Multi-agv distributed scheduling method based on multi-layer coordination mechanism in dynamic environment

Abstract

The application discloses a multi-AGV distributed scheduling method in a dynamic environment based on a multi-layer cooperative mechanism. First, an environment map is constructed and gridded; then, a path planning layer calculates task benefits, adopts a CBBA algorithm for task pre-allocation, and obtains pre-allocated AGVs; a path decision layer utilizes an Actor-Critic algorithm for path decision according to an environment state and a real-time state of the pre-allocated AGVs, and obtains a path planning strategy; the path planning layer performs path re-planning according to the path planning strategy, and obtains path node sequences of the pre-allocated AGVs; path costs of the pre-allocated AGVs are calculated based on the path node sequences; a task allocation layer calculates task benefits according to the path costs of the pre-allocated AGVs, and performs task re-allocation; when the pre-allocated AGVs and the re-allocated AGVs are the same AGV, the task is allocated to the AGV, otherwise, the task is re-allocated; finally, each AGV executes tasks in sequence according to a task sequence. The method solves problems such as scheduling conflicts and poor environment adaptability in a dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of robot scheduling technology, specifically relating to a distributed scheduling method for multiple AGVs in a dynamic environment based on a multi-layer collaborative mechanism. Background Technology

[0002] With the rapid development of intelligent manufacturing, intelligent warehousing, and automated logistics, Automated Guided Vehicles (AGVs), as core automated equipment for cargo handling, logistics sorting, and workshop operations, are gradually replacing manual operations. This not only significantly reduces labor costs but also greatly improves task execution efficiency and accuracy. The core of multi-AGV collaborative scheduling lies in the efficient coordination of task allocation and path planning; the level of collaborative optimization between the two directly determines the overall performance of the scheduling system.

[0003] While existing multi-AGV scheduling methods have incorporated distributed algorithms, reinforcement learning, and pathfinding algorithms, they still suffer from numerous technical shortcomings in terms of adaptability to dynamic environments and global collaborative optimization. Currently, task allocation and path planning are designed as independent modular systems, lacking a deeply coupled closed-loop collaborative mechanism. Path planning results are not fed back to the task allocation process in real time, easily leading to local optima problems such as "reasonable task allocation but excessively high path costs" or "optimal path planning but conflicting task allocations," resulting in insufficient global optimization capabilities. Most scheduling schemes employ single-layer or two-layer algorithm architectures, but fail to achieve layered decoupling and collaborative optimization of "task allocation-path decision-path planning," making it difficult to simultaneously consider the decentralization of distributed task allocation, the intelligence of path decision-making, and the high accuracy of path planning. This results in poor system scalability and adaptability to real-world scenarios. Currently, the integration of distributed optimization algorithms and reinforcement learning has become a research hotspot in AGV scheduling, but it lacks multi-dimensional perception of AGV states and environmental changes, as well as global collaborative optimization capabilities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a distributed scheduling method for multiple AGVs in a dynamic environment based on a multi-layer collaborative mechanism. This method realizes distributed dynamic task allocation, adaptive path decision-making, and high-precision path planning for multiple AGV systems, and solves problems such as scheduling conflicts, poor environmental adaptability, and insufficient global optimization capabilities in dynamic environments. This improves the robustness, real-time performance, and operational efficiency of the multiple AGV scheduling system.

[0005] The present invention solves the aforementioned technical problem by adopting the following technical solution: A distributed scheduling method for multiple AGVs in a dynamic environment based on a multi-layer collaborative mechanism, characterized by the following steps: Step 1: Build an environment map and grid it; Step 2: Perform task pre-assignment; First, the path planning layer determines the tasks to be assigned. Based on the location and real-time location of each AGV, the A* algorithm is used for preliminary path planning to obtain the preliminary path node sequence of each AGV; based on the preliminary path node sequence, the path cost of each AGV is calculated according to the path cost function of equation (1); (1) In the formula, For the first AGV performs tasks Path cost, Adjacent path nodes and Geometric distance between them Path node The neighborhood congestion rate, Path node The amount of congestion penalties in the surrounding area For the first AGV performs tasks The path node sequence at that time , , These are distance weight, congestion weight, and replanning weight, respectively. and ; Then, the task revenue is calculated according to equation (2), and the task allocation layer uses the CBBA algorithm to pre-allocate tasks to obtain pre-allocated AGVs; (2) In the formula, For the first AGV performs tasks The benefits, For the task priority, , The respective The remaining power and total power of each AGV. , and All are weighting coefficients and ; Step 3: The path decision layer uses the Actor-Critic algorithm to make path decisions based on the environmental conditions and the real-time status of the pre-assigned AGVs, and obtains the path planning strategy. Step 4: The path planning layer performs path replanning based on the path planning strategy to obtain the path node sequence of the pre-assigned AGVs; based on the path node sequence of the pre-assigned AGVs, the path cost of the pre-assigned AGVs is calculated using the path cost function. If the path planning strategy is Strategy 1, then the A* algorithm is used for path search to obtain the path node sequence for the pre-assigned AGVs; if the path planning strategy is Strategy 2, then... , , At each time step, the path nodes occupied by other AGVs are treated as prohibited path nodes and removed. Then, the A* algorithm is used for path search to obtain the path node sequence of the pre-assigned AGVs. If the path planning strategy is Strategy 3, then based on Strategy 2, the path nodes in the path cost function are... The neighborhood congestion penalty is forcibly set to 1; Step 5: The task allocation layer calculates the task reward based on the path cost of the pre-allocated AGV and redistributes the task to obtain the redistributed AGV; if the pre-allocated AGV and the redistributed AGV are the same AGV, then the task is... Assign the task to the AGV; otherwise, reassign the AGV as the pre-assigned AGV and repeat steps three through five until the reassigned AGV and the pre-assigned AGV are the same AGV, and then assign the task to the AGV. Assign it to this AGV; Update the environment map, repeat steps 2 to 5, traverse all remaining tasks to be assigned, assign each task to the corresponding AGV, until the set of tasks to be assigned is empty, and obtain the task sequence of each AGV. Step 6: Schedule the AGVs according to the task sequence. The AGVs execute tasks in sequence according to the task sequence. When a new task appears in the task set to be assigned, repeat steps 2 to 5 to assign tasks.

[0006] Furthermore, during the task allocation process using the CBBA algorithm, each AGV broadcasts its own task sequence and revenue to the global database. AGVs bid based on their own task capacity and revenue. When multiple AGVs bid for the same task simultaneously, the AGV revenues are compared, and the task is allocated to the AGV with the highest revenue. If the number of AGVs with the highest revenue is greater than or equal to 2, the number of tasks already allocated to each AGV is further compared, and the task is allocated to the AGV with fewer tasks. If the number of AGVs with fewer tasks is greater than or equal to 2, the remaining battery power of each AGV is further compared, and the task is allocated to the AGV with more remaining battery power.

[0007] Furthermore, the heuristic function of the A* algorithm is: (5) In the formula, Path node To the target path node Heuristic cost estimation Path node With the target path node Geometric distance between them Path node The neighborhood congestion rate, This represents the congestion penalty coefficient.

[0008] Furthermore, the state of the Actor-Critic algorithm includes at least the real-time position and remaining battery power of the pre-assigned AGV, task position, task priority, dynamic obstacle distribution, and road network congestion rate. The action is a path planning strategy, and the reward function is: (3) In the formula, To pre-assign AGVs to perform tasks Instant rewards received To pre-assign AGVs to perform tasks Progress rewards obtained To pre-assign AGVs to perform tasks The path cost.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly improved global collaborative optimization capability: This invention constructs a closed-loop feedback mechanism between layers of "task allocation layer - path decision layer - path planning layer", unifying the upper-layer task benefit evaluation, the middle-layer strategy generation and the lower-layer path cost calculation under the same optimization objective, realizing the transformation from the serial process of "allocation first and planning later" to the collaborative iterative process of "pre-allocation - decision - planning - re-allocation". The path planning layer continuously feeds back the path cost to the task allocation layer and dynamically corrects the task benefits, which can significantly reduce the global efficiency loss caused by local optima, avoid the mismatch problem of "task allocation is feasible but the road network is severely congested" or "the path is shortest but the system time is longer", and make the overall scheduling result closer to the global optimum.

[0010] (2) Strong adaptability to dynamic environment: The neighborhood congestion rate reflects the dynamic environment. By introducing the neighborhood congestion rate into the path low-cost function, the dynamic environment directly affects the task revenue of AGV, which can improve the adaptability of AGV to the dynamic environment.

[0011] (3) High robustness of distributed scheduling: This invention adopts a decentralized distributed negotiation mechanism, in which each AGV can complete task bidding and conflict resolution based on local perception and global broadcast information, avoiding strong dependence on a single central control node. Even if individual AGVs experience communication anomalies, execution delays, or failures, the remaining AGVs can still maintain system operability through consensus negotiation and task release and reallocation mechanisms, avoiding global paralysis caused by "single point failure". At the same time, the distributed structure has natural scalability, which can achieve smooth expansion in terms of the number of AGVs, task density, and site size, meeting the engineering requirements of large-scale cluster scheduling.

[0012] (4) Balancing scheduling accuracy and real-time performance: This invention maintains path-solving accuracy while ensuring real-time response through a collaborative paradigm of "mid-level policy constraints + lower-level fine-grained search". The policy network first obtains the path planning strategy based on the environmental state and the real-time state of the AGV, reducing the path planning search space and suppressing invalid expansion; the path planning layer uses an improved A* algorithm to perform physical-level path solving based on congestion penalty terms and passability constraints, generating a collision-free, landable, high-precision path node sequence. This collaborative paradigm can still stably balance computation speed and accuracy in highly dynamic scenarios, reducing the risk of accuracy degradation or computation timeout common in real-time systems.

[0013] (5) Simultaneous improvement of task allocation rationality and system efficiency: This invention incorporates AGV real-time location, remaining resources, task priority, path cost, and neighborhood congestion rate into the revenue function, forming a multi-factor task evaluation system oriented towards execution feasibility; and reduces conflict propagation through a three-level conflict resolution and chain cancellation mechanism. Compared with allocation methods based solely on distance or static priority, this invention can more effectively suppress task overload, resource idleness, and peak congestion, improve task allocation balance and resource utilization, and further reduce redundant scheduling overhead and invalid mileage under multi-task concurrency conditions, thereby improving the ability to complete tasks per unit time. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the three-layer architecture of the present invention; Figure 2 This invention provides a flowchart of the task allocation process based on the CBBA algorithm. Figure 3 The path decision-making flowchart of this invention. Detailed Implementation

[0015] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail and are not intended to limit the scope of protection of this application.

[0016] like Figure 1-3 As shown, this invention provides a distributed scheduling method for multiple AGVs in a dynamic environment based on a multi-layer collaborative mechanism, comprising the following steps: Step 1: Construct an environment map; The environmental map contains coordinate information for roads, shelving areas, charging points, work sites, and restricted areas. The environmental map is gridded (0.5m x 0.5m grid size), with each grid cell as a node and accessible paths between nodes as edges, and stored in a graph structure. ;in, For a set of nodes, Let it be the set of edges.

[0017] Let the task set be denoted as , For the first One task, The number of tasks is denoted as , and each task includes task location, task type, task priority, and task deadline; the AGV set is denoted as . , For the first One AGV, The number of AGVs is specified, and each AGV includes its real-time location, remaining battery power, and rated battery power; the task sequence is denoted as... , For the first The task sequence of an AGV; Step 2: In the task pre-allocation stage, firstly, the path planning layer uses the A* algorithm to perform preliminary path planning based on the current location of the task to be assigned and the real-time location of each AGV, and obtains the preliminary path node sequence of each AGV; based on the preliminary path node sequence, the path cost of each AGV is calculated according to the path cost function and fed back to the task allocation layer. The path cost function is: (1) In the formula, For the first AGV performs tasks Path cost, Adjacent path nodes and Geometric distance between them Path node The congestion rate in the neighborhood; Path node The neighborhood congestion penalty is set to 1 when the neighborhood congestion rate is greater than the congestion rate threshold, and 0 otherwise. For the first AGV performs tasks The path node sequence generated by the path planning layer. , , These are distance weight, congestion weight, and replanning weight, respectively. ,and ; Then, the task allocation layer pre-allocates tasks based on the path cost of each AGV using an improved CBBA algorithm, resulting in pre-allocated AGVs; specifically, for the current task to be allocated... The revenue of the AGV in performing the task is calculated based on the task revenue function. The task revenue comprehensively considers the path cost, task priority, and remaining battery power of the AGV. Therefore, the task revenue function is: (2) In the formula, For the first AGV performs tasks The benefits, For the first AGV performs tasks Path cost; For the task Priority indicates the urgency of a task; , The respective The remaining power and total power of each AGV. , and All are weighting coefficients, and satisfy the following conditions: ; Each AGV broadcasts its task sequence and rewards globally, and each AGV bids based on its task capacity and rewards. If multiple AGVs bid for the same task simultaneously, it indicates a bidding conflict, which must be resolved through a three-level conflict resolution mechanism. Specifically, first, the AGV rewards are compared, and the task is assigned accordingly. Pre-allocate tasks to the AGV with the highest profit; if the number of AGVs with the highest profit is greater than or equal to 2, then further compare the number of tasks already assigned to each AGV and allocate tasks accordingly. Pre-assign tasks to AGVs with fewer tasks; if the number of AGVs with fewer tasks is greater than or equal to 2, then further compare the remaining battery power of the AGVs and assign tasks accordingly. Pre-allocate to AGVs with more remaining power.

[0018] The core purpose of the pre-allocation phase is to allow each AGV to negotiate consensus through distributed communication, thereby pre-screening obviously unreasonable allocation combinations without consuming a large amount of path search computing power, and providing a deterministic starting point for task assignment for subsequent fine-grained calculations.

[0019] Step 3: The path decision layer uses the Actor-Critic algorithm to make path decisions based on the environmental conditions and the real-time status of the pre-assigned AGVs, and obtains the path planning strategy; the path decision layer then distributes the path planning strategy to the path planning layer. The Actor-Critic algorithm fuses and encodes the environmental state and the pre-allocated AGV's real-time state (including real-time location and remaining battery power) to define the state. At a minimum, this includes the real-time location, remaining battery power, task location, task priority, dynamic obstacle distribution, and road network congestion rate of the pre-assigned AGV; defining the actions. For path planning strategies; Policy networks utilize policy functions According to the status Output Action Value networks utilize state-value functions, based on the state... Output state value Used to evaluate in state The long-term benefits; During the strategy update process, the path planning layer provides real-time feedback on the path cost of the pre-assigned AGVs, and the path decision layer determines the path cost based on the pre-assigned AGVs (…). The path cost is used to calculate the immediate reward; (3) In the formula, To pre-assign AGVs to perform tasks Instant rewards received To pre-assign AGVs to perform tasks Progress rewards obtained To pre-assign AGVs to perform tasks Path cost; If the policy network uses temporal differential error for parameter updates, then: (4) In the formula, for Temporal difference error of time-state value , They are respectively , The value of a state at any given moment for The state at any given moment; This is the discount factor, with a value range of (0,1]. This is the state value correction coefficient, with a value range of (0,1]. Policy network based on temporal difference error By adjusting the weight parameters of the policy function through backpropagation, the selection probability of high-state-value actions can be increased. If the state value of the current action is higher than expected, the probability of selecting the current action needs to be increased; if If the current action yields lower-than-expected benefits, it may lead to long distances, congestion, or replanning, and the probability of choosing the current action needs to be reduced. Value networks aim to minimize the mean square loss of temporal difference error and correct the parameters of the state value function, making the state value assessment more and more in line with the real environment, thereby improving the accuracy of the assessment.

[0020] Step 4: The path planning layer performs path replanning based on the path planning strategy issued by the path decision layer to obtain the path node sequence of the pre-allocated AGVs; based on the path node sequence of the pre-allocated AGVs, the path cost of the pre-allocated AGVs is calculated using the path cost function and fed back to the task allocation layer. If the path planning strategy is Strategy 1, then the A* algorithm is used to search for a feasible path from the starting path node to the target path node, avoiding only static obstacles and path nodes currently occupied by other AGVs, thus obtaining the path node sequence for the pre-assigned AGVs; if the path planning strategy is Strategy 2, then... The path nodes occupied by other AGVs at any given time and , At any given time, the path nodes occupied by other AGVs are designated as prohibited path nodes. After removing prohibited path nodes, the A* algorithm is used for path searching, forcing the pre-assigned AGVs to detour to avoid conflict areas, resulting in a sequence of path nodes for the pre-assigned AGVs. If the path planning strategy is Strategy 3, then based on Strategy 2, the path nodes in the path cost function... The neighborhood congestion penalty is forcibly set to 1, indicating that congestion has occurred in the current constraint space; In this embodiment, the heuristic function of the A* algorithm, based on the geometric distance between path nodes, introduces the neighborhood congestion rate of path nodes to characterize the dynamic environment. Therefore, the heuristic function is: (5) In the formula, Path node To the target path node Heuristic cost estimation Path node With the target path node Geometric distance between them Path node The neighborhood congestion rate, The local congestion rate is 0. This is the congestion penalty coefficient, with a value range of (0,1]. Step 5: The task allocation layer calculates the task reward based on the path cost of the pre-allocated AGV and redistributes the task to obtain the redistributed AGV; if the pre-allocated AGV and the redistributed AGV are the same AGV, then the task is... Assign the task to the AGV; otherwise, use the currently reassigned AGV as the pre-assigned AGV, and repeat steps three through five until the reassigned AGV and the pre-assigned AGV are the same AGV, and then assign the task to the AGV. Assign it to this AGV; For those who have not been assigned a task The AGV must execute a chain reaction of order cancellations, not only removing the task from its own task sequence but also removing tasks located in the same task sequence. All subsequent tasks are released together and returned to the task allocation set; Update the environment map, repeat steps 2 to 5, traverse all remaining tasks to be assigned, assign each task to the corresponding AGV, until the set of tasks to be assigned is empty, and obtain the task sequence of each AGV.

[0021] Step 6: The scheduling system schedules the AGVs according to the task sequence, and the AGVs execute the tasks in sequence. When a new task appears in the task set, steps 2 to 5 are repeated to allocate the task.

[0022] Example This embodiment uses warehousing scenarios of different sizes as examples to verify the effectiveness of the method of the present invention.

[0023] Small-scale warehousing scenario: 20 handling tasks were randomly generated in a 10×10 warehousing scenario, with 3 AGVs. The system allowed a maximum running time of 300 time units. The experimental results are shown in Table 1: Table 1 Experimental results in a small-scale warehousing scenario

[0024] Medium-scale warehouse scenario: 100 handling tasks were randomly generated in a 20×20 warehouse scenario, with 5 AGVs. The system's maximum allowed running time was 2000 time units. The experimental results are shown in Table 2. Table 2 Experimental results for medium-scale warehousing scenarios

[0025] Large-scale warehousing scenario: 300 tasks were randomly generated in a 30×30 warehousing scenario, with 10 AGVs. The system's maximum allowed running time was 3000 time units. The experimental results are shown in Table 3. Table 3 Experimental results in large-scale warehousing scenarios

[0026] The number of replanning attempts represents the number of times a path is replanned. Replanning consumes a significant amount of CPU computing power and causes the AGV to briefly stop in place. Fewer replanning attempts mean lower computing power overhead, and it also indicates that potential high-risk areas have been successfully predicted or avoided during task allocation and path planning, resulting in strong anti-interference capabilities. The methods used in the experiments all have a passive replanning mechanism, which triggers passive replanning when the AGV's waiting time at the same path node exceeds a preset threshold. The method of this invention also has an active replanning mechanism, which triggers active replanning when the pre-assigned AGV and the reassigned AGV are not the same AGV. Congestion refers to the phenomenon where multiple AGVs are assigned to the same area, leading to excessively high vehicle density and reduced traffic efficiency in that area. Fewer congestion triggers mean that the algorithm does not blindly pursue the shortest distance per vehicle, but rather distributes AGVs to different areas through task allocation and path decision-making. In this embodiment, when the number of other AGVs within three grid ranges near the current position of an AGV is greater than or equal to three, congestion is determined to have occurred in that area, and the congestion trigger count is incremented by one. A conflict refers to two or more AGVs about to collide at close range. Conflict avoidance often involves sudden braking and restarting, which consumes a lot of power and time. Conflict avoidance is rare, indicating that the travel trajectories of each AGV are not only staggered in "space" but also in "time". In this embodiment, when the next path node in the path planning of a certain AGV is occupied by another AGV, or when another AGV is about to occupy the node in the next planning step, conflict avoidance is triggered, and the conflict avoidance count is incremented by 1.

[0027] The experimental results show that the method of this invention exhibits optimal performance in warehousing scenarios of different scales. The average time, average travel distance, number of replanning attempts, number of congestion triggers, and number of conflict avoidance triggers are all significantly lower than other methods. The core reason for this is that multi-AGV collaboration is inherently a highly dynamic environment. When AGVs execute tasks, "endogenous congestion" and "moving obstacles" are formed within the system. The greedy allocation + A* algorithm tends to let AGVs grab the task closest to them in a straight line, causing multiple AGVs to frequently cluster in the same area. The CBBA + A* algorithm only considers task priority and remaining AGV power during task allocation, directly allocating tasks to the AGV with the highest benefit. This leads to multiple AGVs, lacking path cost information guidance, tending to choose the same or adjacent high-priority tasks. Although the CBBA algorithm avoids multiple vehicles grabbing the same task through conflict resolution mechanisms, the task sequences of each AGV still contain a large number of spatially adjacent tasks, causing multiple vehicles to frequently enter the same area during subsequent execution, exacerbating local traffic density and traffic conflicts. The task allocation layer of this invention has a global perspective. Combined with the path cost feedback from the decision layer, it can consciously distribute AGVs to different areas during the allocation phase. Simultaneously, the closed-loop collaborative mechanism of "task allocation-path decision-path planning" proposed in this invention can effectively reduce scheduling oscillations caused by dynamic disturbances, avoid local conflicts and congestion, and resolve them in advance through dynamic revenue updates and AC decisions, thereby improving robustness and overall execution efficiency. This enables AGVs to complete tasks efficiently and reliably in complex dynamic environments.

[0028] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A distributed scheduling method for multiple AGVs in a dynamic environment based on a multi-layer collaborative mechanism, characterized in that, Includes the following steps: Step 1: Build an environment map and grid it; Step 2: Perform task pre-assignment; First, the path planning layer determines the tasks to be assigned. Based on the location and real-time location of each AGV, the A* algorithm is used for preliminary path planning to obtain the preliminary path node sequence of each AGV; based on the preliminary path node sequence, the path cost of each AGV is calculated according to the path cost function of equation (1); （1） In the formula, For the first AGV performs tasks Path cost, Adjacent path nodes and Geometric distance between them Path node The neighborhood congestion rate, Path node The amount of congestion penalties in the surrounding area For the first AGV performs tasks The path node sequence at that time , , These are distance weight, congestion weight, and replanning weight, respectively. and ; Then, the task revenue is calculated according to equation (2), and the task allocation layer uses the CBBA algorithm to pre-allocate tasks to obtain pre-allocated AGVs; （2） In the formula, For the first AGV performs tasks The benefits, For the task priority, , The respective The remaining power and total power of each AGV. , and All are weighting coefficients and ; Step 3: The path decision layer uses the Actor-Critic algorithm to make path decisions based on the environmental conditions and the real-time status of the pre-assigned AGVs, and obtains the path planning strategy. Step 4: The path planning layer performs path replanning based on the path planning strategy to obtain the path node sequence of the pre-assigned AGVs; based on the path node sequence of the pre-assigned AGVs, the path cost of the pre-assigned AGVs is calculated using the path cost function. If the path planning strategy is Strategy 1, then the A* algorithm is used for path search to obtain the path node sequence for the pre-assigned AGVs; if the path planning strategy is Strategy 2, then... , , At each time step, the path nodes occupied by other AGVs are treated as prohibited path nodes and removed. Then, the A* algorithm is used for path search to obtain the path node sequence of the pre-assigned AGVs. If the path planning strategy is strategy three, then based on strategy two, the path nodes in the path cost function are... The neighborhood congestion penalty is forcibly set to 1; Step 5: The task allocation layer calculates the task reward based on the path cost of the pre-allocated AGV and redistributes the task to obtain the redistributed AGV; if the pre-allocated AGV and the redistributed AGV are the same AGV, then the task is... Assign the task to the AGV; otherwise, reassign the AGV as the pre-assigned AGV and repeat steps three through five until the reassigned AGV and the pre-assigned AGV are the same AGV, and then assign the task to the AGV. Assign it to this AGV; Update the environment map, repeat steps 2 to 5, traverse all remaining tasks to be assigned, assign each task to the corresponding AGV, until the set of tasks to be assigned is empty, and obtain the task sequence of each AGV. Step 6: Schedule the AGVs according to the task sequence. The AGVs execute tasks in sequence according to the task sequence. When a new task appears in the task set to be assigned, repeat steps 2 to 5 to assign tasks.

2. The distributed scheduling method for multiple AGVs in a dynamic environment based on a multi-layer collaborative mechanism as described in claim 1, characterized in that, In the task allocation process using the CBBA algorithm, each AGV broadcasts its own task sequence and reward to the global database. AGVs bid for tasks based on their own task capacity and reward. When multiple AGVs bid for the same task simultaneously, the AGV rewards are compared, and the task is allocated to the AGV with the highest reward. If the number of AGVs with the highest reward is greater than or equal to 2, the number of tasks already allocated to each AGV is further compared, and the task is allocated to the AGV with fewer tasks. If the number of AGVs with fewer tasks is greater than or equal to 2, the remaining power of each AGV is further compared, and the task is allocated to the AGV with more remaining power.

3. The distributed scheduling method for multiple AGVs in a dynamic environment based on a multi-layer collaborative mechanism as described in claim 1, characterized in that, The heuristic function of the A* algorithm is: （5） In the formula, Path node To the target path node Heuristic cost estimation Path node With the target path node Geometric distance between them Path node The neighborhood congestion rate, This represents the congestion penalty coefficient.

4. The distributed scheduling method for multiple AGVs in a dynamic environment based on a multi-layer collaborative mechanism according to any one of claims 1 to 3, characterized in that, The state of the Actor-Critic algorithm includes at least the real-time position and remaining battery power of the pre-assigned AGV, task position, task priority, dynamic obstacle distribution, and road network congestion rate. The action is a path planning strategy, and the reward function is: （3） In the formula, To pre-assign AGVs to perform tasks Instant rewards received To pre-assign AGVs to perform tasks Progress rewards obtained To pre-assign AGVs to perform tasks The path cost.

Images