Machine learning driven multi-robot path planning method for warehouse logistics

Abstract

The application discloses a kind of warehouse logistics-oriented machine learning driven multi-robot path planning method, it is related to warehouse robot path planning field.The method includes constructing warehouse logistics environment map, generates initial path planning scheme;According to variational inference model, select current destruction heuristic from destruction heuristic set, adaptively determine neighborhood size;Then according to current destruction heuristic and neighborhood size, select part of robots from robot set to form neighborhood, delete the original path of robot in neighborhood, and re-plan the path of robot in neighborhood under the condition that the path of other robots outside neighborhood remains unchanged;If the path set re-planned can reduce the total travel distance of multi-robot, then update the original path with it, otherwise keep the original path;Update variational inference parameters, and repeat the process of destruction heuristic selection, neighborhood generation, path repair and parameter updating until the upper limit of preset running time is reached, output final path planning scheme.

Description

Technical Field

[0001] This invention relates to the field of warehouse robot path planning technology, and more specifically to a machine learning-driven multi-robot path planning method for warehouse logistics. Background Technology

[0002] Warehouse logistics systems typically employ multiple robots to perform tasks such as goods handling, material transfer, and order sorting. When multiple robots operate within the same warehouse logistics environment, it is necessary to plan a set of non-conflicting paths based on the starting and destination positions of each robot, as well as information about the warehouse logistics environment. With the increasing scale of warehouses and the growing number of robot clusters, rapidly generating high-quality path planning solutions within a limited computation time has become a critical problem that urgently needs to be solved in this field.

[0003] To address this problem, researchers proposed an arbitrary-time multi-robot path planning paradigm. This approach no longer aims for a single, one-time output of the final path plan; instead, it quickly returns an initial feasible solution after computation begins and iteratively optimizes the solution's quality over the remaining available time, thus balancing the real-time performance of the system response with the solution's optimization potential. The Anytime method based on large neighborhood search typically selects a subset of robots from the current path set as the robots to be adjusted, deletes their original paths, and replans the paths of the robots to be adjusted while keeping the paths of other robots unchanged. Through multiple rounds of destruction and repair operations, the total travel distance of the multiple robots can be gradually reduced within time constraints, allowing the path planning scheme to progressively improve from a feasible result to a better one.

[0004] However, the existing methods still have the following shortcomings in practical applications: First, existing methods use fixed neighborhood sizes or a limited number of preset neighborhood sizes, making it difficult to flexibly adjust the number of robots to be optimized according to the congestion level, conflict distribution, and search phase of the current path set; second, the adaptability of the neighborhood selection strategy is limited, and when the warehousing and logistics environment changes, robot density increases, or path conflict distribution changes rapidly, the strategy update is prone to lag, leading to getting trapped in local optima; the above problems will cause computing resources to be allocated to path areas with low optimization potential, resulting in a slow improvement in the quality of path planning schemes and a long total travel distance for multiple robots, making it difficult to meet the requirements of large-scale warehousing and logistics systems for efficient collaborative path planning. Summary of the Invention

[0005] In view of at least one of the above technical problems, this application provides a machine learning-driven multi-robot path planning method for warehousing and logistics. It aims to achieve dynamic focusing of computing resources on high-potential optimization areas by using variational inference to perceive and search the state online and adaptively selecting destructive heuristics and neighborhood size. This will continuously and stably reduce the total travel distance of multiple robots under a limited time budget, and realize efficient collaborative path planning for large-scale warehousing and logistics systems.

[0006] The technical solution of this invention is:

[0007] A machine learning-driven multi-robot path planning method for warehousing and logistics, comprising the following steps:

[0008] Step 1: Initialize relevant parameters, including warehousing and logistics environment parameters, robot parameters, variational inference parameters, and upper limit of path planning running time;

[0009] Step 2: Construct a warehouse logistics environment map based on warehouse logistics environment parameters, and determine the starting position and target position of each robot based on robot parameters;

[0010] Step 3: Based on the warehouse logistics environment map, the starting position and target position of each robot, use the priority planning method to plan the initial path of each robot and obtain the initial set of original paths;

[0011] Step 4: Calculate the probability of choosing the destructive heuristic based on the current variational inference parameters, and select the current destructive heuristic from the set of destructive heuristics;

[0012] Step 5: Calculate the neighborhood size conditional distribution based on the current destructive heuristic and the current variational inference parameters, and select the current neighborhood size from the neighborhood size conditional distribution;

[0013] Step 6: Based on the current destruction heuristic and the current neighborhood size, select the robots to be replanned from the robot set to form a neighborhood, and delete the paths of the robots in the neighborhood from the current original path set to obtain the destroyed path set;

[0014] Step 7: Construct a spatiotemporal occupancy table and a safety interval table based on the damaged path set, and replan the robot's path in the neighborhood using a path planning method based on safety intervals to obtain the repaired path set;

[0015] Step 8: Calculate the path cost difference between the robot's current original path set and the repaired path set within the neighborhood;

[0016] Step 9: Update the variational inference parameters based on the path cost difference; determine whether to update the current original path set based on the path cost difference: if the path cost difference is greater than zero, update the current original path set using the repaired path set; if the path cost difference is less than or equal to zero, retain the current original path set.

[0017] Step 10: Determine whether the current running time has reached the upper limit of the path planning running time. If the current running time has not reached the upper limit of the path planning running time, return to step 4; if the current running time has reached the upper limit of the path planning running time, output the current original path set and obtain the multi-robot path planning result for warehousing and logistics.

[0018] Optionally, according to the machine learning-driven multi-robot path planning method, the warehousing and logistics environment parameters include the length of the warehousing and logistics environment. Warehousing and logistics environment width The robot parameters are: a set of passable locations, a set of obstacle locations, and the connection relationships between adjacent passable locations; the robot parameters are expressed as follows: ,in For the number of robots, For the first A robot, The information for each robot is , For the first The starting position of each robot For the first The target position of the robot; the variational inference parameters include the destructive heuristic learning rate. Neighborhood size and learning rate Minimum neighborhood size and maximum neighborhood size .

[0019] Optionally, according to the machine learning-driven multi-robot path planning method, step 2 includes the following steps:

[0020] Determine the set of passable locations and the set of obstacle locations based on warehousing and logistics environment parameters;

[0021] Determine the vertex set of the warehousing and logistics environment map based on the set of accessible locations. ;

[0022] The edge set of the warehousing and logistics environment map is determined based on the connectivity between adjacent accessible locations. ;

[0023] The starting and target vertices of each robot in the warehouse logistics environment map are determined based on the robot parameters.

[0024] Optionally, according to the machine learning-driven multi-robot path planning method described above, the edge set Determined according to the following formula:

[0025]

[0026] In the above formula, For position and location The grid distance between them; when When, it means that two passable positions are adjacent.

[0027] Optionally, according to the machine learning-driven multi-robot path planning method, step 4 includes the following steps:

[0028] Step 4.1: Construct a destructive heuristic set ,in This is a robot-based destruction heuristic used to select robots with large path delays and related robots; A map-based destruction heuristic is used to select robots that pass through intersecting or crowded areas. This is a random destruction heuristic used to randomly select robots;

[0029] Step 4.2: Based on the destructive heuristic variational distribution Calculate the selection probability of each disruptive heuristic in the disruptive heuristic set. ;

[0030] The variational distribution that breaks the heuristic For categorical distribution, it is represented as ,in To disrupt the heuristic variational inference parameters; To disrupt the heuristic choice probability;

[0031] Destroying Heuristics The probability of selection is calculated according to the following formula:

[0032] , ;

[0033] Step 4.3: Obtain the current destruction heuristic by sampling the selection probability of each destruction heuristic.

[0034] Optionally, according to the machine learning-driven multi-robot path planning method, step 5 includes the following steps:

[0035] Step 5.1: Determine the neighborhood size conditional distribution based on the current destructive heuristic; the neighborhood size conditional distribution is a truncated Gaussian distribution:

[0036]

[0037] In the above formula, To choose the destructive heuristic Back Neighborhood Size Conditional distribution; The parameters are used for variational inference of neighborhood size; To break heuristics The corresponding mean neighborhood size; To break heuristics The corresponding standard deviation of the neighborhood size; For the minimum neighborhood size, The maximum neighborhood size; Indicates the interval Upper-truncated Gaussian distribution;

[0038] Step 5.2: Distribute based on neighborhood size conditions The size of the continuous neighborhood is obtained by sampling according to the following formula;

[0039]

[0040] In the above formula, For the first The size of the continuous neighborhood obtained by round-by-round sampling; For the first Round iteration breaks heuristic The corresponding mean neighborhood size; For the first Round iteration breaks heuristic The corresponding standard deviation of the neighborhood size;

[0041] Step 5.3: Integerize the size of the contiguous neighborhood;

[0042] Step 5.4: Apply boundary constraints to the integer neighborhood size to obtain the current neighborhood size.

[0043] Optionally, according to the machine learning-driven multi-robot path planning method, step 6, which involves selecting robots to be replanned from the robot set to form a neighborhood based on the current destruction heuristic and the current neighborhood size, includes:

[0044] If the current destruction heuristic is a robot-based destruction heuristic, then a neighborhood is generated based on path delay; the robot with the largest path delay is selected, and other robots that prevent its path from being shortened are found through random walks and added to the neighborhood set. until the number of robots in the neighborhood is equal to the size of the neighborhood. ;

[0045] If the current destruction heuristic is a map-region-based destruction heuristic, then a neighborhood is generated based on the traversal of the intersection regions; robots that have traversed vertices with a degree greater than 2 in the undirected graph are selected, and then expanded to robots that have traversed nearby intersection vertices using breadth-first search, and these robots are added to the neighborhood set. until the number of robots in the neighborhood is equal to the size of the neighborhood. ;

[0046] If the current destruction heuristic is a random destruction heuristic, then randomly select from all robot sets. One robot joins the neighborhood set. .

[0047] Optionally, according to the machine learning-driven multi-robot path planning method, step 9, updating the variational inference parameters based on the path cost difference, includes the following steps:

[0048] Step 9.1: Calculate the reward value based on the path cost difference. and the reward value Perform normalization to obtain the normalization reward value. ;

[0049] Step 9.2: Update the destructive heuristic variational inference parameters according to the normalized reward value using the following formula;

[0050]

[0051] In the above formula, For the updated destructive heuristic variational inference parameters; To disrupt the heuristic learning rate; Indicates the current number Is the round iteration heuristic broken? , This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. For the first Round iteration breaks heuristic The corresponding variational inference parameters; No. Round iteration breaks heuristic The probability of choosing;

[0052] Step 9.3: Update the neighborhood size variational inference parameters based on the normalized reward value.

[0053] Optionally, according to the machine learning-driven multi-robot path planning method described above, in step 9.1, the reward value is calculated according to the following formula. Perform normalization to obtain the normalization reward value. :

[0054]

[0055] In the above formula, To prevent positive numbers with a denominator of zero; For the first Reward value in each iteration; For the first The original path cost of the robot in the neighborhood during the round of iteration.

[0056] Optionally, according to the machine learning-driven multi-robot path planning method described above, step 9.3, updating the neighborhood size variational inference parameters based on the reward value, includes:

[0057] If the current destruction heuristic is Then the mean neighborhood size is updated according to the following formula:

[0058]

[0059] In the above formula, The updated neighborhood size mean parameter; The learning rate is the neighborhood size; For the first The current neighborhood size in the round of iteration; For the first Round iteration breaks heuristic The corresponding mean neighborhood size; For the first Round iteration breaks heuristic The corresponding standard deviation of the neighborhood size;

[0060] If the current destruction heuristic is The standard deviation of the neighborhood size is then updated according to the following formula:

[0061]

[0062] To ensure that the standard deviation of the neighborhood size is positive, the following constraint is applied:

[0063]

[0064] In the above formula, The standard deviation of the updated neighborhood size. This is the lower bound of the standard deviation.

[0065] The beneficial effects of adopting the above technical solution are as follows:

[0066] (1) It overcomes the problem of inflexible neighborhood size selection and improves the matching degree between the path optimization process and the current search state. This invention uses a variational inference model to learn and reason in real time about the search process, and adaptively selects the neighborhood size. This allows the number of robots involved in each round of destruction and repair operations to be dynamically adjusted according to the search feedback. It can also flexibly adjust the number of robots to be optimized according to the optimization requirements of the current solution. Small neighborhoods are generated when local fine-tuning is needed, and large neighborhoods are generated when structural bottlenecks need to be broken, thereby greatly improving the optimization efficiency of a single destruction-repair operation.

[0067] (2) It overcomes the problem of insufficient targeting of destructive heuristic selection and improves the optimization probability of effective path regions. This invention selects destructive heuristics through variational inference and updates the variational inference parameters according to the improvement of each round of path planning schemes, making it easier for subsequent searches to select destructive heuristics with high improvement potential for the current path set, thereby reducing the waste of computational resources caused by blind or inefficient selection and improving the efficiency of improving the quality of path planning schemes.

[0068] (3) It alleviates the problem of strategy update lag and improves the adaptability to changes in warehousing and logistics tasks. After each round of destruction and repair, the present invention updates the variational inference parameters based on the change in the total cost of the path set, so that the selection strategy can be adjusted with the changes in the current path conflict distribution, robot density and search stage, thereby avoiding the strategy lag problem caused by relying solely on historical training data or statistical feedback, and improving the adaptability to different warehousing and logistics environments and different robot scales.

[0069] (4) The total travel distance of multiple robots is reduced, and the operating efficiency of the warehousing and logistics system is improved. After generating the initial path planning scheme, the present invention continuously optimizes the path set through multiple rounds of destruction, repair and parameter updates, and updates the current path planning scheme when the cost of the repaired path set is lower, so that the better path results can be continuously retained, thereby gradually reducing the total travel distance of multiple robots, reducing detours and waiting, and improving the operating efficiency of cargo handling, material transfer and order sorting operations in the warehousing and logistics system.

[0070] (5) It takes into account both rapid response and continuous optimization capabilities, meeting the path planning requirements of large-scale warehousing and logistics scenarios. The present invention first uses a rapid construction method to generate an executable initial path planning scheme, and then continuously performs local path replanning within a given time budget, so that the system can obtain usable results first, and further improve the quality of the path planning scheme when the computation time allows, thus making it suitable for large-scale warehousing and logistics scenarios with a large number of robots, dense path conflicts, and high requirements for response time. Attached Figure Description

[0071] Figure 1This is a flowchart of a machine learning-driven multi-robot path planning method for warehousing and logistics in this embodiment.

[0072] Figure 2 This is a schematic diagram of a warehouse environment with 100 robots, a width of 30, and a length of 50.

[0073] Figure 3 This is a flowchart illustrating the process of quickly planning the paths of all robots to obtain the initial set of original paths in this embodiment.

[0074] Figure 4 This is a schematic diagram of the path repair process based on safety intervals in this embodiment. Detailed Implementation

[0075] To facilitate understanding of this application, a more comprehensive description of this application will be provided below with reference to the accompanying drawings.

[0076] For multi-robot collaborative path planning in warehousing and logistics, the core idea of ​​this application is as follows: First, a warehousing and logistics environment map is constructed based on the warehousing and logistics environment information, robot starting position information, and robot target position information, and an initial path planning scheme is quickly generated using a priority planning method. Then, a variational inference model is used to select the current destructive heuristic from the set of destructive heuristics, and the neighborhood size is adaptively determined based on the selected destructive heuristic. Next, based on the current destructive heuristic and the neighborhood size, a portion of the robots in the robot set are selected to form a neighborhood, the original paths of the robots in the neighborhood are deleted, and the paths of the robots in the neighborhood are replanned while keeping the paths of other robots outside the neighborhood unchanged. If the replanned path set can reduce the total travel distance of the multi-robots, the current path planning scheme is updated with the replanned path set; otherwise, the current path planning scheme is retained. Finally, the variational inference parameters are updated based on the improvement of the path planning scheme in this round, and the process of destructive heuristic selection, neighborhood generation, path repair, and parameter update is repeated until the preset running time limit is reached, and the final path planning scheme is output.

[0077] Figure 1 This is a flowchart of a machine learning-driven multi-robot path planning method for warehousing and logistics, as described in this embodiment. To enable those skilled in the art to better understand the technical solutions in this specification, a warehouse environment map with 100 robots, a width of 30, and a length of 50 is used as an example to further illustrate the method of this invention. The task scenario in this embodiment is as follows: Figure 2 As shown. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0078] like Figure 1As shown, this embodiment of the machine learning-driven multi-robot path planning method for warehousing and logistics includes the following steps:

[0079] Step 1: Initialize relevant parameters, including warehousing and logistics environment parameters, robot parameters, path planning runtime limit, and variational inference parameters;

[0080] Step 2: Construct a warehousing and logistics environment map based on warehousing and logistics environment parameters. The starting and target positions of each robot are determined based on the robot parameters.

[0081] Step 3: Based on the warehouse logistics environment map, the starting position and target position of each robot, use the priority planning method to quickly plan the path of all robots and obtain the initial set of original paths;

[0082] Step 4: Calculate the probability of choosing the destructive heuristic based on the current variational inference parameters, and select the current destructive heuristic from the set of destructive heuristics;

[0083] Step 5: Calculate the neighborhood size conditional distribution based on the current destructive heuristic and variational inference parameters, and select the current neighborhood size from the neighborhood size conditional distribution;

[0084] Step 6: Based on the current destruction heuristic and the current neighborhood size, select the robots to be replanned from the set of all robots to form a neighborhood, and delete the paths of the robots in the neighborhood from the current original path set to obtain the set of destroyed paths;

[0085] Step 7: Construct a spatiotemporal occupancy table and a safety interval table based on the damaged path set, and replan the robot's path in the neighborhood using a path planning method based on safety intervals to obtain the repaired path set;

[0086] Step 8: Calculate the path cost difference between the robot's current original path set and the repaired path set within the neighborhood;

[0087] Step 9: Update the variational inference parameters based on the path cost difference; determine whether to update the current original path set based on the path cost difference: if the path cost difference is greater than zero, update the current original path set using the repaired path set; if the path cost difference is less than or equal to zero, retain the current original path set.

[0088] Step 10: Determine whether the current running time has reached the upper limit of the path planning running time. If the current running time has not reached the upper limit of the path planning running time, return to step 4 to continue optimizing the path planning scheme; if the current running time has reached the upper limit of the path planning running time, output the current original path set to obtain the multi-robot path planning result for warehousing and logistics.

[0089] In this embodiment, step 1 includes the following steps:

[0090] Step 1.1: Initialize the warehousing and logistics environment parameters, including the length of the warehousing and logistics environment. Warehousing and logistics environment width The set of passable locations, the set of obstacle locations, and the connection relationships between adjacent passable locations;

[0091] Step 1.2: Initialize robot parameters ,in For the number of robots, For the first There are [number] robots, and the information for each robot is: , , For the first The starting position of each robot For the first The target location of the robot; Figure 2 The black circle with a serial number indicates the starting position of the robot with that serial number, and the white dashed circle with a serial number indicates the target position of the robot with that serial number.

[0092] Step 1.3: Set the upper limit for path planning runtime ;

[0093] Step 1.4: Initialize variational inference parameters, including violating the heuristic learning rate. Neighborhood size and learning rate Minimum neighborhood size and maximum neighborhood size .

[0094] In this embodiment, the initialization-related parameters are shown in Table 1.

[0095] Table 1 Examples of Main Parameters

[0096]

[0097] In this embodiment, step 2 includes the following steps:

[0098] Step 2.1: Determine the set of passable locations and the set of obstacle locations based on the warehousing and logistics environment parameters;

[0099] Step 2.2: Determine the set of vertices in the warehousing and logistics environment map based on the set of accessible locations. ;

[0100] Step 2.3: Determine the edge set in the warehousing and logistics environment map based on the connectivity between adjacent accessible locations. ;

[0101] Step 2.4: Determine the starting vertex and target vertex of each robot in the warehouse logistics environment map based on the robot parameters.

[0102] Optionally, the edge set in the warehousing and logistics environment map Determine as follows:

[0103]

[0104] In the above formula, For position and location The grid distance between them; when When, it means that two passable positions are adjacent.

[0105] In this embodiment, reference Figure 3 Step 3 includes the following steps:

[0106] Step 3.1: Input the warehousing and logistics environment map and robot parameters ;

[0107] Step 3.2: Generate a length equal to the number of robots. random priority sequence Assign priorities to each robot according to a priority sequence;

[0108] Furthermore, the robot priority sequence is represented as follows:

[0109]

[0110] In the above formula, For robot priority sequence; The robot with the highest priority; The robot with the lowest priority.

[0111] Step 3.3: Initialize the spacetime occupancy table (Reservation Table) to record the vertices and edges occupied by the planned robot at each time step; further, the spacetime occupancy table includes a vertex occupancy table. Edge Occupancy Table Specifically, it is expressed by the following formula:

[0112]

[0113] In the above formula, For space-time occupancy table, Represents the vertex occupancy table. Represents the edge occupancy table. Indicates position At time step Is it occupied? Representing an edge At time step Is it occupied?

[0114] If position At time step If it is occupied, then:

[0115]

[0116] If the edge At time step If it is occupied, then:

[0117]

[0118] Step 3.4: Select the current robot in sequence according to the robot priority sequence, and plan the path table of the current robot based on the starting position, target position and spatiotemporal occupancy table of the current robot;

[0119] No. The path of a robot It is expressed as follows:

[0120]

[0121] In the above formula, For the first The path of the robot ; For the first A robot at time step Location; For the first The number of steps required for a robot to reach its target location.

[0122] Step 3.5: Write the current conflict-free path of the robot into the spacetime occupancy table; if the robot Obtain Path Then the path will be written into the space-time occupancy table:

[0123]

[0124] in, , This indicates the number of time steps required for the robot to reach the target location;

[0125] Step 3.6: Determine whether all robots have obtained conflict-free paths. If so, output the initial set of original paths. ,in, For the first The initial path of the robot. If not, return to step 3.2 to regenerate the robot priority sequence.

[0126] In this embodiment, step 4 includes the following steps:

[0127] Step 4.1: Construct a destructive heuristic set;

[0128] In this embodiment, the set of destructive heuristics is: ,in This is a robot-based destruction heuristic used to select robots with large path delays and related robots. A map-based destruction heuristic is used to select robots that traverse intersecting or crowded areas. The random destruction heuristic is used to randomly select the robot; the set of destruction heuristics used in this embodiment is shown in Table 2.

[0129] Table 2 Destructive Heuristic Sets

[0130]

[0131] Step 4.2: Calculate the selection probability of each destructive heuristic based on the variational inference parameters of the destructive heuristic;

[0132] In this embodiment, the variational distribution of the selected destructive heuristic is a classification distribution:

[0133]

[0134] In the above formula, To disrupt the heuristic variational distribution; Indicated by The classification distribution of the parameters; To disrupt the heuristic variational inference parameters; To disrupt the probability of heuristic selection;

[0135] Furthermore, the first The probability of a broken heuristic choice is calculated according to the following formula:

[0136] , ;

[0137] In the above formula, To break heuristics The probability of choosing; To break heuristics The corresponding variational inference parameters;

[0138] Step 4.3: Obtain the current destruction heuristic by sampling the selection probability of each destruction heuristic.

[0139] Furthermore, the current destructive heuristic is sampled according to the following formula:

[0140]

[0141] In the above formula, For the first The current breaking heuristic of round iteration selection; For the first The round iteration violates the heuristic variational inference parameters.

[0142] In this embodiment, step 5 includes the following steps:

[0143] Step 5.1: Determine the neighborhood size conditional distribution based on the current destructive heuristic;

[0144] In this embodiment, the neighborhood size conditional distribution is a truncated Gaussian distribution:

[0145]

[0146] In the above formula, To choose the destructive heuristic Back Neighborhood Size Conditional distribution; The parameters are used for variational inference of neighborhood size; To break heuristics The corresponding mean neighborhood size; To break heuristics The corresponding standard deviation of the neighborhood size; For the minimum neighborhood size, The maximum neighborhood size; Indicates the interval Upper-truncated Gaussian distribution;

[0147] Step 5.2: Distribute based on neighborhood size conditions The sampling yields the size of the continuous neighborhood;

[0148] In this embodiment, the size of the contiguous neighborhood is sampled according to the following formula:

[0149]

[0150] In the above formula, For the first The size of the continuous neighborhood obtained by round-by-round sampling;

[0151] Step 5.3: Integerize the size of the contiguous neighborhood;

[0152] Furthermore, integerization is performed according to the following formula:

[0153]

[0154] In the above formula, For the first The current neighborhood size in each iteration. Indicates rounding down;

[0155] Step 5.4: Apply boundary constraints to the integerized neighborhood size to obtain the current neighborhood size.

[0156] Furthermore, the current neighborhood size is determined according to the following formula:

[0157]

[0158] In this embodiment, step 6 includes the following steps:

[0159] Step 6.1: Determine the neighborhood generation method based on the current destruction heuristic;

[0160] Step 6.2: If the current destruction heuristic is a robot-based destruction heuristic, then generate a neighborhood based on path delay; select the robot with the largest path delay, and use a random walk to find other robots that prevent its path from shortening and add them to the neighborhood set. until the number of robots in the neighborhood is equal to the size of the neighborhood. ;

[0161] Step 6.3: If the current destruction heuristic is a map region-based destruction heuristic, then generate a neighborhood based on the crossing regions; select robots that have passed through vertices with a degree greater than 2 in the undirected graph, and expand this to robots that have passed through nearby crossing vertices using breadth-first search, adding these robots to the neighborhood set. until the number of robots in the neighborhood is equal to the size of the neighborhood. .

[0162] Step 6.4: If the current destruction heuristic is a random destruction heuristic, then randomly select from the set of all robots. One robot joins the neighborhood set. ;

[0163] Step 6.5: Delete the robot's path within the neighborhood from the current original path set to obtain the corrupted path set.

[0164] Optionally, the set of destroyed paths We obtain it according to the following formula:

[0165]

[0166] Reference Figure 4 The present invention provides a further detailed description of the solution for repairing damage, wherein step 7 includes the following steps:

[0167] Step 7.1: Establish a repair time-space occupancy table based on the set of damaged paths. For any path in the set of paths after destruction Update and repair the spacetime occupancy table according to the following formula:

[0168]

[0169] Step 7.2: Construct a safety interval table based on the repair time-space occupancy table. :

[0170]

[0171] In the above formula, As vertices The set of safe intervals; For the first One safety interval; As vertices The corresponding number of safety intervals;

[0172] Step 7.3: Generate a repair priority sequence for robots within the neighborhood:

[0173]

[0174] In the above formula, For the first Repair priority sequence in round iteration This indicates the robot with the highest repair priority. For the first Neighborhood size in round iteration;

[0175] Step 7.4: Select the robots to be repaired in sequence according to the repair priority order;

[0176] Step 7.5: Search for a conflict-free path for the robot to be repaired based on the safety interval table;

[0177] Step 7.6: Write the conflict-free path of the robot to be repaired into the repair time-space occupancy table; if the robot Obtain the repair path Then the repair path will be written into the space-time occupancy table:

[0178]

[0179] Step 7.7: Output the set of repair paths. If the neighborhood... If all robots obtain conflict-free paths, then output the set of repair paths. If a robot fails to obtain a conflict-free path, a new repair priority sequence is generated, and the process returns to step 7.3.

[0180] In this embodiment, step 8 includes the following steps:

[0181] Step 8.1: Calculate the current original path cost of the robot within the neighborhood; the current original path cost of the robot within the neighborhood is calculated according to the following formula:

[0182]

[0183] In the above formula, For the first The current original path cost of the robot within the neighborhood during each iteration; For the first The original path length of the robot before repair;

[0184] Step 8.2: Calculate the repair path cost for the robot within the neighborhood;

[0185]

[0186] In the above formula, For the first The cost of the repair path for the robot within the neighborhood during each iteration; For the first The path length after repair by the robot;

[0187] Step 8.3: Calculate the path cost difference based on the original path cost and the repaired path cost. ;

[0188] In this embodiment, step 9, which involves updating the variational inference parameters based on the path cost difference, includes the following steps:

[0189] Step 9.1: Calculate the reward value based on the path cost difference. ;

[0190] Furthermore, the normalized reward value is calculated according to the following formula:

[0191]

[0192] In the above formula, To normalize the reward value, To prevent positive numbers with a denominator of zero;

[0193] Step 9.2: Update the destructive heuristic variational inference parameters based on the normalized reward value; the destructive heuristic variational inference parameters are updated according to the following formula:

[0194]

[0195] In the above formula, For the updated destructive heuristic variational inference parameters, To disrupt the heuristic learning rate, Indicate whether the current breaking heuristic is , This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise.

[0196] Step 9.3: Update the neighborhood size variational inference parameters based on the normalized reward value; if the current destructive heuristic is... Then the mean neighborhood size is updated according to the following formula:

[0197]

[0198] In the above formula, The updated neighborhood size mean parameter, The learning rate is the neighborhood size;

[0199] Furthermore, if the current destruction heuristic is The standard deviation of the neighborhood size is then updated according to the following formula:

[0200]

[0201] To ensure that the standard deviation of the neighborhood size is positive, the following constraint is applied:

[0202]

[0203] In the above formula, The standard deviation of the updated neighborhood size. This is the lower bound of the standard deviation;

[0204] Step 9 describes determining whether to update the current original path set based on the path cost difference: if the path cost difference is greater than zero... Then the repaired path set is used to update the current original path set. If the path cost difference is less than or equal to zero, then... Then the current original set of paths is retained. ;

[0205] In this embodiment, step 10 includes the following steps:

[0206] Step 10.1: Obtain the current running time Determine the current running time Has the maximum runtime for path planning been reached? .

[0207] Step 10.2: If the current running time The maximum runtime for path planning has not been reached. If so, return to step 4 to continue optimizing the path planning scheme;

[0208] Step 10.3: If the current running time Reaching the maximum running time for path planning If the iteration stops, output the latest set of original paths. , of which The final path of the robot is .

[0209] The effects of this invention will be further described below with reference to simulation experiments. The hardware platform for the simulation experiments of this invention is: an Intel Core i7-14700 processor with a main frequency of 5.4GHz and 32GB of RAM; the software platform is the Ubuntu 22.04LTS operating system, implemented in C++17.

[0210] This embodiment verifies the efficiency of the path planning method of the present invention through simulation experiments. Specifically, simulation experiments are conducted on the present invention method and a state-of-the-art existing path planning method based on an adaptive large neighborhood search algorithm. The efficiency of the path planning method of the present invention is verified by comparing the experimental results. The state-of-the-art existing path planning method based on an adaptive large neighborhood search algorithm is the Anytime MAPF method proposed by Li J et al. in "MAPF-LNS2: Fast repairing for multi-agent path finding via large neighborhood search[C] / / Proceedings of the AAAI Conference on Artificial Intelligence. 2022, 36(9): 10256-10265", abbreviated as MAPF-LNS2.

[0211] The simulation experiment scenario is set as follows Figure 2 The warehouse environment shown is 30m wide and 50m long, with each grid cell having a side length of 1m. Gray grid cells represent impassable areas, numbered black circles indicate the starting positions of each robot, and white dashed circles with the same number correspond to the target positions.

[0212] To objectively compare the driving efficiency of the two methods, the experiment recorded the total path length traveled by all robots from the starting position to the target position. In the experiment, the number of robots increased by 100 in increments from 100 to 500, and each experiment was repeated five times. The maximum running time for each experiment was set to 60 seconds. Finally, the average of the five experimental results was taken for statistical analysis, and the improvement rate of path length was calculated. The specific results are shown in Table 3.

[0213] Table 3. Quantitative Analysis of the Path Planning Results of the Invention and Existing Technologies in the Simulation Experiment

[0214]

[0215] As shown in Table 3, the statistical results demonstrate that, compared to the existing MAPF-LNS2 method, the method of this invention exhibits superior performance across different robot numbers. Specifically, the total travel distance generated using the method of this invention is shorter. As the number of robots increases from 100 to 500, the optimization effect of the method of this invention on the total travel distance remains stable, and the path shortening ratio significantly increases. This indicates that the method of this invention can effectively reduce the total travel distance during path planning; under the same time budget, the method of this invention can generate higher-quality path planning schemes, thereby significantly improving the overall efficiency of path planning.

[0216] It should be understood that, inspired by the technical concept of this invention, those skilled in the art can make various improvements or modifications based on the above content without departing from the scope of this invention, and these modifications still fall within the protection scope of this invention.

Claims

1. A machine learning-driven multi-robot path planning method for warehousing and logistics, characterized in that, The method includes the following steps: Step 1: Initialize relevant parameters, including warehousing and logistics environment parameters, robot parameters, variational inference parameters, and upper limit of path planning running time; Step 2: Construct a warehouse logistics environment map based on warehouse logistics environment parameters, and determine the starting position and target position of each robot based on robot parameters; Step 3: Based on the warehouse logistics environment map, the starting position and target position of each robot, use the priority planning method to plan the initial path of each robot and obtain the initial set of original paths; Step 4: Calculate the probability of choosing the destructive heuristic based on the current variational inference parameters, and select the current destructive heuristic from the set of destructive heuristics; Step 5: Calculate the neighborhood size conditional distribution based on the current destructive heuristic and the current variational inference parameters, and select the current neighborhood size from the neighborhood size conditional distribution; Step 6: Based on the current destruction heuristic and the current neighborhood size, select the robots to be replanned from the robot set to form a neighborhood, and delete the paths of the robots in the neighborhood from the current original path set to obtain the destroyed path set; Step 7: Construct a spatiotemporal occupancy table and a safety interval table based on the damaged path set, and replan the robot's path in the neighborhood using a path planning method based on safety intervals to obtain the repaired path set; Step 8: Calculate the path cost difference between the robot's current original path set and the repaired path set within the neighborhood; Step 9: Update the variational inference parameters based on the path cost difference; determine whether to update the current original path set based on the path cost difference: if the path cost difference is greater than zero, update the current original path set using the repaired path set; if the path cost difference is less than or equal to zero, retain the current original path set. Step 10: Determine whether the current running time has reached the upper limit of the path planning running time. If the current running time has not reached the upper limit of the path planning running time, return to step 4. If the current running time reaches the upper limit of the path planning running time, the current original path set is output to obtain the multi-robot path planning result for warehousing and logistics. Step 4 includes the following steps: Step 4.1: Construct a destructive heuristic set ,in This is a robot-based destruction heuristic used to select robots with large path delays and related robots; A map-based destruction heuristic is used to select robots that pass through intersecting or crowded areas. This is a random destruction heuristic used to randomly select robots; Step 4.2: Based on the destructive heuristic variational distribution Calculate the selection probability of each disruptive heuristic in the disruptive heuristic set. ; The variational distribution that breaks the heuristic For categorical distribution, it is represented as ,in To disrupt the heuristic variational inference parameters; To disrupt the heuristic choice probability; Destroying Heuristics The probability of selection is calculated according to the following formula: , ； Step 4.3: Obtain the current destruction heuristic by sampling according to the selection probability of each destruction heuristic; Step 5 includes the following steps: Step 5.1: Determine the neighborhood size conditional distribution based on the current destructive heuristic; the neighborhood size conditional distribution is a truncated Gaussian distribution: ； In the above formula, To choose the destructive heuristic Back Neighborhood Size Conditional distribution; The parameters are used for variational inference of neighborhood size; To break heuristics The corresponding mean neighborhood size; To break heuristics The corresponding standard deviation of the neighborhood size; For the smallest neighborhood size, The maximum neighborhood size; Indicates the interval Upper-truncated Gaussian distribution; Step 5.2: Distribute based on neighborhood size conditions The size of the continuous neighborhood is obtained by sampling according to the following formula; ； In the above formula, For the first The size of the continuous neighborhood obtained by round-by-round sampling; For the first Round iteration breaks heuristic The corresponding mean neighborhood size; For the first Round iteration breaks heuristic The corresponding standard deviation of the neighborhood size; Step 5.3: Integerize the size of the contiguous neighborhood; Step 5.4: Apply boundary constraints to the integerized neighborhood size to obtain the current neighborhood size; Step 6, which involves selecting robots to be replanned from the robot set to form a neighborhood based on the current destruction heuristic and the current neighborhood size, includes: If the current destruction heuristic is a robot-based destruction heuristic, then generate the neighborhood based on path delay; Select the robot with the largest path delay, and use a random walk to find other robots that are preventing its path from being shortened, then add them to the neighborhood set. until the number of robots in the neighborhood is equal to the size of the neighborhood. ; If the current destruction heuristic is a map-region-based destruction heuristic, then a neighborhood is generated based on the traversal of the intersection regions; robots that have traversed vertices with a degree greater than 2 in the undirected graph are selected, and then expanded to robots that have traversed nearby intersection vertices using breadth-first search, and these robots are added to the neighborhood set. until the number of robots in the neighborhood is equal to the size of the neighborhood. ; If the current destruction heuristic is a random destruction heuristic, then randomly select from all robot sets. One robot joins the neighborhood set. .

2. The machine learning-driven multi-robot path planning method according to claim 1, characterized in that, The warehousing and logistics environment parameters include the length of the warehousing and logistics environment. Warehousing and logistics environment width The robot parameters are: a set of passable locations, a set of obstacle locations, and the connection relationships between adjacent passable locations; the robot parameters are expressed as follows: ,in For the number of robots, For the first A robot, The information for each robot is , For the first The starting position of each robot For the first The target position of the robot; the variational inference parameters include the destructive heuristic learning rate. Neighborhood size and learning rate Minimum neighborhood size and maximum neighborhood size .

3. The machine learning-driven multi-robot path planning method according to claim 1, characterized in that, Step 2 includes the following steps: Determine the set of passable locations and the set of obstacle locations based on warehousing and logistics environment parameters; Determine the vertex set of the warehousing and logistics environment map based on the set of accessible locations. ; The edge set of the warehousing and logistics environment map is determined based on the connectivity between adjacent accessible locations. ; The starting and target vertices of each robot in the warehouse logistics environment map are determined based on the robot parameters.

4. The machine learning-driven multi-robot path planning method according to claim 3, characterized in that, The set of edges Determined according to the following formula: ； In the above formula, For position and location The grid distance between them; when When, it means that two passable positions are adjacent.

5. The machine learning-driven multi-robot path planning method according to claim 1, characterized in that, Step 9, which involves updating the variational inference parameters based on the path cost difference, includes the following steps: Step 9.1: Calculate the reward value based on the path cost difference. and the reward value Perform normalization to obtain the normalization reward value. ;in This represents the difference in path costs. Step 9.2: Update the destructive heuristic variational inference parameters according to the normalized reward value using the following formula; ； In the above formula, For the updated destructive heuristic variational inference parameters; To disrupt the heuristic learning rate, Indicates the current number Is the round iteration heuristic broken? ; For the first Round iteration breaks heuristic The corresponding variational inference parameters; No. Round iteration breaks heuristic The probability of choosing; Step 9.3: Update the neighborhood size variational inference parameters based on the normalized reward value.

6. The machine learning-driven multi-robot path planning method according to claim 5, characterized in that, In step 9.1, the reward value is calculated according to the following formula. Perform normalization to obtain the normalization reward value. : ； In the above formula, To prevent positive numbers with a denominator of zero; For the first Reward value in each iteration; For the first The original path cost of the robot in the neighborhood during the round of iteration.

7. The machine learning-driven multi-robot path planning method according to claim 5, characterized in that, Step 9.3, which involves updating the neighborhood size variational inference parameters based on the normalized reward value, includes: If the current destruction heuristic is Then the average neighborhood size is updated according to the following formula: ； In the above formula, The updated neighborhood size mean parameter; The learning rate is the neighborhood size; For the first The current neighborhood size in the round of iteration; For the first Round iteration breaks heuristic The corresponding mean neighborhood size; For the first Round iteration breaks heuristic The corresponding standard deviation of the neighborhood size; If the current destruction heuristic is The standard deviation of the neighborhood size is then updated according to the following formula: ； To ensure that the standard deviation of the neighborhood size is positive, the following constraint is applied: ； In the above formula, This is the lower bound of the standard deviation.

Images