Archive robot scheduling method and device based on ai and digital twinning

Abstract

The application provides an archive robot scheduling method and device based on AI and digital twinning, and belongs to the technical field of robot scheduling. The method comprises the following steps: acquiring a current time sequence knowledge graph; for each robot digital twinning node, taking the robot digital twinning node as a center node, extracting a local subgraph centered on the node, combining a plurality of attribute values of any node in the local subgraph into a feature vector of the node, and performing weighted aggregation on the feature vectors of the nodes in the local subgraph except the center node to obtain an enhanced feature of the center node; determining a to-be-regulated robot digital twinning and a corresponding cooperative regulation instruction based on all the enhanced features; and performing cooperative scheduling on the physical robot corresponding to the to-be-regulated robot digital twinning based on the cooperative regulation instruction to execute a physical archive access operation. The application can accurately represent the correlation between the scheduling objects and accurately perform multi-robot cooperative scheduling.

Description

Technical Field

[0001] This application belongs to the field of robot scheduling technology, and more specifically, relates to a method and apparatus for scheduling archive robots based on AI and digital twins. Background Technology

[0002] In the field of intelligent archives management, with the advancement of archives digitization and intelligent construction, archives storage facilities are gradually adopting robots to complete the automated storage and retrieval of physical archives. Multi-robot collaborative scheduling has become the core technology direction for improving archives storage and retrieval efficiency and realizing intelligent management of storage facilities.

[0003] In existing technologies, most archival robot scheduling solutions rely solely on fixed instructions to control the robot's access to and from archives. This results in problems such as a lack of relational representation and low scheduling efficiency, leading to a lack of accurate data support for scheduling decisions, poor multi-robot collaboration, and issues such as path conflicts, unreasonable task allocation, and low response efficiency for accessing and storing archives. Consequently, these solutions fail to meet the high-efficiency scheduling requirements of smart archives.

[0004] Therefore, there is an urgent need for a method that can accurately characterize the relationships between scheduling objects to achieve precise collaborative scheduling of multiple robots. Summary of the Invention

[0005] This application provides a method and apparatus for scheduling archive robots based on AI and digital twins, in order to solve the technical problems of lack of relationship representation, low scheduling efficiency and difficulty in achieving accurate collaborative scheduling in the prior art, and to achieve the goal of accurately representing the relationship of scheduling objects and accurately coordinating multiple robots.

[0006] A first aspect of this application provides a method for scheduling an archive robot based on AI and digital twins, comprising: Obtain the current temporal knowledge graph, which is constructed based on the digital twins of archives and robots. It is used to represent the state information of each digital twin of archives and each digital twin of robots, as well as the association relationships between digital twins of archives and digital twins of robots, and between digital twins of robots themselves. The association relationships are represented by the dynamic edges in the current temporal knowledge graph. For each robot digital twin node, the robot digital twin node is taken as the central node. A local subgraph centered on this node is extracted. Multiple attribute values ​​of any node in the local subgraph are combined to form the feature vector of that node. The feature vectors of all nodes in the local subgraph except the central node are weighted and aggregated to obtain the enhanced features of the central node. The local subgraph includes the robot digital twin node, its neighboring robot digital twin nodes, and file digital twin nodes that have task or spatial associations. The attribute values ​​of the robot digital twin node include: the current position, remaining battery power, and current action state of the physical robot corresponding to the robot digital twin. The attribute values ​​of the file digital twin node include: the storage location and borrowing status of the physical file corresponding to the file digital twin. Based on all enhanced features, the digital twin of the robot to be controlled and the corresponding collaborative control commands are determined. Based on the coordinated control instructions, the physical robot corresponding to the digital twin of the robot to be controlled is coordinated and scheduled to perform physical file access operations.

[0007] A second aspect of this application provides an AI- and digital twin-based archive robot scheduling device, comprising: The knowledge graph acquisition module is used to acquire the current temporal knowledge graph, which is constructed based on archival digital twins and robot digital twins. It is used to represent the state information of each archival digital twin, the state information of each robot digital twin, and the association relationships between archival digital twins and robot digital twins, as well as between robot digital twins themselves. The association relationships are represented by dynamic edges in the current temporal knowledge graph. The enhanced feature acquisition module is used to extract a local subgraph centered on each robot digital twin node, taking that robot digital twin node as the central node. Multiple attribute values ​​of any node within this local subgraph are combined to form the feature vector of that node. The feature vectors of all nodes within this local subgraph, excluding the central node, are then weighted and aggregated to obtain the enhanced features of the central node. The local subgraph includes the robot digital twin node, its neighboring robot digital twin nodes, and file digital twin nodes with task or spatial associations. The attribute values ​​of the robot digital twin node include: the current position, remaining battery power, and current action state of the physical robot corresponding to the robot digital twin. The attribute values ​​of the file digital twin node include: the storage location and borrowing status of the physical file corresponding to the file digital twin. The module for determining control instructions is used to determine the digital twin of the robot to be controlled and the corresponding collaborative control instructions based on all enhanced features. The collaborative scheduling module is used to coordinate the scheduling of the physical robot corresponding to the digital twin of the robot to be controlled based on collaborative control instructions, so as to perform physical file storage and retrieval operations.

[0008] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described AI-based and digital twin-based file robot scheduling method.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described AI- and digital twin-based archive robot scheduling method.

[0010] The beneficial effects of the AI- and digital twin-based archive robot scheduling method and apparatus provided in this application are as follows: This paper acquires a current temporal knowledge graph based on digital twins of archives and robots. By dynamically representing various relationships through edges, it achieves an integrated and dynamic representation of the full-element status and relationships of physical archives and physical robot scheduling, providing complete basic data for subsequent archive robot scheduling. A local subgraph is extracted centered on the robot's digital twin, and enhanced features of the robot's digital twin are obtained. Multi-dimensional features of the robot itself and surrounding objects are integrated to form feature data that comprehensively reflects the robot's operating status and environment. Based on all enhanced features, the robot's digital twin to be controlled and the collaborative control instructions are determined, making the scheduling more aligned with the actual scheduling scenario and ensuring the accuracy and rationality of scheduling decisions. Based on the collaborative control instructions, the physical robot is scheduled to perform storage and retrieval operations, achieving a seamless connection between the virtual scheduling of the digital twin and the actual execution of the physical robot. This allows multiple archive robots to work collaboratively according to precise instructions, significantly improving robot collaboration and effectively avoiding path conflicts. This embodiment integrates AI and digital twin technology into the archive robot scheduling process, significantly improving the efficiency of archive storage and retrieval operations and robot resource utilization, achieving the effect of accurately representing the relationships between scheduling objects and accurately coordinating the scheduling of multiple robots. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1A flowchart illustrating an AI- and digital twin-based archival robot scheduling method provided in an embodiment of this application; Figure 2 A schematic diagram of the training process of a multi-agent cooperative scheduling model provided in an embodiment of this application; Figure 3 A structural block diagram of an AI- and digital twin-based archive robot scheduling device provided in an embodiment of this application; Figure 4 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0015] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the AI- and digital twin-based archival robot scheduling method provided in this application. The method can be executed by an electronic device and may include: S101: Obtain the current time series knowledge graph.

[0016] The current temporal knowledge graph is constructed based on digital twins of archives and digital twins of robots. It is used to represent the state information of each digital twin of archives, the state information of each digital twin of robots, and the association relationships between digital twins of archives and digital twins of robots, and between digital twins of robots. The association relationships are represented by dynamic edges in the current temporal knowledge graph.

[0017] In this embodiment, the temporal knowledge graph is a structured knowledge network that dynamically updates over time and integrates archival digital twins and robot digital twins. The archival digital twin is a digital mapping of physical archives in virtual space, and the robot digital twin is a digital mapping of a physical archive robot in virtual space. Dynamic edges are the interconnected links between the digital twins, updated in real time as the physical entities' states change. Obtaining the current temporal knowledge graph is a fundamental operation for archival robot scheduling; its core purpose is to acquire data that comprehensively and dynamically reflects the state and relationships of all scheduled objects.

[0018] Specifically, the current temporal knowledge graph is obtained, which contains digital twin nodes corresponding to all physical files and physical robots. The digital twin nodes of the files represent the core states such as storage location (e.g., 3 shelves on the 2nd floor of Area A) and borrowing status (which may include in storage, borrowed, or reserved). The digital twin nodes of the robots represent the core states such as current location (e.g., coordinates X: 10m, Y: 8m in the central area of ​​the warehouse), remaining power (e.g., 75%), and current action status (which may include idle, file retrieval, or movement). The dynamic edges between each node are clearly marked with association relationship types such as "spatial proximity", "task collaboration", and "task assignment", and the attributes of all edges are updated as the physical entity status changes.

[0019] In one embodiment of this application, the current temporal knowledge graph is obtained in the following way: Construct an initial graph structure. In the initial graph structure, each node represents each robot digital twin and each archive digital twin, and each edge represents the initial association relationship between robot digital twin nodes and between robot digital twin nodes and archive digital twin nodes. For each physical robot, the first real-time data of the physical robot is collected at a preset period. The first real-time data includes: the real-time position of the physical robot, the current remaining power, and the current action status. For each physical file, second real-time data of the physical file is collected according to a preset period. The second real-time data includes: the current storage location and borrowing status of the physical file. Based on the first real-time data and the second real-time data, a first association relationship between robot digital twins and a second association relationship between robot digital twins and document digital twins are determined. The first association relationship includes: a spatial proximity relationship determined based on the real-time position of each physical robot and a task collaboration relationship determined based on the current action state of each physical robot. The second association relationship includes: a spatial association relationship determined based on the real-time position of the physical robot and the storage location of the physical document, and a task association relationship determined based on the current action state of the physical robot and the borrowing status of the physical document. Based on each first real-time data, each second real-time data, each first association, and each second association, the initial graph structure is updated to obtain the updated initial graph structure, which serves as the current temporal knowledge graph.

[0020] In this embodiment, the initial graph structure is the basic framework of the time-series knowledge graph. It is a static knowledge network that has not been updated in real time and only contains digital twin nodes and initial association edges. The initial association edges are the carriers of the initial association relationships, used to represent the initial association relationships between nodes. They are generated by mapping the initial association relationships and are used to construct the initial static knowledge network, providing a benchmark for the subsequent updates of dynamic association relationships. The first real-time data is the status monitoring data of the physical robot, including its real-time location, remaining battery power, and current action status. The second real-time data is the status monitoring data of the physical archives, including their storage location and borrowing status. The first association relationship is the association relationship between robot digital twins, specifically divided into spatial proximity relationships (spatial proximity relationships determined based on the real-time location of the robot digital twins) and task collaboration relationships (collaborative operation relationships determined based on the action status of the robot digital twins). The second association relationship is the association relationship between the robot digital twin and the archive digital twin, specifically divided into spatial association relationships (spatial proximity relationships determined based on the location of the robot digital twin and the archive digital twin) and task association relationships (job assignment relationships determined based on the action of the robot digital twin and the borrowing status of the archive digital twin). The initial association relationship is a static relationship such as the association of physical robot operating areas and the association of archive shelf positions with physical robots, which is preset according to the initial layout of the warehouse. The initial association edge is used to represent this initial association relationship. Subsequent initial association edges will be replaced by dynamic association edges (corresponding to dynamic association relationships) generated based on real-time data.

[0021] Specifically, an initial graph structure is first constructed. For example, digital twin nodes are created corresponding to 8 physical robots and 3,000 physical files in the warehouse. Based on the warehouse's zoning layout, initial association edges are preset, such as "Physical robot No. 1 serves files in area A" and "Physical robot No. 2 and physical robot No. 3 have adjacent working areas," forming a static initial framework. Real-time data is collected at preset intervals: for each physical robot, the first real-time data is collected at a frequency of 1 second / time, including real-time position coordinates, the remaining battery percentage collected by the battery management system, and the current action status (e.g., idle, file retrieval, movement, charging, handover). For each physical file, the second real-time data is collected at a frequency of 5 seconds / time, including the storage location of the specific shelf coordinates and the borrowing status (e.g., in storage, borrowed, reserved, in transit). Based on all the collected first and second real-time data, dynamic association relationships are determined.

[0022] For example, regarding the determination of the first association relationship: for spatial proximity relationship, when the Euclidean distance between the real-time position coordinates of two physical robots is less than a first preset threshold (e.g., the first preset threshold is 3 meters), it is determined that there is a spatial proximity relationship between their corresponding robot digital twins, and a "proximity" dynamic edge can be established in the graph; for task collaboration relationship, when two physical robots are currently collaboratively performing the same file access task (e.g., one physical robot hands over a file to another physical robot to complete subsequent archiving), it is determined that there is a task collaboration relationship between their corresponding robot digital twins, and a "collaboration" dynamic edge can be established in the graph.

[0023] Regarding the determination of the second association relationship: For spatial association relationships, when the distance between the real-time location of the physical robot and the storage location of the physical file is less than the second preset threshold (e.g., the second preset threshold is 5 meters), it is determined that there is a spatial association relationship between the robot's digital twin and the file's digital twin, and a "nearby" dynamic edge can be established in the graph; For task association relationships, when the physical robot is currently performing a storage or retrieval task for a physical file (e.g., going to the file shelf, grabbing the physical file, or the physical file being transported by the physical robot), it is determined that there is a task association relationship between the robot's digital twin and the file's digital twin, and a "operating" or "task assigned" dynamic edge can be established in the graph.

[0024] Finally, all real-time data is updated to the attribute fields of the corresponding nodes in the initial graph, the dynamic associations determined above are used to replace the initial association edges, and the initial graph structure is fully updated to obtain the current time-series knowledge graph that reflects the actual state of the current warehouse.

[0025] As can be seen from the above, constructing an initial graph structure lays the framework foundation for the dynamic updating of the temporal knowledge graph, avoiding the resource consumption of repeated construction; collecting real-time data of physical entities at preset intervals ensures the real-time synchronization of the attributes of the temporal knowledge graph nodes with the state of the physical world, improving data timeliness; determining the relationships between physical robots and physical files based on data enables the quantification and accurate identification of these relationships, providing a clear basis for updating graph edges; and obtaining the temporal knowledge graph through dual updates of data and relationships allows the graph to comprehensively and dynamically reflect the actual state of the objects to be scheduled within the warehouse, providing real and effective data support for subsequent scheduling.

[0026] S102: For each robot digital twin node, take the robot digital twin node as the center node, extract a local subgraph centered on the node, combine multiple attribute values ​​of any node in the local subgraph into the feature vector of the node, and perform weighted aggregation on the feature vectors of each node in the local subgraph except the center node to obtain the enhanced feature of the center node.

[0027] The local subgraph includes the robot's digital twin node, its neighboring robot digital twin nodes, and file digital twin nodes that have task or spatial associations. The attribute values ​​of the robot digital twin node include: the current position, remaining battery power, and current action status of the physical robot corresponding to the robot digital twin. The attribute values ​​of the file digital twin node include: the storage location and borrowing status of the physical file corresponding to the file digital twin.

[0028] In this embodiment, a local subgraph refers to a subset of a temporal knowledge graph formed by radiating a single robot digital twin to its associated twins. A feature vector is a sequence of ordered numerical values ​​that transforms multiple attribute values ​​of a node. An enhanced feature is a comprehensive feature formed by fusing the features of the central node itself with the features of surrounding associated nodes. Neighboring robot digital twin nodes refer to robot digital twins that are spatially adjacent.

[0029] Specifically, taking the digital twin of a physical robot (e.g., robot No. 2) in the warehouse as the central node, extract the digital twin nodes of the robot's digital twin node, its neighboring robots No. 1 and No. 4 within a 3-meter radius, and the digital twin nodes of the archives that have task associations (e.g., archive C-12 to be retrieved) or spatial associations (e.g., archive D-09 shelf space within 5 meters) with the robot's digital twin, and form a local subgraph together; standardize all attribute values ​​of each node in this local subgraph (e.g., normalize position coordinates, etc.). After quantization of state values, a 64-dimensional feature vector is formed by combining "position component, state component, and exclusive attribute component" in a fixed dimension. The exclusive component of the robot digital twin node is the remaining power, and the exclusive component of the archive digital twin node is the borrowing status. For the feature vectors of all nodes in the local subgraph except for the digital twin of robot No. 2, preset weights can be set according to experience. Based on the preset weights (such as 0.4 for neighboring robots and 0.6 for related archives), weighted aggregation is performed to finally obtain the enhanced feature vector of the central node of the digital twin of robot No. 2.

[0030] In one embodiment of this application, before extracting a local subgraph centered on the robot digital twin node as the central node, combining multiple attribute values ​​of any node within the local subgraph into the feature vector of that node, and weighted aggregating the feature vectors of all nodes within the local subgraph except the central node to obtain the enhanced features of the central node, the method further includes: For each node in the local subgraph other than the central node, based on the components of the feature vector of the other node and the components of the feature vector of the central node, sub-features between the central node and the other node are extracted; for each other node, if the other node is a neighbor robot digital twin node, the sub-features include position features, task association features, and resource competition features; if the other node is an archive digital twin node, the sub-features include position features and task association features. The positional features are determined based on the positional components in the feature vectors of the other nodes and the positional components in the feature vector of the center node. The task association feature is determined for a digital twin node of an archive based on the borrowing state component in the feature vector of the digital twin node of the archive and the action state component in the feature vector of the central node. It is used to characterize whether there is a task assignment relationship between the central node and the digital twin node of the archive. For a digital twin node of a neighboring robot, it is determined based on the action state component and position component in the feature vector of the digital twin node of the neighboring robot and the central node respectively. It is used to characterize whether there is a task cooperation relationship between the two or whether they serve the same target archive. Resource competition characteristics are determined based on the position and motion state components in the feature vector of the central node, as well as the position and motion state components in the feature vectors of other robot digital twin nodes in the local subgraph besides the central node. They are used to characterize the competition intensity of multiple robot digital twin nodes for the same archive digital twin node or the same target position. Based on the characteristics of each item, calculate the item attention coefficient between the central node and the other nodes respectively; Based on the current task stage and the types of other nodes, the pre-configured fusion coefficient table is queried to obtain the dynamic fusion coefficient corresponding to the current task stage and node type. The fusion coefficient table stores the preset allocation values ​​of the weights of various sub-features corresponding to different node types under different task stages. Based on the attention coefficients of each sub-item and their corresponding dynamic fusion coefficients, the attention weight of the central node to the other nodes is obtained; Specifically, the feature vectors of all nodes within the local subgraph, excluding the central node, are weighted and aggregated to obtain the enhanced features of the central node, including: The enhanced features of the central node are obtained by weighted aggregation based on the feature vectors of each other node and their corresponding attention weights.

[0031] In this embodiment, the sub-features are single-type features extracted from the feature vectors of the central node and other nodes according to their correlation dimensions. Each node's feature vector consists of multiple components corresponding to different attribute dimensions. For example, the feature vector of a robot digital twin node includes components such as position coordinates, remaining battery power, and task status. For each other node, multiple sub-features are extracted by calculating the corresponding components of the same dimension in the central node's feature vector based on the components of each dimension of that other node's feature vector. The sub-feature attention coefficient is a numerical value representing the degree of correlation between the central node and other nodes under a single sub-feature. The fusion coefficient table is a pre-configured three-dimensional lookup table that stores the dynamic fusion coefficients (i.e., feature weight allocation values) corresponding to the task stage, node type, and shared feature type. This fusion coefficient table is based on the historical scheduling data of the physical robot, business operation rules, and domain expert experience and configuration, and is used to dynamically allocate the weights of each sub-feature according to the current scenario. The dynamic fusion coefficient is a weight value representing the importance of the corresponding sub-feature in the current task scenario; the higher the coefficient, the greater the contribution of the sub-feature to the calculation of the central node's attention weight. Attention weight is a numerical value that represents the overall attention paid by the central node to other nodes after integrating the attention coefficients of each component; the task stages include the file retrieval stage, the archiving stage, and the collaborative operation stage.

[0032] The fusion coefficient table in this application is a three-dimensional array Table[Stage, NodeType, FeatureType]. All dynamic fusion coefficients are uniformly limited to the value range of [0.1, 0.9]. The sum of the fusion coefficients corresponding to all feature items under the same task stage and the same node type is always equal to 1, ensuring weight normalization and avoiding weight imbalance.

[0033] The fusion coefficient table has a complete initialization generation process: Step 1: Extract at least 12 months of historical scheduling logs from the archive, and count the frequency of scheduling failures such as path conflicts and task timeouts caused by each feature under different task stages and node types; Step 2: Quantify the basic weight of each feature using the entropy weight method. The lower the feature entropy value, the greater its impact on the scheduling result, and thus a higher basic weight is assigned; Step 3: Introduce a correction coefficient for the archive's exclusive business rules: When storing or retrieving confidential files, increase the coefficient of the file node task association feature by 0.1~0.15; In areas with dense multi-robot operations, increase the coefficient of the robot node resource competition feature by 0.1~0.2; Step 4: Normalize all feature coefficients under the same scenario so that the sum of the coefficients is 1, and generate the initial fusion coefficient table for persistent storage in the scheduling system database.

[0034] During the operation of the scheduling system, the fusion coefficients support adaptive online updates: After all the file storage and retrieval tasks in each batch are completed, the task time, number of robot conflicts, and multi-machine task balance are collected to construct scheduling loss indicators; with minimizing scheduling loss as the optimization objective, the gradient descent method is used to fine-tune the corresponding dimension fusion coefficients. After the update, the values ​​are still constrained to the range of [0.1, 0.9], and the coefficients in the same scenario remain at 1; the updated coefficient table directly overwrites the old data table and is used for the next round of attention weight calculation, realizing automatic iterative optimization with warehouse operation scenarios.

[0035] Specifically, for each node in the local subgraph other than the central node, the node type is determined individually, and based on the components of the feature vectors of the other node and the central node, the sub-features between them are extracted: If the other node is a digital twin node of a neighboring robot, extract location features, task association features, and resource competition features; if the other node is a digital twin node of an archive, extract location features and task association features.

[0036] Based on the numerical values ​​of each sub-feature, the cosine similarity algorithm can be used to calculate the sub-attention coefficients between the central node and other nodes. The coefficients range from 0 to 1, with higher coefficients indicating stronger correlations in that dimension. Depending on the current task stage (e.g., the file retrieval stage) and the types of other nodes (e.g., digital twin nodes), the dynamic fusion coefficients corresponding to each sub-feature under the corresponding task stage and node type are matched in a pre-configured fusion coefficient table. For example, in the file retrieval stage, the dynamic fusion coefficient for the location feature of the digital twin node is 0.6, and the dynamic fusion coefficient for the task association feature is 0.4. Weighted fusion is then performed based on each sub-attention coefficient and its corresponding dynamic fusion coefficient. For example, during the file retrieval stage, the positional attention coefficient for a digital twin node of a certain file is 0.9, and the task-related attention coefficient is 1.0. The calculation 0.9 × 0.6 + 1.0 × 0.4 = 0.94 yields the attention weight of the central node for that other node. Here, 0.9 and 1.0 are instance-based attention coefficients calculated using the cosine similarity algorithm, used only to demonstrate the weighted fusion calculation process; the actual values ​​change dynamically with node status and relationships. If high-priority files are missed in multiple consecutive file retrieval tasks, or the scheduling loss index increases, the system automatically updates the file node fusion coefficients during the file retrieval stage using gradient descent. The positional feature coefficient is adjusted from 0.6 to 0.55, and the task-related feature coefficient is adjusted from 0.4 to 0.45, with the sum remaining 1. The updated coefficients are directly written into the fusion coefficient table for subsequent calculation of attention weights for all nodes. Subsequently, based on the local subgraph, the attribute values ​​of each node are combined into a feature vector, which is then weighted and aggregated with the attention weights of other nodes to finally obtain the enhanced features of the central node.

[0037] As can be seen from the above, extracting corresponding sub-features for different node types can accurately uncover the core correlation dimensions between nodes in the scheduling scenario and avoid interference from irrelevant features; calculating the sub-attention coefficients can quantify the degree of node correlation under a single feature dimension, providing an objective basis for weight calculation; matching the dynamic fusion coefficients with the task stage and node type can realize the scenario-based dynamic adjustment of attention weights, making the weights more in line with actual scheduling needs; and aggregating features on the dynamically fused attention weights can enhance the information of the core related nodes highlighted by the features, improving the accuracy of the features in representing the robot's scheduling state.

[0038] In one embodiment of this application, after obtaining the current time-series knowledge graph, the method further includes: Get the current task queue. The task queue is a set of tasks to be executed that are generated in real time based on the changes in the borrowing status of each physical file. Each task corresponds to a physical file to be accessed, and the tasks in the task queue are sorted from high to low according to their priority. For each task in the current task queue, calculate the task priority based on the request time and urgency identifier of the physical file corresponding to the task. The task priority is used as the attribute value of the digital twin node of the physical file, and updated in the current temporal knowledge graph to obtain the updated temporal knowledge graph. Specifically, for each robot digital twin node, this robot digital twin node is taken as the center node, and a local subgraph centered on this node is extracted, including: Based on the updated temporal knowledge graph, for each robot digital twin node, the robot digital twin node is taken as the central node, and a local subgraph centered on the node is extracted.

[0039] In this embodiment, the current task queue is an ordered set containing all pending file access tasks, generated in real time based on changes in the physical file borrowing status; the task priority is a quantitative value representing the urgency and execution order of the pending tasks; the request time is the time when the user initiates the file access request; the urgency level identifier is an urgency level set according to the file classification or usage requirements, where the file classification is a classification of the file's security level, such as confidential and ordinary. This file classification is the core basis for setting the urgency level, and different classifications correspond to different urgency levels. For example, confidential files correspond to the high urgency level, and ordinary files correspond to the ordinary urgency level.

[0040] Specifically, the physical archives borrowing status is monitored in real time through the archives management system. When the borrowing status changes, such as from in-stock to reserved or borrowed, or from borrowed to returned, a corresponding retrieval, filing, or return task is automatically generated. All unfinished tasks are temporarily stored in the order of their generation time, forming the current task queue. For each task in the current task queue, the priority is calculated using the quantitative formula "Task Priority = Urgency Coefficient × 0.7 + (Current Time - Request Time) / Preset Time Threshold × 0.3". The urgency coefficient is set according to the urgency indicator, which is determined by the archive's security classification: confidential archives correspond to a high urgency level with an urgency coefficient of 1.0, and ordinary archives correspond to a normal urgency level. The urgency coefficient is 0.5; the preset time threshold is 1 hour. For example, a request to retrieve a confidential file initiated at 9:00 on May 20, 2026, will have a priority of 1.0 × 0.7 + 5 / 60 × 0.3 = 0.725 by 9:05, which is higher than a regular file task initiated at the same time. The calculated task priority value (between 0 and 1) is added as a new attribute and updated to the attribute field of the digital twin node of the physical file. This supplements the current temporal knowledge graph with attributes, resulting in an updated temporal knowledge graph. Subsequent extraction of local subgraphs for each robot digital twin node is based on this updated temporal knowledge graph, ensuring that the local subgraph contains the core information of the task priority.

[0041] As can be seen from the above, generating a task queue in real time based on the file borrowing status ensures that no tasks are missed, enabling dynamic and real-time task management. Calculating task priorities by combining request time and urgency allows for scientific and quantitative task sorting, prioritizing high-priority and urgent tasks and improving operational response efficiency. Updating task priorities to a temporal knowledge graph allows for deep integration of graph data with task scheduling requirements, achieving data-driven workflow linkage. Extracting local subgraphs based on the updated graph allows for robot feature analysis combined with current task priorities, making subsequent scheduling decisions more aligned with actual task needs and avoiding resource waste.

[0042] S103: Determine the digital twin of the robot to be controlled and the corresponding collaborative control instructions based on all enhanced features.

[0043] In this embodiment, by comprehensively analyzing the enhanced features of all robot digital twins, the control targets suitable for the current task are selected, and standardized instructions that enable multi-robot collaborative operation are generated. The robot digital twin to be controlled is the one that, based on the enhanced feature analysis, is suitable for the current file digital twin access task and needs to receive scheduling instructions. The collaborative control instructions are standardized scheduling instructions generated for the robot digital twin to be controlled, strongly related to the target file access task, and include action requirements, time nodes, and coordination logic, to achieve coordinated operation of multiple robot digital twins. The enhanced features are obtained by constructing feature vectors for each node in a local subgraph centered on each robot digital twin node and then weighting and aggregating them. These features include the characteristics of the central node as well as the weighted aggregated features of other nodes within the local subgraph containing the central node.

[0044] Specifically, firstly, the requirements for the current file access task to be executed are clearly defined, such as retrieving file C-12 in area A and archiving file B-05 in area B, and this is used as the core of the task to drive subsequent scheduling; secondly, for each robot digital twin, a local subgraph is extracted with it as the central node, and enhanced features are constructed and generated. The enhanced features consist of two parts: the features of the central node itself and the weighted aggregated features of other nodes obtained through weighted aggregation. Finally, the enhanced features of all robot digital twins are quantitatively evaluated through comprehensive analysis. The evaluation dimensions include state adaptability (such as idle state and whether the remaining power meets the task requirements), spatial adaptability (such as the distance to the target file), and cooperative adaptability (such as the situation of operation conflicts with other robots). Combined with the actual task requirements of current file storage and retrieval, robot digital twins No. 2 and No. 4 are selected as the objects to be controlled. At the same time, cooperative control instructions are generated for the two: "Robot digital twin No. 2 departs from its current position, goes to file shelf C-12 in area A to perform file retrieval operation, and goes to the outbound area to wait after file retrieval is completed; Robot digital twin No. 4 moves 3 meters to the side of robot digital twin No. 2 to cooperate in avoiding path conflicts and remains in standby state, and will connect the archiving task after robot digital twin No. 2 completes file retrieval."

[0045] In one embodiment of this application, the determination of the digital twin of the robot to be controlled and the corresponding collaborative control instructions based on all enhanced features is obtained through a trained multi-agent collaborative scheduling model. The trained multi-agent cooperative scheduling model was obtained through the following training method: Initialize multiple agents in a simulation environment that includes digital twins of archives and digital twins of robots, with each agent corresponding to a digital twin of a robot; In the simulation environment, multiple rounds of task simulation are performed. At each time step, a local subgraph with each robot digital twin node as the center node is obtained. The attribute values ​​of each node in the local subgraph are combined into their respective feature vectors. The feature vectors of each node in the local subgraph except the center node are weighted and aggregated to obtain the enhanced features of each robot digital twin node, which serve as the current state of each agent. Each intelligent agent outputs simulated scheduling actions based on its own policy network and the current state. The simulated scheduling actions include the movement, file retrieval, archiving, avoidance, or multi-machine collaborative actions of the robot digital twin. After the robot digital twin performs simulated scheduling actions, it calculates the immediate reward based on the constrained reward function, and obtains the joint state of the next time step after the current time step ends. The joint state is a global state set composed of the enhanced features of all robot digital twin nodes. The constrained reward function includes task completion reward, path conflict penalty, and task balance reward. The current state, simulated scheduling action, immediate reward, and next joint state are combined into a quadruple as experience data storage. The experience data is used to iteratively optimize the agent's policy network. Training stops when the cumulative reward converges, resulting in the trained multi-agent cooperative scheduling model.

[0046] In this embodiment, by simulating multiple rounds of tasks in a simulation environment, the model learns the optimal collaborative scheduling strategy, ultimately achieving accurate determination of the controlled object and instructions. The multi-agent collaborative scheduling model is an intelligent model based on multi-agent reinforcement learning training, used to realize multi-robot collaborative scheduling decisions. An agent is an intelligent decision-making unit corresponding to a single robot digital twin in the simulation environment. The decision network is an algorithm network within the agent used to output scheduling actions based on the current state. The simulated scheduling action is the virtual operation action output by the agent and executed by the robot digital twin in the simulation environment. The constrained reward function is a quantization function containing reward and penalty terms, used to guide model optimization. The immediate reward is the single reward value calculated by the constrained reward function after the robot digital twin executes the simulated scheduling action. The joint state is the global state set composed of the enhanced features of all robot digital twins. The simulation environment is a virtual simulation environment that replicates the actual physical warehouse layout and operating rules. The time step is the smallest time unit in the simulation environment. The empirical data is the "state, action, reward, next state" quadruple data recording the agent training process. The cumulative reward is the sum of rewards from multiple rounds of task simulation.

[0047] In this embodiment, during the training phase, the enhanced features of the robot digital twin nodes corresponding to each agent are used as input to output simulated scheduling actions; during the application phase, the enhanced features of the robot digital twin nodes are used as model input to output the robot digital twin to be controlled and the corresponding collaborative control instructions.

[0048] like Figure 2 As shown, Figure 2 This is a schematic diagram of the training process of the multi-agent cooperative scheduling model in this application embodiment. The training process is performed according to the following steps: Specifically, in a simulation environment that replicates the actual physical warehouse layout (e.g., 8 robots, 3000 files, three work areas A / B / C) and work rules (e.g., file retrieval / archiving processes, path avoidance rules), 8 intelligent agents can be initialized, each uniquely corresponding to a robot digital twin. The simulation environment executes 500 rounds of task simulation. At each time step (1 second), a local subgraph centered on each robot digital twin node is acquired. Feature vectors for each node are constructed and weighted and aggregated to obtain enhanced features, which are then used as the current state of each intelligent agent. This differs from traditional independent deep Q-networks (Deep Q-networks). This application addresses the issue of dynamic non-stationarity in multi-robot random games by improving the DQN training architecture through a single-agent independent training method. Specifically, it establishes a globally shared joint experience replay pool, storing the quadruples (local state, simulated action, immediate single-agent reward, and global joint state) generated by all robot agents in this pool. During each gradient update, batch sampling of multi-agent joint experience samples synchronously updates the DQN policy network of all agents, enabling each agent to synchronously perceive changes in other robot policies during training. This eliminates the non-stationarity caused by dynamic changes in environmental transition probabilities and satisfies the mathematical conditions for DQN iterative convergence. Each agent, based on its own DQN decision network, outputs the Q-value (action value) corresponding to the simulated scheduling action according to the enhanced features of the current local subgraph, and selects the optimal simulated scheduling action based on the Q-value. For example, simulated scheduling actions include moving, retrieving files, archiving, and avoiding obstacles. After the robot digital twin executes the simulated scheduling action, it calculates the immediate reward based on the constrained reward function. For example, the immediate reward calculation can be +10 points for retrieving files, -20 points for path conflicts, and +5 points for task balancing. An additional global collaborative balancing penalty term is added to the constrained reward function: when multiple robots have path conflicts or task allocation imbalances, the global reward is deducted uniformly, forcing each agent to actively consider the global multi-machine operation status when making local decisions based solely on local subgraph observations. At the same time, resource competition features are introduced when calculating the sub-features of the local subgraph to quantify the competition intensity of other robots for the same file or space, so that the local features naturally carry cross-robot global correlation information, making up for the deficiency of a single robot only collecting local observations and lacking global information, and ensuring the effectiveness of multi-robot collaborative scheduling.

[0049] After the current time step ends, the enhanced features of all robot digital twins are acquired to form the joint state for the next time step. The global joint state data, including the current state, simulated scheduling actions, immediate rewards, and the next time step, are formed into a quadruple. This quadruple serves as the experience optimization material for model training, not as the input data for model decision-making. The quadruple is stored in a globally shared joint experience pool. When the experience pool reaches 100,000 data entries, a batch of multi-agent joint experience samples are randomly sampled to simultaneously optimize the global agent decision network using synchronous gradient descent. Training stops when the model's cumulative reward curve shows no significant fluctuations (convergence) for 20 consecutive rounds, resulting in the trained multi-agent cooperative scheduling model. When determining the robot digital twin to be regulated and the corresponding cooperative regulation instructions based on all enhanced features, the enhanced features are input into the multi-agent cooperative scheduling model, which then automatically outputs the cooperative regulation instructions.

[0050] The system establishes a one-to-one mapping generation logic between coordinated control commands, simulated scheduling actions, and the robot's digital twin state. For different types of simulated scheduling actions, the system combines the robot's current state (e.g., real-time location, task status, remaining battery power, surrounding environment) to generate specific, directly executable coordinated control commands. An example mapping relationship is as follows: If the simulated scheduling action is to retrieve or archive files, a retrieval or archiving instruction is generated, containing the target file shelf coordinates, the optimal work path, and the work sequence requirements. If the simulated scheduling action is to move or avoid obstacles, a move or avoidance instruction is generated, containing the target position coordinates, the obstacle avoidance path, and the travel speed constraints. If the simulated scheduling action is multi-machine collaboration, a collaborative work instruction is generated, containing the collaborative task division, work sequence, interaction rules, and conflict avoidance strategies. Through the above mapping logic, the abstract scheduling actions output by the model are transformed into specific control instructions that the physical robot can directly execute, realizing a complete closed loop from scheduling decision to physical execution.

[0051] The DQN decision network is the core component for mapping the state to actions of a single agent. Its network structure employs a multi-layer fully connected architecture consisting of an input layer, multiple hidden fully connected layers, and an output layer. For example, the network structure is as follows: the input dimension matches the dimension of the robot's digital twin's one-dimensional augmented feature vector (e.g., 128 dimensions); the dimensions of the three hidden fully connected layers are 256, 128, and 64 dimensions respectively; each hidden layer uses the ReLU activation function to introduce non-linear characteristics, enabling deep feature extraction, dimensional transformation, and associated feature mining of the one-dimensional augmented feature vector; the output layer dimension matches the dimension of the simulated scheduling action space, for example, 5 dimensions, corresponding to five types of actions: movement, fetching, archiving, avoidance, and multi-machine collaboration. A linear activation function can be used to output the Q-values ​​corresponding to each simulated scheduling action of the agent in the current state; the agent then... - Greedy strategy (initial) The value can be set to 0.9 based on experience, and then linearly decayed to 0.05 during training. The action with the largest Q value is selected as the simulated scheduling action for the final output.

[0052] As can be seen from the above, initializing the agent in a simulation environment and replicating the actual scheduling scenario allows the model training to be highly consistent with actual applications, reducing the adaptation cost of model deployment; using the enhanced features of the robot digital twin as the agent's state allows the agent to accurately perceive its own and the surrounding environment's scheduling information, improving the comprehensiveness of state representation; simulating scheduling actions covering the robot digital twin's full operational behavior allows the model to learn complete scheduling strategies; constrained reward functions take into account task completion, operational compliance, and resource balance, guiding the model to learn efficient and orderly collaborative scheduling strategies; iterative optimization through empirical data until reward convergence allows the model to possess stable optimal decision-making capabilities. Based on this model, determining the control objects and instructions can significantly improve the accuracy and efficiency of scheduling decisions, realizing intelligent collaborative scheduling of multiple robot digital twins.

[0053] S104: Based on the collaborative control instructions, perform collaborative scheduling of the physical robot corresponding to the digital twin of the robot to be controlled, so as to perform physical file storage and retrieval operations.

[0054] In this embodiment, the physical robot is an intelligent hardware device that actually performs file storage and retrieval operations within the warehouse. It has a unique correspondence with the robot's digital twin, ensuring the accuracy of command issuance. Through a one-to-one mapping between the virtual and physical worlds, the collaborative control commands from the virtual layer to the robot's digital twin are transformed into actual operations that the physical robot can recognize and execute, achieving a seamless transition from virtual scheduling to physical execution.

[0055] Specifically, a pre-defined mapping table can be used to identify the digital twins of robots 2 and 4, which correspond to physical robots numbered R-2 and R-4 in the warehouse, respectively. The generated collaborative control instructions are converted into standardized control codes (such as movement instruction codes and robotic arm operation codes) that can be recognized by the physical robot control system. The control codes are then sent to the main control modules of physical robots R-2 and R-4 according to the correspondence. After receiving the instruction, physical robot R-2 starts its movement module and moves to the C-12 file shelf in area A according to the preset path. It then uses its robotic arm to complete the file retrieval operation. Physical robot R-4 simultaneously performs movement and avoidance actions. The warehouse positioning system and robot feedback module monitor the execution status of the instructions to ensure that the operation is completed accurately.

[0056] In one embodiment of this application, during the process of a physical robot performing a physical file access operation, the method further includes: Real-time monitoring of the physical operating status of each physical robot, including the current remaining battery power and the actual execution trajectory; The actual execution trajectory of the physical robot is compared with the simulated predicted trajectory of the corresponding digital twin of the robot, and the trajectory deviation is calculated. When the trajectory deviation is greater than the preset deviation threshold, the expected pose at the corresponding time point in the simulation prediction trajectory is extracted, and the movement command in the collaborative control command is corrected with the expected pose as the target. The corrected movement command is then sent to the corresponding physical robot for execution. When it is detected that the current remaining power of any physical robot is less than the preset charging threshold, a charging task is generated. Based on the attribute values ​​of other robot digital twins in the current temporal knowledge graph, the unfinished tasks of the physical robot are reassigned to other available physical robots, and the corresponding task association edges in the current temporal knowledge graph are updated.

[0057] In this embodiment, during the physical robot's operation, real-time monitoring, deviation correction, and low battery handling ensure the accuracy and continuity of the operation, avoiding task delays or failures caused by trajectory deviations or equipment power outages; the collaborative control commands include movement commands, operation execution commands, and collaborative avoidance commands. The physical operating state refers to the real-time hardware state of the physical robot during actual operation. The actual execution trajectory is the actual movement path of the physical robot when performing the task. The simulation predicted trajectory is the optimal operation movement path of the physical robot generated in the simulation environment based on the scheduling decision results output by the robot digital twin based on the multi-agent cooperative scheduling model. Its generation logic is as follows: taking the enhanced features of the robot digital twin as input, the trained multi-agent cooperative scheduling model outputs cooperative control instructions, extracts the path planning information in the instructions, and reproduces the simulation predicted trajectory of the robot digital twin in the simulation environment. This trajectory is completely matched with the actual task, real-time state, and warehouse environment constraints of the physical robot and serves as a virtual benchmark for the actual execution trajectory of the physical robot. The trajectory deviation is the spatial deviation value between the actual execution trajectory and the simulation predicted trajectory. The preset deviation threshold is the maximum acceptable value of the trajectory deviation. The desired pose is the spatial position and attitude that the robot digital twin should be in at the corresponding time point in the simulation predicted trajectory. The preset charging threshold is the critical value of the remaining power of the physical robot that needs to be charged.

[0058] Specifically, during the physical robot's file retrieval operations, the warehouse positioning system collects the robot's actual execution trajectory in real time, at a frequency of 0.5 seconds per instance. The battery management system collects the robot's remaining battery power, forming physical operating status data. The actual execution trajectory of the physical robot is compared point-by-point with the simulated predicted trajectory of the corresponding robot's digital twin, and the trajectory deviation is calculated using the Euclidean distance algorithm. The collaborative control instructions include job execution instructions for the physical robot to perform core file retrieval and archiving operations, and collaborative avoidance instructions for multiple physical robots to coordinate and avoid path conflicts and maintain safe operating distances.

[0059] This application's document retrieval robot uses a two-wheeled differential AGV chassis. The chassis's fixed hardware parameters are uniformly calibrated: wheelbase L, single drive wheel radius r; the robot's global planar pose is uniformly denoted as a two-dimensional attitude vector. =( , , ), , Using Cartesian coordinates, This represents the robot's heading angle. The desired pose alone cannot directly drive the motors; a complete conversion process from the target pose to the motor execution signal is required through inverse kinematics and PID closed-loop control. The conversion steps are as follows: Step 1: Solving for the pose difference. Let the robot's current actual pose be... Simulation outputs the desired pose Calculate coordinate deviation , , angle deviation Fixed control cycle .

[0060] In this embodiment, the linear velocity of the vehicle body is solved using the global coordinate linear average rate. This simplified calculation method is suitable for routine trajectory correction scenarios in archive storage facilities with small pose deviations, minimal heading angle deviations, and near-straight-line approach to the target point. However, it is not suitable for situations involving in-situ rotation, large-angle turns, curve tracking, or extremely large heading angles between the target and the vehicle. For large-deviation turning situations, the system automatically breaks down the correction into multiple control cycles to avoid amplifying the tracking error of a single large deviation. To maintain computational simplicity, this application directly calculates the speed based on global Cartesian coordinate deviations without converting to the robot's body coordinate system to decouple translation and rotation. This is suitable for business scenarios involving low-speed file retrieval and small-amplitude trajectory corrections, meeting the trajectory correction accuracy requirements of the archive storage facility.

[0061] Step 2: Calculate the target wheel speed using inverse kinematics of the chassis. Solve for the target linear velocity of the vehicle body using coordinate deviation. Target angular velocity of the vehicle body Then, the target angular velocities of the left and right drive wheels are solved using the inverse differential wheel formula: , ,in The target angular velocity of the left wheel, The target angular velocity is represented by the right wheel, in rad / s.

[0062] Step 3: Incremental PID speed closed-loop control generates motor drive signals. Real-time reading of the actual angular velocities of the left and right wheels from the robot encoder is performed. , Calculate the speed deviation between the two wheels respectively. , An incremental PID controller is used for closed-loop adjustment of speed deviation, outputting control increments. Before the equipment leaves the factory, the linear mapping calibration of motor angular velocity and PWM duty cycle is completed, and a data table corresponding to angular velocity and output duty cycle is established. The PID output control increment is converted into a PWM pulse signal with the corresponding duty cycle. This pulse signal is a low-level execution command that the chassis motor can directly recognize. The independent speed inner loop of each wheel can suppress speed disturbances caused by voltage fluctuations, wheel slippage, and wheel wear. The incremental PID controller has no integral saturation problem, making it suitable for low-speed AGV operation scenarios in archive warehouses.

[0063] Step 4: Send the corrected movement command containing the PWM control parameters of the left and right wheels to the robot's main control board to drive the two-wheel motors to complete the pose tracking.

[0064] For example, when physical robot No. 3 performs the file retrieval task, its current actual pose is X: 10.7m, Y: 8.5m. The deviation between the actual trajectory and the predicted trajectory reaches 8 cm. Assuming the preset deviation threshold is 5 cm, this deviation is greater than the preset deviation threshold, and it is necessary to extract the expected pose at the corresponding time point in the simulated predicted trajectory. For example, the extracted expected pose is X: 10.5m, Y: 8.2m, angle: 0°. Using this expected pose as the target, the underlying conversion is completed through the above-mentioned differential chassis kinematics inverse kinematics and PID closed-loop algorithm. The upper-level target pose is converted into PWM control signals for the left and right wheel motors, encapsulated to obtain the corrected movement command, and sent to physical robot No. 3 in real time. The corresponding numerical calculation example: chassis wheel track L = 0.4m, wheel radius r = 0.1m. After inverse kinematics calculation, the target angular velocity of the left and right wheels is 7.21 rad / s. The PID controller outputs PWM pulses with a 65% duty cycle to drive the motors to adjust to the target coordinates at a constant speed. At the same time, the corrected actual execution trajectory will be updated to the state attributes of the corresponding robot digital twin to ensure that the simulation trajectory of the robot digital twin is consistent with the actual operating state of the physical robot in real time, providing an accurate virtual benchmark for subsequent scheduling decisions. For example, if the actual remaining battery power of physical robot No. 6 drops to 15%, and assuming the preset charging threshold is 20%, then the actual remaining battery power is less than the preset charging threshold. Therefore, a charging task needs to be generated immediately for this physical robot, instructing it to go to the smart charging station in the warehouse. Simultaneously, based on the current time-series knowledge graph, if physical robots No. 1 and No. 5 are found to be idle, the unfinished B-10 file archiving task of physical robot No. 6 can be reassigned to physical robot No. 1, and the current time-series knowledge graph can be updated: the task association edge between physical robot No. 6 and the B-10 file can be deleted, and a new task association edge between physical robot No. 1 and the B-10 file can be added. The state attributes of the corresponding robot digital twins are also updated synchronously: the task state of the digital twin of robot No. 6 is updated to "charging," and the task state of the digital twin of robot No. 1 is updated to "executing the B-10 file archiving task," ensuring that the state of the robot digital twins is completely synchronized with the actual working state of the physical robots, achieving a closed-loop digital twin that is consistent between the virtual and physical worlds.

[0065] As can be seen from the above, real-time monitoring of the core operating status of physical robots can dynamically control the operation process, promptly detect operational deviations and equipment anomalies, and provide real-time and accurate basis for subsequent handling; comparing virtual and real trajectories and calculating and correcting trajectory deviations can ensure that the actual operation of physical robots conforms to scheduling instructions, improve the accuracy of operations, and avoid equipment collisions and task delays caused by trajectory deviations; monitoring the actual remaining power of physical robots and generating charging tasks can promptly charge low-power physical robots, avoiding operation interruptions caused by equipment power failures; and based on graph reassignment of incomplete tasks and updating task-related edges, seamless task connection can be achieved, ensuring the continuity of file access operations, while keeping the temporal knowledge graph synchronized with the actual operation status, providing accurate data for subsequent scheduling decisions.

[0066] In one embodiment of this application, after performing collaborative scheduling of the physical robot corresponding to the digital twin of the robot to be controlled based on collaborative control instructions to execute physical file access operations, the method further includes: Real-time collection of actual operating data after the corresponding physical robot performs the physical file storage and retrieval operation. The actual operating data includes the current position, remaining battery power and current action status of the physical robot corresponding to the robot digital twin, and the storage location and borrowing status of the physical file corresponding to the file digital twin. Based on actual operational data, the state information and dynamic edge representation of the corresponding robot digital twin nodes and archive digital twin nodes in the current temporal knowledge graph are updated to obtain the temporal knowledge graph of the next state of the current temporal knowledge graph.

[0067] In this embodiment, by collecting actual operational data after the physical robot performs its tasks, the temporal knowledge graph for the next state is updated in real time. This ensures that the graph always reflects the latest state of the scheduled objects within the warehouse, providing accurate and effective data support for the next scheduling decision. The actual operational data refers to the latest state data of the physical robot and the physical files after the physical robot completes the file storage and retrieval operation, covering the physical robot's location, battery level, and action status, as well as the storage location and borrowing status of the physical files.

[0068] Specifically, after the coordinated scheduling of physical robots is completed based on coordinated control commands, and after the physical robots have completed the storage and retrieval operation of physical files, the actual operating data of the corresponding objects is collected in real time: for physical robot No. 5, which performed the file retrieval operation, its latest current position is collected as the warehouse exit area entrance (X: 15m, Y: 2m), remaining battery power is 75%, and current action status is idle; for physical file A-05 corresponding to the storage and retrieval operation, its latest storage location is collected as the temporary storage shelf in the exit area, and the borrowing status is borrowed; based on the collected actual operating data, the current... The temporal knowledge graph is updated as follows: the latest data of physical robot No. 5 is updated to the attribute value field of the corresponding robot digital twin node, and the latest data of physical file A-05 is updated to the attribute value field of the corresponding file digital twin node; at the same time, the graph association relationship is updated according to the state change: the "task assignment" association edge between physical robot No. 5 and physical file A-05 is deleted, and the "spatial proximity" association edge between physical robot No. 5 and outbound area is added. After the update is completed, the temporal knowledge graph of the next state of the current temporal knowledge graph is obtained, which is used for subsequent scheduling decisions.

[0069] As can be seen from the above, collecting actual operational data from physical robots and physical archives can comprehensively obtain the latest status of objects after storage and retrieval operations, providing an accurate and complete data source for graph updates and avoiding information lag; updating the information of robot digital twin nodes allows the graph to accurately reflect the latest operational status of physical robots, providing a reliable basis for subsequent task allocation of physical robots; updating the information of archive digital twin nodes allows the graph to reflect the storage and borrowing status of physical archives in real time, avoiding information errors in task scheduling; updating the dynamic edge relationships of the graph allows the relationships between scheduled objects to highly match the actual relationships, improving the accuracy of graph representation; and forming a temporal knowledge graph of the next state through data updates can realize closed-loop iteration of scheduling data, ensuring that every scheduling decision is based on the latest actual state of the warehouse, guaranteeing the effectiveness and accuracy of scheduling decisions.

[0070] Corresponding to the AI- and digital twin-based archive robot scheduling method in the above embodiment, Figure 3 This is a structural block diagram of an AI- and digital twin-based archive robot scheduling device according to an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 3 The AI- and digital twin-based archive robot scheduling device 20 includes: a knowledge graph acquisition module 21, an enhanced feature acquisition module 22, a control instruction determination module 23, and a collaborative scheduling module 24.

[0071] The knowledge graph acquisition module 21 is used to acquire the current temporal knowledge graph, which is constructed based on the digital twins of archives and robots. It is used to represent the state information of each digital twin of archives, the state information of each digital twin of robots, and the association relationships between digital twins of archives and digital twins of robots, and between digital twins of robots. The association relationships are represented by the dynamic edges in the current temporal knowledge graph. The enhanced feature acquisition module 22 is used to extract a local subgraph centered on each robot digital twin node, taking the robot digital twin node as the central node. Multiple attribute values ​​of any node within the local subgraph are combined to form the feature vector of that node. The feature vectors of all nodes within the local subgraph except the central node are then weighted and aggregated to obtain the enhanced features of the central node. The local subgraph includes the robot digital twin node, its neighboring robot digital twin nodes, and file digital twin nodes with task or spatial associations. The attribute values ​​of the robot digital twin node include: the current position, remaining battery power, and current action state of the physical robot corresponding to the robot digital twin. The attribute values ​​of the file digital twin node include: the storage location and borrowing status of the physical file corresponding to the file digital twin. The control instruction determination module 23 is used to determine the digital twin of the robot to be controlled and the corresponding collaborative control instructions based on all enhanced features; The collaborative scheduling module 24 is used to perform collaborative scheduling of the physical robot corresponding to the digital twin of the robot to be controlled based on collaborative control instructions, so as to perform physical file storage and retrieval operations.

[0072] In one embodiment of this application, the current temporal knowledge graph is obtained in the following way: Construct an initial graph structure. In the initial graph structure, each node represents each robot digital twin and each archive digital twin, and each edge represents the initial association relationship between robot digital twin nodes and between robot digital twin nodes and archive digital twin nodes. For each physical robot, the first real-time data of the physical robot is collected at a preset period. The first real-time data includes: the real-time position of the physical robot, the current remaining power, and the current action status. For each physical file, second real-time data of the physical file is collected according to a preset period. The second real-time data includes: the current storage location and borrowing status of the physical file. Based on the first real-time data and the second real-time data, a first association relationship between robot digital twins and a second association relationship between robot digital twins and document digital twins are determined. The first association relationship includes: a spatial proximity relationship determined based on the real-time position of each physical robot and a task collaboration relationship determined based on the current action state of each physical robot. The second association relationship includes: a spatial association relationship determined based on the real-time position of the physical robot and the storage location of the physical document, and a task association relationship determined based on the current action state of the physical robot and the borrowing status of the physical document. Based on each first real-time data, each second real-time data, each first association, and each second association, the initial graph structure is updated to obtain the updated initial graph structure, which serves as the current temporal knowledge graph.

[0073] In one embodiment of this application, the AI- and digital twin-based archive robot scheduling device 20 further includes an attention weight determination module, used for: For each node in the local subgraph other than the central node, based on the components of the feature vector of the other node and the components of the feature vector of the central node, sub-features between the central node and the other node are extracted; for each other node, if the other node is a neighbor robot digital twin node, the sub-features include position features, task association features, and resource competition features; if the other node is an archive digital twin node, the sub-features include position features and task association features. The positional features are determined based on the positional components in the feature vectors of the other nodes and the positional components in the feature vector of the center node. The task association feature is determined for a digital twin node of an archive based on the borrowing state component in the feature vector of the digital twin node of the archive and the action state component in the feature vector of the central node. It is used to characterize whether there is a task assignment relationship between the central node and the digital twin node of the archive. For a digital twin node of a neighboring robot, it is determined based on the action state component and position component in the feature vector of the digital twin node of the neighboring robot and the central node respectively. It is used to characterize whether there is a task cooperation relationship between the two or whether they serve the same target archive. Resource competition characteristics are determined based on the position and motion state components in the feature vector of the central node, as well as the position and motion state components in the feature vectors of other robot digital twin nodes in the local subgraph besides the central node. They are used to characterize the competition intensity of multiple robot digital twin nodes for the same archive digital twin node or the same target position. Based on the characteristics of each item, calculate the item attention coefficient between the central node and the other nodes respectively; Based on the current task stage and the types of other nodes, the pre-configured fusion coefficient table is queried to obtain the dynamic fusion coefficient corresponding to the current task stage and node type. The fusion coefficient table stores the preset allocation values ​​of the weights of various sub-features corresponding to different node types under different task stages. Based on the attention coefficients of each sub-item and their corresponding dynamic fusion coefficients, the attention weight of the central node to the other nodes is obtained; Specifically, the enhanced feature acquisition module, when performing weighted aggregation of the feature vectors of all nodes within the local subgraph except the center node to obtain the enhanced features of the center node, is used for: The enhanced features of the central node are obtained by weighted aggregation based on the feature vectors of each other node and their corresponding attention weights.

[0074] In one embodiment of this application, the AI- and digital twin-based archive robot scheduling device 20 further includes an updated temporal knowledge graph acquisition module, used for: Get the current task queue. The task queue is a set of tasks to be executed that are generated in real time based on the changes in the borrowing status of each physical file. Each task corresponds to a physical file to be accessed, and the tasks in the task queue are sorted from high to low according to their priority. For each task in the current task queue, calculate the task priority based on the request time and urgency identifier of the physical file corresponding to the task. The task priority is used as the attribute value of the digital twin node of the physical file, and updated in the current temporal knowledge graph to obtain the updated temporal knowledge graph. Specifically, the enhanced feature acquisition module, when extracting a local subgraph centered on each robot digital twin node as the central node, is used for: Based on the updated temporal knowledge graph, for each robot digital twin node, the robot digital twin node is taken as the central node, and a local subgraph centered on the node is extracted.

[0075] In one embodiment of this application, the control instruction determination module 23 is specifically used for: The digital twin of the robot to be controlled and the corresponding collaborative control instructions are determined based on all enhanced features, which is obtained through a trained multi-agent collaborative scheduling model. The module for obtaining the trained multi-agent cooperative scheduling model is used for: Initialize multiple agents in a simulation environment that includes digital twins of archives and digital twins of robots, with each agent corresponding to a digital twin of a robot; In the simulation environment, multiple rounds of task simulation are performed. At each time step, a local subgraph with each robot digital twin node as the center node is obtained. The attribute values ​​of each node in the local subgraph are combined into their respective feature vectors. The feature vectors of each node in the local subgraph except the center node are weighted and aggregated to obtain the enhanced features of each robot digital twin node, which serve as the current state of each agent. Each intelligent agent outputs simulated scheduling actions based on its own policy network and the current state. The simulated scheduling actions include the movement, file retrieval, archiving, avoidance, or multi-machine collaborative actions of the robot digital twin. After the robot digital twin performs simulated scheduling actions, it calculates the immediate reward based on the constrained reward function, and obtains the joint state of the next time step after the current time step ends. The joint state is a global state set composed of the enhanced features of all robot digital twin nodes. The constrained reward function includes task completion reward, path conflict penalty, and task balance reward. The current state, simulated scheduling action, immediate reward, and joint state of the next time step are combined into a quadruple as experience data storage. The policy network of the agent is iteratively optimized using the experience data. Training stops when the cumulative reward converges, resulting in the trained multi-agent cooperative scheduling model.

[0076] In one embodiment of this application, the collaborative scheduling module 24 is further configured to: During the physical robot's execution of physical file access operations, Real-time monitoring of the physical operating status of each physical robot, including actual remaining battery power and actual execution trajectory; The actual execution trajectory of the physical robot is compared with the simulated predicted trajectory of the corresponding digital twin of the robot, and the trajectory deviation is calculated. When the trajectory deviation is greater than the preset deviation threshold, the expected pose at the corresponding time point in the simulation prediction trajectory is extracted, and the movement command in the collaborative control command is corrected with the expected pose as the target. The corrected movement command is then sent to the corresponding physical robot for execution. When it is detected that the actual remaining power of any physical robot is less than the preset charging threshold, a charging task is generated. Based on the attribute values ​​of other robot digital twins in the current temporal knowledge graph, the unfinished tasks of that physical robot are reassigned to other available physical robots, and the corresponding task association edges in the current temporal knowledge graph are updated.

[0077] In one embodiment of this application, the AI- and digital twin-based archive robot scheduling device 20 further includes a temporal knowledge graph acquisition module for the next state, used for: Real-time collection of actual operating data after the corresponding physical robot performs the physical file storage and retrieval operation. The actual operating data includes the current position, remaining battery power and current action status of the physical robot corresponding to the robot digital twin, and the storage location and borrowing status of the physical file corresponding to the file digital twin. Based on actual operational data, the state information and dynamic edge representation of the corresponding robot digital twin nodes and archive digital twin nodes in the current temporal knowledge graph are updated to obtain the temporal knowledge graph of the next state of the current temporal knowledge graph.

[0078] See Figure 4 , Figure 4 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 4 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 3 The functions of the knowledge graph acquisition module 21, the enhanced feature acquisition module 22, the control instruction determination module 23, and the collaborative scheduling module 24 are shown.

[0079] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0080] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0081] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0082] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation method described in the AI ​​and digital twin-based archive robot scheduling method provided in the embodiments of this application, or they can execute the implementation method of the electronic device described in the embodiments of this application, which will not be repeated here.

[0083] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0084] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., provided on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0085] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0086] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0087] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces or units, or they may be electrical, mechanical, or other forms of connection.

[0088] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0089] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0090] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for scheduling archive robots based on AI and digital twins, characterized in that, include: Obtain the current temporal knowledge graph, which is constructed based on archival digital twins and robot digital twins, and is used to represent the state information of each archival digital twin, the state information of each robot digital twin, and the association relationships between archival digital twins and robot digital twins, and between robot digital twins; the association relationships are represented by dynamic edges in the current temporal knowledge graph; For each robot digital twin node, the robot digital twin node is taken as the central node, and a local subgraph centered on the central node is extracted. Multiple attribute values ​​of any node in the local subgraph are combined to form the feature vector of the node. The feature vectors of all nodes in the local subgraph except the central node are weighted and aggregated to obtain the enhanced features of the central node. The local subgraph includes the robot digital twin node, the neighboring robot digital twin nodes of the robot digital twin node, and the archive digital twin nodes that have task association or spatial association. The attribute values ​​of the robot digital twin node include: the current position, remaining battery power, and current action status of the physical robot corresponding to the robot digital twin; the attribute values ​​of the file digital twin node include: the storage location and borrowing status of the physical file corresponding to the file digital twin. Based on all the aforementioned enhanced features, the digital twin of the robot to be controlled and the corresponding collaborative control instructions are determined. Based on the coordinated control instructions, the physical robot corresponding to the digital twin of the robot to be controlled is coordinated and scheduled to perform physical file access operations.

2. The archival robot scheduling method based on AI and digital twins as described in claim 1, characterized in that, The current time-series knowledge graph was obtained through the following methods: Construct an initial graph structure, in which each node represents each robot digital twin and each archive digital twin, and each edge represents the initial association relationship between robot digital twin nodes and between robot digital twin nodes and archive digital twin nodes; For each physical robot, the first real-time data of the physical robot is collected at a preset period. The first real-time data includes: the real-time position of the physical robot, the current remaining power, and the current action status. For each physical file, second real-time data of the physical file is collected according to a preset period. The second real-time data includes: the current storage location and borrowing status of the physical file. Based on the first real-time data and the second real-time data, a first association relationship between robot digital twins and a second association relationship between robot digital twins and document digital twins are determined. The first association relationship includes: a spatial proximity relationship determined based on the real-time position of each physical robot and a task collaboration relationship determined based on the current action state of each physical robot. The second association relationship includes: a spatial association relationship determined based on the real-time position of the physical robot and the storage location of the physical document, and a task association relationship determined based on the current action state of the physical robot and the borrowing status of the physical document. Based on each of the first real-time data, each of the second real-time data, each of the first association relationships, and each of the second association relationships, the initial graph structure is updated to obtain the updated initial graph structure, which serves as the current temporal knowledge graph.

3. The archive robot scheduling method based on AI and digital twins as described in claim 1, characterized in that, Before extracting a local subgraph centered on the robot's digital twin node, using this node as the central node, combining multiple attribute values ​​of any node within this local subgraph into its feature vector, and then weighted and aggregating the feature vectors of all nodes within this local subgraph except the central node to obtain the enhanced features of the central node, the process further includes: For each node in the local subgraph other than the central node, based on each component of the feature vector of the other node and each component of the feature vector of the central node, sub-features between the central node and the other node are extracted; for each other node, if the other node is a neighbor robot digital twin node, the sub-features include position features, task association features, and resource competition features; if the other node is an archive digital twin node, the sub-features include position features and task association features. The positional features are determined based on the positional components in the feature vectors of the other nodes and the positional components in the feature vectors of the central node. The task association feature is determined for a digital twin node of an archive based on the borrowing state component in the feature vector of the digital twin node of the archive and the action state component in the feature vector of the central node, and is used to characterize whether there is a task assignment relationship between the central node and the digital twin node of the archive; for a digital twin node of a neighboring robot, it is determined based on the action state component and position component in the feature vector of the digital twin node of the neighboring robot and the central node, and is used to characterize whether there is a task cooperation relationship between the two or whether they serve the same target archive. The resource competition feature is determined based on the position and motion state components in the feature vector of the central node, as well as the position and motion state components in the feature vectors of other robot digital twin nodes in the local subgraph besides the central node. It is used to characterize the competition intensity of multiple robot digital twin nodes for the same archive digital twin node or the same target position. Based on each of the aforementioned features, the attention coefficients between the central node and the other nodes are calculated respectively. Based on the current task stage and the types of other nodes, the pre-configured fusion coefficient table is queried to obtain the dynamic fusion coefficient corresponding to the current task stage and node type; the fusion coefficient table stores the preset allocation values ​​of the weights of various sub-features corresponding to different node types under different task stages. Based on the attention coefficients of each sub-item and their corresponding dynamic fusion coefficients, the attention weight of the central node to the other nodes is obtained; The step of weighted aggregation of the feature vectors of all nodes within the local subgraph except for the central node to obtain the enhanced features of the central node includes: The enhanced features of the central node are obtained by weighted aggregation based on the feature vectors of the other nodes and their corresponding attention weights.

4. The archive robot scheduling method based on AI and digital twins as described in claim 1, characterized in that, After obtaining the current time-series knowledge graph, the following is also included: Obtain the current task queue, which is a set of tasks to be executed generated in real time based on the changes in the borrowing status of each physical file. Each task corresponds to a physical file to be accessed, and the tasks in the task queue are sorted from high to low according to task priority. For each task in the current task queue, calculate the task priority based on the request time and urgency identifier of the physical file corresponding to the task. The task priority is used as the attribute value of the digital twin node of the physical file, and updated in the current time-series knowledge graph to obtain the updated time-series knowledge graph. Specifically, for each robot digital twin node, this robot digital twin node is taken as the center node, and a local subgraph centered on this node is extracted, including: Based on the updated temporal knowledge graph, for each robot digital twin node, the robot digital twin node is taken as the central node, and a local subgraph centered on the node is extracted.

5. The archival robot scheduling method based on AI and digital twins as described in claim 1, characterized in that, The determination of the digital twin of the robot to be controlled and the corresponding collaborative control instructions based on all the enhanced features is obtained through a trained multi-agent collaborative scheduling model. The trained multi-agent cooperative scheduling model is obtained through the following training method: Multiple intelligent agents are initialized in a simulation environment that includes the digital twin of the archive and the digital twin of the robot, with each intelligent agent corresponding to a digital twin of the robot; In the simulation environment, multiple rounds of task simulation are performed. At each time step, a local subgraph with each robot digital twin node as the central node is obtained. The attribute values ​​of each node in the local subgraph are combined into their respective feature vectors. The feature vectors of each node in the local subgraph other than the central node are weighted and aggregated to obtain the enhanced features of each robot digital twin node, which serve as the current state of each agent. Each intelligent agent is controlled to output simulated scheduling actions based on its own policy network and the current state. The simulated scheduling actions include the movement, file retrieval, archiving, avoidance, or multi-machine collaborative actions of the robot digital twin. After the robot digital twin executes the simulated scheduling action, an immediate reward is calculated based on a constrained reward function, and after the current time step ends, the joint state of the next time step is obtained; the joint state is a global state set composed of the enhanced features of all robot digital twin nodes; the constrained reward function includes a task completion reward, a path conflict penalty, and a task balance reward. The current state, the simulated scheduling action, the immediate reward, and the joint state of the next time step are combined into a quadruple as experience data storage. The experience data is used to iteratively optimize the agent's policy network. Training stops when the cumulative reward converges, resulting in the trained multi-agent cooperative scheduling model.

6. The archival robot scheduling method based on AI and digital twins as described in claim 1, characterized in that, During the process of the physical robot performing physical file access operations, the method further includes: Real-time monitoring of the physical operating status of each physical robot, including the actual remaining battery power and the actual execution trajectory; The actual execution trajectory of the physical robot is compared with the simulated predicted trajectory of the corresponding digital twin of the robot, and the trajectory deviation is calculated. When the trajectory deviation is greater than the preset deviation threshold, the expected pose at the corresponding time point in the simulation prediction trajectory is extracted, and the movement command in the collaborative control command is corrected with the expected pose as the target. The corrected movement command is then sent to the corresponding physical robot for execution. When it is detected that the actual remaining power of any physical robot is less than the preset charging threshold, a charging task is generated. Based on the attribute values ​​of other robot digital twins in the current time-series knowledge graph, the unfinished tasks of the physical robot are reassigned to other available physical robots, and the corresponding task association edges in the current time-series knowledge graph are updated.

7. The archive robot scheduling method based on AI and digital twins as described in claim 1, characterized in that, After the step of coordinating the scheduling of the physical robot corresponding to the digital twin of the robot to be controlled based on the coordinated control instructions to perform physical file access operations, the method further includes: The actual operating data of the physical robot after performing the storage and retrieval operation of the physical file is collected in real time. The actual operating data includes the current position, remaining power and current action status of the physical robot corresponding to the robot digital twin, and the storage location and borrowing status of the physical file corresponding to the file digital twin. Based on the actual operating data, the state information and dynamic edge representation of the corresponding robot digital twin nodes and archive digital twin nodes in the current temporal knowledge graph are updated to obtain the temporal knowledge graph of the next state of the current temporal knowledge graph.

8. A file robot scheduling device based on AI and digital twins, characterized in that, include: The knowledge graph acquisition module is used to acquire the current temporal knowledge graph, which is constructed based on archival digital twins and robot digital twins. It is used to represent the state information of each archival digital twin, the state information of each robot digital twin, and the association relationships between archival digital twins and robot digital twins, and between robot digital twins themselves. The association relationships are represented by dynamic edges in the current temporal knowledge graph. An enhanced feature acquisition module is used to extract a local subgraph centered on each robot digital twin node, taking the robot digital twin node as the central node. Multiple attribute values ​​of any node within the local subgraph are combined to form the feature vector of that node. The feature vectors of all nodes within the local subgraph except the central node are then weighted and aggregated to obtain the enhanced features of the central node. The local subgraph includes the robot digital twin node, its neighboring robot digital twin nodes, and archive digital twin nodes with task or spatial associations. The attribute values ​​of the robot digital twin node include: the current position, remaining battery power, and current action status of the physical robot corresponding to the robot digital twin; the attribute values ​​of the file digital twin node include: the storage location and borrowing status of the physical file corresponding to the file digital twin. The control instruction determination module is used to determine the digital twin of the robot to be controlled and the corresponding collaborative control instructions based on all the enhanced features. The collaborative scheduling module is used to perform collaborative scheduling of the physical robot corresponding to the digital twin of the robot to be controlled based on the collaborative control instructions, so as to perform physical file storage and retrieval operations.

9. The archive robot scheduling device based on AI and digital twin as described in claim 8, characterized in that, The current time-series knowledge graph was obtained through the following methods: Construct an initial graph structure, in which each node represents each robot digital twin and each archive digital twin, and each edge represents the initial association relationship between robot digital twin nodes and between robot digital twin nodes and archive digital twin nodes; For each physical robot, the first real-time data of the physical robot is collected at a preset period. The first real-time data includes: the real-time position of the physical robot, the current remaining power, and the current action status. For each physical file, second real-time data of the physical file is collected according to a preset period. The second real-time data includes: the current storage location and borrowing status of the physical file. Based on the first real-time data and the second real-time data, a first association relationship between robot digital twins and a second association relationship between robot digital twins and document digital twins are determined. The first association relationship includes: a spatial proximity relationship determined based on the real-time position of each physical robot and a task collaboration relationship determined based on the current action state of each physical robot. The second association relationship includes: a spatial association relationship determined based on the real-time position of the physical robot and the storage location of the physical document, and a task association relationship determined based on the current action state of the physical robot and the borrowing status of the physical document. Based on each of the first real-time data, each of the second real-time data, each of the first association relationships, and each of the second association relationships, the initial graph structure is updated to obtain the updated initial graph structure, which serves as the current temporal knowledge graph.

10. The archive robot scheduling device based on AI and digital twin as described in claim 8, characterized in that, Also includes: The attention weight determination module is used for: For each node in the local subgraph other than the central node, based on each component of the feature vector of the other node and each component of the feature vector of the central node, sub-features between the central node and the other node are extracted; for each other node, if the other node is a neighbor robot digital twin node, the sub-features include position features, task association features, and resource competition features; if the other node is an archive digital twin node, the sub-features include position features and task association features. The positional features are determined based on the positional components in the feature vectors of the other nodes and the positional components in the feature vectors of the central node. The task association feature is determined for a digital twin node of an archive based on the borrowing state component in the feature vector of the digital twin node of the archive and the action state component in the feature vector of the central node, and is used to characterize whether there is a task assignment relationship between the central node and the digital twin node of the archive; for a digital twin node of a neighboring robot, it is determined based on the action state component and position component in the feature vector of the digital twin node of the neighboring robot and the central node, and is used to characterize whether there is a task cooperation relationship between the two or whether they serve the same target archive. The resource competition feature is determined based on the position and motion state components in the feature vector of the central node, as well as the position and motion state components in the feature vectors of other robot digital twin nodes in the local subgraph besides the central node. It is used to characterize the competition intensity of multiple robot digital twin nodes for the same archive digital twin node or the same target position. Based on each of the aforementioned features, the attention coefficients between the central node and the other nodes are calculated respectively. Based on the current task stage and the types of other nodes, the pre-configured fusion coefficient table is queried to obtain the dynamic fusion coefficient corresponding to the current task stage and node type; the fusion coefficient table stores the preset allocation values ​​of the weights of various sub-features corresponding to different node types under different task stages. Based on the attention coefficients of each sub-item and their corresponding dynamic fusion coefficients, the attention weight of the central node to the other nodes is obtained; The enhanced feature acquisition module, when performing weighted aggregation of the feature vectors of all nodes within the local subgraph except the center node to obtain the enhanced features of the center node, is specifically used for: The enhanced features of the central node are obtained by weighted aggregation based on the feature vectors of the other nodes and their corresponding attention weights.

Images