Warehouse location dynamic picking path collaborative optimization method based on AI algorithm

Abstract

The invention discloses a warehouse location dynamic picking path collaborative optimization method based on an AI algorithm, and relates to the technical field of warehouse picking path planning. A scheduling server collects data of electronic positioning tags installed on shelf layers, point cloud data of laser radars deployed at channel nodes and inertial navigation module data of AGV logistics vehicles; and collecting operation habit data of a picker of the manual picking terminal APP. According to the invention, through multi-source data fusion and in combination with an obstacle identification technology of a Transform attention mechanism, the type of a dynamic obstacle can be accurately identified, and real-time updating of a space-time occupied grid can dynamically reflect a regional passage probability, so that the adaptability of the system to a complex storage environment is improved, and path interruption caused by environment change is reduced. By means of intelligent scheduling, the utilization rate of the storage space is increased, the dependence on skilled picking personnel is reduced through personalized path planning, the training period of new employees is shortened, and an extensible efficient solution is provided for a large-scale storage scene.

Description

Technical Field

[0001] This invention relates to the field of warehouse picking route planning technology, and in particular to a collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms. Background Technology

[0002] Current warehouse picking route planning technology presents a hybrid pattern of traditional algorithms dominating and partial AI pilot projects.

[0003] The mainstream solutions still rely on traditional methods such as S-shaped traversal and genetic algorithms. For example, when a daily chemical warehouse adopts a fixed shelf layout, pickers need to cross the warehouse more than 5 times a day, and the proportion of repeated paths is high.

[0004] Some leading companies have already introduced basic AI applications: a certain technology company's WMS system optimizes warehouse location allocation through AI algorithms, shortening picking paths; another warehouse reduces repetition rates through intelligent path planning. In the field of robotic picking, reinforcement learning technology is beginning to be implemented, such as AGV robots based on the DQN algorithm that can achieve dynamic obstacle avoidance, but these are mostly concentrated in single-device control scenarios.

[0005] Traditional systems rely on static path planning, which cannot cope with order fluctuations and changes in warehouse location. During a Black Friday event at a certain retailer, the distribution of best-selling products across five areas caused a sharp drop in picking efficiency, exposing a critical problem with the WMS system's heatmap update delay. Even with the improved Dijkstra algorithm reducing planning time to 0.8 seconds, it still struggles to handle real-time order peaks during major e-commerce promotions.

[0006] Human and robotic pickers frequently encounter path conflicts. Data from a 3C warehouse shows that unoptimized task allocation resulted in 15 high-frequency conflict points, leading to a high rate of employee fatigue complaints. The existing system lacks a global coordination mechanism, resulting in frequent overlap and waste in the working areas of AGV robots and human pickers.

[0007] Warehousing companies face challenges in system integration, with data on warehouse location management, order processing, and equipment control being difficult to exchange. One e-commerce warehouse experienced an operational incident where 5,000 refrigerators vanished due to a mismatch between inventory data and the picking system, resulting in direct losses exceeding 3 million yuan.

[0008] Storage location allocation and route planning are often separate processes. In one clothing warehouse, casual pants and T-shirts were assigned to different picking waves, resulting in a surge in repeat visits to the same area, and it was not uncommon for best-selling products to be placed in the corner of the warehouse. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms. The technical solution adopted is as follows: The AI-based method for collaborative optimization of dynamic picking routes in warehouse locations includes the following steps: Step 1: The scheduling server collects data from the electronic positioning tags installed on the shelf, the point cloud data of the lidar deployed at the aisle nodes, the inertial navigation module data of the AGV logistics vehicle, and the operation habit data of the pickers in the manual picking terminal APP. Step 2: Construct a warehouse topology map based on shelf coordinate data. Nodes are the intersections of storage location aisles, edges are reachable paths, the area is divided into grids, the attributes of each grid are labeled, and the map is stored as a raster map. The LiDAR point cloud data is transformed into obstacle outlines, and the obstacle type is identified by combining the Transformer attention mechanism; a spatiotemporal occupancy grid is constructed: the occupancy probability is marked for each grid. Update storage location labels by combining real-time inventory data with order picking status; Data on picker operational habits was used to extract core characteristics of picker habits. Step 3: Based on the correlation between goods and the distance between storage locations, the DBSCAN algorithm is used to cluster picking orders into task packages; Step 4: Calculate the comprehensive score based on the order urgency coefficient, total weight, and the dispersion of the involved storage locations. The decision-making agent calls the MADRL model and inputs the task package features, AGV status, and core features of the picker's habits. The MADRL model outputs the allocation scheme of AGV logistics vehicle picking tasks and manual picking tasks with the optimization objectives of minimizing the total task completion time and balancing equipment load. Step 5: For AGV picking tasks, the improved A* algorithm is used to search for the shortest path on the static topology graph; for manual picking tasks, an S-shaped and hotspot-first hybrid path is generated based on the location heat index.

[0010] Optionally, in step 1, UWB positioning technology is used to install electronic positioning tags on each row of shelves, and the scheduling server collects the shelf coordinate data output by the electronic positioning tags in real time. LiDAR is installed at the channel nodes. Point cloud data is filtered in real time by edge computing nodes, downsampled using the PCL library, compressed, and then transmitted to the scheduling server. The AGV logistics vehicle is equipped with a MEMS inertial navigation sensor, which communicates with the scheduling server at set time intervals to output a motion status message. The manual picking terminal has a built-in Beidou positioning module and motion sensor, and the scheduling server collects the operation sequence data and human-computer interaction data of the manual picking terminal APP.

[0011] Optionally, nodes in the topology map are aisle intersections, storage location endpoints, and entrances / exits; the physical distance of an edge is calculated based on the Euclidean distance between the coordinates of the two points, and the passage cost coefficient of an edge is calculated using the base cost, slope coefficient, and congestion coefficient; the grid map uses a dynamic grid size and is divided into ordinary storage areas, high-density shelving areas, and main aisle areas.

[0012] Optionally, in step 2, the 3D coordinate feature matrix of the LiDAR point cloud is input, and the spatial correlation features are captured by an 8-layer self-attention network with the Transformer attention mechanism. The obstacle type probability distribution is output, and the boundary is extracted by the minimum bounding rectangle algorithm for successfully identified obstacles. A spatiotemporal occupancy grid is constructed based on the set spatiotemporal granularity to cover the spatiotemporal range of a future set time. The spatiotemporal occupancy probability of each grid is calculated, and the occupancy probability of each grid is marked. Storage location labels include pending picking, occupied, and abnormal; The core characteristics of pickers' habits include familiarity with the area, preference for different levels of inventory, operational rhythm, route preference, and sensitivity to disturbances.

[0013] Optionally, area familiarity is calculated by dividing the area picking time by the total time; floor preference is obtained by subtracting the low floor picking percentage from the high floor picking percentage; operation rhythm is obtained by dividing the average barcode scanning interval by the average barcode scanning value per person in the area. If it is less than or equal to 1, the output operation rhythm is fast, otherwise the dual output operation rhythm is slow; if the value of left turn number minus right turn number is positive, the output path tends to favor left turns, otherwise the output path tends to favor right turns; interference sensitivity is obtained by dividing the number of abnormal operations by the number of regular operations.

[0014] Optionally, step 3 includes the following steps: Step 31: Extract the co-occurrence frequency of goods in historical order data and construct a goods association matrix: Let the set of goods be... The formula for calculating the correlation degree of goods is: ; in To simultaneously include goods and , It includes goods Total number of orders It includes goods The total number of orders; Step 32: Extract the coordinates of the cargo location based on the raster map constructed in Step 2.

[0015] The formula for calculating the overall distance between storage locations is: ; in It is the combined distance between storage location k and storage location l. It is the Euclidean distance between storage location k and storage location l. This is the passage cost coefficient between storage location k and storage location l, which is taken from the edge attributes of the topological graph. It is the Euclidean distance weighting coefficient. It is the toll cost weighting coefficient; Step 33: The DBSCAN algorithm clusters picking orders using the order set as input. Each order contains a sequence of storage locations. ; Step 34, for the order If there are at least 5 orders within the defined ε-neighborhood, it is defined as a core object. Unclassified orders are randomly selected. If an order is a core object, a new cluster is created. All reachable orders within the ε-neighborhood are recursively merged. This process is repeated until all orders are classified. Step 34: The DBSCAN algorithm outputs a set of task packets. Each task package meets the following requirements: the number of orders contained is less than or equal to 10, the diameter of the storage location distribution is less than or equal to 20m, and the average total correlation degree is greater than or equal to 0.5.

[0016] Optionally, the method for calculating the overall score in step 4 is to first calculate the order urgency factor. Total weight of the task package Storage location dispersion And calculate the overall score. ; ; in It is the base emergency value. It is the remaining time coefficient. , This is the standard processing time. This is the actual remaining time. It's about order complexity. , It includes the number of cargo spaces. These are the corresponding weights. It is the total minimum inventory unit quantity. These are the corresponding weights; The following formula is used to calculate the dispersion of cargo locations. : ; This represents the maximum distance between storage locations within the task package. This represents the average spacing between cargo locations in the area. Overall score The calculation formula is: .

[0017] Optionally, the cost function of the improved A* algorithm in step 5. The calculation formula is: Cost function ; in It is the actual cost from the starting point to node n, obtained by multiplying the distance by the toll cost coefficient. This is the estimated cost from node n to the destination, calculated using Manhattan distance. The improved A* algorithm supplements this calculation with an improved term: It is a dynamic obstacle penalty item. It is the penalty coefficient. It is the probability of grid occupancy; It is an energy consumption coefficient penalty item: It's the slope of the path. This is the distance of the path segment; The path smoothing process uses B-spline curves to optimize the turning angle, ensuring that the AGV's turning angle is less than 30°.

[0018] Optionally, in step 5, the location heat index is first calculated during the path planning for manual picking tasks. , ;in This refers to the daily picking frequency of location k. It is the number of days since the most recent picking at location k; if A value greater than or equal to 0.6 is marked as a hotspot area, and otherwise it is marked as a normal area; The path generation rule is: hotspot areas use a hotspot optimization strategy, and ordinary areas use an S-shaped route; Hotspot optimization strategy refers to... Arrange the storage locations in descending order to generate a sequence. A greedy algorithm is used to connect adjacent hotspots.

[0019] Optionally, in step 5, when planning the path for the manual picking task, a candidate path set is generated, and each path contains a turning sequence. If the turn value is greater than 0, it is a left turn, and if the turn value is less than 0, it is a right turn. The initial path weight is calculated, and the path weight is adjusted according to the picker's path tendency. The standard path length is calculated, and the compression rate is determined based on the picker's operation rhythm. The compressed path length is calculated based on the compression rate, and the planned path with the shortest compressed path length is selected as the output planned path.

[0020] In summary, the present invention has at least one of the following beneficial technical effects: This invention provides a collaborative optimization method for dynamic picking routes in warehouses based on AI algorithms. Through multi-source data fusion, it achieves comprehensive perception of the warehouse environment, equipment status, and personnel habits. Combined with obstacle recognition technology based on the Transformer attention mechanism, it can accurately identify dynamic obstacle types. The real-time update of the spatiotemporal occupancy grid can dynamically reflect the access probability of an area, improving the system's adaptability to complex warehouse environments and reducing path interruptions caused by environmental changes.

[0021] The DBSCAN-based order clustering strategy divides task packages into packages based on both cargo relevance and location distance, increasing the concentration of similar tasks in locations and significantly reducing cross-regional travel. The MADRL model dynamically allocates tasks through comprehensive scores, enabling collaborative scheduling of AGVs and manual picking. This shortens the total task completion time, improves equipment load balancing, and avoids overloading or idleness of individual devices.

[0022] The improved A* algorithm provides AGVs with dynamic obstacle avoidance paths based on a static topology map. By combining energy consumption coefficients and obstacle penalty terms, the AGV's travel distance is reduced, and energy consumption is lowered. The manual picking path is based on a hybrid strategy using the location heat index, combined with personalized adjustments to the picker's habits, thus improving manual picking efficiency and reducing picker fatigue.

[0023] Intelligent scheduling through AI algorithms improves warehouse space utilization, and collaborative operations between AGVs and humans reduce the cost of processing a single order. Personalized route planning reduces reliance on skilled pickers, shortens the training cycle for new employees, and lowers the overall system maintenance cost, providing a scalable and efficient solution for large-scale warehousing scenarios. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms, as described in this invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings.

[0026] This invention discloses a collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms.

[0027] Reference Figure 1 Example 1, a collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms, includes the following steps: Step 1: The scheduling server collects data from the electronic positioning tags installed on the shelf, the point cloud data of the lidar deployed at the aisle nodes, the inertial navigation module data of the AGV logistics vehicle, and the operation habit data of the pickers in the manual picking terminal APP. Step 2: Construct a warehouse topology map based on shelf coordinate data. Nodes are the intersections of storage location aisles, edges are reachable paths, the area is divided into grids, the attributes of each grid are labeled, and the map is stored as a raster map. The LiDAR point cloud data is transformed into obstacle outlines, and the obstacle type is identified by combining the Transformer attention mechanism; a spatiotemporal occupancy grid is constructed: the occupancy probability is marked for each grid. Update storage location labels by combining real-time inventory data with order picking status; Data on picker operational habits was used to extract core characteristics of picker habits. Step 3: Based on the correlation between goods and the distance between storage locations, the DBSCAN algorithm is used to cluster picking orders into task packages; Step 4: Calculate the comprehensive score based on the order urgency coefficient, total weight, and the dispersion of the involved storage locations. The decision-making agent calls the MADRL model and inputs the task package features, AGV status, and core features of the picker's habits. The MADRL model outputs the allocation scheme of AGV logistics vehicle picking tasks and manual picking tasks with the optimization objectives of minimizing the total task completion time and balancing equipment load. Step 5: For AGV picking tasks, the improved A* algorithm is used to search for the shortest path on the static topology graph; for manual picking tasks, an S-shaped and hotspot-first hybrid path is generated based on the location heat index.

[0028] By adopting the above technical solutions and fusing multi-source data, a comprehensive perception of the warehouse environment, equipment status, and personnel habits can be achieved. Combined with obstacle recognition technology based on the Transformer attention mechanism, dynamic obstacle types (such as temporary stacking of goods, personnel passage, etc.) can be accurately identified. Real-time updates of the spatiotemporal occupancy grid dynamically reflect the area's passage probability, improving the system's adaptability to complex warehousing environments and reducing path interruptions caused by environmental changes.

[0029] The DBSCAN-based order clustering strategy divides task packages into two dimensions: cargo relevance and location distance. This increases the concentration of similar tasks in the location pool and significantly reduces cross-regional travel. The MADRL model dynamically allocates tasks based on a comprehensive score (urgency factor, weight, and dispersion), enabling collaborative scheduling between AGVs and manual picking. This shortens the total task completion time, improves equipment load balancing, and avoids overloading or idleness of any single device.

[0030] The improved A* algorithm provides AGVs with dynamic obstacle avoidance paths based on a static topology map. By combining energy consumption coefficients and obstacle penalty terms, the AGV's travel distance and energy consumption are reduced. The manual picking path is based on a hybrid strategy of location heat index (S-shaped and hotspot priority), combined with personalized adjustments based on picker habits (area familiarity, path preference, etc.), which improves manual picking efficiency and reduces picker fatigue (e.g., left-turn preference paths reduce turning time, and fast-paced adaptation paths shorten ineffective travel).

[0031] The real-time dynamic adjustment mechanism enables the system to respond quickly to emergencies, with emergency order processing time controlled within 5 seconds, and order fulfillment on-time rate increased to over 98%. Simultaneously, dynamic updates to location tags ensure that inventory data is synchronized with actual status in real time, reducing picking error rates.

[0032] Intelligent scheduling through AI algorithms improves warehouse space utilization, and collaborative operations between AGVs and humans reduce the cost of processing a single order. Personalized route planning reduces reliance on skilled pickers, shortens the training cycle for new employees, and lowers the overall system maintenance cost, providing a scalable and efficient solution for large-scale warehousing scenarios.

[0033] In Example 2, in step 1, UWB positioning technology is used to install electronic positioning tags on each row of shelves, and the scheduling server collects the shelf coordinate data output by the electronic positioning tags in real time. LiDAR is installed at the channel nodes. Point cloud data is filtered in real time by edge computing nodes, downsampled using the PCL library, compressed, and then transmitted to the scheduling server. The AGV logistics vehicle is equipped with a MEMS inertial navigation sensor, which communicates with the scheduling server at set time intervals to output a motion status message. The manual picking terminal has a built-in Beidou positioning module and motion sensor, and the scheduling server collects the operation sequence data and human-computer interaction data of the manual picking terminal APP.

[0034] In Example 3, the nodes in the topology map are aisle intersections, storage location endpoints, and entrances / exits; the physical distance of the edges is calculated based on the Euclidean distance between the coordinates of the two points, and the passage cost coefficient of the edges is calculated using the basic cost, slope coefficient, and congestion coefficient; the grid map uses a dynamic grid size and is divided into ordinary storage area, high-density shelving area, and main aisle area.

[0035] By adopting the above technical solution and using UWB positioning technology to collect the coordinates of the shelves, the positioning accuracy is high. The real-time collected shelf coordinate data provides a millimeter-level reference for topology map construction, ensuring that the location information of the goods is accurate and reducing path planning deviations caused by positioning errors.

[0036] After real-time edge computing filtering (removing environmental noise points) and PCL library downsampling (preserving key feature points) of the LiDAR point cloud at the channel nodes, the data volume can be compressed by 60%-70%, reducing transmission bandwidth usage and decreasing the data processing pressure on the scheduling server. Simultaneously, the preprocessed point cloud data retains key obstacle contour information, providing high-quality input for obstacle recognition using the Transformer attention mechanism, improving recognition accuracy to over 95% and shortening dynamic obstacle response time.

[0037] The MEMS inertial navigation sensor on the AGV outputs motion status messages at set intervals (such as 100ms), which can capture parameters such as speed, acceleration, and heading angle in real time. Combined with UWB shelf positioning, it can achieve centimeter-level trajectory correction of the AGV, avoid cumulative positioning errors, improve the autonomous navigation accuracy of the AGV, and control the path offset within 5 centimeters.

[0038] The Beidou positioning (1-meter level outdoors and 0.5-meter level indoor auxiliary positioning) of the manual picking terminal, combined with motion sensors, can accurately record the picker's movement trajectory, shelf layer operation, barcode scanning interval and other data, providing fine-grained data support for extracting core features such as area familiarity and layer preference, thereby improving the adaptability of personalized path planning.

[0039] In Example 4, in step 2, the 3D coordinate feature matrix of the LiDAR point cloud is input, and the spatial correlation features are captured by an 8-layer self-attention network with Transformer attention mechanism. The obstacle type probability distribution is output, and the boundary is extracted for successfully identified obstacles using the minimum bounding rectangle algorithm. A spatiotemporal occupancy grid is constructed based on the set spatiotemporal granularity to cover the spatiotemporal range of a future set time. The spatiotemporal occupancy probability of each grid is calculated, and the occupancy probability of each grid is marked. Storage location labels include pending picking, occupied, and abnormal; The core characteristics of pickers' habits include familiarity with the area, preference for different levels of inventory, operational rhythm, route preference, and sensitivity to disturbances.

[0040] Example 5: Area familiarity is calculated by dividing the area picking time by the total time; floor preference is obtained by subtracting the low floor picking percentage from the high floor picking percentage; operation rhythm is obtained by dividing the average barcode scanning interval by the average barcode scanning value of people in the area. If it is less than or equal to 1, the output operation rhythm is fast, otherwise the dual output operation rhythm is slow; if the value of the left turn number minus the right turn number is positive, the output path tends to prefer left turns, otherwise the output path tends to prefer right turns; interference sensitivity is obtained by dividing the number of abnormal operations by the number of normal operations.

[0041] By adopting the above technical solution and using an 8-layer self-attention network to process the 3D coordinate feature matrix of the LiDAR point cloud, the spatial correlation features of obstacles (such as personnel between shelves and differences in the shape of stacked goods) can be effectively captured. Compared with traditional CNN networks, the accuracy of identifying multiple types of obstacles (static shelves, moving personnel, temporary stacks of goods) in complex scenes is improved, and the obstacle type classification accuracy is high. The output probability distribution provides a reference for the system's decision confidence.

[0042] For successfully identified obstacles, the minimum bounding rectangle algorithm is used to extract the boundaries, which can control the accuracy of the description of the obstacle's spatial occupancy within ±10cm. Compared with polygon fitting methods, the amount of boundary data is reduced, providing lightweight input for path planning algorithms (such as improved A*). For example, for personnel who suddenly enter the passage, a 2m×0.8m rectangular restricted area can be quickly generated, allowing the AGV to start decelerating and avoiding the obstacle 3 meters in advance, reducing the risk of collision.

[0043] Based on a set spatiotemporal granularity (e.g., 10s time granularity, 0.5m×0.5m spatial granularity), a spatiotemporal occupancy grid covering the next 3-5 minutes is constructed. The occupancy probability (0-1) of each grid can quantify and predict the traffic risk in the area. For example, if the grid at the intersection of the main passage has an occupancy probability greater than or equal to 0.7 after 10s, the system will guide the AGV to detour in advance, reducing the number of path adjustments and shortening the average task completion time.

[0044] The occupancy probability comprehensively considers factors such as historical traffic data, real-time obstacle locations, and equipment movement trajectories. Compared with static grid maps, it can more accurately reflect dynamic congestion (such as AGV convergence and manual picking clusters), making the prediction accuracy of high-probability occupancy areas reach over 85%, thus avoiding the system falling into the planning trap of being optimal at present but congested in the future.

[0045] Example 6, step 3 includes the following steps: Step 31: Extract the co-occurrence frequency of goods in historical order data and construct a goods association matrix: Let the set of goods be... The formula for calculating the correlation degree of goods is: ; in To simultaneously include goods and , It includes goods Total number of orders It includes goods The total number of orders; Step 32: Extract the coordinates of the cargo location based on the raster map constructed in Step 2.

[0046] The formula for calculating the overall distance between storage locations is: ; in It is the combined distance between storage location k and storage location l. It is the Euclidean distance between storage location k and storage location l. This is the passage cost coefficient between storage location k and storage location l, which is taken from the edge attributes of the topological graph. It is the Euclidean distance weighting coefficient. It is the toll cost weighting coefficient; Step 33: The DBSCAN algorithm clusters picking orders using the order set as input. Each order contains a sequence of storage locations. ; Step 34, for the order If there are at least 5 orders within the defined ε-neighborhood, it is defined as a core object. Unclassified orders are randomly selected. If an order is a core object, a new cluster is created. All reachable orders within the ε-neighborhood are recursively merged. This process is repeated until all orders are classified. Step 34: The DBSCAN algorithm outputs a set of task packets. Each task package meets the following requirements: the number of orders contained is less than or equal to 10, the diameter of the storage location distribution is less than or equal to 20m, and the average total correlation degree is greater than or equal to 0.5.

[0047] By employing the above technical solution, an association matrix is ​​constructed based on the co-occurrence frequency of goods in historical orders. A formulaic association degree is then used, which, compared to traditional empirical clustering grouping by category, can more accurately capture implicit associations (such as complementary products or bundled sales). For example, the association degree between milk powder and baby bottles can be quantified to 0.75 using co-occurrence data, improving the clustering accuracy of such high-frequency associated orders and reducing cross-regional picking caused by mixing unrelated goods. The association degree value is normalized to the [0,1] range, making the association degrees of goods of different categories and sales volumes comparable, avoiding distortion of association degrees due to sales volume differences (such as best-selling items versus long-tail products). After normalization, goods with low sales volume but strong associations (such as printers and specific ink cartridges) can be effectively clustered, improving the processing efficiency of such orders and reducing duplicate paths caused by missed associations.

[0048] The comprehensive distance formula for storage locations breaks through the limitations of traditional straight-line distances by taking into account actual passage resistance such as aisle slope and congestion coefficient. For example, the passage cost coefficient for aisles in high-density racking areas can be set to 1.5 (1.0 for ordinary aisles), making distance calculations more closely match the actual passage difficulty of AGVs or manual labor, and reducing the deviation between the theoretical optimality of path planning and actual feasibility.

[0049] The core object is defined by having at least 5 orders within its ε-neighborhood (combined with dual thresholds of correlation ≥ 0.6 and distance ≤ 5m). Compared to algorithms like K-means that require pre-setting the number of clusters, this approach can adapt to fluctuations in order volume (such as a surge in orders during promotional periods). Core object-driven recursive clustering can automatically merge spatially concentrated and closely related orders, increasing the location clustering of similar task packages and reducing the number of cross-region picking operations.

[0050] The process of randomly selecting uncategorized orders until all orders are categorized avoids clustering bias caused by order order order sequence, ensuring that small-batch or urgent orders are not marginalized. For example, if a single urgent order meets the criteria for a core object, it can be grouped into a smaller task package for priority processing, thus shortening the response time for urgent orders.

[0051] In Example 7, the method for calculating the overall score in step 4 is to first calculate the order urgency coefficient. Total weight of the task package Storage location dispersion And calculate the overall score. ; ; in It is the base emergency value. It is the remaining time coefficient. , This is the standard processing time. This is the actual remaining time. It's about order complexity. , It includes the number of cargo spaces. These are the corresponding weights. It is the total minimum inventory unit quantity. These are the corresponding weights; The following formula is used to calculate the dispersion of cargo locations. : ; This represents the maximum distance between storage locations within the task package. This represents the average spacing between cargo locations in the area. Overall score The calculation formula is: .

[0052] Example 8, the cost function of the improved A* algorithm in step 5 The calculation formula is: Cost function ; in It is the actual cost from the starting point to node n, obtained by multiplying the distance by the toll cost coefficient. This is the estimated cost from node n to the destination, calculated using Manhattan distance. The improved A* algorithm supplements this calculation with an improved term: It is a dynamic obstacle penalty item. It is the penalty coefficient. It is the probability of grid occupancy; It is an energy consumption coefficient penalty item: It's the slope of the path. This is the distance of the path segment; The path smoothing process uses B-spline curves to optimize the turning angle, ensuring that the AGV's turning angle is less than 30°.

[0053] By adopting the above technical solution, the order urgency coefficient formula achieves precise priority ranking through three dimensions: the basic urgency value (EV) distinguishes order types (express / urgent / normal), clarifying inherent priority; the remaining time coefficient (RTC = standard processing time / actual remaining time) dynamically reflects time urgency, avoiding the backlog of expired orders; and the order complexity combines the number of storage locations and SKUs (weighted coefficient) to prevent small but complex orders from being overlooked. Compared to the traditional method of sorting only by time, the urgency coefficient calculation improves the accuracy of order priority judgment and results in a high on-time fulfillment rate for urgent orders.

[0054] The location dispersion formula quantifies the spatial dispersion of task packages, reflecting the picking difficulty better than simply looking at the total distance. For example, given two task packages, each containing 5 locations, if task package A has MD=15m and AD=10m, the system will prioritize assigning task package B to manual picking, thus improving manual walking efficiency.

[0055] The comprehensive scoring formula achieves multi-objective optimization through weight allocation: urgency (0.5) ensures timeliness priority and avoids delays; reciprocal of weight (0.2) guides AGVs to prioritize heavy goods (reducing manual workload); reciprocal of dispersion (0.3) promotes spatial concentration of task packages.

[0056] This mechanism makes the task allocation scheme of the MADRL model more in line with actual needs, improves the load balancing of equipment, and enhances the efficiency of human-machine collaboration.

[0057] In Example 9, step 5, the location heat index is first calculated in the path planning for manual picking tasks. , ;in This refers to the daily picking frequency of location k. It is the number of days since the most recent picking at location k; if A value greater than or equal to 0.6 is marked as a hotspot area, and otherwise it is marked as a normal area; The path generation rule is: hotspot areas use a hotspot optimization strategy, and ordinary areas use an S-shaped route; Hotspot optimization strategy refers to... Arrange the storage locations in descending order to generate a sequence. A greedy algorithm is used to connect adjacent hotspots.

[0058] In Example 10, if in step 5, when planning the path for the manual picking task, a candidate path set is generated, and each path contains a turning sequence greater than 0 for left turns and less than 0 for right turns, the initial path weight is calculated, and the path weight is adjusted according to the picker's path tendency; the standard path length is calculated, and the compression rate is determined based on the picker's operation rhythm; the compressed path length is calculated based on the compression rate, and the planned path with the shortest compressed path length is selected as the output planned path.

[0059] By adopting the above technical solution, the location popularity index formula achieves dual-dimensional evaluation: 70% weight is given to daily frequency to ensure that high-frequency picking locations (such as best-selling products) are marked first; 30% weight is given to freshness by counting back the interval days to avoid missing locations with a recent surge in demand (such as promotional new products).

[0060] Compared to the traditional method of classifying by frequency alone, the accuracy of heat identification is improved, the coverage of storage locations in hot spots is high, and the number of back-and-forth picking caused by missed hot spots is reduced.

[0061] The system automatically divides areas into hotspots and ordinary areas based on their popularity index, enabling it to plan routes tailored to the characteristics of different areas. Hotspot areas (such as the daily necessities section) account for 80% of picking needs; adopting targeted strategies can reduce unnecessary walking. Ordinary areas (such as areas with slow-moving goods) are covered efficiently using a fixed pattern to avoid wasting resources.

[0062] Dynamic updates to regional divisions (refreshed daily based on data) improve the alignment between route planning and actual freight flow changes.

[0063] Hotspot areas are sorted in descending order of H(k) to generate a location sequence. A greedy algorithm is then used to connect adjacent hotspots (prioritizing locations ≤3m apart). This approach is superior to random picking. The path length in hotspot areas is shortened; the proportion of time pickers spend in hotspot areas is reduced, minimizing repeated trips.

[0064] In general areas, an S-shaped route is used to traverse the aisles, combined with the "prioritize same-side locations" rule, which reduces the number of aisle crossings. For areas with sparsely distributed locations, the S-shaped route ensures no omissions and achieves 100% picking completeness, while avoiding the time wasted by backtracking routes.

[0065] By combining the corner sequence of the candidate path set with the picker's path preference adjustment weight (e.g., increasing the weight of the left turn segment for those who prefer to turn left): the proportion of turns that conform to habits increases and the turning time is reduced; Pickers' operational fluency improved, and the error rate caused by unfamiliar turning decreased.

[0066] Efficiency and fatigue balance in operation rhythm adaptation are achieved by adjusting path length based on operation rhythm: Fast-paced pickers have improved path compression rates and shortened task completion times; Slow-paced pickers retain buffer paths, reducing operational stress and extending continuous work hours.

[0067] The compression strategy achieves this by eliminating detours and merging nearby storage locations, ensuring 100% storage location access completeness while balancing efficiency and accuracy.

[0068] The following specific embodiments illustrate the implementation principle of the present invention: Taking a large e-commerce warehousing center as an example, the warehouse has a total area of ​​approximately 5,000 square meters, including a general storage area, a high-density shelving area, and a main aisle area. It processes an average of 8,000 orders per day and is equipped with 15 AGV logistics vehicles and 30 pickers. The specific implementation process of this technical solution is as follows: I. Data Collection Phase: Shelf positioning: UWB electronic positioning tags are installed on each row of shelves. The scheduling server collects tag data in real time to obtain the precise coordinates of the shelves (accuracy up to 20cm), such as the coordinates of the shelves in area A as (X:12.5,Y:8.3,Z:0) and the coordinates of the shelves in area B as (X:25.1,Y:10.2,Z:0), etc.

[0069] Environmental perception: Ten LiDARs are deployed at the intersection of warehouse aisles. Point cloud data is filtered by edge nodes and then downsampled and compressed through the PCL library (reducing the data volume by 65%) before being transmitted to the server for obstacle identification.

[0070] Equipment status: The AGV is equipped with a MEMS inertial navigation sensor, which sends motion status messages (including speed, heading angle, etc.) to the server every 100ms; the manual picking terminal has a built-in Beidou positioning module to record the picker's movement trajectory and barcode scanning operation data (such as barcode scanning interval, shelf layer selection, etc.).

[0071] II. Data Processing Stage: Map construction: A topology map is constructed based on the rack coordinates. Nodes include aisle intersections (such as C1, C2), storage location endpoints (such as A1-01, B3-05), and entrances and exits. The passage cost coefficient of the edge is calculated by combining the basic distance, slope (slope coefficient of 1.2 for high-density rack areas) and real-time congestion. The grid map is dynamically adjusted in size according to the area (1m×1m for ordinary areas and 0.5m×0.5m for high-density areas).

[0072] Obstacle recognition: The LiDAR point cloud is processed by an 8-layer Transformer self-attention network to identify obstacles such as temporary cargo piles and pedestrians. The boundary is extracted using the minimum bounding rectangle (e.g., the boundary of a pedestrian obstacle is 2m×0.8m) and marked to the spatiotemporal occupancy grid (temporal granularity 10s, spatial granularity 0.5m×0.5m). For example, the occupancy probability of the grid near node C1 is 0.8 after 10s.

[0073] Location and personnel characteristics: Location labels are updated in real time to "pending picking", "occupied" or "abnormal" (e.g., A1-03 is marked as abnormal due to damaged goods); characteristics are extracted from the picking staff's operation data, such as picking staff Mr. Li's area familiarity with area A is 0.8, he prefers to turn left and has a fast operating pace, while Mr. Wang prefers low shelves and has high sensitivity to interference.

[0074] III. Order Clustering Stage: The system extracts co-occurrence patterns of goods from historical orders, clusters strongly correlated products such as shampoo and conditioner, and mobile phones and phone cases, and combines this with the distance between storage locations (e.g., the distance between storage locations in area A is ≤5m) to generate task packages using the DBSCAN algorithm. Task package T1 contains 10 orders, involving warehouses concentrated in area A (with a distribution diameter of 15m), with an average total correlation of 0.7, and includes products such as shampoo and conditioner; Task package T2 contains 8 orders, involving storage locations in area B (distribution diameter 18m), with an average total correlation of 0.6, and includes mobile phone accessories.

[0075] IV. Task Allocation Phase: Calculate the overall score of the task packages: T1 has an urgency coefficient of 0.8 (including 3 rush orders), a total weight of 50kg, a dispersion of 0.6, and an overall score of 0.72; T2 has an urgency coefficient of 0.5, a total weight of 30kg, a dispersion of 0.5, and an overall score of 0.61.

[0076] The decision-making agent invokes the MADRL model, taking into account task package features, AGV status (e.g., AGV1 has 80% remaining battery and 0 load) and picker features, and outputs an allocation scheme: AGV1 is responsible for T2 (lighter weight, lower urgency), while AGV2 is responsible for another urgent task package; Master Li (fast-paced, familiar with Area A) is assigned to T1, and Master Wang (prefers lower levels) is assigned to Area C task packages.

[0077] V. Path Planning Phase: AGV Path: AGV1 plans its path based on the improved A* algorithm, avoiding grids with an occupancy probability > 0.7 and prioritizing smooth routes (reducing energy consumption by 12%). The path is: from the starting point to C2 to B2 channel to the target storage location, with all turning angles ≤ 30°.

[0078] Manual path: Mr. Li's T1 task involves the following areas: hot spots (popularity index 0.7) such as A1-01 and A1-02 are arranged in descending order of popularity, and a greedy algorithm is used to connect adjacent locations; ordinary areas use an S-shaped route to reduce the number of crossings (from 8 to 4).

[0079] The system generates three candidate routes. After adjusting the weights based on Mr. Li's left-turn preference, the system selects the route with a left-turn percentage of 70%, compressing the length from the original 60m to 51m (compression rate of 15%) to match his fast-paced driving habits.

[0080] VI. Dynamic Adjustment and Effects: When the lidar detects a sudden accumulation of goods in channel B2 (occupancy probability 0.9), the system immediately replans the path for AGV1, which increases the detour time by 5 seconds but avoids blockage; when an emergency order (highest priority) is inserted before T1, the path recalculation takes 4 seconds.

[0081] After implementation, the total warehouse picking time was reduced by 32%, AGV energy consumption was reduced by 15%, the manual picking error rate was reduced to 0.3%, the training cycle for new employees was shortened from 2 weeks to 1 week, and the cost per order was reduced by 20%.

[0082] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms, characterized in that, Includes the following steps: Step 1: The scheduling server collects data from the electronic positioning tags installed on the shelf, the point cloud data of the lidar deployed at the aisle nodes, the inertial navigation module data of the AGV logistics vehicle, and the operation habit data of the pickers in the manual picking terminal APP. Step 2: Construct a warehouse topology map based on shelf coordinate data. Nodes are the intersections of storage location aisles, edges are reachable paths, the area is divided into grids, the attributes of each grid are labeled, and the map is stored as a raster map. The LiDAR point cloud data is converted into obstacle outlines, and the obstacle type is identified by combining the Transformer attention mechanism. Construct a spatiotemporal occupancy grid: label the occupancy probability of each grid cell; Update storage location labels by combining real-time inventory data with order picking status; Data on picker operational habits was used to extract core characteristics of picker habits. Step 3: Based on the correlation between goods and the distance between storage locations, the DBSCAN algorithm is used to cluster picking orders into task packages; Step 4: Calculate the comprehensive score based on the order urgency coefficient, total weight, and the dispersion of the involved storage locations. The decision-making agent calls the MADRL model and inputs the task package features, AGV status, and core features of the picker's habits. The MADRL model outputs the allocation scheme of AGV logistics vehicle picking tasks and manual picking tasks with the optimization objectives of minimizing the total task completion time and balancing equipment load. Step 5: For AGV picking tasks, the improved A* algorithm is used to search for the shortest path on the static topology graph; for manual picking tasks, an S-shaped and hotspot-first hybrid path is generated based on the location heat index.

2. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 1, characterized in that, In step 1, UWB positioning technology is used to install electronic positioning tags on each row of shelves, and the scheduling server collects the shelf coordinate data output by the electronic positioning tags in real time. LiDAR is installed at the channel nodes. Point cloud data is filtered in real time by edge computing nodes, downsampled using the PCL library, compressed, and then transmitted to the scheduling server. The AGV logistics vehicle is equipped with a MEMS inertial navigation sensor, which communicates with the scheduling server at set time intervals to output a motion status message. The manual picking terminal has a built-in Beidou positioning module and motion sensor, and the scheduling server collects the operation sequence data and human-computer interaction data of the manual picking terminal APP.

3. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 2, characterized in that, The nodes in the topology map are aisle intersections, storage location endpoints, and entrances / exits; the physical distance of the edges is calculated based on the Euclidean distance between the coordinates of the two points, and the passage cost coefficient of the edges is calculated using the base cost, slope coefficient, and congestion coefficient; the grid map uses a dynamic grid size and is divided into ordinary storage areas, high-density shelving areas, and main aisle areas.

4. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 3, characterized in that, In step 2, the 3D coordinate feature matrix of the LiDAR point cloud is input, and the spatial correlation features are captured by an 8-layer self-attention network with Transformer attention mechanism. The obstacle type probability distribution is output, and the boundary is extracted by the minimum bounding rectangle algorithm for successfully identified obstacles. A spatiotemporal occupancy grid is constructed based on the set spatiotemporal granularity to cover the spatiotemporal range of a future set time. The spatiotemporal occupancy probability of each grid is calculated, and the occupancy probability of each grid is marked. Storage location labels include pending picking, occupied, and abnormal; The core characteristics of pickers' habits include familiarity with the area, preference for different levels of inventory, operational rhythm, route preference, and sensitivity to disturbances.

5. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 4, characterized in that, Area familiarity is calculated by dividing the area picking time by the total time; floor preference is obtained by subtracting the low floor picking percentage from the high floor picking percentage; operation rhythm is obtained by dividing the average barcode scanning interval by the average barcode scanning value per person in the area. If it is less than or equal to 1, the output operation rhythm is fast, otherwise the dual output operation rhythm is slow; if the value of left turn number minus right turn number is positive, the output path tends to favor left turns, otherwise the output path tends to favor right turns; interference sensitivity is obtained by dividing the number of abnormal operations by the number of normal operations.

6. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 5, characterized in that, Step 3 includes the following steps: Step 31: Extract the co-occurrence frequency of goods in historical order data and construct a goods association matrix: Let the set of goods be... The formula for calculating the correlation degree of goods is: ； in To simultaneously include goods and , It includes goods Total number of orders It includes goods The total number of orders; Step 32: Extract the coordinates of the cargo location based on the raster map constructed in Step 2. ; The formula for calculating the overall distance between storage locations is: ; in It is the combined distance between storage location k and storage location l. It is the Euclidean distance between storage location k and storage location l. This is the passage cost coefficient between storage location k and storage location l, which is taken from the edge attributes of the topological graph. It is the Euclidean distance weighting coefficient. It is the toll cost weighting coefficient; Step 33: The DBSCAN algorithm clusters picking orders using the order set as input. Each order contains a sequence of storage locations. ; Step 34, for the order If there are at least 5 orders within the defined ε-neighborhood, it is defined as a core object. Unclassified orders are randomly selected. If an order is a core object, a new cluster is created. All reachable orders within the ε-neighborhood are recursively merged. This process is repeated until all orders are classified. Step 34: The DBSCAN algorithm outputs a set of task packets. Each task package meets the following requirements: the number of orders contained is less than or equal to 10, the diameter of the storage location distribution is less than or equal to 20m, and the average total correlation degree is greater than or equal to 0.

5.

7. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 6, characterized in that, The method for calculating the overall score in step 4 is to first calculate the order urgency factor. Total weight of the task package Storage location dispersion And calculate the overall score. ; ； in It is the base emergency value. It is the remaining time coefficient. , This is the standard processing time. This is the actual remaining time. It's about order complexity. , It includes the number of cargo spaces. These are the corresponding weights. It is the total minimum inventory unit quantity. These are the corresponding weights; The following formula is used to calculate the dispersion of cargo locations. : ; This represents the maximum distance between storage locations within the task package. This represents the average spacing between cargo locations in the area. Overall Score The calculation formula is: 。 8. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 7, characterized in that, The cost function of the improved A* algorithm in step 5 The calculation formula is: Cost function ; in It is the actual cost from the starting point to node n, obtained by multiplying the distance by the toll cost coefficient. This is the estimated cost from node n to the destination, calculated using Manhattan distance. The improved A* algorithm supplements this calculation with an improved term: It is a dynamic obstacle penalty item. It is the penalty coefficient. It is the probability of grid occupancy; It is an energy consumption coefficient penalty item: It's the slope of the path. This is the distance of the path segment; The path smoothing process uses B-spline curves to optimize the turning angle, ensuring that the AGV's turning angle is less than 30°.

9. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 8, characterized in that, In step 5, the location heat index is first calculated during the path planning for manual picking tasks. , ;in This refers to the daily picking frequency of location k. It is the number of days since the most recent picking at location k; if A value greater than or equal to 0.6 is marked as a hotspot area, and otherwise it is marked as a normal area; The path generation rule is: hotspot areas use a hotspot optimization strategy, and ordinary areas use an S-shaped route; Hotspot optimization strategy refers to... Arrange the storage locations in descending order to generate a sequence. A greedy algorithm is used to connect adjacent hotspots.

10. The collaborative optimization method for dynamic picking routes in warehouse locations based on AI algorithms according to claim 9, characterized in that, If in step 5, when planning the path for the manual picking task, a set of candidate paths is generated, and each path contains a turn sequence greater than 0 for left turns and less than 0 for right turns, the initial path weight is calculated, and the path weight is adjusted according to the picker's path preference. Calculate the standard path length, adjust and determine the compression rate based on the picker's operating rhythm, calculate the compressed path length based on the compression rate, and select the shortest compressed path length as the output planned path.

Images