurately simulate the geometric constraints, physical characteristics, and dynamic changes of the construction site. This allows the reinforcement learning model to undergo millions of rounds of trial and error training in a zero-cost, zero-risk virtual environment, learning robust policies in various edge scenarios. This effectively solves the problems of difficult sample acquisition, high trial and error costs, and uncontrollable safety risks encountered in direct training on real construction sites. Furthermore, by designing a reward function, the multi-objective optimization problem is transformed into a single-objective reward maximization problem. This reward function simultaneously considers five dimensions: construction progress, work efficiency, safety conflicts, energy consumption costs, and disturbance control. The weight coefficients can be adjusted according to specific project needs, allowing the trained strategy to flexibly switch between different project priorities, solving the problem that traditional scheduling methods struggle to balance multi-objective optimization. Finally, by setting up a feedback learning module and limiting its collection of actual execution results to form feedback samples and periodically incrementally training the deep reinforcement learning model, the scheduling strategy can continuously evolve with the accumulation of construction site data. For example, when a problem is discovered... When certain types of conflicts occur frequently in real-world environments but are not adequately simulated in simulations, incremental training can adjust strategies to better address real-world scenarios. When projects enter different construction phases and constraints change, strategies can adaptively adjust to maintain optimality, solving the problem of performance fixation and inability to adapt to dynamic changes on-site after deployment in traditional scheduling systems. The aforementioned algorithm architecture and feedback mechanism together constitute a complete technical link from offline training to online fine-tuning, and from simulation environments to real-world scenarios. Each link is connected through a closed-loop data flow, forming an endogenous driving force for continuous strategy optimization, significantly improving the platform's adaptability and performance ceiling in long-term operation.

[0031] Optionally, it also includes an augmented reality human-machine collaboration unit: used to receive scheduling instructions generated by the decision scheduling module and convert them into augmented reality guidance information to push to the on-site personnel terminal, including personnel positioning and tracking module, dynamic safety fence generation module and instruction generation and push module;

[0032] Personnel location and tracking module: used to track the location of personnel on site in real time and update it to the dynamic layer of the digital twin model;

[0033] Dynamic safety fence generation module: used to generate dynamic safety fences in real time based on the movement trajectory of the smart capsule, and trigger graded warnings when personnel enter;

[0034] Instruction generation and push module: Used to automatically generate and push augmented reality guidance information that includes work point markers, operation steps, and safety tips.

[0035] By adopting the above technical solution, the augmented reality human-machine collaboration unit receives the scheduling instructions generated by the decision-making and scheduling module, converts them into augmented reality guidance information, and pushes them to the on-site personnel terminals. Traditional construction instruction transmission relies on walkie-talkies, paper work orders, or handheld terminal text displays, resulting in low information transmission efficiency and ambiguity. This technical solution transforms scheduling instructions into visual guidance information overlaid on the real scene, enabling personnel to intuitively understand task requirements, significantly reducing communication costs and operational error rates. It solves the technical problems of unintuitive instruction transmission and low collaboration efficiency in human-machine hybrid operations. Furthermore, by setting up a personnel positioning and tracking module and limiting its use to real-time tracking of on-site personnel positions and updating them to the dynamic layer of the digital twin model, [the following is implied:] ... Personnel are incorporated into the global state perception system as a dynamic element, on par with the intelligent capsule pods. In traditional construction management, personnel location information is isolated from the equipment scheduling system, resulting in scheduling decisions failing to fully consider the impact of personnel distribution on operational safety. This technical solution maps personnel locations to a digital twin model in real time, enabling the adaptive scheduling unit to treat personnel as dynamic obstacles and avoid them during path planning and conflict prediction, or proactively schedule intelligent capsule pods to the personnel's location for collaborative work. By setting up a dynamic safety fence generation module, which generates dynamic safety fences in real time based on the intelligent capsule pod's movement trajectory and triggers tiered warnings when personnel enter, a proactive safety protection mechanism is constructed. Dynamic safety fences differ from traditional static ones. The isolation zone, whose shape and position change in real time with the movement of the smart capsule, can precisely encompass the working range and movement path of the smart capsule. When the spatial relationship between the personnel's position and the safety fence changes, the system triggers a tiered warning, ensuring personnel safety while avoiding the large amount of idle work space caused by overly conservative fixed isolation zones, thus solving the technical problem of balancing safety and efficiency in human-machine shared spaces. Through the setting of the instruction generation and push module, it automatically generates augmented reality guidance information containing work point markers, operation step instructions, and safety tips, integrating multi-dimensional information into a single visual interface. The work point markers use spatial anchoring technology to accurately indicate the physical locations requiring human intervention, and the operation step instructions use text, arrows, and other methods to guide the process. The operation process is displayed in the form of heads and animations, which reduces the requirements for personnel experience. Safety prompts display the location of the smart capsule, the boundary of the danger zone and the evacuation route in real time, so that personnel can perceive the surrounding risks at any time. This AR guidance interface with multi-information fusion allows even newcomers to quickly adapt to the complex collaborative work environment and improves the flexibility of human resource allocation. The above modules are connected through a closed loop of data flow. Personnel positioning data updates the digital twin model, dynamic safety fences are generated based on the trajectory of the smart capsule, and AR commands integrate task requirements and safety information. Together, they form a complete human-machine collaborative link from environmental perception, risk warning to task execution, which significantly improves the safety and efficiency of human-machine mixed operations in confined spaces.

[0036] Optionally, in the data fusion and construction unit, the geometry layer is used to store three-dimensional spatial coordinates, dimensions, and topological relationships, while the physical layer is used to store material properties, structural loads, and equipment performance parameters.

[0037] The pluggable task module is selected from one or more of the following modules: detection module, spraying module, reinforcement module and handling module; the standardized interface includes a standardized mechanical interface and an electrical interface; and the scheduling instructions include one or more of the following: task allocation instructions, path adjustment instructions, avoidance coordination instructions and cooperation instructions.

[0038] By adopting the above technical solution, the geometric layer is used to store three-dimensional spatial coordinates, dimensions, and topological relationships, clarifying the digital twin model's ability to geometrically describe physical space. The three-dimensional spatial coordinates provide an absolute positional reference for the intelligent capsule's positioning and path planning. Dimensional information is used to determine whether the intelligent capsule can pass through narrow passages. Topological relationships describe the connectivity between various construction areas, providing a graph theory basis for global path search. This structured storage of geometric information enables the adaptive scheduling unit to accurately determine spatial feasibility during path planning, solving the technical problem that relying solely on two-dimensional coordinates cannot handle the constraints of passage in narrow spaces. By limiting the physical layer to store material properties, structural loads, and equipment performance parameters, the physical characteristics of the construction object are incorporated into the digital twin model. Material properties affect the applicable construction techniques and parameters, structural load information is used to determine whether the work area can bear the weight of the intelligent capsule and its equipment, and equipment performance parameters constrain the execution of the intelligent capsule. The introduction of physical information, such as task types and operational scope, enables scheduling decisions to consider not only spatial feasibility but also the technical feasibility of task execution, avoiding the assignment of tasks beyond the capabilities of the intelligent capsule pod and solving the problem of matching task allocation with equipment capabilities. By limiting the selection of pluggable task modules to one or more of detection, spraying, reinforcement, and handling modules, the specific implementation form of the functional heterogeneity of the intelligent capsule pod is clarified. The standardized design of these functional modules allows for the rapid replacement of task modules on the same basic power platform according to the needs of the construction stage, improving equipment utilization, reducing project equipment procurement costs, and solving the problem of high idle rates of single-function equipment in traditional construction. By limiting standardized interfaces, including standardized mechanical and electrical interfaces, the technical basis for pluggable module replacement is clarified. The mechanical interface ensures the rigid connection and precise positioning of the module with the power platform, while the electrical interface realizes power supply, data communication, and control signal transmission. This dual-interface design allows for module replacement without the need for specialized tools and complex debugging, enabling rapid completion on-site and improving equipment relocation efficiency. It also solves the problems of difficult and impractical modular equipment replacement. By limiting scheduling commands to one or more of the following: task allocation commands, path adjustment commands, obstacle avoidance coordination commands, and collaborative commands, the adaptive scheduling unit clarifies multiple control dimensions of the intelligent capsule. The combined use of multiple command types enables the scheduling unit to perform refined behavioral control of the intelligent capsule, adapting to the complex and ever-changing needs of the construction site and solving the problem that a single command type cannot cover diverse control requirements.

[0039] Secondly, this invention provides an intelligent scheduling method for urban renewal, which utilizes the intelligent construction platform for urban renewal as described in the first aspect and employs the following technical solutions to achieve the invention's objective:

[0040] The intelligent scheduling method for urban renewal includes the following steps:

[0041] Steps for building a digital twin model: Obtain multi-source data of the urban renewal area, and build a four-layer digital twin model of the construction scene, including a geometric layer, a physical layer, a semantic layer, and a dynamic layer. The dynamic layer maps the location of the smart capsule warehouse, personnel location, environmental monitoring data, and dynamic obstacle data in real time.

[0042] The steps for real-time dynamic layer updates are as follows: Real-time sensing of the surrounding environment of the smart capsule and construction of a local occupation grid map and transmission of data, and real-time updating of the dynamic layer;

[0043] Steps for extracting state features: Extract system state features from the digital twin model at the current moment to form a high-dimensional state vector;

[0044] The steps for generating scheduling instructions are as follows: input the high-dimensional state vector into the pre-trained deep reinforcement learning model, output the collaborative scheduling strategy, generate scheduling instructions, and send them to the corresponding smart capsule warehouses.

[0045] Execution of scheduling instructions: The smart capsule pod executes the specified construction procedures according to the scheduling instructions through the installed standardized interface.

[0046] By adopting the above technical solution, a four-layer digital twin model of the construction scenario is constructed by acquiring multi-source data through the steps of building a digital twin model, including geometric, physical, semantic, and dynamic layers. This provides a unified data foundation for subsequent scheduling decisions. This method integrates multi-source heterogeneous data from urban renewal areas, such as BIM models, 3D point clouds, and underground pipeline data, into a single model framework, solving the fragmentation problems of scattered, inconsistent, and difficult-to-integrate data in traditional construction management. The dynamic layer maps the location of the smart capsule warehouse, personnel location, environmental data, and obstacle data in real time, ensuring that the state information on which decisions are based is always in line with the current situation. The system maintains synchronization with the actual scene, resolving the decision-making inaccuracies caused by information lag. By updating the dynamic layer in real time, it perceives the environment around the intelligent capsule and constructs a local occupancy grid map, achieving the fusion of micro-environmental information and macro-scene model. The local occupancy grid map details the distribution of obstacles around the intelligent capsule, providing a high-resolution environmental representation for accurate local path planning and obstacle avoidance. This information is updated to the dynamic layer in real time, ensuring that the digital twin model includes not only global static information but also dynamically changing local details, solving the problem that the global model's accuracy is insufficient to support fine-grained obstacle avoidance. By extracting state features, the system can... The system state features are extracted from the digital twin model of the previous moment to form a high-dimensional state vector. This transforms the complex construction site state into a feature representation that can be processed by a deep reinforcement learning model. This step achieves the extraction of key information from the perceived data, reduces the input dimension of the subsequent decision-making module, and retains the core information required for decision-making. This solves the problems of computational explosion and feature redundancy caused by directly inputting the original perceived data. In the step of generating scheduling instructions, the high-dimensional state vector is input into the pre-trained deep reinforcement learning model, which outputs a collaborative scheduling strategy and generates scheduling instructions to be issued to the corresponding smart capsules. This realizes intelligent decision-making based on global state perception. The deep reinforcement learning model learns the optimal scheduling strategy under complex constraints through offline training and can quickly output decisions in the current state. This solves the problems of long computation time and inability to respond to dynamic changes in real time in traditional optimization methods. The scheduling instructions are directly issued to the smart capsules, realizing an end-to-end closed loop from perception to execution. In the step of executing scheduling instructions, the smart capsules execute the specified construction procedures according to the installed standardized interfaces, completing the transformation from virtual decision-making to physical execution. This solves the problems of the scheduling system and execution equipment being independent of each other and poor information transmission, and provides systematic methodological support for intelligent scheduling of urban renewal construction.

[0047] Optionally, in the step of constructing the digital twin model, the semantic layer is used to store semantic attribute information in the urban renewal area. The semantic attribute information includes regional functional attributes, time constraint attributes, security attributes, and marked anisotropic constraint areas. The anisotropic constraint areas include pedestrian traffic directions, directional prohibition of entry into cultural relic protection areas, and traffic directions in narrow passages.

[0048] Between the step of extracting state features and the step of generating scheduling instructions, there is also a step of constructing a spatiotemporal joint potential field: based on the semantic attribute information stored in the semantic layer and the data mapped in real time in the dynamic layer, a spatiotemporal joint potential field describing the risk distribution of the job space is constructed, and the gradient or potential field value of the spatiotemporal joint potential field is integrated into the high-dimensional state vector.

[0049] The spatiotemporal joint potential field includes a static potential field and a dynamic potential field. The static potential field is constructed based on the regional functional attributes in the semantic layer, and its potential field strength is direction-dependent. The dynamic potential field is constructed and updated in real time based on the location data of the smart capsules, and a direction modulation factor is introduced based on the anisotropic constraint region to characterize the mutual avoidance requirements between smart capsules. Its repulsion strength varies with the angle between the smart capsules and the constraint direction. The repulsion strength function of the dynamic potential field is: ,in, Let q be the Euclidean distance between the current smart capsule and the obstacle smart capsule, and let q represent the pose vector of the current smart capsule in the configuration space, which includes position coordinates and orientation angle information. The radius of influence of the pre-set repulsive force. The relative speed between the current smart capsule and the obstacle smart capsule. The maximum relative speed is preset, η is the first gain coefficient, α is the second gain coefficient, and θ is the current motion direction angle of the smart capsule relative to the obstacle smart capsule. β represents the main constraint direction marked by the semantic layer, and β is the third gain coefficient.

[0050] By adopting the above technical solutions, and by limiting the semantic layer storage of semantic attribute information in the digital twin model construction step to include regional functional attributes, temporal constraint attributes, safety attributes, and anisotropic constraint areas of the labels, the construction rules and spatial constraints unique to urban renewal scenarios are integrated into the methodology in a structured form. These constraints are the essential characteristics that distinguish narrow spaces in urban renewal from open industrial environments. Explicitly modeling them as inputs for subsequent potential field construction provides a rule-based foundation for solving the problem of orderly spatial passage. By setting a step to construct a spatiotemporal joint potential field between the state feature extraction step and the scheduling instruction generation step, and limiting it to constructing a spatiotemporal joint potential field describing the risk distribution of the work space based on semantic attribute information and real-time data from the dynamic layer, a potential field-guided intermediate representation layer is introduced. This potential field integrates static rules with dynamic information to form a continuous quantitative description of the risk level of the work space, enabling subsequent reinforcement learning decisions to be optimized based on risk distribution rather than simple binary obstacle judgments. By integrating the gradient or potential field value of the spatiotemporal joint potential field into the high-dimensional state vector, the potential field information and reinforcement learning state are integrated. The deep integration of space enables deep reinforcement learning models to autonomously learn optimal strategies under potential field constraints during training, rather than simply following preset potential field rules. Simultaneously, the potential field information provides rich prior knowledge for reinforcement learning, solving the problem that pure reinforcement learning methods require extensive trial and error to learn basic avoidance behavior in complex constrained environments. By defining a spatiotemporal joint potential field, including static and dynamic potential fields, the static potential field is constructed based on regional functional attributes to express fixed rule constraints, while the dynamic potential field is constructed and updated in real-time based on the location data of the smart capsules. Furthermore, a direction modulation factor is introduced based on anisotropic constraint regions to express the dynamic avoidance needs between moving entities. This decomposition architecture optimizes the allocation of computational resources and provides a precise mathematical description with speed adaptation and direction selectivity. In the formula, the distance term ensures basic avoidance, the velocity term enables dynamic response, and the direction term satisfies anisotropic constraints. The synergy of these three terms allows the potential field to accurately reflect the complex interaction relationships between smart capsules in urban renewal scenarios. The adjustable gain coefficient allows the potential field model to adapt to the characteristics of different construction scenarios.

[0051] Optionally, it also includes augmented reality human-machine collaboration steps: receiving the generated scheduling instructions, converting them into augmented reality guidance information and pushing them to the on-site personnel terminals. The augmented reality guidance information includes work point markings, operation step guidance and safety prompts. At the same time, it tracks the on-site personnel's location in real time and updates it to the dynamic layer of the digital twin model. It generates dynamic safety fences in real time based on the movement trajectory of the smart capsule and triggers graded warnings when personnel enter.

[0052] By adopting the above technical solution, and by setting up augmented reality human-machine collaboration steps to receive generated scheduling instructions and convert them into augmented reality guidance information to push to on-site personnel terminals, abstract digital instructions are transformed into visual guidance that personnel can intuitively understand. This method allows scheduling instructions, originally only for the smart capsule, to simultaneously serve on-site personnel, achieving synchronization of human-machine task information. The guidance information includes work point markers, operation step guidance, and safety prompts, covering the core information elements required for personnel to perform tasks, solving the problems of task information asymmetry and personnel's lack of understanding of the smart capsule's intent in human-machine collaborative operations. By limiting the augmented reality guidance information to include work point markers, operation step guidance, and safety prompts, multi-dimensional information fusion is achieved. By tracking the on-site personnel's location in real time and updating it to the dynamic layer of the digital twin model, personnel are incorporated as dynamic elements into the global state perception system, enabling scheduling decisions to perceive personnel distribution in real time and fully consider personnel factors when planning paths and allocating tasks. Location data also provides a basis for the generation of dynamic safety fences, realizing two-way perception between humans and machines. By generating dynamic safety fences in real time based on the movement trajectory of the smart capsule and triggering graded warnings when personnel enter, an active safety protection mechanism is constructed. The dynamic safety fence updates in real time with the movement of the smart capsule, accurately covering dangerous areas. The graded warning mechanism adopts differentiated responses based on the spatial relationship between personnel and the fence, avoiding efficiency losses due to over-conservatism. Warning information can be directly presented to personnel through the AR interface, solving the problem of traditional safety protection relying on personnel self-awareness and delayed response. Scheduling commands drive the operation of smart capsules and trigger AR guidance information. Personnel location is fed back to the dynamic layer in real time, affecting subsequent scheduling decisions. The dynamic safety fence is updated in real time based on the trajectory of the smart capsule and the location of personnel, forming a safety closed loop. The steps are interconnected through data flow, forming a complete methodology from intelligent scheduling to human-machine collaboration, significantly improving the safety, efficiency, and collaboration of human-machine hybrid operations at urban renewal construction sites.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] 1. By setting up data fusion and construction units, a four-layer digital twin model comprising geometric, physical, semantic, and dynamic layers was built, achieving multi-dimensional digital mapping of urban renewal construction scenarios. This provides a real-time and accurate state awareness foundation for subsequent scheduling decisions. By setting up a cluster of intelligent capsule pods, the modularity and functional reconfigurability of construction execution terminals were achieved. The environmental perception and communication module constructs a local occupancy grid map in real time and transmits data, providing the platform with micro-environmental information surrounding the intelligent capsule pods. Pluggable task modules are quickly replaced through standardized interfaces, allowing the same intelligent capsule pod to be flexibly configured with different functions according to construction process requirements, improving equipment utilization and scenario adaptability. Furthermore, by setting up an adaptive scheduling unit… The state space construction module extracts the system state features at the current moment from the digital twin model to form a high-dimensional state vector, transforming the complex construction site state into a feature representation that can be processed by a deep reinforcement learning model. The decision-making and scheduling module deploys a pre-trained deep reinforcement learning model, outputs a collaborative scheduling strategy based on the high-dimensional state vector, and generates scheduling instructions to be sent to the intelligent capsule, realizing intelligent decision-making based on global state perception. This technical solution effectively solves the technical problem of balancing real-time performance and global optimality in multi-machine collaborative scheduling under complex scenarios through a deeply coupled modular design, achieving a synergistic improvement in construction organization efficiency and safety, enabling refined control of the construction process, thereby improving the construction efficiency of urban renewal projects.

[0055] 2. By defining the functional attributes, temporal constraints, and security attributes of the semantic layer storage area, the construction rules, time windows, and safety requirements specific to urban renewal scenarios are integrated into the digital twin model in a structured form. By adding a conflict prediction and potential field construction module, and defining its construction based on semantic attribute information and real-time data from the dynamic layer to build a spatiotemporal joint potential field, a conflict prediction mechanism based on the potential field method is introduced. The spatiotemporal joint potential field integrates static rules with dynamic information, forming a continuous description of the risk distribution in the work space. This allows potential conflicts to no longer rely on discrete threshold judgments, but rather on the continuous quantification of risk levels through potential field values, improving the sensitivity and accuracy of conflict prediction. By using the gradient or potential field value of the spatiotemporal joint potential field as input features of the high-dimensional state vector, deep integration of potential field information and deep reinforcement learning models is achieved. This technical solution uses potential field features as... As part of the state space of the reinforcement learning model, it enables the model to autonomously learn the optimal strategy under potential field constraints during training, rather than simply following preset potential field rules. This fusion architecture of potential field guidance and reinforcement learning optimization retains the intuitive expressive ability of the potential field method for spatial constraints while leveraging the decision-making advantages of reinforcement learning in multi-objective optimization. It solves the technical problems of difficulty in handling complex collaborative tasks by relying solely on the potential field method and the difficulty in training reinforcement learning in sparse reward environments. The deep reinforcement learning model is trained to learn and output collaborative scheduling strategies under potential field constraints, ensuring that the generated scheduling instructions conform to spatial safety constraints. This reduces the computational cost of posterior conflict detection and correction, improves the real-time performance and executability of scheduling decisions, and enables refined control of the construction process, thereby improving the construction efficiency of urban renewal projects.

[0056] 3. By defining a spatiotemporal joint potential field, comprising a static potential field and a dynamic potential field, the static potential field is constructed based on the regional functional attributes in the semantic layer. It characterizes the risk distribution generated by fixed obstacles and rule constraints. This potential field does not change over time or only changes with the construction phase and can be pre-calculated and stored. The dynamic potential field is constructed and updated in real-time based on the location data of the smart capsule pods. It characterizes the mutual avoidance needs between mobile smart capsule pods and between smart capsule pods and mobile personnel. This potential field changes in real-time with the movement of the smart capsule pods and requires frequent updates. This decomposed architecture optimizes the allocation of computational resources while ensuring the accuracy of the potential field description, solving the technical problem of excessive computational overhead caused by the need for full real-time calculation of a single potential field. By defining the direction dependence of the potential field strength of the static potential field, it introduces the ability to express the special constraints of urban renewal scenarios, surpassing the traditional potential field. In traditional methods, the repulsive potential field generated by obstacles is usually isotropic, meaning the repulsion strength is the same from the center of the obstacle in all directions. However, in the narrow spaces of urban renewal, the prohibition requirements in some directions are much stronger than in others. Direction-dependent potential fields can accurately express this asymmetric constraint, allowing smart capsules to move more naturally along feasible directions when planning their paths, avoiding dead zones or oscillations caused by isotropic potential fields. By defining the repulsion strength function of the dynamic potential field as a function, relative velocity is considered. The distance term in the formula represents the basic repulsion strength that increases sharply as the distance decreases, meeting the basic requirements for safe avoidance. The velocity modulation term ensures that when two smart capsules are moving towards each other with a large relative velocity, the repulsion strength increases accordingly, thereby triggering avoidance behavior in advance and solving the problem of delayed response in dynamic environments caused by traditional static potential field methods.

[0057] 4. By defining semantic attribute information, including marked anisotropic constraint regions, and integrating urban renewal-specific constraints such as pedestrian traffic direction, directional prohibition in cultural relic protection areas, and traffic direction in narrow passages into the digital twin model in a structured form, these directional constraints are essential characteristics that distinguish urban renewal scenarios from open industrial environments. Explicitly modeling them is a key technology for solving orderly passage in narrow spaces. By defining a dynamic potential field based on anisotropic constraint regions and making its repulsion strength change with the angle between the smart capsule and the constraint direction, a mathematical characterization of directional constraints is achieved. Traditional isotropic potential fields cannot distinguish the degree of risk in different directions, which can lead to the smart capsule being constrained. In constrained regions, rigid or irregular behavior can occur. This technical solution introduces a direction modulation factor, enabling the potential field to sense the degree of deviation between the movement direction and the constraint direction, and adjust the repulsion strength accordingly to guide the intelligent capsule pod to move along the compliant direction. By expanding the repulsion strength function of the dynamic potential field, the repulsion strength increases when the movement direction is consistent with the main constraint direction, encouraging the pod to maintain that direction; when the movement direction is opposite to the main constraint direction, the repulsion strength decreases, suppressing reverse movement; and when the movement direction is perpendicular to the main constraint direction, a neutral effect is generated. This direction-selective potential field modulation allows the intelligent capsule pod to autonomously form an orderly flow in the constrained region, avoiding congestion and conflict caused by disorderly competition.

[0058] 5. By limiting the use of a multi-agent deep deterministic policy gradient algorithm, each intelligent capsule is treated as an agent, learning a distributed cooperative policy. An advanced reinforcement learning framework suitable for multi-agent continuous control scenarios is introduced, employing a centralized training and distributed execution architecture. During the training phase, global information guides the optimization of each agent's policy, while during the execution phase, each agent relies only on local observations for decision-making. This architecture ensures both the global optimality of multi-agent cooperation and meets the low-latency decision-making requirements of real-time scheduling, solving the technical problems of excessive computational latency in traditional centralized scheduling in large-scale cluster scenarios and the difficulty in ensuring global coordination in distributed scheduling. A reward function is designed to address these issues. The algorithm transforms the multi-objective optimization problem into a single-objective reward maximization problem. This reward function considers five dimensions simultaneously: construction progress, work efficiency, safety conflicts, energy consumption costs, and noise pollution control. The weight coefficients of each factor can be adjusted according to the specific needs of the project, enabling the trained strategy to flexibly switch between different project priorities. This solves the problem that traditional scheduling methods struggle to balance multi-objective optimization. The aforementioned algorithm architecture and feedback mechanism together constitute a complete technical chain from offline training to online fine-tuning, and from simulation environment to real-world scenario. The links are connected through a closed-loop data flow, forming an endogenous driving force for continuous strategy optimization, significantly improving the platform's adaptability and performance ceiling in long-term operation. Attached Figure Description

[0059] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0060] Figure 1 This is a system block diagram of the urban renewal intelligent construction platform according to an embodiment of the present invention;

[0061] Figure 2 This is a flowchart of the intelligent scheduling method for urban renewal according to an embodiment of the present invention. Detailed Implementation

[0062] The following will be based on embodiments of the present invention. Figure 1 and Figure 2 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0063] Example 1: This example discloses an intelligent construction platform for urban renewal, referring to... Figure 1 It includes a data fusion and construction unit, an intelligent capsule cluster, an adaptive scheduling unit, and an augmented reality human-machine collaboration unit.

[0064] Data fusion and construction unit: used to acquire multi-source data of urban renewal areas and construct a four-layer digital twin model of construction scenarios, including geometric, physical, semantic and dynamic layers;

[0065] The dynamic layer maps in real time the location data of the smart capsule, personnel location data, environmental monitoring data, and dynamic obstacle data. The semantic layer stores semantic attribute information in the urban renewal area, including regional functional attributes, time constraint attributes, safety attributes, and marked anisotropic constraint areas. Anisotropic constraint areas include personnel passage directions, directional prohibition of entry in cultural relic protection areas, and passage directions in narrow passages. The geometric layer stores three-dimensional spatial coordinates, dimensions, and topological relationships. The physical layer stores material properties, structural loads, and equipment performance parameters.

[0066] In this embodiment, the smart capsule warehouse status data specifically includes: location, speed, battery level, and current task progress, etc. Personnel location data is obtained in real time through the UWB positioning tag on the safety helmet, and environmental monitoring data specifically includes: noise, dust, wind speed, light, etc. Dynamic obstacle data is obtained through AI recognition of newly added material stacking by on-site cameras.

[0067] Intelligent capsule cluster: consists of multiple functionally heterogeneous intelligent capsules, each of which integrates an environmental perception and communication module and a pluggable task module.

[0068] Environmental perception and communication module: used to perceive the surrounding environment in real time, build a local occupancy grid map, and transmit data.

[0069] In this embodiment, the environmental perception and communication module is specifically a stereo vision camera, a 16-line LiDAR, or an ultrasonic sensor array, etc.

[0070] In this embodiment, the environmental perception and communication module is used to transmit data such as the current waiting status of the smart capsule, the estimated completion time, the list of executable process types, the operation accuracy index, the operation radius, the movement speed, the battery life and sensor configuration.

[0071] Pluggable task modules: These modules can be replaced and installed on the smart capsule via standardized interfaces to perform specified construction procedures.

[0072] The pluggable task module is selected from one or more of the following modules: detection module, spraying module, reinforcement module and handling module; the standardized interface includes a standardized mechanical interface and an electrical interface; and the scheduling instructions include one or more of the following: task allocation instructions, path adjustment instructions, avoidance coordination instructions and cooperation instructions.

[0073] In this embodiment, the detection module is specifically equipped with an infrared thermal imager, a hammer impact tester, and a 3D laser scanner; the spraying module is specifically equipped with a high-pressure airless sprayer and a film thickness detection sensor; the reinforcement module is specifically equipped with a miniature drilling machine, a grouting pump, and a weld flaw detector; and the handling module is specifically equipped with a lightweight robotic arm and a vision-guided grasping system.

[0074] Adaptive Scheduling Unit: Used to generate scheduling instructions based on deep reinforcement learning, including a state space construction module, a conflict prediction and potential field construction module, a decision scheduling module, and a feedback learning module.

[0075] State space construction module: used to extract the system state features at the current moment from the digital twin model and form a high-dimensional state vector.

[0076] Conflict prediction and potential field construction module: Based on semantic attribute information and real-time mapping of smart capsule warehouse location data, personnel location data, environmental monitoring data and dynamic obstacle data in the dynamic layer, it constructs a spatiotemporal joint potential field to describe the risk distribution of the work space. The gradient or potential field value of the spatiotemporal joint potential field is used as the input feature of the high-dimensional state vector. The deep reinforcement learning model is trained to learn and output the collaborative scheduling strategy under the potential field constraint.

[0077] The spatiotemporal joint potential field includes a static potential field and a dynamic potential field. The static potential field is constructed based on the regional functional attributes in the semantic layer, and its potential field strength is direction-dependent. The dynamic potential field is constructed and updated in real time based on the location data of the smart capsules, and a direction modulation factor is introduced based on the anisotropic constraint region to characterize the mutual avoidance requirements between smart capsules. Its repulsion strength varies with the angle between the smart capsules and the constraint direction. The repulsion strength function of the dynamic potential field is: ,in, Let q be the Euclidean distance between the current smart capsule and the obstacle smart capsule, and let q represent the pose vector of the current smart capsule in the configuration space, which includes position coordinates and orientation angle information. The radius of influence of the pre-set repulsive force. The relative speed between the current smart capsule and the obstacle smart capsule. The maximum relative speed is preset, η is the first gain coefficient, α is the second gain coefficient, and θ is the current motion direction angle of the smart capsule relative to the obstacle smart capsule. β represents the main constraint direction marked by the semantic layer, and β is the third gain coefficient.

[0078] Decision scheduling module: Deployed with a pre-trained deep reinforcement learning model, which is used to output a collaborative scheduling strategy based on a high-dimensional state vector, generate scheduling instructions and send them to the corresponding smart capsules;

[0079] The deep reinforcement learning model employs a multi-agent deep deterministic policy gradient algorithm, treating each smart capsule as an agent to learn a distributed cooperative policy. The training process includes: constructing a high-fidelity simulation environment based on a digital twin model, and designing a reward function R= ×Progress Rewards+ ×Efficiency Rewards- ×Conflict Punishment- ×Energy Penalty- × Disturbance punishment, among which As the first weighting coefficient, This is the second weighting coefficient. The third weighting coefficient, It is the fourth weighting coefficient. The fifth weighting coefficient is used for offline training in a simulation environment, with online fine-tuning of model parameters based on real feedback data.

[0080] In this embodiment, the progress reward in the reward function is specifically: a positive reward is given for completing the task objective; the efficiency reward is specifically: a positive reward is given for shortening the total construction period; the conflict penalty is specifically: a negative penalty is given for collisions or violations of safety distances; the energy consumption penalty is specifically: a negative penalty is given for unnecessary empty runs or waiting; and the disturbance penalty is specifically: a negative penalty is given for working during residents' rest periods.

[0081] Feedback learning module: This module collects the actual execution results of the intelligent capsule after executing the scheduling instructions, forms feedback samples, and periodically uses the feedback samples to incrementally train the deep reinforcement learning model to achieve adaptive optimization of the scheduling strategy.

[0082] Augmented Reality Human-Machine Collaboration Unit: This unit receives scheduling instructions generated by the decision-making and scheduling module and converts them into augmented reality guidance information to be pushed to on-site personnel terminals. It includes a personnel positioning and tracking module, a dynamic safety fence generation module, and an instruction generation and push module.

[0083] Personnel location and tracking module: used to track the location of personnel on site in real time and update it to the dynamic layer of the digital twin model.

[0084] Dynamic safety fence generation module: used to generate dynamic safety fences in real time based on the movement trajectory of the smart capsule, and trigger graded warnings when personnel enter.

[0085] Instruction generation and push module: Used to automatically generate and push augmented reality guidance information that includes work point markers, operation steps, and safety tips.

[0086] In this embodiment, for procedures requiring manual intervention, augmented reality guidance information is automatically generated and pushed to the AR glasses or mobile terminals of on-site personnel via a 5G network. The augmented reality guidance information specifically includes: work location markings: displaying the precise location requiring manual operation overlaid on the real-world image; operation step guidance: displaying the operation process in the form of text, arrows, and animations; and safety prompts: displaying the current location of the smart capsule, the boundary of the danger zone, and suggested evacuation routes.

[0087] In this embodiment, the scheduling instructions specifically include: the task execution sequence for each smart capsule, the estimated start and end time of each task, path planning for transfers between smart capsules, dynamic adjustment of smart capsule speed, local path replanning, task priority arbitration, and temporary waiting instructions.

[0088] In this embodiment, the specific types of conflicts include: inter-smart capsule path intersection conflict, smart capsule safe distance conflict between smart capsule and personnel conflict, smart capsule entering restricted area conflict, and task time window exceeding limit conflict.

[0089] Workflow Example 1: Comprehensive renovation of the exterior facade, pipelines, and landscape of old residential communities.

[0090] For the old residential area built in the 1980s, there are 8 six-story residential buildings. The project includes three parts: facade renovation, rainwater and sewage pipe network renovation and landscape improvement. The construction period is required to be 120 days, and the residents must live in the area normally during the construction period.

[0091] Platform Deployment:

[0092] Digital twin modeling: A 3D model of the community was acquired using drone oblique photography, imported from the original architectural drawings, and a digital twin model was constructed including each building, each household window, and underground pipelines. Semantic annotations were used to identify: construction areas, resident access routes, protected objects such as transformer boxes / ancient trees, and quiet nighttime periods.

[0093] Intelligent capsule warehouse cluster configuration: Detection warehouse × 4: for detecting hollow areas in exterior walls and pipeline detection; Spraying warehouse × 6: for spraying primer / topcoat on exterior walls; Cleaning warehouse × 2: for cleaning exterior walls; Reinforcement warehouse × 2: for local structural reinforcement; Handling warehouse × 2: for material transfer.

[0094] Scheduling process:

[0095] Week 1: The platform prioritizes buildings 1-8 based on resident complaint history and the degree of building damage, generating inspection plans accordingly. Four inspection units operate in parallel, from 6:00 AM to 10:00 PM daily, automatically returning to their charging stations after 10:00 PM.

[0096] Week 2: Inspection data is uploaded to the platform. AI automatically generates a map marking the hollow areas. The platform groups the hollow areas by area, generates reinforcement task packages, and assigns them to reinforcement warehouses. At the same time, cleaning operations are initiated for areas that have passed inspection.

[0097] Weeks 3-8: Spraying operations are in full swing. The platform dynamically adjusts the spraying plan based on real-time wind speed sensor data: when the wind speed is >5m / s, it automatically suspends high-rise spraying and switches to lower floors or to buildings on the leeward side.

[0098] Dynamic Event Response: At 14:30 on a certain day, while Detection Module 3 was operating on the north side of Building 5, the sensors detected an abnormal underground cavity. The platform immediately: suspended all operations in the area, marked the danger zone in the digital twin model, dispatched Detection Module 4 carrying a ground-penetrating radar to conduct a re-measurement, automatically notified the project manager and pushed the re-measurement data, and after confirming the cavity, generated a reinforcement plan and dispatched the reinforcement module to handle it.

[0099] Implementation results: The actual construction period was 98 days, 22 days ahead of schedule; 3 complaints were received from residents during the construction period, compared to an average of 25 for similar projects; no safety accidents occurred.

[0100] Workflow Example 2: Restoration of stone slab roads in historical districts + micro-renovation of facades.

[0101] A historical and cultural district needs to undergo stone pavement repair, facade renovation of buildings along the street, and undergrounding of pipelines. The district is required to protect its historical features to the greatest extent possible, and will be partially open to the public during the construction period.

[0102] Special restrictions: Mechanical work is prohibited within 3 meters of the protected historical buildings. The stone slab road must be restored using the original techniques. Some procedures need to be completed manually. The street is open during the day, and the main construction period is from 22:00 to 6:00.

[0103] Platform Deployment:

[0104] Digital twin: A street block model with millimeter-level precision is obtained using 3D laser scanning, and the boundaries of each protected building and each stone slab that needs to be preserved are manually marked.

[0105] Smart capsule customization:

[0106] Ultra-thin inspection chamber: used for detecting voids under stone slabs; Miniature excavation chamber: used for excavating pipeline trenches; Antique-style spraying chamber: used for facade repair, with a built-in database of historical building color schemes.

[0107] Scheduling strategy:

[0108] Night mode: 22:00-6:00, the platform schedules all workable warehouses to work together, prioritizing the completion of main tasks.

[0109] Daytime mode: 6:00-22:00, only low-noise, small-volume testing chambers are scheduled to operate at the edge of open areas.

[0110] Cultural heritage protection area: Set hard constraint boundaries in the digital twin model. Any smart capsule will automatically trigger deceleration within 3 meters of the boundary, stop within 2 meters, and trigger bidirectional emergency stop and report within 1 meter.

[0111] Implementation results: The renovation was completed on schedule while ensuring the normal daytime opening of the block, with zero contact and zero damage to the cultural heritage buildings.

[0112] Workflow Example 3: Functional upgrade of large public buildings, internal renovation + construction without interruption of business.

[0113] A large shopping mall needs to upgrade its internal fire protection system and renovate its public areas. The mall will remain open for business during the construction period.

[0114] Special constraints: The construction area and business area are dynamically divided, alternating between the business area during the day and the construction area at night. The area has high and dense pedestrian traffic, posing extremely high safety risks. Material transportation must pass through the business area, and must avoid peak customer traffic times.

[0115] Platform Deployment:

[0116] Dynamic zoning: The digital twin model is linked to the mall's business hours, marking the business area as a "restricted area" during the day and switching to a "construction area" at night.

[0117] Obstacle avoidance in crowded scenes: The intelligent capsule cabin's sensing unit has been upgraded to a 4D imaging radar, which can identify pedestrians' intentions and predict their movement trajectories in dense crowds.

[0118] AR human-machine collaboration: For processes that need to be carried out at the edge of the business area, the platform pushes real-time monitoring images and early warning information to security personnel through AR glasses.

[0119] Results: The mall remained open during the renovation period, with no pedestrian-machine collisions or accidents, and customer traffic was not significantly affected.

[0120] The implementation principle of the urban renewal intelligent construction platform in this embodiment is as follows:

[0121] By setting up data fusion and construction units to build a four-layer digital twin model including a geometric layer, a physical layer, a semantic layer, and a dynamic layer, a multi-dimensional digital mapping of urban renewal construction scenarios is realized. Among them, the dynamic layer maps the location data of smart capsule warehouses, personnel location data, environmental monitoring data, and dynamic obstacle data in real time, providing a real-time and accurate status perception basis for subsequent scheduling decisions, and solving the problems of information lag and lack of transparency of on-site status in traditional construction management.

[0122] By defining the functional attributes, time constraints, and safety attributes of the semantic layer storage area, the construction rules, time windows, and safety requirements unique to urban renewal scenarios are integrated into the digital twin model in a structured form. The functional attributes define the functions of different areas, such as construction areas, material areas, and personnel passages, providing a basis for determining area types for subsequent conflict prediction. The time constraints define the time windows during which each area is allowed to operate, such as avoiding residents' rest periods, enabling scheduling decisions to automatically comply with disturbance control requirements. The safety attributes define safety rules such as safe distance thresholds and no-entry boundaries, providing a rule basis for the generation of dynamic safety fences.

[0123] The geometry layer stores three-dimensional spatial coordinates, dimensions, and topological relationships, clarifying the digital twin model's ability to geometrically describe physical space. The three-dimensional spatial coordinates provide an absolute positional reference for the intelligent capsule's positioning and path planning. The dimensional information is used to determine whether the intelligent capsule can pass through narrow passages. The topological relationships describe the connectivity between various construction areas, providing a graph theory basis for global path search. The structured storage of this geometric information enables the adaptive scheduling unit to accurately determine spatial feasibility when planning paths, solving the technical problem that relying solely on two-dimensional coordinates cannot handle the constraints of passage through narrow spaces.

[0124] By defining a physical layer for storing material properties, structural loads, and equipment performance parameters, the physical characteristics of the construction object are incorporated into the digital twin model. Material properties influence the available construction techniques and parameters, structural load information is used to determine whether the work area can withstand the weight of the smart capsule and its equipment, and equipment performance parameters constrain the types of tasks and work scope that the smart capsule can perform. The introduction of this physical information enables scheduling decisions to consider not only spatial feasibility but also the technical feasibility of task execution, avoiding the assignment of tasks beyond the capabilities of the smart capsule to the equipment, and solving the problem of matching task allocation with equipment capabilities.

[0125] By setting up a cluster of intelligent capsule pods and limiting each intelligent capsule pod to integrate an environmental perception and communication module and a pluggable task module, the modularity and functional reconfigurability of the construction execution terminal are realized. The environmental perception and communication module constructs a local occupancy grid map in real time and transmits data, providing the platform with micro-environmental information around the intelligent capsule pod. The pluggable task module can be quickly replaced through standardized interfaces, allowing the same intelligent capsule pod to be flexibly configured with different functions according to the needs of the construction process, thereby improving equipment utilization and scenario adaptability.

[0126] By limiting the pluggable task modules to one or more of the detection, spraying, reinforcement, and handling modules, the specific implementation form of the functional heterogeneity of the intelligent capsule pod is clarified. The standardized design of these functional modules enables the same basic power platform to quickly replace task modules according to the needs of the construction stage, thereby improving equipment utilization, reducing project equipment procurement costs, and solving the problem of high idle rate of single-function equipment in traditional construction.

[0127] By defining standardized interfaces, including standardized mechanical and electrical interfaces, the technical basis for pluggable module replacement is clarified. The mechanical interface ensures the rigid connection and precise positioning of the module with the power platform, while the electrical interface enables power supply, data communication, and control signal transmission.

[0128] This dual-interface design allows the module replacement process to be completed quickly on the construction site without the need for specialized tools and complex debugging, improving the efficiency of equipment relocation and solving the problems of difficult and impractical modular equipment replacement.

[0129] By setting up an adaptive scheduling unit, the state space construction module extracts the system state features at the current moment from the digital twin model to form a high-dimensional state vector, transforming the complex construction site state into a feature representation that can be processed by a deep reinforcement learning model. The decision scheduling module deploys a pre-trained deep reinforcement learning model, outputs a collaborative scheduling strategy based on the high-dimensional state vector, and generates scheduling instructions to be sent to the intelligent capsule warehouse, thus realizing intelligent decision-making based on global state perception.

[0130] By adding a conflict prediction and potential field construction module and limiting it to constructing a spatiotemporal joint potential field based on semantic attribute information and real-time data from the dynamic layer, a conflict prediction mechanism based on the potential field method is introduced. The spatiotemporal joint potential field integrates static rules and dynamic information to form a continuous description of the risk distribution in the work space. This makes potential conflicts no longer rely on discrete threshold judgments, but instead continuously quantify the degree of risk through potential field values, thereby improving the sensitivity and accuracy of conflict prediction.

[0131] By using the gradient or potential value of the spatiotemporal joint potential field as the input feature of the high-dimensional state vector, the deep integration of potential field information and deep reinforcement learning model is achieved. In traditional methods, the potential field method usually outputs obstacle avoidance instructions as an independent module, which is independent of the decision module of this invention. This technical solution uses the potential field feature as part of the state space of the reinforcement learning model, so that the model can learn the optimal strategy under the potential field constraint during the training process, rather than simply following the preset potential field rules.

[0132] This integrated architecture of potential field guidance and reinforcement learning optimization retains the intuitive ability of the potential field method to express spatial constraints, while leveraging the decision-making advantages of reinforcement learning in multi-objective optimization. It solves the technical problems that relying solely on the potential field method makes it difficult to handle complex collaborative tasks and that relying solely on reinforcement learning makes it difficult to train in sparse reward environments.

[0133] Deep reinforcement learning models are trained to learn and output collaborative scheduling strategies under potential field constraints, which makes the generated scheduling instructions conform to spatial safety constraints, reduces the computational cost of posterior conflict detection and correction, improves the real-time performance and executability of scheduling decisions, enables refined control of the construction process, and thus improves the construction efficiency of urban renewal.

[0134] By defining a spatiotemporal joint potential field, including a static potential field and a dynamic potential field, the potential field construction task is decomposed into two components with clear physical meanings. The static potential field is constructed based on the regional functional attributes in the semantic layer and is used to characterize the risk distribution generated by fixed obstacles and rule constraints. This potential field does not change with time or only changes with the construction stage, and can be pre-calculated and stored to reduce the burden of real-time computing.

[0135] The dynamic potential field is constructed and updated in real time based on the location data of the smart capsules. It is used to characterize the mutual avoidance needs between mobile smart capsules and between smart capsules and mobile personnel. This potential field changes in real time with the movement of the smart capsules and needs to be updated frequently. This decomposition architecture optimizes the allocation of computing resources while ensuring the accuracy of the potential field description, and solves the technical problem of excessive computing power overhead caused by the need to calculate the entire potential field in real time.

[0136] By defining the direction-dependent potential field strength of the static potential field, the ability to express the special constraints of urban renewal scenarios is introduced. In the traditional potential field method, the repulsive potential field generated by obstacles is usually isotropic, that is, the repulsive strength is the same from the center of the obstacle to all directions. However, in the narrow space of urban renewal, the prohibition requirements in some directions, such as towards the building wall, are much stronger than in other directions, such as along the alleyway. The direction-dependent potential field can accurately express this asymmetric constraint, so that the smart capsule can move more naturally along the feasible direction when planning the path, avoiding dead zones or oscillation problems caused by isotropic potential fields.

[0137] By defining the repulsion strength function of the dynamic potential field as , the relative velocity is taken into consideration. The distance term in the formula represents the basic repulsion strength that increases sharply as the distance decreases, which meets the basic requirements of safe avoidance. The velocity modulation term makes the repulsion strength increase accordingly when two smart capsules move towards each other and the relative velocity is large, thereby triggering avoidance behavior in advance and solving the problem of the traditional static potential field method's delayed response in dynamic environments.

[0138] The gain coefficients η and α in the above formula are adjustable parameters that can be optimized through simulation experiments or online learning. This allows the potential field model to adapt to the avoidance characteristics requirements of different construction scenarios, improves the scenario adaptability and generalization ability of the technical solution, enables refined control of the construction process, and thus improves the construction efficiency of urban renewal.

[0139] By limiting semantic attribute information, it also includes anisotropic constraint areas marked, and integrates urban renewal-specific constraints such as pedestrian traffic direction, directional prohibition of entry into cultural relic protection areas, and traffic direction of narrow passages into the digital twin model in a structured form. The pedestrian traffic direction constraint means that in densely populated areas, smart capsule warehouses should try to keep in line with the direction of pedestrian flow or reduce reverse crossing.

[0140] Directional restrictions in cultural relic protection zones mean that approaching the cultural relic itself in certain directions is strictly prohibited, while movement parallel to the boundary of the protection zone may be permitted. The direction of passage in narrow passages means that in narrow alleys, the first-in-first-out or specific direction of passage should be followed to avoid two-way congestion. These directional constraints are the essential characteristics that distinguish urban renewal scenarios from open industrial environments, and explicitly modeling them is the key technology for solving the orderly passage in narrow spaces.

[0141] By constructing a dynamic potential field based on an anisotropic constraint region and making its repulsion strength change with the angle between the smart capsule and the constraint direction, a mathematical characterization of directional constraints is achieved. Traditional isotropic potential fields cannot distinguish the degree of risk in different directions, which may cause the smart capsule to behave rigidly or violate regulations in the constrained region. However, this technical solution introduces a direction modulation factor, which enables the potential field to sense the degree of deviation between the movement direction and the constraint direction and adjust the repulsion strength accordingly to guide the smart capsule to move in the compliant direction.

[0142] By expanding the repulsion intensity function of the dynamic potential field, the repulsion intensity increases when the direction of motion is consistent with the main constraint direction, encouraging the maintenance of that direction; the repulsion intensity decreases when the direction of motion is opposite to the main constraint direction, suppressing reverse motion; and a neutral effect is generated when the direction of motion is perpendicular to the main constraint direction. This direction-selective potential field modulation enables the intelligent capsule to autonomously form an orderly flow in the constrained area, avoiding congestion and conflict caused by disorderly competition.

[0143] The aforementioned directional modulation mechanism, along with the potential field features, serves as the state vector input and forms a synergy with online fine-tuning in reinforcement learning. In the early stages of training, the directional modulation potential field provides prior guidance for the reinforcement learning model, accelerating convergence. In the later stages of training, the reinforcement learning model can optimize the β coefficient through online fine-tuning, making the directional modulation intensity adaptable to the measured traffic efficiency data on-site. This achieves refined control over the construction process, thereby improving the construction efficiency of urban renewal projects.

[0144] By employing a multi-agent deep deterministic policy gradient algorithm, each intelligent capsule is treated as an agent, learning a distributed cooperative policy. An advanced reinforcement learning framework suitable for multi-agent continuous control scenarios is introduced, adopting a centralized training and distributed execution architecture. During the training phase, global information is used to guide the optimization of each agent's policy, while during the execution phase, each agent relies only on local observations to make decisions. This architecture ensures both the global optimality of multi-agent cooperation and meets the requirement of low-latency decision-making in real-time scheduling. It solves the technical problems of excessive computational latency in traditional centralized scheduling in large-scale cluster scenarios and the difficulty in ensuring global coordination in distributed scheduling.

[0145] By defining the training process of the deep reinforcement learning model, a high-fidelity simulation environment based on the digital twin model is constructed, creating a technical path of virtual training and real execution. The simulation environment, built on the digital twin model, can accurately simulate the geometric constraints, physical characteristics, and dynamic changes of the construction site. This allows the reinforcement learning model to undergo millions of rounds of trial and error training in a zero-cost, zero-risk virtual environment, learning robust strategies in various edge scenarios. This effectively solves the problems of difficulty in obtaining samples, high trial and error costs, and uncontrollable safety risks faced when training directly on real construction sites.

[0146] By designing a reward function, the multi-objective optimization problem is transformed into a single-objective reward maximization problem. This reward function considers five dimensions simultaneously: construction progress, work efficiency, safety conflicts, energy consumption costs, and noise pollution control. The weight coefficients of each factor can be adjusted according to the specific needs of the project, enabling the trained strategy to flexibly switch between different project priorities. This solves the problem that traditional scheduling methods cannot take into account multi-objective optimization.

[0147] By setting up a feedback learning module and limiting its collection of actual execution results to form feedback samples, and by periodically incrementally training the deep reinforcement learning model, the scheduling strategy can continuously evolve with the accumulation of construction site data. For example, when it is found that a certain type of conflict occurs frequently in the actual environment but is not fully simulated in the simulation, incremental training can adjust the strategy to better cope with the real scenario. When the project enters different construction stages and the constraints change, the strategy can adaptively adjust to maintain optimality, solving the problem of traditional scheduling systems having fixed performance after deployment and being unable to adapt to dynamic changes on site.

[0148] The aforementioned algorithm architecture and feedback mechanism together constitute a complete technical link from offline training to online fine-tuning, and from simulation environment to real-world scenario. The links are connected through a closed-loop data flow, forming an endogenous driving force for continuous strategy optimization, which significantly improves the platform's adaptability and performance ceiling in long-term operation.

[0149] By limiting the scheduling instructions to one or more of the following: task allocation instructions, path adjustment instructions, avoidance coordination instructions, and collaboration instructions, the adaptive scheduling unit clarifies multiple control dimensions of the intelligent capsule pod. The combined use of multiple types of instructions enables the scheduling unit to perform refined behavior control of the intelligent capsule pod, adapt to the complex and ever-changing needs of the construction site, and solve the problem that a single type of instruction cannot cover diverse control needs.

[0150] The augmented reality human-machine collaboration unit receives scheduling instructions generated by the decision-making and scheduling module, converts them into augmented reality guidance information, and pushes them to the on-site personnel terminals. Traditional construction instruction transmission relies on walkie-talkies, paper work orders, or handheld terminal text displays, which is inefficient and prone to ambiguity. This technical solution transforms scheduling instructions into visual guidance information superimposed on the real scene, enabling personnel to intuitively understand task requirements, significantly reducing communication costs and operational error rates, and solving the technical problems of unintuitive instruction transmission and low collaboration efficiency in human-machine hybrid operations.

[0151] By setting up a personnel positioning and tracking module and limiting it to the dynamic layer of the digital twin model for real-time tracking of on-site personnel locations, personnel are incorporated into the global state perception system as dynamic elements on par with the intelligent capsule pods. In traditional construction management, personnel location information is isolated from the equipment scheduling system, resulting in scheduling decisions that cannot fully consider the impact of personnel distribution on operational safety. This technical solution maps personnel locations to the digital twin model in real time, enabling the adaptive scheduling unit to avoid personnel as dynamic obstacles when performing path planning and conflict prediction, or to proactively schedule intelligent capsule pods to the personnel's location for collaborative work.

[0152] By setting up a dynamic safety fence generation module, it generates dynamic safety fences in real time based on the movement trajectory of the smart capsule and triggers graded warnings when personnel enter, thus constructing an active safety protection mechanism. The dynamic safety fence is different from the traditional static isolation area. Its shape and position change in real time with the movement of the smart capsule, and it can accurately cover the working range and movement path of the smart capsule.

[0153] When the spatial relationship between personnel location and safety fence changes, the system triggers a tiered warning, which not only ensures personnel safety but also avoids a large amount of idle work space caused by overly conservative fixed isolation zones, thus solving the technical problem of balancing safety and efficiency in human-machine shared spaces.

[0154] By setting up an instruction generation and push module, it automatically generates augmented reality guidance information that includes work point markers, operation step instructions, and safety tips. This integrates multi-dimensional information into a single visual interface. The work point markers use spatial anchoring technology to precisely indicate the physical locations requiring human intervention. The operation step instructions display the operation process in the form of text, arrows, and animations, reducing the requirements for personnel experience. The safety tips display the location of the smart capsule, the boundary of the danger zone, and the evacuation route in real time, enabling personnel to perceive the surrounding risks at any time. This multi-information integrated AR guidance interface allows even newcomers to quickly adapt to the complex collaborative work environment and improves the flexibility of human resource allocation.

[0155] The modules mentioned above are connected through a closed-loop data flow. Personnel positioning data updates the digital twin model, dynamic safety fences are generated based on the trajectory of the smart capsule, and AR commands integrate task requirements and safety information. Together, they form a complete human-machine collaborative link from environmental perception and risk warning to task execution, which significantly improves the safety and efficiency of human-machine hybrid operations in narrow spaces.

[0156] The modules are tightly coupled through data flow and control flow, forming a complete technical closed loop from environmental perception, state representation, intelligent decision-making to instruction execution. Compared with the existing architecture where the scheduling system and execution equipment are independent, this technical solution effectively solves the technical problem of balancing real-time performance and global optimization in multi-machine collaborative scheduling under complex scenarios through deeply coupled modular design. It achieves a synergistic improvement in construction organization efficiency and safety, enables refined management and control of the construction process, and thus improves the construction efficiency of urban renewal.

[0157] Example 2: This example discloses an intelligent scheduling method for urban renewal, referring to... Figure 2 The steps include: constructing a digital twin model (S1), updating the dynamic layer in real time (S2), extracting state features (S3), constructing a spatiotemporal joint potential field (S4), generating scheduling instructions (S5), executing scheduling instructions (S6), and augmented reality human-machine collaboration (S7).

[0158] S1. Steps for building a digital twin model: Obtain multi-source data of the urban renewal area and build a four-layer digital twin model of the construction scene, including geometric, physical, semantic and dynamic layers.

[0159] The dynamic layer maps the location of the smart capsule warehouse, personnel location, environmental monitoring data, and dynamic obstacle data in real time. The semantic layer stores semantic attribute information in the urban renewal area, including regional functional attributes, time constraint attributes, safety attributes, and marked anisotropic constraint areas. Anisotropic constraint areas include personnel passage direction, directional prohibition of entry in cultural relic protection areas, and passage direction of narrow passages. The geometric layer stores three-dimensional spatial coordinates, dimensions, and topological relationships. The physical layer stores material properties, structural loads, and equipment performance parameters.

[0160] S2, Real-time Update of Dynamic Layer: Real-time perception of the surrounding environment of the smart capsule, construction of a local occupation grid map and transmission of data, and real-time update of the dynamic layer.

[0161] S3. Extracting State Features: Extract system state features from the digital twin model at the current moment to form a high-dimensional state vector.

[0162] S4. Steps for constructing a spatiotemporal joint potential field: Based on the semantic attribute information stored in the semantic layer and the data mapped in real time in the dynamic layer, construct a spatiotemporal joint potential field that describes the risk distribution of the job space, and integrate the gradient or potential field value of the spatiotemporal joint potential field into the high-dimensional state vector.

[0163] The spatiotemporal joint potential field includes a static potential field and a dynamic potential field. The static potential field is constructed based on the regional functional attributes in the semantic layer, and its potential field strength is direction-dependent. The dynamic potential field is constructed and updated in real time based on the location data of the smart capsules, and a direction modulation factor is introduced based on the anisotropic constraint region to characterize the mutual avoidance requirements between smart capsules. Its repulsion strength varies with the angle between the smart capsules and the constraint direction. The repulsion strength function of the dynamic potential field is: ,in, Let q be the Euclidean distance between the current smart capsule and the obstacle smart capsule, and let q represent the pose vector of the current smart capsule in the configuration space, which includes position coordinates and orientation angle information. The radius of influence of the pre-set repulsive force. The relative speed between the current smart capsule and the obstacle smart capsule. The maximum relative speed is preset, η is the first gain coefficient, α is the second gain coefficient, and θ is the current motion direction angle of the smart capsule relative to the obstacle smart capsule. β represents the main constraint direction marked by the semantic layer, and β is the third gain coefficient.

[0164] S5. Step to generate scheduling instructions: Input the high-dimensional state vector into the pre-trained deep reinforcement learning model, output the collaborative scheduling strategy, generate scheduling instructions and send them to the corresponding smart capsule warehouses.

[0165] S6. Execution of scheduling instructions: The intelligent capsule pod executes the specified construction procedures according to the scheduling instructions through the installed standardized interface.

[0166] S7. Augmented Reality Human-Machine Collaboration Steps: Receive the generated scheduling instructions, convert them into augmented reality guidance information and push them to the on-site personnel terminals. The augmented reality guidance information includes work point markings, operation step guidance and safety prompts. At the same time, track the on-site personnel's location in real time and update it to the dynamic layer of the digital twin model. Generate dynamic safety fences in real time based on the movement trajectory of the smart capsule and trigger graded warnings when personnel enter.

[0167] The implementation principle of the intelligent scheduling method for urban renewal in this embodiment is as follows:

[0168] By constructing a digital twin model, multi-source data is obtained and a four-layer digital twin model of the construction scenario, including geometric, physical, semantic, and dynamic layers, is built. This provides a unified data foundation for subsequent scheduling decisions. The method integrates multi-source heterogeneous data of urban renewal areas, such as BIM models, 3D point clouds, and underground pipeline data, into the same model framework, solving the fragmentation problem of scattered, different formats, and difficult comprehensive application of data in traditional construction management.

[0169] The dynamic layer maps the location of the smart capsule, personnel location, environmental data, and obstacle data in real time, ensuring that the status information on which decision-making depends is always synchronized with the actual situation on site, thus solving the problem of inaccurate decision-making caused by information lag.

[0170] By updating the dynamic layer in real time to perceive the environment around the smart capsule and constructing a local occupancy grid map, the integration of micro-environmental information and macro-scene model is achieved. The local occupancy grid map describes in detail the distribution of obstacles around the smart capsule, providing a high-resolution environmental representation for accurate local path planning and obstacle avoidance. This information is updated to the dynamic layer in real time, so that the digital twin model not only contains global static information, but also dynamically changing local details, solving the problem that the accuracy of the global model is insufficient to support fine obstacle avoidance.

[0171] By extracting state features, system state features are extracted from the digital twin model at the current moment to form a high-dimensional state vector. This transforms the complex construction site state into a feature representation that can be processed by a deep reinforcement learning model. This step achieves the extraction of key information from the perception data, reduces the input dimension of the subsequent decision-making module, and retains the core information required for decision-making. It solves the problems of computational explosion and feature redundancy caused by direct input of the original perception data.

[0172] By limiting the semantic layer storage of semantic attribute information, including regional functional attributes, time constraint attributes, security attributes, and marked anisotropic constraint areas, the construction rules and spatial constraints unique to urban renewal scenarios are integrated into the methodology in a structured form. These constraints are the essential characteristics that distinguish narrow spaces in urban renewal from open industrial environments. By explicitly modeling them as inputs for subsequent potential field construction, a rule-based foundation is provided for solving the problem of orderly spatial passage.

[0173] By setting a step of constructing a spatiotemporal joint potential field between the step of extracting state features and the step of generating scheduling instructions, and limiting it to constructing a spatiotemporal joint potential field describing the risk distribution of the job space based on semantic attribute information and real-time data of the dynamic layer, a potential field-guided intermediate representation layer is introduced. This potential field integrates static rules and dynamic information to form a continuous quantitative description of the risk level of the job space, enabling subsequent reinforcement learning decisions to be optimized based on risk distribution rather than simple binary obstacle judgment.

[0174] By incorporating the gradient or potential value of the spatiotemporal joint potential field into a high-dimensional state vector, a deep fusion of potential field information and the reinforcement learning state space is achieved. This fusion enables deep reinforcement learning models to autonomously learn the optimal policy under potential field constraints during training, rather than simply following pre-defined potential field rules. Simultaneously, the potential field information provides rich prior knowledge for reinforcement learning, solving the problem that pure reinforcement learning methods require extensive trial and error to learn basic avoidance behaviors in complex constrained environments.

[0175] By defining a spatiotemporal joint potential field, including a static potential field and a dynamic potential field, the static potential field is constructed based on the functional attributes of the region to express fixed rule constraints, while the dynamic potential field is constructed and updated in real time based on the location data of the smart capsule, and a direction modulation factor is introduced based on the anisotropic constraint region to express the dynamic avoidance requirements between moving entities. This decomposition architecture optimizes the allocation of computing resources and provides a precise mathematical description with speed adaptation and direction selectivity.

[0176] In the formula, the distance term ensures basic avoidance, the velocity term enables dynamic response, and the direction term satisfies anisotropic constraints. The synergy of these three terms enables the potential field to accurately reflect the complex interaction between intelligent capsules in urban renewal scenarios. The gain coefficient is adjustable, allowing the potential field model to adapt to the characteristics of different construction scenarios.

[0177] By generating scheduling instructions, a high-dimensional state vector is input into a pre-trained deep reinforcement learning model, which outputs a collaborative scheduling strategy and generates scheduling instructions to be sent to the corresponding smart capsules. This achieves intelligent decision-making based on global state awareness. The deep reinforcement learning model learns the optimal scheduling strategy under complex constraints through offline training and can quickly output decisions in the current state. This solves the problems of long computation time and inability to respond to dynamic changes in real time in traditional optimization methods. The scheduling instructions are directly sent to the smart capsules, realizing an end-to-end closed loop from awareness to execution.

[0178] By executing scheduling instructions, the intelligent capsule warehouse performs designated construction procedures according to the standardized interfaces installed, completing the transformation from virtual decision-making to physical execution. This solves the problems of the scheduling system and execution equipment being independent of each other and poor information transmission, providing systematic methodological support for the intelligent scheduling of urban renewal construction.

[0179] By setting up augmented reality human-machine collaboration steps to receive generated scheduling instructions and convert them into augmented reality guidance information to push to on-site personnel terminals, abstract digital instructions are transformed into visual guidance that personnel can intuitively understand. This method enables scheduling instructions that were originally only for smart capsule warehouses to also serve on-site personnel, achieving synchronization of human-machine task information.

[0180] The guidance information includes work site markings, operation procedure instructions, and safety tips, covering the core information elements required for personnel to perform tasks. It solves the problems of information asymmetry in human-machine collaborative operations and personnel's lack of understanding of the intentions of the intelligent capsule warehouse.

[0181] By limiting augmented reality guidance information to include work site markers, operation procedure instructions, and safety tips, a multi-dimensional information fusion presentation is achieved.

[0182] By tracking the location of personnel on site in real time and updating it to the dynamic layer of the digital twin model, personnel are incorporated as dynamic elements into the global state perception system. This enables scheduling decisions to perceive personnel distribution in real time, fully consider personnel factors when planning paths and allocating tasks, and at the same time, personnel location data also provides a basis for judgment in the generation of dynamic safety fences, realizing two-way perception between humans and machines.

[0183] By generating dynamic safety fences in real time based on the movement trajectory of the smart capsule and triggering graded warnings when personnel enter, an active safety protection mechanism is constructed. The dynamic safety fences are updated in real time as the smart capsule moves, accurately covering dangerous areas.

[0184] The tiered early warning mechanism adopts differentiated responses based on the spatial relationship between people and the fence, avoiding efficiency losses due to excessive conservatism. Early warning information can be directly presented to people through an AR interface, solving the problem of traditional security protection relying on people's self-awareness and having a delayed response.

[0185] The dispatching instructions drive the operation of the smart capsule pods and trigger AR guidance information. The personnel location is fed back to the dynamic layer in real time, affecting subsequent dispatching decisions. The dynamic safety fence is updated in real time based on the trajectory of the smart capsule pods and the personnel location, forming a safety closed loop. The steps are interconnected through data flow, which together constitute a complete methodology from intelligent dispatching to human-machine collaboration, significantly improving the safety, efficiency and collaboration of human-machine hybrid operations at urban renewal construction sites.

[0186] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An intelligent construction platform for urban renewal, characterized in that, Includes the following modules: Data fusion and construction unit: used to acquire multi-source data of urban renewal areas and construct a four-layer digital twin model of construction scenarios, including geometric, physical, semantic and dynamic layers. The dynamic layer maps the location data of smart capsule warehouses, personnel location data, environmental monitoring data and dynamic obstacle data in real time. Intelligent capsule cluster: consists of multiple functionally heterogeneous intelligent capsules, each of which integrates an environmental perception and communication module and a pluggable task module; Environmental perception and communication module: used to perceive the surrounding environment in real time, build a local occupancy grid map, and transmit data; Pluggable task modules: These modules can be replaced and installed on the smart capsule via standardized interfaces to perform specified construction procedures. Adaptive scheduling unit: used to generate scheduling instructions based on deep reinforcement learning, including a state space construction module and a decision scheduling module; State space construction module: used to extract the system state features at the current moment from the digital twin model to form a high-dimensional state vector; Decision scheduling module: Deployed with a pre-trained deep reinforcement learning model, which is used to output a collaborative scheduling strategy based on a high-dimensional state vector, generate scheduling instructions and send them to the corresponding smart capsules.

2. The intelligent construction platform for urban renewal according to claim 1, characterized in that: In the data fusion and construction unit, the semantic layer is used to store semantic attribute information in the urban renewal area. The semantic attribute information includes regional functional attributes, time constraint attributes, and security attributes. The adaptive scheduling unit also includes a conflict prediction and potential field construction module: based on semantic attribute information and real-time mapped intelligent capsule location data, personnel location data, environmental monitoring data and dynamic obstacle data in the dynamic layer, it constructs a spatiotemporal joint potential field to describe the risk distribution of the work space, and uses the gradient or potential field value of the spatiotemporal joint potential field as the input feature of the high-dimensional state vector. The deep reinforcement learning model is trained to learn and output a collaborative scheduling strategy under the potential field constraint.

3. The intelligent construction platform for urban renewal according to claim 2, characterized in that: In the conflict prediction and potential field construction module, the spatiotemporal joint potential field includes a static potential field and a dynamic potential field. The static potential field is constructed based on the regional functional attributes in the semantic layer, and its potential field strength is direction-dependent. The dynamic potential field is constructed and updated in real time based on the location data of the smart capsules, and is used to characterize the mutual avoidance requirements between smart capsules. The repulsion strength function of the dynamic potential field is: ,in, Let q be the Euclidean distance between the current smart capsule and the obstacle smart capsule, and let q represent the pose vector of the current smart capsule in the configuration space, which includes position coordinates and orientation angle information. The radius of influence of the pre-set repulsive force. The relative speed between the current smart capsule and the obstacle smart capsule. The maximum relative velocity is preset, η is the first gain coefficient, and α is the second gain coefficient.

4. The intelligent construction platform for urban renewal according to claim 3, characterized in that: In the data fusion and construction unit, the semantic attribute information also includes marked anisotropic constraint regions, which include personnel passage direction, directional prohibition of entry into cultural relic protection areas, and passage direction of narrow passages; In the conflict prediction and potential field construction module, a dynamic potential field is also constructed based on the anisotropic constraint regions marked by the semantic layer to reflect the anisotropic constraints in the urban renewal scenario. The repulsion strength varies with the angle between the smart capsule and the constraint direction. The repulsion strength function of the dynamic potential field is further updated as follows: Where θ is the current intelligent capsule's direction of motion relative to the obstacle intelligent capsule. β represents the main constraint direction marked by the semantic layer, and β is the third gain coefficient.

5. The intelligent construction platform for urban renewal according to claim 1, characterized in that: In the decision-making and scheduling module, a multi-agent deep deterministic policy gradient algorithm is used, treating each intelligent capsule as an agent to learn a distributed collaborative strategy. The training process of the deep reinforcement learning model includes: constructing a high-fidelity simulation environment based on a digital twin model, and designing a reward function R= ×Progress Rewards+ ×Efficiency Rewards- ×Conflict Punishment- ×Energy Penalty- × Disturbance punishment, among which As the first weighting coefficient, This is the second weighting coefficient. The third weighting coefficient, It is the fourth weighting coefficient. The fifth weight coefficient is used for offline training in a simulation environment, and the model parameters are fine-tuned online based on real feedback data. The adaptive scheduling unit also includes a feedback learning module: used to collect the actual execution results of the intelligent capsule after executing the scheduling instructions, form feedback samples, and periodically use the feedback samples to incrementally train the deep reinforcement learning model to achieve adaptive optimization of the scheduling strategy.

6. The intelligent construction platform for urban renewal according to claim 1, characterized in that: It also includes an augmented reality human-machine collaboration unit: used to receive scheduling instructions generated by the decision scheduling module and convert them into augmented reality guidance information to push to the on-site personnel terminal, including personnel positioning and tracking module, dynamic safety fence generation module and instruction generation and push module; Personnel location and tracking module: used to track the location of personnel on site in real time and update it to the dynamic layer of the digital twin model; Dynamic safety fence generation module: used to generate dynamic safety fences in real time based on the movement trajectory of the smart capsule, and trigger graded warnings when personnel enter; Instruction generation and push module: Used to automatically generate and push augmented reality guidance information that includes work point markers, operation steps, and safety tips.

7. The intelligent construction platform for urban renewal according to claim 1, characterized in that: In the data fusion and construction unit, the geometry layer is used to store three-dimensional spatial coordinates, dimensions and topological relationships, and the physical layer is used to store material properties, structural loads and equipment performance parameters. The pluggable task module is selected from one or more of the following modules: detection module, spraying module, reinforcement module and handling module; the standardized interface includes a standardized mechanical interface and an electrical interface; and the scheduling instructions include one or more of the following: task allocation instructions, path adjustment instructions, avoidance coordination instructions and cooperation instructions.

8. An intelligent scheduling method for urban renewal, employing the intelligent construction platform for urban renewal as described in any one of claims 1-7, characterized in that, Includes the following steps: Steps for building a digital twin model: Obtain multi-source data of the urban renewal area, and build a four-layer digital twin model of the construction scene, including a geometric layer, a physical layer, a semantic layer, and a dynamic layer. The dynamic layer maps the location of the smart capsule warehouse, personnel location, environmental monitoring data, and dynamic obstacle data in real time. The steps for real-time dynamic layer updates are as follows: Real-time perception of the surrounding environment of the smart capsule cabin, construction of a local occupation grid map and transmission of data, and real-time updates of the dynamic layer; Steps for extracting state features: Extract system state features from the digital twin model at the current moment to form a high-dimensional state vector; The steps for generating scheduling instructions are as follows: input the high-dimensional state vector into the pre-trained deep reinforcement learning model, output the collaborative scheduling strategy, generate scheduling instructions, and send them to the corresponding smart capsule warehouses. Execution of scheduling instructions: The smart capsule pod executes the specified construction procedures according to the scheduling instructions through the installed standardized interface.

9. The intelligent scheduling method for urban renewal according to claim 8, characterized in that: In the step of constructing the digital twin model, the semantic layer is used to store semantic attribute information in the urban renewal area. The semantic attribute information includes regional functional attributes, time constraint attributes, security attributes, and marked anisotropic constraint areas. The anisotropic constraint areas include pedestrian traffic directions, directional prohibition of entry into cultural relic protection areas, and traffic directions in narrow passages. Between the step of extracting state features and the step of generating scheduling instructions, there is also a step of constructing a spatiotemporal joint potential field: based on the semantic attribute information stored in the semantic layer and the data mapped in real time in the dynamic layer, a spatiotemporal joint potential field describing the risk distribution of the job space is constructed, and the gradient or potential field value of the spatiotemporal joint potential field is integrated into the high-dimensional state vector. The spatiotemporal joint potential field includes a static potential field and a dynamic potential field. The static potential field is constructed based on the regional functional attributes in the semantic layer, and its potential field strength is direction-dependent. The dynamic potential field is constructed and updated in real time based on the location data of the smart capsules, and a direction modulation factor is introduced based on the anisotropic constraint region to characterize the mutual avoidance requirements between smart capsules. Its repulsion strength varies with the angle between the smart capsules and the constraint direction. The repulsion strength function of the dynamic potential field is: ,in, Let q be the Euclidean distance between the current smart capsule and the obstacle smart capsule, and let q represent the pose vector of the current smart capsule in the configuration space, which includes position coordinates and orientation angle information. The radius of influence of the pre-set repulsive force. The relative speed between the current smart capsule and the obstacle smart capsule. The maximum relative speed is preset, η is the first gain coefficient, α is the second gain coefficient, and θ is the current motion direction angle of the smart capsule relative to the obstacle smart capsule. β represents the main constraint direction marked by the semantic layer, and β is the third gain coefficient.

10. The intelligent scheduling method for urban renewal according to claim 8, characterized in that: It also includes augmented reality human-machine collaboration steps: receiving generated scheduling instructions, converting them into augmented reality guidance information and pushing them to on-site personnel terminals. The augmented reality guidance information includes work point markings, operation step guidance and safety prompts. At the same time, it tracks the on-site personnel's location in real time and updates it to the dynamic layer of the digital twin model. It generates dynamic safety fences in real time based on the movement trajectory of the smart capsule and triggers graded warnings when personnel enter.

Images