An agv cluster control system design platform and method

Abstract

The present application belongs to the field of cluster control, and particularly relates to an AGV cluster control system design platform and method, to solve the technical problem of poor robustness and self-adaptive ability of existing planning strategies in complex environments. The AGV cluster control method comprises the following steps: S1, starting from the space-time position of the task end point, performing reverse space-time search to generate a path set for the AGV from the current position to the task end point; in the node expansion process of the reverse space-time search: the expansion priority is determined by solving the gradient of the composite potential field; the transfer cost between nodes is calculated by using a resource bidding mechanism, the resource bidding mechanism determines the use right of the node through the bidding evaluation of each AGV; the expansion branch with the risk of circular waiting is identified and pruned; S2, according to the path set generated in S1, a main path and multiple alternative paths are determined for the AGV. If the AGV encounters a real-time conflict during task execution, the alternative path can be spliced into the main path to form a new path, without the need for global re-planning.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of cluster control, and particularly relates to an AGV cluster control system design platform and method. BACKGROUND

[0002] In a complex warehouse or factory environment, multiple AGVs need to share limited path resources and collaboratively complete a large number of carrying tasks. Existing AGV path planning methods are mostly based on classical graph search algorithms, such as the A* algorithm. Such methods perform well in handling single AGV path optimization in static environments, but have obvious limitations in high-density and high-concurrency cluster scenarios. When expanding nodes, the traditional A* algorithm usually only evaluates the estimated cost to the end point with a heuristic function, and is difficult to handle dynamic obstacles caused by the movement of other AGVs, which can easily lead to path conflicts and frequent waiting, reducing the overall efficiency of the cluster.

[0003] Strategies based on time windows or multi-agent path finding, such as conflict search, avoid collisions by reserving path resources in the time-space dimension for AGVs. However, existing strategies usually only focus on path length or travel time, and fail to comprehensively consider the urgency of tasks, the flexibility of path selection, and the impact of different paths on future traffic patterns, resulting in suboptimal resource allocation. Moreover, when dealing with complex interaction scenarios, it is easy to cause circular waiting or deadlock problems. Existing deadlock prevention or resolution mechanisms are often conservative or respond slowly. When unexpected congestion or unexpected situations occur in actual operation, AGVs lack flexible alternative solutions, and completely re-planning paths not only has large computational overhead, but also causes task interruption and system fluctuations. Therefore, there is an urgent need for a more intelligent cluster control method to cope with dynamic and complex environments. SUMMARY

[0004] The application provides an AGV cluster control system design platform and method to solve the technical problem of poor robustness and adaptability of existing planning strategies in complex environments.

[0005] In a first aspect, the application provides an AGV cluster control method, comprising the following steps:

[0006] S1, obtaining a preset grid map, current states of all AGVs, and task information to be executed; for the task assigned to the AGV, performing reverse space-time search from the space-time position of the task end point to generate a path set for the AGV from the current position to the task end point; during the node expansion process of the reverse space-time search:

[0007] The expansion priority is determined by solving the gradient of a composite potential field, which is superimposed by a static map potential field and a dynamic coupling potential field, wherein the dynamic coupling potential field is calculated based on the predicted trajectories of all AGVs except the current searching AGV;

[0008] The transfer cost between nodes is calculated by using a resource bidding mechanism, and the use right of the node is determined by the bidding evaluation of each AGV, wherein the bidding evaluation is determined according to the task priority and the path flexibility, and the path flexibility is used to quantify the diversion cost of the path after losing the current expansion node;

[0009] An instantaneous dependency graph between AGVs is constructed, and a causal chain analysis is performed to identify and prune the expansion branches with a risk of circular waiting, wherein the nodes in the instantaneous dependency graph are all AGVs, and the edges represent the resource dependency relationship between AGVs;

[0010] S2, according to the path set generated in S1, a main path and multiple alternative paths are determined for the AGV, and the alternative paths are used to dynamically splice a new execution path in response to real-time conflicts during execution.

[0011] Further, a reverse space-time search is performed in a three-dimensional space-time grid, wherein the space-time nodes in the three-dimensional space-time grid are composed of two-dimensional space grid coordinates (x, y) and discrete time steps t; the transfer between space-time nodes represents that the AGV moves from one grid to an adjacent grid or stays in place within a unit time.

[0012] Further, the process of determining the expansion priority by solving the gradient of the composite potential field includes the following steps:

[0013] Constructing a static map potential field: setting the potential field value of the grid where the fixed obstacle is located as a preset high value, and setting the potential field value of the passable grid as zero;

[0014] For AGVs other than the current searching AGV, the predicted trajectories of the AGVs within a future T time window are obtained, the nodes passed through by the predicted trajectories are modeled as repulsive sources, and the dynamic potential fields of the AGVs are constructed, and the potential field value of the dynamic potential field decays according to a Gaussian function as the space-time distance increases;

[0015] All the dynamic potential fields are superimposed to form a dynamic coupling potential field, and the dynamic coupling potential field is superimposed with the static map potential field to form a composite potential field;

[0016] The composite potential field gradient of the current expansion node to its neighbor nodes is calculated, and the neighbor node corresponding to the direction with the fastest gradient descent is selected as the highest priority expansion node.

[0017] Further, the process of calculating the transfer cost between space-time nodes by using the resource bidding mechanism includes the following steps:

[0018] For each AGV, the path branch of the AGV submits a bid value for the space-time node to be expanded according to the following formula : ;

[0019] wherein, is a normalized task priority, is a normalized value of path elasticity, and are preset weight coefficients;

[0020] The path elasticity is an increment of the cost of the path to change the route from the last node of the path after losing the current space-time node relative to the cost of the current optimal path;

[0021] The path branch with the highest bid value obtains the occupation right of the space-time node, and the transfer cost is inversely proportional to the highest bid value.

[0022] Further, the process of identifying and pruning the expansion branch with the risk of circular waiting includes the following steps:

[0023] At each search time step, a directed graph is created with all AGVs as nodes;

[0024] If one expansion branch of AGV A plans to occupy a space-time node, and the space-time node has been occupied by the current optimal candidate path of AGV B, a directed edge from A to B is created in the graph, indicating the resource dependence of A on B;

[0025] A depth-first search algorithm is used to detect whether there is a loop in the directed graph;

[0026] If the introduction of an expansion branch leads to the formation of a loop, it is determined that the expansion branch has the risk of circular waiting, and the expansion branch is pruned.

[0027] Further, in S2, the process of determining a main path and multiple candidate paths includes the following steps:

[0028] In the path set, select a path with the minimum cumulative transfer cost as the main path, and the cumulative transfer cost is the sum of all transfer costs on a path;

[0029] In the remaining paths of the path set, filter out paths with a node coincidence degree lower than a preset threshold and a cumulative transfer cost not higher than a preset multiple of the main path cost;

[0030] Arrange the filtered paths in ascending order of cumulative transfer cost, and select the top N paths as candidate paths.

[0031] In a second aspect, the present application provides an AGV cluster control system design platform, comprising:

[0032] The path set generation module is configured to acquire a preset grid map, current states of all AGVs, and task information to be executed; for a task assigned to an AGV, a reverse space-time search is performed from a space-time position of a task end point to generate a path set for the AGV from a current position to the task end point; during a node expansion process of the reverse space-time search:

[0033] The expansion priority is determined by solving a gradient of a composite potential field, and the composite potential field is formed by superimposing a static map potential field and a dynamic coupling potential field, wherein the dynamic coupling potential field is calculated based on predicted trajectories of all AGVs except a currently searched AGV;

[0034] The transfer cost between nodes is calculated by using a resource bidding mechanism, and the resource bidding mechanism determines the use right of a node through bidding evaluation of each AGV, wherein the bidding evaluation is determined according to a task priority and path elasticity, and the path elasticity is used to quantify a rerouting cost of a path after losing a currently expanded node;

[0035] An instantaneous dependency graph between AGVs is constructed, and a cause-effect chain analysis is performed to identify and prune an expanded branch with a risk of circular waiting, wherein nodes in the instantaneous dependency graph are all AGVs, and edges represent resource dependency relationships between AGVs;

[0036] The new path generation module determines a main path and multiple alternative paths for the AGV according to the generated path set, and the alternative paths are used to dynamically splice a new execution path in response to real-time conflicts during execution.

[0037] Further, the reverse space-time search is performed in a three-dimensional space-time grid, wherein a space-time node in the three-dimensional space-time grid is composed of a two-dimensional space grid coordinate (x, y) and a discrete time step t; and a transfer between space-time nodes represents that an AGV moves from one grid to an adjacent grid or stays in place within a unit time.

[0038] Further, when the expansion priority is determined by solving the gradient of the composite potential field, the static map potential field is constructed: a potential field value of a grid where a fixed obstacle is located is set as a preset high value, and a potential field value of a passable grid is set as zero;

[0039] For AGVs other than the currently searched AGV, predicted trajectories of the AGVs within a future T time window are acquired, nodes passed through by the predicted trajectories are modeled as repulsive sources, and respective dynamic potential fields are constructed, and a potential field value of the dynamic potential field decays according to a Gaussian function with an increase of a space-time distance;

[0040] All the dynamic potential fields are superimposed to form a dynamic coupling potential field, and the dynamic coupling potential field is superimposed with the static map potential field to form the composite potential field;

[0041] The compound potential field gradient of the current expansion node to its neighbor nodes is calculated, and the neighbor node corresponding to the direction of the fastest gradient descent is selected as the highest priority expansion node.

[0042] Further, when determining a main path and multiple alternative paths, in the path set, a path with the minimum cumulative transfer cost is selected as the main path, and the cumulative transfer cost is the sum of all transfer costs on the path;

[0043] In the remaining paths of the path set, paths with a node coincidence degree lower than a preset threshold and a cumulative transfer cost not higher than a preset multiple of the main path cost are screened out;

[0044] The screened paths are arranged in ascending order of cumulative transfer cost, and the top N paths are selected as the alternative paths.

[0045] The beneficial effects are: compared with the prior art, the static map potential field and the dynamic coupling potential field calculated based on all AGV predicted trajectories are superimposed to form a compound potential field, the node expansion priority in the reverse space-time search is generated through the compound potential field, so that the path planning can avoid potential traffic congestion areas in the future. At the same time, the transfer cost between nodes is calculated by using the resource bidding mechanism, the bidding evaluation in the resource bidding mechanism is determined according to the task priority and the path flexibility, the limited path resources can be more reasonably allocated to AGVs with high priority or limited path selection, through building an instantaneous dependence graph and performing causal chain analysis, the expansion branch with a risk of circular waiting can be pruned in the planning stage, avoiding the occurrence of deadlock phenomenon, and ensuring the smooth operation of the AGV cluster. In the present application, the path set includes a main path and multiple alternative paths, if the AGV encounters a real-time conflict during task execution, the alternative paths can be spliced into the main path to form a new path, without global re-planning. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The flowchart of the AGV cluster control method is shown in the figure.

[0047] Figure 2 The schematic diagram for identifying the expansion branch with a risk of circular waiting is shown in the figure. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application, and those skilled in the art should know that the embodiments described below are part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0049] The embodiments of the AGV cluster control method provided by the present application are:

[0050] As Figure 1 and Figure 2 shown, an AGV cluster control method, comprising the following steps:

[0051] S1, obtaining a preset grid map, current states of all AGVs, and task information to be executed; for a task assigned to an AGV, performing reverse space-time search from a space-time position of a task end point to generate a path set for the AGV from a current position to the task end point; in a node expansion process of the reverse space-time search:

[0052] an expansion priority is determined by solving a gradient of a composite potential field, the composite potential field being superimposed by a static map potential field and a dynamic coupling potential field, wherein the dynamic coupling potential field is calculated based on predicted trajectories of all AGVs except a current search AGV;

[0053] a transfer cost between nodes is calculated using a resource bidding mechanism, the resource bidding mechanism determining a use right of a node through bidding evaluation of each AGV, wherein the bidding evaluation is determined according to a task priority and a path elasticity, and the path elasticity is used to quantify a diversion cost of a path after losing a current expansion node;

[0054] an instantaneous dependency graph between AGVs is constructed, and a cause-effect chain analysis is performed to identify and prune an expansion branch with a risk of circular waiting, wherein nodes in the instantaneous dependency graph are all AGVs, and edges represent resource dependency relationships between AGVs.

[0055] In an optional embodiment, the reverse space-time search is performed in a three-dimensional space-time grid, wherein a space-time node in the three-dimensional space-time grid is composed of a two-dimensional space grid coordinate (x, y) and a discrete time step t; a transfer between space-time nodes represents that an AGV moves from one grid to an adjacent grid or stays in place in a unit time.

[0056] Assume that the entire warehouse floor is divided into multiple 1m x 1m grids, each of which corresponds to a unique two-dimensional coordinate (x, y); time is divided into multiple discrete time steps t in seconds. The node with coordinates (10, 25, 30) indicates that the AGV is located at the 10th row and 25th column grid position at the 30th second. In this way, the dynamic motion problem of the AGV is converted into a pathfinding problem in a three-dimensional space-time grid. When the AGV expands from the space-time node (10, 25, 30), the subsequent nodes represent the positions that the AGV can reach at the next time (i.e., 31 seconds). The subsequent space-time nodes include staying in place (node (10, 25, 31)), or moving to the adjacent four grids: space-time node (9, 25, 31), space-time node (11, 25, 31), space-time node (10, 24, 31), and space-time node (10, 26, 31). Each basic action of the AGV, including movement and waiting, is modeled as an edge in the three-dimensional space-time grid.

[0057] In an optional embodiment, the process of determining the expansion priority by solving the gradient of the composite potential field includes the following steps:

[0058] Constructing a static map potential field: setting the potential field value of the grid where the fixed obstacle is located to a preset high value, and setting the potential field value of the passable grid to zero;

[0059] For AGVs other than the current search AGV, obtain their predicted trajectories within a future T time window, model the nodes passed by the predicted trajectories as repulsive sources, construct their respective dynamic potential fields, and the potential field value of the dynamic potential field decays according to a Gaussian function as the space-time distance increases;

[0060] Superimpose all the dynamic potential fields to form a dynamic coupling potential field, and superimpose it with the static map potential field to form a composite potential field;

[0061] Calculate the composite potential field gradient from the current expansion node to its neighbor nodes, select the neighbor node corresponding to the direction with the fastest gradient descent as the highest priority expansion node.

[0062] For example, set the potential field value of the grid where the shelf or column in the warehouse is located to a maximum value (such as 10000), to ensure that the AGV avoids these fixed obstacles when performing path search, and set the potential field value of all passable grids to zero.

[0063] Dynamic coupling potential field is used to handle mutual avoidance among AGVs. Suppose AGV B predicts that it will pass the grid with coordinates (15, 40) at time 50, a repulsive force center will be generated at the spatiotemporal node (15, 40, 50). The strength of the repulsive force field decreases with the distance in space and time, for example, the potential field value at node (15, 40, 50) is 500, while at a location 1 second later or 1 meter away, its value may decay to 200. When there are multiple other AGVs, their respective dynamic potential fields together form a dynamic coupling potential field, which is added to the static potential field to form a composite potential field that changes over time.

[0064] In path search, when AGV A needs to decide the next step, it calculates the potential field change from the current expanded node to each adjacent node. If the potential field value decreases by 50 when moving to adjacent node A, and the potential field value decreases by 200 when moving to adjacent node B, the direction of adjacent node B is considered to be the direction with the fastest gradient descent, and the highest expansion priority is given to guide AGV A to move towards adjacent node B.

[0065] In an optional embodiment, the process of calculating the transition cost between spatiotemporal nodes using a resource bidding mechanism includes the following steps:

[0066] For the spatiotemporal node to be expanded, each AGV's path branch submits a bid estimate according to the following formula

[0067] wherein, is the normalized task priority, is the normalized value of path flexibility, and are preset weight coefficients;

[0068] Path flexibility is the incremental cost of the path to reroute from its previous node after losing the current spatiotemporal node relative to the cost of the current optimal path;

[0069] The path branch with the highest bid estimate obtains the right to occupy the spatiotemporal node, and the transition cost is inversely proportional to the highest bid estimate.

[0070] Specifically, when two or more AGVs' candidate paths all plan to pass the grid with coordinates (22, 35) at time 60, they need to bid for the spatiotemporal node (22, 35, 60). For example, AGV A is executing an urgent replenishment task, and its task priority is 0.9; while AGV B is executing a regular inventory task, and its is 0.3. At the same time, the location of AGV A is the entrance of a narrow passage, and if it cannot pass through the spatiotemporal node, its rerouting cost is huge, and the path flexibility​​ higher, 0.8; AGV B is in the open area, with a very small cost of detour, and a flexible path lower, 0.2.

[0071] Further assume the weights are 0.7, 0.3, then the competitive value of AGV A is 0.87. The competitive value of AGV B is 0.27. Since is greater than , AGV A wins the right to occupy the space-time node. In the path search, the cost of AGV A moving from the last space-time node to the winning space-time node will be set to a small value inversely proportional to 0.87, encouraging its path planning algorithm to eventually confirm this path. AGV B, on the other hand, will fail to bid and regard this node as a high-cost or impassable obstacle, thus seeking alternative paths. The weights , are selected based on experience or expert knowledge. For example, if it is necessary to ensure that high-priority tasks can pass through, a higher value can be set for ; if more emphasis is placed on global efficiency and avoiding AGVs getting stuck in detours, the value of can be increased.

[0072] In an optional embodiment, the process of identifying and pruning extended branches with the risk of circular waiting includes the following steps:

[0073] At each search time step, a directed graph is created with all AGVs as nodes;

[0074] If an extended branch of AGV A plans to occupy a space-time node that has already been occupied by the current optimal candidate path of AGV B, a directed edge from A to B is created in the graph, representing A's resource dependence on B;

[0075] A depth-first search algorithm is used to detect whether there is a loop in the directed graph;

[0076] If the introduction of an extended branch will lead to the formation of a loop, it is determined that the extended branch has the risk of circular waiting, and the extended branch is pruned.

[0077] A typical scenario of deadlock is that AGV A waits for AGV B to give way, while AGV B is also waiting for AGV A to give way. To avoid this situation, risk assessment is performed at each step of path planning. For example, if Figure 2As shown, at one point in time, the system has determined that AGV B's path needs to wait for AGV C to clear a certain intersection, so there is an edge from B to C in the instantaneous dependency graph. If AGV A is planning a path and considers occupying a time-space node that is currently occupied by AGV B's path, a new edge from A to B will be created in the instantaneous dependency graph. Before adding this edge, a check will be made: if there is still a dependency edge from C to A, then after adding the edge from A to B, a loop will be formed, with A pointing to B, B pointing to C, and C pointing to A. A depth-first search algorithm can detect the existence of the loop. Once a loop is detected, it is determined that this expansion step for AGV A has a very high risk of deadlock, and the expansion branch will be pruned from the time-space search, leaving AGV A to explore other paths that do not lead to a circular wait.

[0078] S2, according to the path set generated in S1, determining a main path and multiple alternative paths for the AGV, the alternative paths being used to dynamically splice new execution paths in response to real-time conflicts during execution.

[0079] In an optional embodiment, the process of determining a main path and multiple alternative paths includes the following steps:

[0080] In the path set, select a path with the minimum cumulative transition cost as the main path, the cumulative transition cost being the sum of all transition costs on a path;

[0081] In the remaining paths of the path set, filter out paths that have a node coincidence degree lower than a preset threshold with the main path and a cumulative transition cost not higher than a preset multiple of the main path cost;

[0082] Arrange the filtered paths in ascending order of cumulative transition cost, and select the top N paths as alternative paths.

[0083] In an optional embodiment, after multi-branch search, several effective paths leading to the target are found for the AGV. The path with the lowest cumulative transfer cost is selected as the main path. Assuming that the total cost of path one is 150 units, the total cost of path two is 160 units, the total cost of path three is 190 units, and the total cost of path four is 210 units, path one is selected as the main path, and the AGV will preferentially execute path one. The preset threshold of the node coincidence degree is 60%, and the preset multiple is 1.5 times. The coincidence nodes of path two and the main path one reach 70%, indicating that most of the two routes are overlapped, and if the main path is blocked, path two is likely to be unable to pass through, so path two is eliminated. The coincidence degree of path three and the main path is 40%, which is lower than the preset threshold, and the cost of path three is 190, which is lower than 1.5 times (225) of the cost of the main path, so path three becomes a qualified alternative path. The cost of path four is 210, which also meets the requirement of being lower than 1.5 times of the cost of the main path, and the coincidence degree of path four and the main path is only 20%, so path four also becomes a qualified alternative path. Assuming that 2 alternative paths are required, the qualified path three and path four are sorted according to the cumulative transfer cost to obtain an alternative path list. When the main path encounters unexpected obstacles during execution, the AGV can seamlessly switch to the alternative path three with the lowest cumulative transfer cost to continue executing the task.

[0084] When in use, when the AGV travels along the main path to the node P, if the vehicle-mounted sensor detects a temporary obstacle in front, the AGV reports a conflict. At this time, it is not necessary to re-perform global planning, but to find a path from the alternative paths that does not pass through the conflict node P and has a common node O before the P point of the current main path, and then splice the part of the alternative path from the node O to the path executed by the AGV to form a new conflict-free execution path and issue the new conflict-free execution path.

[0085] The specific embodiments of the AGV cluster control system design platform provided by the application are as follows:

[0086] The AGV cluster control system design platform comprises a path set generation module and a new path generation module.

[0087] The path set generation module is used for acquiring a preset grid map, current states of all AGVs, and task information to be executed; for a task assigned to an AGV, a reverse space-time search is performed from a space-time position of a task terminal point to generate a path set from a current position to the task terminal point for the AGV; during node expansion of the reverse space-time search: an expansion priority is determined by solving a gradient of a composite potential field, the composite potential field is obtained by superimposing a static map potential field and a dynamic coupling potential field, and the dynamic coupling potential field is calculated based on predicted trajectories of all AGVs except the current search AGV;

[0088] The transfer cost between nodes is calculated by using a resource bidding mechanism, and the resource bidding mechanism determines the use right of the node through the bidding evaluation of each AGV, wherein the bidding evaluation is determined according to the task priority and path elasticity, and the path elasticity is used to quantify the rerouting cost of the path after losing the current extended node.

[0089] A transient dependency graph between AGVs is constructed, and a causal chain analysis is performed to identify and prune the extended branch with a risk of circular waiting, wherein the nodes in the transient dependency graph are all AGVs, and the edges represent the resource dependency relationship between AGVs.

[0090] The new path generation module determines a main path and multiple alternative paths for the AGV according to the generated path set, and the alternative paths are used to dynamically splice a new execution path in response to real-time conflicts during execution.

[0091] In an optional embodiment, a reverse space-time search is performed in a three-dimensional space-time grid, wherein the space-time nodes in the three-dimensional space-time grid are composed of two-dimensional space grid coordinates (x, y) and discrete time steps t; the transfer between space-time nodes represents that the AGV moves from one grid to an adjacent grid or stays in place in a unit of time.

[0092] In an optional embodiment, when the extension priority is determined by solving the gradient of the composite potential field, a static map potential field is constructed: the potential field value of the grid where the fixed obstacle is located is set to a preset high value, and the potential field value of the passable grid is set to zero.

[0093] For AGVs other than the current search AGV, the predicted trajectory of the AGV in a future T time window is obtained, the nodes passed through by the predicted trajectory are modeled as repulsive sources, and a dynamic potential field of each AGV is constructed, and the potential field value of the dynamic potential field decays according to a Gaussian function as the space-time distance increases.

[0094] All dynamic potential fields are superimposed to form a dynamic coupling potential field, and the dynamic coupling potential field is superimposed with the static map potential field to form a composite potential field.

[0095] The composite potential field gradient of the current extended node to its neighbor nodes is calculated, and the neighbor node corresponding to the direction with the fastest gradient descent is selected as the highest priority extended node.

[0096] In an optional embodiment, when the resource bidding mechanism is used to calculate the transfer cost between space-time nodes, for the space-time node to be extended, the path branch of each AGV submits a bidding evaluation according to the following formula : ;

[0097] wherein, is the normalized task priority, is the normalized value of the path elasticity, and preset weight coefficient;

[0098] path elasticity is the increment of the cost of the path after losing the current node to be expanded relative to the cost of the current optimal path;

[0099] The path branch with the highest bidding value obtains the occupation right of the space-time node, and the transfer cost is inversely proportional to the highest bidding value.

[0100] In an optional embodiment, when identifying and pruning the expansion branch with the risk of circular waiting, a directed graph with all AGVs as nodes is created at each search time step;

[0101] If one of the expansion branches of AGV A plans to occupy a space-time node, and the space-time node has been occupied by the current optimal candidate path of AGV B, a directed edge from A to B is created in the graph, indicating the resource dependence of A to B.

[0102] A depth-first search algorithm is used to detect whether there is a loop in the directed graph;

[0103] If the introduction of a certain expansion branch will lead to the formation of a loop, it is determined that the expansion branch has the risk of circular waiting, and the expansion branch is pruned.

[0104] In an optional embodiment, when determining a main path and multiple alternative paths, in the path set, the path with the minimum cumulative transfer cost is selected as the main path, and the cumulative transfer cost is the sum of all transfer costs on a path;

[0105] In the remaining paths of the path set, paths with a node coincidence degree lower than a preset threshold and a cumulative transfer cost not higher than a preset multiple of the main path cost are screened out;

[0106] The screened paths are arranged in ascending order of cumulative transfer cost, and the top N paths are selected as alternative paths.

[0107] In addition, in the description of the present specification, the meaning of "a plurality of" is at least two, such as two, three or more, etc., unless otherwise explicitly and specifically limited.

Claims

1. An AGV cluster control method, characterized by, The method comprises the following steps: S1, obtaining a preset grid map, current states of all AGVs, and task information to be executed; for a task assigned to an AGV, performing reverse space-time search from a space-time position of a task end point to generate a path set for the AGV from a current position to the task end point; in a node expansion process of the reverse space-time search: constructing a static map potential field: setting a potential field value of a grid where a fixed obstacle is located as a preset high value, and setting a potential field value of a passable grid as zero; for AGVs other than a currently searched AGV, obtaining a predicted trajectory of each AGV within a T time window in the future, modeling nodes passed through by the predicted trajectory as repulsive force sources, and constructing a respective dynamic potential field, a potential field value of the dynamic potential field being attenuated according to a Gaussian function as a space-time distance increases; superimposing all the dynamic potential fields to form a dynamic coupling potential field, and superimposing the dynamic coupling potential field with the static map potential field to form a composite potential field; calculating a composite potential field gradient of a current expansion node to neighbor nodes of the current expansion node, and selecting a neighbor node corresponding to a direction in which the composite potential field gradient decreases most quickly as a highest priority expansion node; For the to-be-expanded spatiotemporal node, the path branch of each AGV submits a bidding evaluation according to the following formula : ; is the normalized task priority, is the normalized value of path elasticity, and are preset weight coefficients; the path elasticity is the increment of the cost of the path from the previous node of the path to the current optimal path cost after losing the current to-be-expanded node; a path branch with a highest bidding evaluation obtains an occupation right of a space-time node, and a transfer cost is a value inversely proportional to the highest bidding evaluation; a resource bidding mechanism determines a use right of a node according to bidding evaluations of the AGVs, wherein the bidding evaluations are determined according to task priorities and path flexibility, and the path flexibility is used to quantify a diversion cost of a path after losing a current expansion node; constructing an instantaneous dependence graph between the AGVs, and performing causal chain analysis to identify and prune an expansion branch with a risk of circular waiting, wherein nodes in the instantaneous dependence graph are all the AGVs, and edges represent resource dependence relationships between the AGVs; S2, determining a main path and multiple candidate paths for the AGV according to the path set generated in S1, and using the candidate paths to dynamically splice a new execution path in response to real-time conflicts in an execution process.

2. The AGV swarm control method of claim 1, wherein, The reverse space-time search is performed in a three-dimensional space-time grid, wherein a space-time node in the three-dimensional space-time grid is composed of a two-dimensional space grid coordinate (x, y) and a discrete time step t; and a transfer between the space-time nodes represents that an AGV moves from one grid to an adjacent grid or stays in the original grid in a unit time.

3. The AGV swarm control method of claim 1, wherein, The process of identifying and pruning the expansion branch with the risk of circular waiting comprises the following steps: at each search time step, a directed graph with all the AGVs as nodes is created; if an expansion branch of AGV A plans to occupy a space-time node, and the space-time node has been occupied by a current optimal candidate path of AGV B, a directed edge from A to B is created in the graph, representing that A depends on B for resources; a depth-first search algorithm is used to detect whether a loop exists in the directed graph; if the introduction of an expansion branch leads to the formation of a loop, it is determined that the expansion branch has the risk of circular waiting, and the expansion branch is pruned.

4. The AGV swarm control method according to any one of claims 1 to 3, characterized in that, In S2, the process of determining the main path and the multiple candidate paths comprises the following steps: in the path set, a path with a minimum cumulative transfer cost is selected as the main path, and the cumulative transfer cost is a sum of all transfer costs on a path; In the remaining paths of the path set, paths with a node coincidence degree lower than a preset threshold and an accumulated transfer cost not higher than a preset multiple of the main path cost are screened out; The screened paths are arranged in ascending order of accumulated transfer cost, and the top N paths are selected as candidate paths.

5. An AGV swarm control system design platform, characterized in that, It comprises: The path set generation module is used for acquiring a preset grid map, current states of all AGVs, and task information to be executed; For the task assigned to the AGV, a reverse space-time search is performed from the space-time position of the task end point to generate a path set for the AGV from the current position to the task end point; in the node expansion process of the reverse space-time search: A static map potential field is constructed: the potential field value of the grid where the fixed obstacle is located is set to a preset high value, and the potential field value of the passable grid is set to zero; For other AGVs except the currently searched AGV, the predicted trajectory of each AGV in a future T time window is acquired, the nodes passed through by the predicted trajectory are modeled as repulsive sources, and a dynamic potential field of each AGV is constructed, and the potential field value of the dynamic potential field decays according to a Gaussian function as the space-time distance increases; All dynamic potential fields are superimposed to form a dynamic coupling potential field, which is superimposed with the static map potential field to form a composite potential field; the composite potential field gradient of the current expansion node to its neighbor nodes is calculated, and the neighbor node corresponding to the direction with the fastest gradient descent is selected as the highest priority expansion node; For the to-be-expanded spatiotemporal node, the path branch of each AGV submits a bidding evaluation according to the following formula : ; is the normalized task priority, is the normalized value of path elasticity, and are preset weight coefficients; the path elasticity is the increment of the cost of the path from the previous node of the path to the current optimal path cost after losing the current to-be-expanded node. The path branch with the highest bidding evaluation obtains the occupation right of the space-time node, and the transfer cost is inversely proportional to the highest bidding evaluation; The resource bidding mechanism determines the use right of the node through the bidding evaluation of each AGV, wherein the bidding evaluation is determined according to the task priority and the path flexibility, and the path flexibility is used to quantify the diversion cost of the path after losing the current expansion node; An instantaneous dependence graph between AGVs is constructed, and a causal chain analysis is performed to identify and prune the expansion branches with a risk of circular waiting, wherein the nodes in the instantaneous dependence graph are all AGVs, and the edges represent the resource dependence relationship between AGVs; The new path generation module determines a main path and multiple candidate paths for the AGV according to the generated path set, and the candidate paths are used to dynamically splice a new execution path in response to real-time conflicts during execution.

6. The AGV swarm control system design platform of claim 5, wherein, The reverse space-time search is performed in a three-dimensional space-time grid, wherein the space-time node in the three-dimensional space-time grid is composed of a two-dimensional space grid coordinate (x, y) and a discrete time step t; the transfer between space-time nodes represents that the AGV moves from one grid to an adjacent grid or stays in place in unit time.

7. The AGV swarm control system design platform of claim 5, wherein, When the expansion priority is determined by solving the gradient of the composite potential field, a static map potential field is constructed: the potential field value of the grid where the fixed obstacle is located is set to a preset high value, and the potential field value of the passable grid is set to zero; For other AGVs except the currently searched AGV, the predicted trajectory of each AGV in a future T time window is acquired, the nodes passed through by the predicted trajectory are modeled as repulsive sources, and a dynamic potential field of each AGV is constructed, and the potential field value of the dynamic potential field decays according to a Gaussian function as the space-time distance increases; All dynamic potential fields are superimposed to form a dynamic coupling potential field, which is superimposed with the static map potential field to form a composite potential field; The compound potential field gradient of the current expansion node to its neighbor nodes is calculated, and the neighbor node corresponding to the direction of the fastest gradient descent is selected as the highest priority expansion node.

8. The AGV swarm control system design platform of claim 5 or 6 or 7, wherein, When determining a main path and multiple alternative paths, in the path set, a path with the minimum cumulative transition cost is selected as the main path, and the cumulative transition cost is the sum of all transition costs on the path; In the remaining paths of the path set, paths with a node coincidence degree lower than a preset threshold and a cumulative transition cost not higher than a preset multiple of the main path cost are screened out; The screened paths are arranged in ascending order of cumulative transition cost, and the top N paths are selected as the alternative paths.

Images