An AGV manufacturing resource dynamic scheduling method based on Agentic AI multi-agent arrangement

Abstract

The application relates to the technical field of intelligent manufacturing, in particular to an AGV manufacturing resource dynamic scheduling method based on Agentic AI multi-agent arrangement, which comprises the following steps: a total control arrangement agent receives a scheduling task request and generates standardized scheduling task information; a state sensing agent acquires an AGV real-time state and generates a resource state snapshot, and a task analysis and priority judgment agent determines a task priority; a scheduling execution agent performs hard constraint checking to screen candidate AGVs, generates a conflict-free candidate scheduling scheme through space-time conflict detection and resolution, and determines a target scheduling scheme through multi-target comprehensive evaluation; after the total control arrangement agent issues the scheme, a real-time state monitoring and execution feedback agent monitors the execution, detects an exception, determines a rescheduling range according to the exception type, and triggers rescheduling only for the affected part. Through the multi-agent arrangement and closed-loop scheduling mechanism, the application realizes fast response and local adjustment under dynamic constraints, and improves the real-time performance and rationality of multi-AGV collaborative scheduling.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a dynamic scheduling method for AGV manufacturing resources based on Agentic AI multi-agent orchestration. Background Technology

[0002] With the rapid development of intelligent manufacturing, production logistics systems are showing a trend towards automation and flexibility. To improve the efficiency and reliability of AGV scheduling, traditional technologies mainly employ methods such as rule-based scheduling, mathematical programming, multi-agent path planning, and deep reinforcement learning.

[0003] Static rule-based scheduling and heuristic algorithms rely on preset rules for scheduling strategies. When temporary order insertions, equipment failures, or path congestion occur in the workshop, manual reconfiguration is required to adjust the scheduling, resulting in insufficient dynamic response capabilities. Methods based on mathematical programming and multi-agent path planning build models around task assignment and path collision avoidance. While they can generate optimal solutions during normal operation, changes in the field state require re-solving or renegotiation, leading to long convergence times and deadlocks. The significant rescheduling delay between state awareness and solution updates prevents a smooth, uninterrupted transition, impacting the continuity and stability of production logistics. Data-driven methods such as deep reinforcement learning improve the adaptability of scheduling strategies to complex scenarios to some extent. However, as the number of AGVs, task concurrency, and field constraints increase, the difference between the training environment and the actual manufacturing site becomes increasingly significant, leading to a decline in model generalization ability. Furthermore, these methods rely on large amounts of high-quality training data and specific workshop topologies. When process changes, map node adjustments, or safety controls occur, data must be re-collected and retrained, significantly increasing time and maintenance costs. Furthermore, the aforementioned scheduling methods are all based on a centralized or fixed hierarchical architecture. When the abnormal events faced by the system exceed its designed processing range, even if the impact range of each abnormality is different, the entire scheduling scheme may still fail, making it difficult to ensure the continuous and stable operation of the production logistics system.

[0004] In real-world scenarios, critical tasks such as emergency order insertion and fault response have extremely high real-time requirements, while routine material delivery tasks have lower response time requirements. However, once the architecture of traditional scheduling systems is determined, it is difficult to change. All tasks are processed using a uniform process, leading to the system being configured with resources according to the highest requirements, resulting in over-configuration of design resources and increased costs. At the same time, existing methods lack unified orchestration and closed-loop management of task parsing, status awareness, solution generation, and execution feedback, making it difficult to trace and quantitatively evaluate scheduling results. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a dynamic scheduling method for AGV manufacturing resources based on Agentic AI multi-agent orchestration. The multi-agent system includes a central control orchestration agent, a task parsing and priority determination agent, a state awareness agent, a scheduling and execution agent, and a real-time state monitoring and execution feedback agent. The method comprises the following steps: S1: The overall control and orchestration intelligent agent receives the AGV scheduling task request, calls the task parsing and priority determination intelligent agent to perform semantic parsing and standardization processing on the AGV scheduling task request, and generates standardized scheduling task information containing task attributes and constraints. S2: The state-aware agent acquires the real-time status of the AGV and map topology information, and generates a real-time resource status snapshot; the task parsing and priority determination agent determines the task priority based on the urgency of the task and time constraints. S3: The scheduling execution agent performs hard constraint verification on the status attributes and capability attributes of candidate AGVs based on the real-time resource status snapshot, and selects a set of candidate AGVs that meet the task execution conditions. S4: The scheduling execution agent generates a candidate path for each AGV in the candidate AGV set, converts the candidate path into a spatiotemporal occupancy sequence containing nodes, road segments and resource occupancy time intervals, detects spatiotemporal conflicts based on the spatiotemporal occupancy sequence, and generates a conflict-free candidate scheduling scheme by performing conflict resolution processing. S5: The scheduling execution agent performs a multi-objective comprehensive evaluation of the conflict-free candidate scheduling schemes based on task priorities, determines the target scheduling scheme, and outputs it. S6: The overall control orchestration agent issues the target scheduling scheme, and the real-time status monitoring and execution feedback agent monitors the execution process and collects abnormal events; when an abnormal event is detected, the overall control orchestration agent determines the rescheduling scope according to the type of abnormality, retains the unaffected part of the scheduling results, and triggers the rescheduling process only for the affected part to update the scheduling scheme.

[0006] Preferably, in step S1, standardized scheduling task information is generated, including: Identify the source of AGV scheduling task requests and convert task requests of different formats into a unified task data structure; Extract the task number, task type, material identifier, source location, target location, required completion time, emergency status, load requirements, and process constraints from the task request; The source and target locations are matched with map topology nodes, and the material parameters and AGV capability parameters are processed to unify and normalize data units to generate the standardized scheduling task information.

[0007] Preferably, in step S2, determining the task priority includes: Obtain the urgency indicator for the task and the remaining time until the required completion time; When the emergency flag is set to the preset emergency value, or when the remaining time is less than the first time threshold, the task will be classified as the highest priority. When an emergency is marked as non-urgent and the remaining time is between the first and second time thresholds, the task is classified as medium priority. When an emergency flag is set to a non-urgent value and the remaining time is greater than the second time threshold, the task will be classified as a normal priority.

[0008] Preferably, in step S2, generating the real-time resource status snapshot includes: Obtain the AGV number, current position, travel speed, remaining battery power, rated load, current load, vehicle size, online status, fault status, running status, current task, and task completion rate for each AGV; The workshop map is represented as a topology map consisting of passable nodes and connecting road segments, and the length, permitted driving direction, traffic capacity, speed limit information, restricted status, congestion status and real-time occupancy status of each connecting road segment are obtained. Obtain the resource occupancy status of loading / unloading points, buffer zones, intersections, narrow passages, and charging areas, and calculate the estimated release time of the corresponding resources based on the AGV position, AGV speed, and current task; Add timestamps to AGV status data, map topology data, and resource usage data. When the difference between the timestamp of any data and the current scheduling time exceeds the preset data validity period, the corresponding data is re-acquired. Based on the data that has passed the time validity verification, a real-time resource status snapshot is generated, which includes AGV availability, vehicle capacity, path reachability, path congestion level, and resource occupancy time interval.

[0009] Preferably, in step S3, hard constraint verification is performed on the AGV's state attributes and capability attributes, including: Verify at least one of the following: online status, fault status, order-accepting status, load capacity, vehicle size, task type adaptability, and path reachability of the AGV. AGVs that meet the hard constraint verification items are added to the initial candidate AGV set; when any verification item is not met, the corresponding AGV is removed.

[0010] Preferably, in step S4, detecting spatiotemporal conflicts based on the spatiotemporal occupancy sequence includes: Overlap and compare the spatiotemporal occupancy sequences of different AGVs in both time and space dimensions; When different AGVs are detected occupying the same node, the same road segment in opposite directions, the same narrow passage, or the same loading / unloading resource within the overlapping time interval, it is determined to be a conflict; The resolution process includes: determining the resource occupancy order based on task priority, and generating a conflict-free path by using methods such as waiting, detouring, speed adjustment, or replacing candidate AGVs.

[0011] Preferably, in step S4, generating the conflict-free candidate scheduling scheme includes: Based on the task start point, task end point and map topology, generate at least one candidate path for each candidate AGV to avoid restricted areas and congested areas. Based on the candidate AGV's expected travel speed, the length of each road segment, the loading and unloading operation time, and the intersection passage time, each candidate path is converted into a node occupation time interval, a road segment occupation time interval, and a loading and unloading resource occupation time interval. Compare the node occupancy time interval, road segment occupancy time interval, and loading / unloading resource occupancy time interval with the resource reservation information corresponding to the issued scheduling plan; For node conflicts, road segment conflicts, meeting conflicts, or loading and unloading resource conflicts, the priority order of resource occupation is determined according to task priority, and at least one of the following methods is used to resolve the conflict: delaying entry into conflicting resources, adjusting driving speed, selecting alternative paths, and replacing candidate AGVs. After the conflict is resolved, spatiotemporal conflict detection is performed again, and the candidate AGVs, candidate paths and resource occupancy time intervals for which no spatiotemporal conflict was detected are combined into the conflict-free candidate scheduling scheme.

[0012] Preferably, in step S5, the multi-objective comprehensive evaluation includes: Calculate the evaluation indicators for the task completion time, path length, energy consumption, path congestion, low power risk, and load balancing deviation of the conflict-free candidate scheduling schemes respectively. The weights of each evaluation indicator are dynamically adjusted based on task priority: for high-priority tasks, the weight of task completion time is increased; for medium-priority tasks, the weights of task completion time and resource utilization efficiency are balanced; for ordinary-priority tasks, the weights of energy consumption and load balancing are increased. The scheme with the lowest overall cost is selected as the target scheduling scheme based on the weighted results.

[0013] Preferably, the scheme with the minimum overall cost is selected as the target scheduling scheme based on the weighted result, including: The number is determined according to the following formula. The combined cost of conflict-free candidate scheduling schemes :

[0014] in, The estimated completion time for the normalization task. To normalize the path length, To normalize the projected energy consumption, To normalize the level of path congestion, To normalize the risk of low battery levels, To normalize the AGV load balancing deviation, , , , , , The weights are the corresponding evaluation indicators, and the sum of the weights is 1.

[0015] Preferably, in step S6, determining the rescheduling scope based on the exception type includes: When the abnormal event is path congestion or path blockage, only the local path of the affected AGV is replanned, while the assignment results of other AGVs remain unchanged. When the abnormal event is equipment failure or low battery, cancel the resource reservation of the abnormal AGV and re-match the AGV resources for the affected task. When an abnormal event is the insertion of a new task or a change in task time limit, the task priority is re-divided, and resource matching and path planning are performed on the affected task set.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention unifies the scheduling process through a central control and scheduling agent, and incorporates task parsing, state perception, resource matching, path planning and execution feedback into the same scheduling link, which solves the problems of scattered information and scheduling lag in traditional AGV scheduling, and ensures scheduling efficiency and scheme reliability.

[0017] (2) This invention generates real-time resource status snapshots through state-aware intelligent agents, schedules execution intelligent agents to perform hard constraint verification and generate conflict-free candidate scheduling schemes, and realizes multi-agent collaborative processing according to role division, thereby improving the on-site adaptability of resource matching and path planning, and reducing the risk of multi-vehicle path conflicts and resource deadlock.

[0018] (3) The present invention monitors the execution process through real-time status monitoring and execution feedback intelligent agents. The overall control and scheduling intelligent agent determines the rescheduling scope according to the anomaly type and triggers the rescheduling process only for the affected part, forming a scheduling closed loop from perception to decision-making to execution feedback, thereby ensuring the system's rapid response capability under dynamic constraints, avoiding the computational overhead and execution disturbance caused by full rescheduling, and having high reliability and flexibility. Attached Figure Description

[0019] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0020] Figure 1 This is a flowchart of a dynamic scheduling method for AGV manufacturing resources based on Agentic AI multi-agent orchestration, according to the present invention. Figure 2 This is a schematic diagram of the architecture of the AGV manufacturing resource dynamic scheduling system based on Agentic AI multi-agent orchestration in an embodiment of the present invention; Figure 3 This is a schematic diagram of the collaborative decision-making process for AGV scheduling tasks in an embodiment of the present invention; Figure 4 This is a schematic diagram of the AGV scheduling execution closed-loop feedback and rescheduling process in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. Obviously, the described embodiments are only some, not all, of the embodiments described in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application.

[0022] Those skilled in the art will understand that, unless otherwise stated, the singular forms “a” and “an” used herein, and “the”, may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0023] First embodiment: "Agentic AI" refers to an intelligent software entity with autonomous perception, decision-making, and execution capabilities. It can proactively complete specific scheduled sub-tasks according to predetermined orchestration logic within a multi-agent collaborative framework, including but not limited to task parsing, state perception, path planning, conflict resolution, and execution feedback. Unlike conventional passive scripts or single algorithm modules, Agentic AI agents possess the ability to dynamically adjust their behavior based on real-time environmental states and can accept unified scheduling and collaborative instructions from the central orchestration agent.

[0024] Please see Figure 1 and Figure 2 As shown, this embodiment provides a dynamic scheduling method for AGV manufacturing resources based on Agentic AI multi-agent orchestration. The multi-agent system includes a central control orchestration agent, a task parsing and priority determination agent, a state awareness agent, a scheduling execution agent, and a real-time state monitoring and execution feedback agent. The central control orchestration agent unifies the orchestration and scheduling process, while the task parsing and priority determination agent, the state awareness agent, the scheduling execution agent, and the real-time state monitoring and execution feedback agent collaboratively process tasks, integrating task semantic parsing, AGV state awareness, task priority determination, resource matching, path planning, and execution feedback into the same scheduling link. This invention establishes a dynamic generation and closed-loop feedback mechanism from task requests to executable scheduling schemes, thereby improving the real-time performance, rationality, and traceability of multi-AGV collaborative scheduling. Addressing AGV material handling and production line delivery scenarios in smart factories, manufacturing tasks are often presented in the form of MES work orders, warehouse scheduling tasks, and inter-process material handling requests. These tasks are prone to issues such as inconsistent field formats, missing emergency information, ambiguous constraints, and unpredictable dynamic changes on-site. This leads to coupling and conflict between task priority determination, AGV resource matching, and path planning, making it difficult to explicitly identify and systematically resolve these pain points. This invention introduces the Agentic AI multi-agent orchestration system, decomposing the complex AGV scheduling process into multi-stage collaborative tasks, including the following steps: S1: The central control and orchestration agent receives the AGV scheduling task request, calls the task parsing and priority determination agent to perform semantic parsing and standardization processing on the AGV scheduling task request, and generates standardized scheduling task information containing task attributes and constraints.

[0025] Preferably, in step S1, standardized scheduling task information is generated, including: Identify the source of AGV scheduling task requests and convert task requests of different formats into a unified task data structure. Specifically, the AGV scheduling task requests received by the central control and orchestration agent can come from the Manufacturing Execution System (MES), the warehouse management system, or the manual scheduling terminal. Extract the task number, task type, material identifier, source location, target location, required completion time, emergency status, load requirements, and process constraints from the task request; The source and target locations are matched with map topology nodes, and the data units of material parameters and AGV capacity parameters are unified and normalized to generate standardized scheduling task information.

[0026] For example, the central control and orchestration agent receives a scheduling task request from the MES system: "Transport material A from warehouse 1 to assembly line 3, with a requirement to complete within 30 minutes." The task parsing and priority determination agent extracts fields from this request, identifying the task objective as "transportation of material A," the constraint as "complete within 30 minutes," the starting point as "warehouse 1," the destination as "assembly line 3," the task type as "material transfer," and the urgency flag as "non-urgent." The extracted results are then organized into standardized scheduling task information in a unified format.

[0027] In a preferred implementation, the source and target locations are matched with node identifiers in the map topology; the data units of material weight and size are unified and normalized to min-max for AGV capacity parameters; and the required completion time is converted to time information under a unified time base. This ensures that subsequent scheduling agents can perform resource matching and path planning based on consistent data standards. The normalization process maps each parameter value to the [0,1] interval according to the following formula: x is the original parameter value. and The standardized scheduling task information is generated by taking the minimum and maximum values ​​of the same type of parameter within a preset statistical period.

[0028] S2: The state-aware agent acquires the real-time status of the AGV and map topology information, and generates a real-time resource status snapshot; the task parsing and priority determination agent determines the task priority based on the urgency of the task and time constraints.

[0029] Preferably, in step S2, generating a real-time resource status snapshot includes: Obtain the AGV number, current position, travel speed, remaining battery power, rated load, current load, vehicle size, online status, fault status, running status, current task, and task completion rate for each AGV; The workshop map is represented as a topology map consisting of passable nodes and connecting road segments, and the length, permitted driving direction, traffic capacity, speed limit information, restricted status, congestion status and real-time occupancy status of each connecting road segment are obtained. Obtain the resource occupancy status of loading / unloading points, buffer zones, intersections, narrow passages, and charging areas, and calculate the estimated release time of the corresponding resources based on the AGV position, AGV speed, and current task; Add timestamps to AGV status data, map topology data, and resource usage data. When the difference between the timestamp of any data and the current scheduling time exceeds the preset data validity period, the corresponding data is re-acquired to ensure that the data on which scheduling decisions are based is timely. Based on the data that has passed the time validity verification, a real-time resource status snapshot is generated, which includes AGV availability, vehicle capacity, path reachability, path congestion level, and resource occupation time interval. Specifically, based on the real-time occupation status, traffic capacity, and the ratio of the number of currently occupied AGVs to the upper limit of traffic capacity of each connecting road segment, the road segment congestion index is calculated, and a real-time resource status snapshot is generated based on the road segment congestion index.

[0030] Continuing with the aforementioned manufacturing workshop scenario, the state-aware intelligent agent acquires the real-time status of all AGVs in the workshop: AGV-03 is located near warehouse 2, with a speed of 1.2 m / s, 85% battery, rated load of 500 kg, current load of 0 kg, online status, no fault status, and currently idle; AGV-07 is located near warehouse 1, with a speed of 1.2 m / s, 92% battery, rated load of 500 kg, current load of 0 kg, online status, no fault status, and currently idle; AGV-12 is located in the B area aisle, with a speed of 0.8 m / s, 45% battery, rated load of 500 kg, current load of 350 kg, online status, no fault status, currently executing task T-1025, and task completion rate of 60%. Meanwhile, the state-aware agent obtains map topology and path congestion information: the current traffic flow on the main road in area A is moderate, there is slight congestion in the second alleyway in area B, and there are two passable paths from warehouse 1 to assembly line 3.

[0031] Preferably, in step S2, determining the task priority includes: Obtain the urgency indicator for the task and the remaining time until the required completion time; When the emergency flag is a preset emergency value, or the remaining time is less than a first time threshold, the task is assigned the highest priority. Specifically, let the emergency flag of the i-th task be Ei, the emergency flag of emergency tasks be 1, and the emergency flag of non-emergency tasks be 0. Based on the time difference between the required completion time of the i-th task and the current scheduling time, the remaining completion time H of the i-th task is determined. i When Ei=1 or Hi≤4h, the i-th task is classified as level P1; When an emergency is marked as a non-emergency value and the remaining time is between the first and second time thresholds, the task is classified as medium priority. Specifically, when Ei=0 and 4h<Hi≤24h, the i-th task is classified as level P2. When an urgent task is marked as non-urgent and the remaining time is greater than the second time threshold, the task is classified as a normal priority. Specifically, when Ei=0 and Hi>24h, the i-th task is classified as P3 level. For multiple tasks with the same priority, they are sorted from earliest to latest according to their required completion time; when the required completion time is the same, they are sorted from earliest to latest according to their task generation time.

[0032] For example, the urgency flag for the material A transportation task is E=0, and the remaining time between the completion time and the task generation time is H=25 minutes ≈ 0.42 hours, satisfying the condition H≤4h. The task parsing and priority determination agent classifies this task as P1 level. Meanwhile, another routine material delivery task within the workshop has an urgency flag E=0 and a remaining time H=18 hours, satisfying the condition 4h<H≤24h. This task is classified as P2 level.

[0033] Calculated based on the time difference between the task's required completion time and the current scheduling time: , Standard Priority Determined according to the following rules: , Among them, P1 is the highest priority task, P2 is the medium-high priority task, and P3 is the normal priority task. Task priority is used to influence the candidate AGV selection order, path resource usage order, and multi-objective comprehensive evaluation weight. In one optional implementation, task priority is used to influence the candidate AGV selection order, path resource reservation order, and the weight of evaluation indicators in multi-objective comprehensive evaluation. After a new task is inserted, the task time limit changes, equipment failure occurs, or a path blocking event occurs, the remaining completion time of unfinished tasks is recalculated, and the corresponding task priorities and task execution order are updated.

[0034] In this embodiment, task priority determination accuracy is used as one of the core evaluation indicators to assess whether the system's automatic determination of task priorities conforms to preset rules. The system output priority is denoted as... Standard priority is denoted as , judgment function The value is 1 if the condition is met, and 0 otherwise. The formula for calculating the accuracy of task priority determination is: , in, Determine accuracy based on task priority. This represents the total number of valid task samples. Valid task samples refer to tasks generated by the system and entering the AGV scheduling process within the statistical period. Tasks canceled before scheduling, tasks with missing test data fields that cannot be calculated, and tasks manually reassigned and whose original system output cannot be restored are not included in the valid samples. When the indicator reaches the target value.

[0035] S3: The scheduling execution agent performs hard constraint verification on the status attributes and capability attributes of candidate AGVs based on real-time resource status snapshots, and selects a set of candidate AGVs that meet the task execution conditions.

[0036] Preferably, in step S3, hard constraint verification is performed on the AGV's state attributes and capability attributes, including: Hard constraint verification includes verifying at least one of the following: AGV online status, fault status, order-accepting status, load capacity, vehicle size, task type adaptability, and path reachability. AGVs that satisfy the hard constraint verification items are added to the initial candidate AGV set. When any verification item is not satisfied, the corresponding AGV is removed. Specifically, for the i-th task and the j-th AGV, online status verification value, fault-free status verification value, order-acceptable status verification value, load matching verification value, vehicle size matching verification value, task type matching verification value, and path reachability verification value are determined respectively. Each verification value is set to 1 when the corresponding condition is met, and 0 otherwise. The product of each verification value is used as the feasibility indicator between the i-th task and the j-th AGV. When the feasibility indicator is 1, the j-th AGV is added to the initial candidate AGV set of the i-th task; when the feasibility indicator is 0, the j-th AGV is removed from the initial candidate AGV set.

[0037] In one specific implementation, the scheduling agent calculates the estimated empty-run time for each AGV in the initial candidate AGV set from its current position to the task start point, and determines the available order-accepting time after the AGV finishes its current task. The initial candidate AGV set is sorted based on the estimated empty-run time, available order-accepting time, remaining battery power, current task load, and task priority. A preset number of AGVs are selected from the sorted initial candidate AGV set to form a candidate AGV set. If no AGV satisfies the hard constraints, the task is placed in the scheduling queue, or the resources of a low-priority task currently being executed are re-matched.

[0038] For example, continuing with the aforementioned manufacturing workshop scenario, the state-aware agent obtains the real-time status of all AGVs in the workshop: AGV-03 is located near warehouse 2, with a speed of 1.2 m / s, 85% battery, rated load of 500 kg, current load of 0 kg, online status, no fault status, and currently idle; AGV-07 is located near warehouse 1, with a speed of 1.2 m / s, 92% battery, rated load of 500 kg, current load of 0 kg, online status, no fault status, and currently idle; AGV-12 is located in aisle B, with a speed of 0.8 m / s, 45% battery, rated load of 500 kg, current load of 350 kg, online status, no fault status, currently executing task T-1025, and task completion rate of 60%. The scheduling and execution agent performs hard constraint verification on the above AGVs respectively, and adds AGV-03 and AGV-07, which meet all verification items, to the candidate AGV set.

[0039] S4: The scheduling execution agent generates a candidate path for each AGV in the candidate AGV set, converts the candidate path into a spatiotemporal occupancy sequence containing nodes, road segments and resource occupancy time intervals, detects spatiotemporal conflicts based on the spatiotemporal occupancy sequence, and generates a conflict-free candidate scheduling scheme by performing conflict resolution processing.

[0040] Please see Figure 3 As shown, in step S4, a conflict-free candidate scheduling scheme is generated, including: The scheduling and execution agent generates at least one candidate path for each candidate AGV that avoids restricted and blocked areas, based on the task start point, task end point and map topology. Based on the candidate AGV's expected travel speed, the length of each road segment, the loading and unloading operation time, and the intersection passage time, each candidate path is converted into a node occupation time interval, a road segment occupation time interval, and a loading and unloading resource occupation time interval. Compare the node occupancy time interval, road segment occupancy time interval, and loading / unloading resource occupancy time interval with the resource reservation information corresponding to the issued scheduling plan; For node conflicts, road segment conflicts, meeting conflicts, or loading and unloading resource conflicts, the priority order of resource occupation is determined according to task priority, and at least one of the following methods is used to resolve the conflict: delaying entry into conflicting resources, adjusting driving speed, selecting alternative paths, and replacing candidate AGVs. After conflict resolution, spatiotemporal conflict detection is performed again, and candidate AGVs, candidate paths, and resource occupancy time intervals for which no spatiotemporal conflicts were detected are combined into conflict-free candidate scheduling schemes. If spatiotemporal conflicts still exist after the above resolution process, the current candidate AGV-path combination is marked as invalid, and the next AGV is selected according to the initial candidate AGV set order to re-execute path planning and conflict detection until a valid scheme is obtained or the candidate set is traversed. If all candidate schemes are invalid, the current task is placed in the scheduling queue and the central control orchestration agent is notified.

[0041] Preferably, in step S4, detecting spatiotemporal conflicts based on the spatiotemporal occupancy sequence includes: Overlap and compare the spatiotemporal occupancy sequences of different AGVs in both time and space dimensions; When different AGVs are detected occupying the same node, the same road segment in opposite directions, the same narrow passage, or the same loading / unloading resource within an overlapping time interval, it is determined to be a conflict. Specifically, in spatiotemporal conflict detection, when different AGVs occupy the same node within an overlapping time interval, it is determined to be a node conflict; when different AGVs occupy the same road segment in opposite directions within an overlapping time interval, it is determined to be a road segment conflict; when different AGVs enter a narrow passage that does not allow two-way traffic within an overlapping time interval, it is determined to be a meeting conflict; and when different AGVs occupy the same loading / unloading resource within an overlapping time interval, it is determined to be a loading / unloading resource conflict. The conflict resolution process includes: determining the resource occupancy order based on task priority, and generating conflict-free paths by waiting, detouring, speed adjustment, or replacing candidate AGVs.

[0042] In one alternative implementation, the execution order of the conflict resolution process is as follows: First, try to delay entering the conflicting resource, that is, stagger the occupied intervals in the time dimension; when delay cannot resolve the conflict, try to adjust the AGV's travel speed to change the time of arrival at the conflict node; when speed adjustment still cannot resolve the conflict, select an alternative path to bypass the conflict area; when the alternative path is unavailable or too costly, replace the candidate AGV and re-execute path planning and conflict detection.

[0043] S5: The scheduling execution agent performs a multi-objective comprehensive evaluation of conflict-free candidate scheduling schemes based on task priorities, determines the target scheduling scheme, and outputs it.

[0044] Preferably, in step S5, the multi-objective comprehensive evaluation includes: The scheduling agent calculates evaluation metrics for conflict-free candidate scheduling schemes, including task completion time, path length, energy consumption, path congestion, low power risk, and load balancing deviation. Based on the maximum and minimum values ​​of each evaluation metric among all conflict-free candidate scheduling schemes, the metrics are normalized to obtain the corresponding normalized evaluation values.

[0045] The weights of each evaluation indicator are dynamically adjusted according to task priority: for high-priority tasks (P1), the weight of task completion time is increased; for medium-priority tasks (P2), the weights of task completion time and resource utilization efficiency are balanced; for ordinary-priority tasks (P3), the weights of energy consumption and load balancing are increased. Conflict-free candidate scheduling schemes are sorted in ascending order of comprehensive cost, and the conflict-free candidate scheduling scheme with the lowest comprehensive cost that passes the consistency checks of task constraints, vehicle capacity constraints, and spatiotemporal conflict constraints is determined as the target scheduling scheme.

[0046] For example, for the material A transportation task (P1 level), the scheduling execution agent evaluates the candidate solutions of AGV-03 and AGV-07. The candidate solution of AGV-07 has a lower overall cost and passes the consistency check. Therefore, AGV-07 is determined as the target AGV, and the corresponding equipment assignment, path planning, and timing arrangement are output.

[0047] The scheme with the lowest overall cost is selected as the target scheduling scheme based on the weighted results.

[0048] Preferably, the scheme with the minimum overall cost is selected as the target scheduling scheme based on the weighted result, including: The number is determined according to the following formula. The combined cost of conflict-free candidate scheduling schemes : , in, The estimated completion time for the normalization task. To normalize the path length, To normalize the projected energy consumption, To normalize the level of path congestion, To normalize the risk of low battery levels, To normalize the AGV load balancing deviation, , , , , , The weights are assigned to the corresponding evaluation indicators, and the sum of all weights is 1. Specifically, the highest priority (P1) is: , , , , , Medium priority (P2): , , , , , Normal priority (P3): , , , , , The scheduling scheme must meet the following constraints: each task is assigned to one AGV for execution, the weight of materials carried by the AGV does not exceed its rated load capacity, and the real-time flow of any path does not exceed the capacity limit of that path.

[0049] In this embodiment, personalized material scheduling speed is used as one of the core evaluation indicators to evaluate the system's response efficiency in generating scheduling plans and issuing tasks for personalized orders, customized materials, specific process materials, or customized processed parts. Let the generation time of the i-th personalized material scheduling request be denoted as . The system completes the generation of the scheduling plan or the issuance of tasks in the following time. The scheduling time for this task is: , The specified completion time threshold is denoted as . The default statistical threshold can be set to 5 or 10 minutes, and the specific value can be set according to the size of the workshop.

[0050] The speed target achievement result for a single task is recorded as : , Calculation formula: , in, To achieve the target speed for personalized material scheduling, To determine the total number of tasks for effective personalized material scheduling. If If so, then the indicator reaches the target value.

[0051] S6: The central control orchestration agent issues the target scheduling scheme, and the real-time status monitoring and execution feedback agent monitors the execution process and collects abnormal events. When an abnormal event is detected, the central control orchestration agent determines the rescheduling scope according to the type of abnormality, retains the unaffected part of the scheduling results, and triggers the rescheduling process only for the affected part to update the scheduling scheme.

[0052] like Figure 4 As shown, the overall process includes six stages: scheme issuance, AGV execution, real-time monitoring, anomaly detection, feedback control and rescheduling triggering, and scheme update output.

[0053] In this embodiment, the real-time status monitoring and execution feedback agent compares the AGV's actual position, actual speed, remaining battery power, task completion rate, path occupancy status, and loading / unloading status with the planned position, planned arrival time, planned battery consumption, and resource occupancy time interval in the target scheduling scheme. A path anomaly event is generated when the AGV cannot enter the target road segment within the planned time interval, the target road segment is marked as blocked, or the path travel time exceeds the preset travel time; a vehicle anomaly event is generated when the AGV is marked as faulty, offline, or its remaining battery power is lower than the battery power required to complete the current task; and a task anomaly event is generated when a new task is received or the task completion time changes.

[0054] Preferably, in step S6, determining the rescheduling scope based on the exception type includes: When the abnormal event is path congestion or path blockage, only the local path of the affected AGV is replanned, while the assignment results of other AGVs remain unchanged. When the abnormal event is equipment failure or low battery, cancel the resource reservation of the abnormal AGV and re-match the AGV resources for the affected task. When an abnormal event is the insertion of a new task or a change in task time limit, the task priority is re-divided, and resource matching and path planning are performed on the affected task set.

[0055] For example, AGV-07 is performing a material transport task along path P1. When it reaches the vicinity of intersection X-03, the real-time status monitoring and execution feedback agent detects a sudden congestion ahead on the main road in area A. This anomaly is determined to be a path anomaly event, and the central control and scheduling agent triggers a local path replanning. Based on the current position of AGV-07 and the task endpoint, the scheduling execution agent searches for an alternative path and selects a route from X-03 via a branch road in area B to assembly line 3. This route is 15 meters longer than the original path but avoids the congested area, and the estimated completion time is increased by 3 minutes. The updated scheduling plan is then issued for execution after passing a consistency check.

[0056] This embodiment also includes a scheduling effect evaluation step. Based on system task records, AGV status records, path planning records, and scheduling logs, statistics are performed on the accuracy of task priority determination, the rationality of task allocation in the multi-AGV collaborative scheduling model, and the speed of personalized material scheduling. The accuracy of task priority determination is used to represent the proportion of tasks whose system output priority is consistent with the standard priority rules. When this indicator reaches 95% or more, the target value is reached. The rationality of task allocation in the multi-AGV collaborative scheduling model is used to represent the proportion of tasks that meet the conditions of AGV availability, load or size matching, task type matching, path reachability, and no task conflicts. When this indicator reaches 90% or more, the target value is reached. The speed of personalized material scheduling is used to represent the proportion of tasks that complete the generation of scheduling schemes or task issuance within a specified time threshold. When this indicator reaches 90% or more, the target value is reached.

[0057] Through the orderly collaboration of the aforementioned Agentic AI multi-agent, this embodiment successfully transforms multi-source heterogeneous AGV scheduling data into a unified scheduling decision scheme, establishing a mapping link from task requests to executable scheduling schemes. This effectively solves the technical problems of difficulty in unified utilization of multi-source data, difficulty in systematically resolving multi-vehicle path conflicts and timing conflicts, and difficulty in quickly responding to dynamic changes such as temporary order insertions and equipment anomalies. Under multiple constraints, it generates an efficient scheduling scheme that fits the needs of the manufacturing site, significantly improving the efficiency of production logistics collaboration and the dynamic adaptability of the manufacturing system.

[0058] Second Embodiment Based on the same concept, this embodiment also provides a computer device, including a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of an AGV manufacturing resource dynamic scheduling method based on Agentic AI multi-agent orchestration as described in the embodiment.

[0059] Based on the same concept, the present invention also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of a dynamic scheduling method for AGV manufacturing resources based on Agentic AI multi-agent orchestration as described in any of the embodiments.

[0060] It is understood that, regarding the aforementioned AGV manufacturing resource dynamic scheduling method based on Agentic AI multi-agent orchestration, if all components are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer server or a network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0061] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0062] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration, characterized in that, The multi-agent system includes a central control and orchestration agent, a task parsing and priority determination agent, a state-aware agent, a scheduling and execution agent, and a real-time state monitoring and execution feedback agent, and includes the following steps: S1: The overall control and orchestration intelligent agent receives the AGV scheduling task request, calls the task parsing and priority determination intelligent agent to perform semantic parsing and standardization processing on the AGV scheduling task request, and generates standardized scheduling task information containing task attributes and constraints. S2: The state-aware agent acquires the real-time status of the AGV and map topology information, and generates a real-time resource status snapshot; the task parsing and priority determination agent determines the task priority based on the urgency of the task and time constraints. S3: The scheduling execution agent performs hard constraint verification on the status attributes and capability attributes of candidate AGVs based on the real-time resource status snapshot, and selects a set of candidate AGVs that meet the task execution conditions. S4: The scheduling execution agent generates a candidate path for each AGV in the candidate AGV set, converts the candidate path into a spatiotemporal occupancy sequence containing nodes, road segments and resource occupancy time intervals, detects spatiotemporal conflicts based on the spatiotemporal occupancy sequence, and generates a conflict-free candidate scheduling scheme by performing conflict resolution processing. S5: The scheduling execution agent performs a multi-objective comprehensive evaluation of the conflict-free candidate scheduling schemes based on task priorities, determines the target scheduling scheme, and outputs it. S6: The overall control orchestration agent issues the target scheduling scheme, and the real-time status monitoring and execution feedback agent monitors the execution process and collects abnormal events; when an abnormal event is detected, the overall control orchestration agent determines the rescheduling scope according to the type of abnormality, retains the unaffected part of the scheduling results, and triggers the rescheduling process only for the affected part to update the scheduling scheme.

2. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 1, characterized in that, In step S1, standardized scheduling task information is generated, including: Identify the source of AGV scheduling task requests and convert task requests of different formats into a unified task data structure; Extract the task number, task type, material identifier, source location, target location, required completion time, emergency status, load requirements, and process constraints from the task request; The source and target locations are matched with map topology nodes, and the material parameters and AGV capability parameters are processed to unify and normalize data units to generate the standardized scheduling task information.

3. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 1, characterized in that, In step S2, the task priority is determined, including: Obtain the urgency indicator for the task and the remaining time until the required completion time; When the emergency flag is set to the preset emergency value, or when the remaining time is less than the first time threshold, the task will be classified as the highest priority. When an emergency is marked as non-urgent and the remaining time is between the first and second time thresholds, the task is classified as medium priority. When an emergency flag is set to a non-urgent value and the remaining time is greater than the second time threshold, the task will be classified as a normal priority task.

4. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 1, characterized in that, In step S2, generating the real-time resource status snapshot includes: Obtain the AGV number, current position, travel speed, remaining battery power, rated load, current load, vehicle size, online status, fault status, running status, current task, and task completion rate for each AGV; The workshop map is represented as a topology map consisting of passable nodes and connecting road segments, and the length, permitted driving direction, traffic capacity, speed limit information, restricted status, congestion status and real-time occupancy status of each connecting road segment are obtained. Obtain the resource occupancy status of loading / unloading points, buffer zones, intersections, narrow passages, and charging areas, and calculate the estimated release time of the corresponding resources based on the AGV position, AGV speed, and current task; Add timestamps to AGV status data, map topology data, and resource usage data. When the difference between the timestamp of any data and the current scheduling time exceeds the preset data validity period, the corresponding data is re-acquired. Based on the data that has passed the time validity verification, a real-time resource status snapshot is generated, which includes AGV availability, vehicle capacity, path reachability, path congestion level, and resource occupancy time interval.

5. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 1, characterized in that, In step S3, hard constraint verification is performed on the AGV's state attributes and capability attributes, including: Verify at least one of the following: online status, fault status, order-accepting status, load capacity, vehicle size, task type adaptability, and path reachability of the AGV. AGVs that meet the hard constraint verification items are added to the initial candidate AGV set; when any verification item is not met, the corresponding AGV is removed.

6. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 1, characterized in that, In step S4, spatiotemporal conflicts are detected based on the spatiotemporal occupancy sequence, including: Overlap and compare the spatiotemporal occupancy sequences of different AGVs in both time and space dimensions; When different AGVs are detected occupying the same node, the same road segment in opposite directions, the same narrow passage, or the same loading / unloading resource within the overlapping time interval, it is determined to be a conflict; The resolution process includes: determining the resource occupancy order based on task priority, and generating a conflict-free path by using methods such as waiting, detouring, speed adjustment, or replacing candidate AGVs.

7. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 1, characterized in that, In step S4, generating the conflict-free candidate scheduling scheme includes: Based on the task start point, task end point and map topology, generate at least one candidate path for each candidate AGV to avoid restricted areas and congested areas. Based on the candidate AGV's expected travel speed, the length of each road segment, the loading and unloading operation time, and the intersection passage time, each candidate path is converted into a node occupation time interval, a road segment occupation time interval, and a loading and unloading resource occupation time interval. Compare the node occupancy time interval, road segment occupancy time interval, and loading / unloading resource occupancy time interval with the resource reservation information corresponding to the issued scheduling plan; For node conflicts, road segment conflicts, meeting conflicts, or loading and unloading resource conflicts, the priority order of resource occupation is determined according to task priority, and at least one of the following methods is used to resolve the conflict: delaying entry into conflicting resources, adjusting driving speed, selecting alternative paths, and replacing candidate AGVs. After the conflict is resolved, spatiotemporal conflict detection is performed again, and the candidate AGVs, candidate paths and resource occupancy time intervals for which no spatiotemporal conflict was detected are combined into the conflict-free candidate scheduling scheme.

8. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 1, characterized in that, In step S5, the multi-objective comprehensive evaluation includes: Calculate the evaluation indicators for the task completion time, path length, energy consumption, path congestion, low power risk, and load balancing deviation of the conflict-free candidate scheduling schemes respectively. The weights of each evaluation indicator are dynamically adjusted based on task priority: for high-priority tasks, the weight of task completion time is increased; for medium-priority tasks, the weights of task completion time and resource utilization efficiency are balanced; for ordinary-priority tasks, the weights of energy consumption and load balancing are increased. The scheme with the lowest overall cost is selected as the target scheduling scheme based on the weighted results.

9. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 8, characterized in that, Based on the weighted results, the scheme with the minimum overall cost is selected as the target scheduling scheme, including: The number is determined according to the following formula. The combined cost of conflict-free candidate scheduling schemes : in, The estimated completion time for the normalization task. To normalize the path length, To normalize the projected energy consumption, To normalize the level of path congestion, To normalize the risk of low battery levels, To normalize the AGV load balancing deviation, , , , , , The weights are the corresponding evaluation indicators, and the sum of the weights is 1.

10. The method for dynamic scheduling of AGV manufacturing resources based on Agentic AI multi-agent orchestration according to claim 1, characterized in that, In step S6, the rescheduling scope is determined based on the exception type, including: When the abnormal event is path congestion or path blockage, only the local path of the affected AGV is replanned, while the assignment results of other AGVs remain unchanged. When the abnormal event is equipment failure or low battery, cancel the resource reservation of the abnormal AGV and re-match the AGV resources for the affected task. When an abnormal event is the insertion of a new task or a change in task time limit, the task priority is re-divided, and resource matching and path planning are performed on the affected task set.

Images