A warehouse management system based on dynamic programming picking path

Abstract

The present application relates to the technical field of intelligent warehouse management, in particular to a warehouse management system based on dynamic planning picking path. In view of the technical problems that the dynamic planning of the prior art is prone to state space explosion in multi-block warehouse, and the static planning is prone to congestion deadlock, the system comprises a digital twin preprocessing module, a bit mask dynamic planning engine, a dynamic weight adjustment module and a collaborative control module. The system blocks the exponential divergence of state index through connectivity state relaxation and bit mask dimensionality reduction, realizes millisecond-level path generation, and introduces a space congestion matrix and a dynamic cost penalty to trigger bottom high-frequency re-planning and access order inversion. The present application completely eliminates congestion blind area and greatly improves warehouse throughput, and is mainly used in high-density multi-agent warehouse environment.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent warehousing management, and particularly to a warehousing management system based on a dynamic programming picking path. Background Art

[0002] In modern warehousing management and logistics distribution systems, with the explosive growth of e-commerce and the popularization of omnichannel retail, warehousing operations are facing the dual challenges of a large number of fragmented orders and extremely high timeliness. Order picking operations are the most critical, time-consuming, and cost-intensive links in the entire warehousing operation life cycle. To reduce the ineffective physical movement and detours of pickers or autonomous mobile robots shuttling between different storage locations, the optimization of the underlying path algorithm has become the core means to reconstruct the picking path and improve throughput. The classic path generation algorithm based on dynamic programming can construct a globally optimal traversal loop in linear time for a fixed channel structure, laying the theoretical foundation for path optimization.

[0003] However, with the expansion of the scale of intelligent warehousing, the widespread adoption of discrete storage strategies, and the geometric increase in the density of automation equipment, the existing warehousing path planning and management system based on traditional dynamic programming faces insurmountable underlying technical defects in industrial applications.

[0004] The core technical problems that this application actually needs to solve are as follows: First, in complex multi-block warehousing scenarios, the precise dynamic programming algorithm faces the problem of "state space explosion". Modern logistics center networks include multiple horizontal main channels and vertical secondary channels. To ensure the absolute connectivity in the recursive stage of the algorithm, traditional dynamic programming must record all combinations of channel access states. The number of valid states in the multi-block state space diverges exponentially or even factorially, resulting in a sharp increase in memory consumption and calculation delay, making it difficult to support the real-time high-frequency order dispatching requirements of seconds, and it is extremely easy to crash due to exhausted computing power. Second, in a multi-agent concurrent physical environment, static global path planning is prone to falling into "congestion blind spots" and systematic deadlocks. The underlying state transition cost of traditional algorithms only depends on the static two-dimensional physical distance matrix, and fails to internalize the real-time traffic flow density, queuing waiting expectation, and kinematic start-stop overhead of heavy-load agents in the dynamic environment as continuous congestion penalty weights. This leads the system to frequently guide a large number of agents to the hot channels with the shortest static distance, triggering serious dynamic resource competition, local queuing interference, or even complete head-on deadlocks, ultimately resulting in a cliff-like decline in the actual throughput capacity of the system. Summary of the Invention

[0005] The main objective of this invention is to provide a warehouse management system based on dynamic planning picking paths to solve the problem of "state space explosion" computing power bottleneck caused by strict tracking of physical connectivity in complex multi-block warehouse scenarios. At the same time, it solves the defects of static global path planning that easily leads to systemic deadlock and severe congestion in multi-agent concurrent operation environments, and realizes adaptive planning that deeply couples computing power compression and dimensionality reduction with real-time traffic feedback. To achieve the above objectives, the present invention provides the following technical solution: a warehouse management system based on dynamic programming picking paths, comprising: a data acquisition and digital twin preprocessing module, used to map the physical warehouse environment into a topological directed graph model, receive order data streams, perform order clustering and kinematic parameter binding to generate a picking point set, and simultaneously establish a bijective mapping relationship between topological graph nodes and bitmask indices; a bitmask dynamic programming engine based on state relaxation, connected to the data acquisition and digital twin preprocessing module, used to compress the dynamic picking subset into bitmask states, and perform connectivity state relaxation in the search domain, merging multiple physical entry and exit nodes into relaxed supernodes, and then using low-level bit operations to execute state transition equations to derive the globally optimal macroscopic picking sequence in polynomial time complexity; congestion perception and... The dynamic weight adjustment module, connected to the state-relaxation-based bitmask dynamic programming engine, is used to collect the coordinates and velocity vectors of the work entities in real time to calculate the spatiotemporal flow density, generate a dynamic congestion index based on the capacity limit, and reconstruct a dynamic edge weight matrix that integrates real-time traffic density through a continuous nonlinear mapping function. The dynamic edge weight matrix is ​​then injected into the state transition cost of the dynamic programming engine at high frequency. The multi-agent deadlock prevention and cooperative control module, connected to both the state-relaxation-based bitmask dynamic programming engine and the congestion perception and dynamic weight adjustment module, is used to apply mutex control commands in real time on restricted road sections, synchronously trigger a cost hot update mechanism to inject maximum value penalties, and induce subsequent agents to perform local trajectory replanning and order reversal while retaining the original unfinished task framework.

[0006] As a preferred embodiment of the present invention, the data acquisition and digital twin preprocessing module internally includes a multi-dimensional road network topology directed graph construction unit and a kinematics perception and order wave analysis unit; the multi-dimensional road network topology directed graph construction unit is used to construct a static directed topology graph model and assign physical distance length, geometric width attributes, and theoretical capacity upper limit to a single directed edge; the kinematics perception and order wave analysis unit is used to establish a bijective mapping function. The binary bit sequence number in the bitmask is matched one-to-one with the corresponding physical node identifier in the topology diagram, and the hardware kinematic parameters of the device undertaking the task are bound together.

[0007] As a preferred embodiment of the present invention, the kinematic sensing and order wave analysis unit is further configured to: when the number of picking points in a single wave... Exceeding the system's preset single-time enumerable threshold When this occurs, a hierarchical task decomposition strategy is triggered, dividing the original set of picking points into several groups based on spatial proximity and block affiliation, with each group containing no more than [number of elements]. The subtask groups are divided into subtask groups. Each subtask group independently solves the optimal access sequence and then completes the splicing through lightweight inter-group connection sorting, thereby ensuring that the dynamic programming solution state is within the range of hardware computing power.

[0008] As a preferred technical solution of the present invention, the bitmask dynamic programming engine based on state relaxation includes: a bitmask state dimensionality reduction definition unit, a state space connectivity relaxation and aggregation unit, a core dynamic programming state transition execution unit, and a connectivity restoration and path repair unit.

[0009] As a preferred embodiment of the present invention, the state space connectivity relaxation and aggregation unit introduces a many-to-one dimensionality reduction mapping function. This will connect multiple physical nodes located at the edge of the same logical block that are equivalent in terms of macroscopic path direction decision-making. At the logical level, they are merged and mapped into abstract relaxed supernodes. The system stores its unique identifier as a hash key in the dynamic programming dictionary to prevent the exponential explosion of node dimensions. When constructing the macroscopic transfer cost matrix between relaxed supernodes, the system takes the minimum shortest path cost between the physical nodes merged into two different relaxed supernodes as the macroscopic transfer cost between those two supernodes. and Its macroeconomic transfer cost is defined as ,in Topology graph From physical nodes To physical node The shortest path distance. The minimum value strategy ensures that the cost generated in the macroscopic stage is a lower bound of the actual reachable cost, so that when the subsequent connectivity restoration and path repair stages complete the physical connection path based on this lower bound, the final path cost is close to the global optimum.

[0010] As a preferred embodiment of the present invention, the core dynamic programming state transition execution unit adopts a bottom-up triple loop mechanism for optimal substructure iterative update, and filters effective pick points and unvisited targets through bit operations. The core state transition equation it triggers is: in, This is the current mask. To ensure the current picking point index is valid, Index unvisited points for the target to be explored. and This is the identifier of the corresponding physical node in the topology graph. This refers to the dynamic integrated path cost.

[0011] As a preferred embodiment of the present invention, the connectivity restoration and path repair unit is configured to perform a constrained shortest path search in the local neighborhood space along the generated preliminary macroscopic path array, physically connect and complete the connectivity gaps caused by relaxation aggregation, and recalculate and eliminate the redundant back-visits in the local range by calling dynamic programming again.

[0012] As a preferred embodiment of the present invention, the congestion perception and dynamic weight adjustment module internally includes a real-time spatiotemporal flow density monitoring unit, a dynamic congestion weight coefficient calculation unit, and a cost function reconstruction and hot update feedback unit; the dynamic congestion weight coefficient calculation unit calculates the congestion index coefficient based on the number of real-time active operation entities on the directed edge and the theoretical capacity. The penalty weight coefficient is calculated using a continuous nonlinear mapping function. Its mapping logic is: when Greater than or equal to the warning threshold At that time, activate the exponential penalty: in This is a hyperparameter for congestion sensitivity.

[0013] As a preferred embodiment of the present invention, the cost function reconstruction and hot update feedback unit sets the dynamic edge weights as follows: Furthermore, through a double buffering mechanism and atomic pointer swapping operation, the shortest path cost that incorporates dynamic edge weights and the expected queuing waiting loss are written into the activity cost matrix buffer for dynamic programming addressing and reading, thus eliminating the risk of read-write concurrency tearing.

[0014] As a preferred embodiment of the present invention, the multi-agent deadlock prevention and cooperative control module includes a narrow channel mutex scheduling and deadlock prevention unit and a local trajectory replanning and order reversal unit. When the narrow channel mutex scheduling and deadlock prevention unit applies a locking control command, it synchronously calls the event-driven injection interface to force the passage cost of the locked segment to a maximum value and triggers an atomic switch. The local trajectory replanning and order reversal unit extracts the mask state of the unfinished picking points of subsequent agents, re-triggers optimization calculation on the activity cost matrix containing the maximum value penalty, and outputs a new route to avoid the locked channel or adjusts the access sequence of unvisited picking targets.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention introduces a connectivity state relaxation mechanism into the search domain of dynamic programming, actively abandoning the blind pursuit of all physical connectivity. It intelligently aggregates microscopic physical nodes located at the edges of the same logical block into macroscopic relaxed superstates, directly preventing the exponential divergence of the number of states as the number of blocks increases. Simultaneously, this invention fully introduces bitmasking technology to represent the access state of the picking subset, completely replacing redundant set mapping operations and transforming massive data queries and updates into low-level binary bitwise operations within a single clock cycle. Furthermore, a hierarchical task decomposition strategy constrains the number of picking points to within a preset threshold, ensuring that while the theoretical time complexity of a single dynamic programming solution is exponential, the actual computational scale is strictly limited to a fixed, engineering-controllable overhead. This core engine's fundamental innovation compresses the number of effective retrieval states in each planning stage by at least two to three orders of magnitude, achieving millisecond-level optimal picking path generation even in environments with extremely limited computing resources. This makes multi-objective dynamic programming truly feasible for low-cost commercial deployment in ultra-large-scale warehouses.

[0016] This invention constructs a dynamic weighted high-frequency penalty and feedback closed-loop mechanism based on a spatiotemporal congestion matrix, internalizing physical congestion and kinematic constraints as first-order continuous variables into the cost transfer term of global dynamic programming. The system can quantify the traffic density of road network segments in real time and derive a continuously smooth dynamic congestion penalty coefficient. When the local density exceeds a critical threshold, the road network edge weights can be updated in milliseconds without restarting coarse-grained programming, immediately triggering lightweight re-entry trajectory calculation. This mechanism can flexibly guide subsequent agents in progress to actively detour around congestion hotspots or temporarily change the order of picking point visits, achieving system-level automatic traffic load balancing. This fundamentally eliminates local queuing interference and systemic deadlock under multi-agent concurrent operations, significantly reducing average order response time and overall resource idle rate. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall architecture of a warehouse management system based on dynamic planning picking paths, provided for an embodiment of the present invention.

[0018] Figure labeling: 100, System overall architecture; 101, Data acquisition and digital twin preprocessing module; 101.1, Multi-dimensional road network topology directed graph construction unit; 101.2, Kinematic perception and order wave analysis unit; 102, Bitmask dynamic programming engine based on state relaxation; 102.1, Bitmask state dimensionality reduction definition unit; 102.2, State space connectivity relaxation and aggregation unit; 102.3, Core dynamic programming state transition execution unit; 102.4, Connectivity restoration and path repair unit; 103, Congestion perception and dynamic weight adjustment module; 103.1, Real-time spatiotemporal flow density monitoring unit; 103.2, Dynamic congestion weight coefficient calculation unit; 103.3, Cost function reconstruction and hot update feedback unit; 104, Multi-agent deadlock prevention and cooperative control module; 104.1, Narrow channel mutex lock scheduling and deadlock prevention unit; 104.2, Local trajectory replanning and order inversion unit. Detailed Implementation

[0019] To make the technical solution, design intent, and achieved technical effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to illustrate the technical solution of the present invention and not to limit it. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the protection scope of the present invention.

[0020] like Figure 1 As shown, this invention provides a warehouse management system based on dynamic planning picking paths, designated System Architecture 100. This system, serving as a purely software-driven and cloud-edge collaborative intelligent warehouse management hub, is deployed on a cloud server cluster with high throughput and concurrent processing capabilities or on localized edge computing nodes. At the physical layer, the system communicates downwards with various heterogeneous execution devices via an industrial-grade wireless network, including autonomous mobile robots, unmanned forklifts, automated guided vehicles, and wearable manual picking terminals. The network protocol used for communication can be an industrial-grade protocol such as 5G private network or Wi-Fi 6, meeting the requirements for low latency and high reliability. Upwards, the system connects and interacts with enterprise-level enterprise resource planning systems or higher-level warehouse management systems via standard application programming interfaces (APIs) for data exchange and command interaction.

[0021] Within the overall system architecture 100, the system is logically divided into four highly decoupled core functional sub-modules that interact via a high-speed memory bus and message queue: data acquisition and digital twin preprocessing module 101, a state-relaxed bitmask dynamic programming engine 102, congestion perception and dynamic weight adjustment module 103, and multi-agent deadlock prevention and collaborative control module 104. These modules maintain clear decoupling at functional boundaries while forming a tight closed-loop linkage at the data flow and control flow levels. The internal operating mechanisms, algorithm models, and logical control flows of each module are explained below.

[0022] See Figure 1 The data acquisition and digital twin preprocessing module 101 serves as a bridge connecting the physical warehousing world and the virtual algorithm space. Its core task is to map the complex physical warehousing environment with high fidelity into a graph theory model that software algorithms can directly traverse and compute, and to complete the parsing and clustering of high-concurrency order tasks. The data acquisition and digital twin preprocessing module 101 contains two collaborative functional units.

[0023] The multidimensional road network topology directed graph construction unit 101.1 in the data acquisition and digital twin preprocessing module 101 actively reads the high-precision design drawings of the physical warehouse, the coordinate data matrix of the shelf space layout, and the access rule configuration table of specific areas during the system initialization phase. The system strictly defines all passable longitudinal aisles, transverse connecting aisles, physical intersections, and key dwelling points within the warehouse as a set of directed vertices in graph theory. A set of directed edges is defined as a continuous spatial line segment that allows legal physical passage between vertices. Based on this, a static directed topological graph model reflecting the physical environment is constructed. This directed graph model is not simply a topological logical connection; each edge... During initialization, it is assigned a static physical distance length calculated based on Euclidean geometry or the Manhattan rule. (here) This characterizes the physical length of a single directed edge (distinct from the shortest path distance between nodes, which will be defined later), while also incorporating the geometric width attribute and theoretical capacity limit of the road segment. The pre-setting of capacity attributes provides an indispensable baseline data foundation for the density calculation of the subsequent congestion perception and dynamic weight adjustment module 103. In an optional implementation, the system can also use a pre-processed full-source shortest path algorithm to offline calculate the basic static shortest path distance matrix between all arbitrary pairs of nodes during the initialization phase, and lock this matrix in the server's high-speed cache segment to avoid redundant pathfinding overhead during runtime.

[0024] The kinematics sensing and order wave analysis unit 101.2 in the data acquisition and digital twin preprocessing module 101 is responsible for listening to and receiving fragmented order pools from the upstream warehouse management system through the application programming interface. Upon receiving the order data stream, the system triggers batch processing logic based on spatial clustering, intelligently merging multiple fine-grained orders that are geographically close and related in category into high-density picking wave tasks. This clustering process comprehensively considers the spatial distribution density of the corresponding storage locations for each product in the order, the rated load capacity of the vehicles, and timeliness constraints, ensuring that each wave task is spatially compact and load-saturated. After wave division, the system queries and accurately matches the three-dimensional coordinates of the real physical storage location corresponding to the unique identifier of each product in the batch from the global product master data table, projecting it losslessly into a directed graph. A set of destination nodes that must be visited ,in The set represents the total number of picking points that a single agent needs to visit for a task assigned to it. For vertex set A subset. To facilitate bitmask index addressing in the subsequent state-relaxed bitmask dynamic programming engine 102, the system synchronously establishes a bijective mapping function. Each binary bit sequence number in the bitmask With topology graph The corresponding physical node identifier in A strict one-to-one correspondence is established. It is particularly noteworthy that when dispatching tasks, the kinematic perception and order wave analysis unit 101.2 binds the current load weight of the robot or device receiving the task to its hardware kinematic parameters. These kinematic parameters include maximum acceleration, full-load braking distance, turning radius, etc., thereby forming a personalized physical constraint model specific to that robot. This step completely abandons the coarse assumption in traditional algorithms that treat all moving bodies as uniformly moving points, enabling the cost function referenced in subsequent path planning to truly reflect the differentiated motion characteristics of different vehicles under different load states.

[0025] Furthermore, after completing wave clustering, the kinematic perception and order wave analysis unit 101.2 is also responsible for assessing the engineering feasibility of the picking point set size and performing necessary grouping. Since the core algorithm of the state-relaxation-based bitmask dynamic programming engine 102 adopts a bitmask dynamic programming paradigm similar to solving the traveling salesman problem, its time complexity is O(n log n). The size of the state space varies with the number of picking points. Present Level growth, belonging to the category of... The exponential time complexity. When Exceeding the system's preset single-time enumerable threshold (This threshold is determined by the available memory capacity of the deployment environment and the target computation latency. On industrial-grade servers equipped with large-capacity memory,) The typical value range is 20 to 25), and the kinematic perception and order wave analysis unit 101.2 will automatically trigger a hierarchical task decomposition strategy. This strategy will decompose the original picking point set. Based on spatial proximity and block affiliation, it is divided into several elements, each with a quantity not exceeding [number missing]. Subtask groups Because the number of picking points in each subtask group is strictly limited to... Within this range, the state space size for a single dynamic programming solution is [not specified]. In engineering terms, it is definite and bounded (with...). For example, the required computation is approximately The basic operations (which can be completed in milliseconds to tens of milliseconds on modern multi-core processors) transform the theoretical exponential complexity into a controllable fixed computational overhead in engineering practice. The state-relaxed bitmask dynamic programming engine 102 independently solves for the optimal access sequence within each subtask group. Subsequently, the kinematic perception and order wave analysis unit 101.2 determines the execution order between subtask groups through a lightweight inter-group connection sorting process, ultimately concatenating the optimal sequences within each group into a complete access path covering all picking points. This hierarchical strategy ensures that regardless of the number of picking points in the original wave, the system can always keep the actual state space size of each dynamic programming solution process within the limits of hardware resources.

[0026] The state-relaxed bitmask dynamic programming engine 102 is the core computational module upon which this system achieves ultra-fast path generation. By deeply reconstructing the state transition logic and storage format of the dynamic programming algorithm, the state-relaxed bitmask dynamic programming engine 102 establishes a low-latency computational paradigm for handling complex multi-block tasks. The state-relaxed bitmask dynamic programming engine 102 internally contains four tightly connected functional units.

[0027] The bitmask state reduction definition unit 102.1 in the state-relaxed bitmask dynamic programming engine 102 addresses the inherent efficiency bottleneck of traditional set theory programming implementations when managing dynamically changing access subsets. Traditional schemes typically rely on advanced data structures such as red-black trees and hash sets to represent and manage constantly changing access subsets consisting of dozens of picking points. These structures suffer from severe memory fragmentation and clock cycle overhead due to node traversal pointer jumps. The bitmask state reduction definition unit 102.1 forcibly compresses the current access state of the subset into a fixed-length bitmask integer supported by native computer hardware. Specifically, the system assigns an unsigned integer variable as a... Its bit width depends on the number of picking points in the current subtask group. (After hierarchical task decomposition by the kinematic perception and order wave analysis unit 101.2,) It has been constrained to not exceed Select within the range, when Use a 32-bit integer when the value is no more than 32. For values ​​exceeding 32, a 64-bit integer is used. If the index in the task is... Picking points (which are on the topology map) The corresponding physical node identifier is ,index From 0 to If an integer variable has already been physically or logically accessed by the agent, then the bitwise access of the integer variable will be terminated using bitwise operations. The bit position is set to 1; otherwise, if the point has not been visited, it remains 0. Based on this, the system defines a highly compact two-dimensional dynamic programming state space structure in memory. ,in The range of values ​​is to , The range of values ​​is , representing the index number in the set of picking points. The total size of this two-dimensional array is . ,exist No more than Under these constraints, its required memory space is within the range of available resources for industrial-grade servers (within...). For example, The memory size is approximately 3.2 gigabytes, which is perfectly feasible for servers equipped with large amounts of memory. The logical semantics are: in the path exploration currently being tracked by the system, the binary representation of the set of pick points that have been visited is: And the index of the last pick point in the exploration sequence is At that time, the system accumulates and calculates the global minimum path cost. This data structure design reduces the state addressing and update operations, which originally required complex object comparisons, to the lowest level of binary operations.

[0028] The state space connectivity relaxation and aggregation unit 102.2 in the bitmask dynamic programming engine 102 based on state relaxation is specifically designed to address the challenge of the rapidly expanding state space in the node tracking dimension of dynamic programming in large multi-block warehouses. In multi-block warehouses containing multiple lateral main channels, if the point-to-point physical connectivity constraints are strictly followed, the number of entry and exit points when traversing different blocks is numerous. In traditional precise dynamic programming modeling, the number of intermediate node entry and exit state combinations that need to be tracked in each solution stage is extremely high. Following the recursive relation ,in The number of blocks, and the initial boundary conditions are: (At block zero, there is only one trivial start and end state.) (For a single block, this corresponds to two traversal directions entering from both ends of the channel). Based on this, it can be deduced that... , , , As can be seen, this sequence exhibits an extremely steep, super-exponential growth trend with the increase in the number of blocks. This explosion in node tracking dimensions, superimposed on the bitmask representation... The individual subset states collectively constitute the overall computational burden of traditional exact dynamic programming schemes, rendering them impractical for large-scale warehouses. It is particularly important to note that the above recursive relation... This describes the node tracking dimension explosion problem in traditional exact dynamic programming, caused by the strict maintenance of cross-block connectivity, where the dimension is independent of the bitmask. Subset dimension.

[0029] To address this problem, the State-Space Connectivity Relaxation and Aggregation Unit 102.2 innovatively implements a connectivity relaxation algorithm. Its core idea is that during the intermediate recursive phase of dynamic programming's cross-block decision-making, the system actively suppresses and shields the fine-grained tracking of the exact entry and exit nodes of the path's physical connectivity. The system introduces a many-to-one dimensionality reduction mapping function. ,in Topology graph This function considers multiple underlying physical nodes located at the edge of a block that have a very similar impact on the macroscopic path decision. At the logical level, the mapping is merged into an abstract relaxed supernode. The system stores the supernode's unique identifier as a hash key in the dynamic programming state space dictionary.

[0030] After partitioning the relaxed supernodes, the system needs to construct a macroscopic transfer cost matrix between them. Since each relaxed supernode contains multiple physical entry and exit nodes, the transfer cost between two different supernodes is no longer a single, fixed value, but depends on the set of path distances between their respective internal physical nodes. To address this issue, the system adopts a minimum strategy to construct the macroscopic cost matrix: for any two relaxed supernodes... and Its macroeconomic transfer cost is defined as ,in Topology graph From physical nodes To physical node The shortest path distance can be obtained directly from the full-source shortest path distance matrix pre-calculated during the system initialization phase and resident in the cache. The reason for adopting the minimum value strategy is that the goal of the macro-stage is to determine the optimal access order between picking points, rather than the precise physical path trajectory; taking the minimum value can ensure that the macro-level transfer cost is a lower bound of the actual reachable cost, so that the access sequence generated by dynamic programming at the macro level has the global optimal approach characteristic; when the connectivity restoration and path repair unit 102.4 completes the physical connection path at the micro level, the actual path cost will increase slightly based on this lower bound, but since the distance difference between physical nodes within a supernode is usually much smaller than the macro-level path cost across blocks, the total cost of the final path can still closely approximate the global optimal solution.

[0031] By executing the aforementioned dimensionality reduction mapping and minimum cost construction logic, the number of effective states that need to be passed backward in the node tracking dimension during each solution stage is reduced from the previously uncontrollable exponential fission with the number of blocks (as mentioned above). And continued to grow rapidly), compressed to The increase in complexity is within the range of magnitude. This complexity bound describes the compression effect on the node tracking dimension; dynamic programming still maintains [stability / preservation] on a subset dimension. The state size, and because The hierarchical task decomposition strategy of the kinematic perception and order wave analysis unit 101.2 is already constrained by the task decomposition strategy. Within this range, the size of the subset dimension is manageable in engineering. The combined effect of the two dimensions brings precise solutions, which were previously completely infeasible in large-scale multi-block repositories due to node tracking dimension explosion, back to a practically computable level. In one specific implementation, the granularity of the relaxed supernode partitioning can be adaptively adjusted according to the block size and picking point density of the repository: when the number of blocks is small, supernodes can be aggregated by channel port; when the number of blocks increases significantly, supernodes can be further aggregated by the entire block, to achieve a balance between path accuracy and computational efficiency.

[0032] The core dynamic programming state transition execution unit 102.3 in the state-relaxed bitmask dynamic programming engine 102 is the engine core that triggers computational power. Before the computation engine starts, the system initializes all state values ​​in the dynamic programming table to maximum values ​​to represent unexplored states. For the agent's initial residence point, if it is in the selection point set... The logical index in is Then the initial state of this starting point is: .

[0033] The state table filling derivation is performed bottom-up using a triple loop mechanism. The outer control loop iterates through all theoretically possible set state mask values ​​in ascending order of numerical value, i.e., iterates from the integer 1 to... This order naturally guarantees that when calculating the current state, the optimal solutions for the preceding states corresponding to all smaller subsets have been absolutely determined. The middle-level control loop iterates through the mask currently being evaluated. In the table, the corresponding binary bit is 1, which represents the valid current picking point index. This indicates the end of a valid sequence. The inner control loop utilizes bitwise operations. Perform a rapid mask comparison to filter out all unvisited target picking points in the current mask whose corresponding bits are still 0. .

[0034] In the innermost layer of the triple loop, the system iteratively updates the optimal substructure by triggering the following core state transition equation: In this equation, and Picking point index and The bijective mapping function established by the kinematic sensing and order wave analysis unit 101.2 After transformation, in the topology graph The corresponding physical node identifier. In the current time sequence Below, from physical nodes To physical node The dynamic integrated path cost is not a fixed constant determined during system initialization, but a dynamic value calculated and injected in real time by the congestion perception and dynamic weight adjustment module 103. Its generation mechanism will be detailed later. It is this design that introduces the state transition process of dynamic programming from the static computation domain to the time-varying dynamic domain, so that the path output by the algorithm naturally carries the perception information of real-time traffic conditions.

[0035] The connectivity restoration and path repair unit 102.4 in the state-relaxed bitmask dynamic programming engine 102 performs physical executability verification and correction on the initial path processed by the relaxation strategy. Since the state-space connectivity relaxation and aggregation unit 102.2 aggregates multiple physical nodes into relaxed supernodes in the intermediate stage, the state transitions completed by dynamic programming at the supernode abstraction level may exhibit connectivity gaps—a lack of direct physical paths between adjacent access nodes—when backtracked and expanded into a sequence of specific physical nodes. This gap is not a logical error in the dynamic programming state transition equation itself, but rather a result of the inherent granularity difference between the relaxation abstraction and the physical instances. Multiple physical entrances and exits within a supernode are considered equivalent in the macro-decision stage, but a unique physical connection path needs to be selected in the micro-execution stage. To bridge this gap and ensure the physical executability of the path, the connectivity restoration and path repair unit 102.4 automatically injects local path repair logic into the algorithm pipeline after the dynamic programming backtracks to generate the initial macro-path array.

[0036] Its working mechanism is as follows: the system follows the generated node sequence, utilizing the original rigorous physical road network diagram. A forward connectivity scan is performed. Once a connectivity gap is detected between adjacent nodes (i.e., no direct, valid directed edge connects the two points), the system triggers an anomaly detection mechanism. At the location of the gap, the system utilizes the original topology graph. Within the local neighborhood space, a constrained shortest path search is performed, inserting missing physically connected path segments into the node sequence. If local repair introduces repeated traversal of already visited nodes, the system automatically detects and eliminates this redundant revisit by re-invoking the core dynamic programming state transition execution unit 102.3 for local secondary calculations, incorporating the repair constraints within the local area. Since this local reconstruction process involves a very small spatial scale and a limited number of nodes, its computation time is negligible in the overall solution cycle. The system repeatedly executes the above scanning and local correction process until the final output selection sequence possesses complete physical continuity.

[0037] The congestion perception and dynamic weight adjustment module 103 endows the dynamic programming system with the ability to globally and adaptively perceive the physical environment, upgrading path planning from static geometric calculation to spatiotemporal dynamic optimization. The congestion perception and dynamic weight adjustment module 103 contains three functional units.

[0038] The real-time spatiotemporal flow density monitoring unit 103.1 in the congestion perception and dynamic weight adjustment module 103 constructs a high-frequency polling guardian thread independent of the main business process. The system collects the physical coordinates of all moving or stationary work entities within the entire warehouse space in a very short time period by parsing the real-time telemetry data stream uploaded by the autonomous mobile robot's built-in positioning system, or the heartbeat data packets continuously sent by manually driven forklifts and personnel's handheld terminals. With the velocity vector of the current motion In a typical implementation, the data acquisition period is set to one frame every 100 milliseconds, which can be adaptively adjusted according to the warehouse's operational density and the system's computing power. The system then displays a pre-built two-dimensional warehouse topology map in the background. Directed edges in The mapping transforms the data into a series of two-dimensional spatial polygon bounding boxes with physical widths. Using computational geometry algorithms to determine if a point lies within a polygon, the system can determine this at any time cross-section. Accurately count the positions located on a specific directed edge The exact number of active agents within the corresponding physical boundary. .

[0039] The dynamic congestion weight coefficient calculation unit 103.2 in the congestion perception and dynamic weight adjustment module 103, combined with the maximum physical design capacity preset for each edge in the multi-dimensional road network topology directed graph construction unit 101.1, is used in this module. The dynamic congestion index coefficient of the directed edge is continuously calculated in each polling cycle and is defined as follows: Based on this continuously fluctuating density index, the system derives and calculates the penalty weight coefficient representing the current road surface condition using a continuous nonlinear mapping function. To ensure the numerical continuity and smooth transition characteristics of the penalty function in the congestion critical region, and to avoid frequent oscillations and jumps in the vehicle path caused by a step change at the threshold, its specific mapping calculation logic is set as follows: when The value is less than the warning threshold parameter set by the system. When the system determines that the tunnel is currently unobstructed, no penalty is applied, and the setting is... ;when The value is greater than or equal to the warning threshold. At that time, the system determined that the alleyway showed signs of congestion and immediately activated the penalty calculation: The above formula is in The value at that location is This precisely aligns with the fixed value of 1.0 below the threshold, ensuring the continuity of the penalty function across the entire domain. When the congestion index continues to rise and far exceeds the threshold, the exponential growth characteristic causes the penalty coefficient to amplify rapidly, reflecting the system's strong aversion to severe congestion. The constants in the formula... This is a congestion sensitivity hyperparameter that can be dynamically assigned via a configuration file, used to control the algorithm's responsiveness to congestion. In a typical implementation configuration, the warning threshold... Set to 0.8, hyperparameter The initial value can be set to 5.0. Maintenance personnel can calibrate and fine-tune this parameter based on the actual warehouse aisle width, vehicle size, and historical congestion frequency data.

[0040] The cost function reconstruction and hot update feedback unit 103.3 in the congestion perception and dynamic weight adjustment module 103 performs low-level logical reconstruction of the comprehensive path cost between nodes called by the state transition of the bitmask-based dynamic programming engine 102 based on state relaxation. In traditional static path planning, the transition cost between any two picking points is usually taken from the topology graph. The pre-calculated static shortest path distance remains constant throughout system operation. The key change achieved by the cost function reconstruction and hot update feedback unit 103.3 lies in: [the following text appears to be incomplete and requires further context: "topology graph..."] Each directed edge The cost of passage, from its static physical distance Reconstructed into dynamic edge weights that incorporate real-time congestion penalties The calculation formula is as follows: Based on this, any two picking point nodes and Comprehensive path cost between Defined as in On a dynamically weighted graph with edge weights, from node To the node The shortest path cost is obtained, and then the estimated queuing waiting cost based on the agent's kinematic parameters is added: in Indicates from node To the node The set of all legal paths is summed and all directed edges along that path are traversed. This definition allows for the dynamic adjustment of edge weights due to congestion in the road network. When the congestion increases sharply, the shortest path solution will automatically bypass the congested segment and select an alternative path that is physically farther away but has a lower congestion penalty, so that the cost between nodes can truly reflect the actual travel cost under the current spatiotemporal conditions. This is based on the physical kinematic parameters of the agent and the estimated queue length along the route calculated in the kinematic perception and order wave analysis unit 101.2. The general calculation logic is as follows: First, the system estimates the expected travel time of each segment on the shortest path from node i to node j based on the agent's current load state and its kinematic parameters (maximum speed, acceleration, braking distance), i.e., the dynamically weighted length of each segment divided by the estimated average travel speed of the agent under the corresponding load; then, for each congested segment along the path, the expected queuing waiting time is calculated by multiplying the currently detected number of queuing agents by the estimated average travel time of a single agent. At the engineering implementation level, due to the number of picking points... Finite and topological graph The scale is determined during warehouse design. The cost function reconstruction and hot update feedback unit 103.3 can recalculate only the local paths affected by the congestion weight changes in each update cycle using the incremental shortest path algorithm, instead of performing a full-source shortest path search of the entire graph every time, thereby keeping the update time within the order of magnitude that matches the polling cycle.

[0041] At the end of each timed calculation cycle, the cost function reconstruction and hot update feedback unit 103.3 uses a double-buffering mechanism to write the updated cost matrix into the high-speed shared memory area. Specifically, the system pre-allocates two cost matrix buffers of equal size in memory. The cost function reconstruction and hot update feedback unit 103.3 always writes the latest data to the currently inactive backup buffer. After writing, it switches the backup buffer to the active buffer through an atomic pointer swap operation, while releasing the original active buffer as the target for the next round of writing. When performing state transition addressing, the core dynamic programming state transition execution unit 102.3 always reads cost data from the currently active buffer. Since the reading end and the writing end operate on different physical memory areas at any given time, this mechanism completely eliminates the risk of data tearing caused by read-write concurrency without the need for mutex locks, ensuring that the cost matrix read by the core dynamic programming state transition execution unit 102.3 is always a logically complete and self-consistent snapshot.

[0042] In addition to the regular update channel driven by the aforementioned timed period, the cost function reconstruction and hot update feedback unit 103.3 also exposes an event-driven synchronization injection interface, which is specifically for the multi-agent deadlock prevention and cooperative control module 104 to call immediately when it detects sudden events such as mutex lock conflicts. When the narrow channel mutex lock scheduling and deadlock prevention unit 104.1 activates the mutex lock of a restricted edge, at the same logical moment when the locking control command is applied, the narrow channel mutex lock scheduling and deadlock prevention unit 104.1 synchronously calls the event-driven injection interface of the cost function reconstruction and hot update feedback unit 103.3, forcibly sets the penalty cost of the locked edge segment to a maximum value close to infinity, and immediately triggers an additional write and atomic switch operation of the buffer, so that the maximum value penalty takes effect instantly in the active buffer. This event-driven injection mechanism completely bypasses the 100-millisecond timing period, ensuring that the cost matrix read by any subsequent agent when triggering replanning necessarily contains the maximum penalty information of the locked edge segment, thereby completely eliminating the risk of timing breakage that may occur between the mutex event and the asynchronous refresh of the cost matrix.

[0043] Through the dual-channel collaborative mechanism of conventional timed updates and event-driven instant injection, when a road segment in a severely congested state or a segment blocked by a mutex lock is introduced into the dynamic programming recursive equation for evaluation, the path cost of passing through the road segment will be much higher than that of the alternative route that is physically farther but unobstructed. This directly induces the intelligent agent cluster to actively divert traffic and avoid congestion at the algorithm principle level.

[0044] To more intuitively demonstrate the actual effect of the aforementioned cost function reconstruction mechanism in system operation, taking the warehouse road network state at a certain time point as an example, the following derivation and analysis of the real-time mapping between congestion penalty weight and state transition cost is given. Assume the system detects four representative arc segments at a certain moment, and the congestion sensitivity hyperparameter... Set to 5.0, warning threshold The value is set to 0.8. The static physical distance of arc segment Aisle-A1 is 15.0 meters. Currently, one vehicle is detected moving within this area, while the initial saturation capacity of this road segment is set to 4 vehicles. The congestion index is then calculated. If this value is less than the warning threshold of 0.8, the system determines the condition to be unobstructed and sets [the appropriate parameters]. The dynamic edge weight of this edge segment is The static distance of arc segment Aisle-B2 is 20.0 meters. Two vehicles were detected, and the capacity threshold is 4 vehicles. Still below the warning level Dynamic edge weights are While the static distance of the cross passage (Cross-C1) is only 12.0 meters, it detected the movement of 5 vehicles, which is also the capacity threshold of 5 vehicles. The warning threshold has been exceeded, and penalty calculation has been activated. Dynamic edge weights rise to The static distance of the Aisle-D4 arc segment was only 10.0 meters, yet it detected three vehicles moving in a narrow passage with a capacity of only two vehicles. Severe overloading Dynamic edge weights surged to Therefore, it can be seen that Cross-C1 and Aisle-D4, which are originally the shortest in terms of physical distance at a purely geometric level, experience a sharp increase in congestion penalty weight coefficients due to the exponential amplification caused by the instantaneous influx of traffic. This leads to a dramatic inflation of the overall cost of paths passing through these segments on the dynamic weighted graph. When the core dynamic programming state transition execution unit 102.3 calculates the shortest path cost between picking points on this dynamic weighted graph, the globally optimal decision will automatically shift the path towards Aisle-A1 or Aisle-B2, which have ample free-flowing redundancy capacity.

[0045] The multi-agent deadlock prevention and cooperative control module 104 addresses the physical conflicts and deadlock problems that may arise when multiple agents work concurrently in a confined space. The multi-agent deadlock prevention and cooperative control module 104 internally contains two cooperative functional units.

[0046] The narrow-channel mutex scheduling and deadlock prevention unit 104.1 in the multi-agent deadlock prevention and cooperative control module 104 specifically addresses the deadlock risk in single-lane warehouse layouts with numerous extremely narrow channels or where parallel passing between two vehicles is impossible. The narrow-channel mutex scheduling and deadlock prevention unit 104.1 works by pre-marking edges in the graph model with restricted or narrow attributes. A mutex mechanism is introduced to prevent deadlock. When a certain agent in the system... The path is planned by the state-relaxed bitmask dynamic programming engine 102 and is about to cross the boundary in the physical world to enter the restricted edge. At that time, the narrow channel mutex scheduling and deadlock prevention unit 104.1 is immediately awakened, and the system checks the edge in the logical network. and the opposite side A high-priority state locking control command is applied synchronously at the global level. Simultaneously, the narrow-channel mutex scheduling and deadlock prevention unit 104.1 synchronously calls the event-driven synchronization injection interface exposed by the cost function reconstruction and hot update feedback unit 103.3, forcibly setting the passage cost of the locked edge segment and its reverse edge to a near-infinite maximum value, and immediately triggering an additional atomic switch in the buffer, ensuring that the maximum value penalty is reflected in the activity cost matrix at the same logical moment the locking command takes effect. This locking state will persist until the agent... Once the restricted road segment is completely traversed and the occupied space is released, the narrow passage mutex scheduling and deadlock prevention unit 104.1 will again call the event-driven injection interface to restore the penalty cost of that segment to the normal dynamically calculated value. In warehouse layouts with many restricted edges, the granularity of the mutex can be further refined to the segmented intervals within the road segment to reduce the impact of locking on the overall system throughput efficiency.

[0047] The local trajectory replanning and order reversal unit 104.2 in the multi-agent deadlock prevention and cooperative control module 104 handles the path adjustment problem of subsequent agents during the activation of the mutex lock. When another independent agent subsequently enters the system... When the underlying dynamic programming solution process attempts to traverse a locked channel, because the narrow channel mutex scheduling and deadlock prevention unit 104.1 has set the penalty cost of the edge segment to a maximum value through the event-driven injection interface at the same time as applying the lock, the cost matrix in the active buffer necessarily contains the locking information. The bitmask dynamic programming engine 102 based on state relaxation then... When re-executing the state transition solution for the remaining unvisited picking point set, all route schemes requiring passage through the locked passage will be automatically excluded mathematically. Specifically, the local trajectory replanning and order inversion unit 104.2 extracts the agent. The current task mask state, i.e., the set of unfinished picking points corresponding to the bits that have not yet been set to 1, is represented by the agent. The node corresponding to the current physical location is taken as the new starting point. On the latest cost matrix, which already includes locking penalties, the solution process of the core dynamic programming state transition execution unit 102.3 is retried. Due to the extremely low computational latency characteristics provided by the bitmask architecture, this replanning process can be completed within milliseconds. The output of the replanning may manifest in two forms at the physical execution level: firstly, if the topology graph... If a detour exists, the algorithm will automatically output an alternative physical route to avoid the locked passage; secondly, if no physical detour is possible due to warehouse terrain limitations, the algorithm will naturally adjust the access order of unvisited picking targets during the solution process. Taking a specific scenario as an example, if the agent... The original plan was to visit picking point A, picking point B, and picking point C in sequence, but the only path leading to picking point A was disrupted by the intelligent agent. When a location is locked and its path cost approaches infinity, the core dynamic programming state transition execution unit 102.3, when resolving the remaining unvisited set, automatically excludes the option to visit picking point A first by comparing the minimum values ​​of the state transition equations. Instead, it outputs a new sequence to proceed to picking point B or picking point C first, and then returns to visit picking point A after the lock is released. Throughout this process, the meaning of the bitmask remains consistent: a bit value of 0 represents unvisited, and a bit value of 1 represents visited. The system changes the optimal access order output by the dynamic programming solution, rather than semantically altering the bitmask.

[0048] The following explains the complete working principle of this system, namely the full-process control logic from order receipt to the completion of a complete warehouse physical delivery lifecycle. This process is strictly executed in a five-stage sequential closed loop.

[0049] During the global environment digital twin initialization and task access phase, before the official start of each day or each large-scale logistics operation wave, the core thread of the warehouse management system loads and wakes up the data acquisition and digital twin preprocessing module 101. After the multi-dimensional road network topology directed graph construction unit 101.1 starts, the system reads the preset configuration file and transforms the physical warehouse environment into a weighted directed topology graph model through graph theory modeling. During this process, the system preloads static shelf physical dimensions, aisle boundary attributes, and the maximum passage capacity of each segment. It then offline calculates the basic static shortest path distance matrix between any two physical nodes using a pre-processed full-source shortest path algorithm, locking this matrix permanently in the server's cache. The raw order data stream continuously generated by the upper-level warehouse management system, containing product demand details and location coordinates, is pushed into the buffer queue of the kinematics perception and order wave analysis unit 101.2. Within the kinematics perception and order wave analysis unit 101.2, the system triggers clustering and batch optimization logic, intelligently reorganizing fragmented orders into multiple high-density picking wave tasks and establishing a bijective mapping relationship between picking point indices and topology node identifiers. If the total number of picking points in a certain wave of tasks exceeds the system's preset single-batch enumerable threshold... The kinematic perception and order wave analysis unit 101.2 automatically further divides it into several elements with a number not exceeding [a certain threshold]. The system precisely binds and distributes the storage coordinate set and mapping table corresponding to each subtask group to the task stack of the corresponding target agent.

[0050] During the dynamic initialization and dimensionality reduction construction phase of the relaxed mask state space, for each active agent within the system assigned a picking point sub-task group, the bitmask state dimensionality reduction definition unit 102.1 of the state-relaxed bitmask dynamic programming engine 102 immediately intervenes. The system automatically extracts the physical coordinate point data of each sub-task group and maps them in the logical memory space according to the mapping function. Strictly mapped to linear integer index values, numbered consecutively from 0 to... The system initializes and sets the binary bitmask representation register for the current task sequence. Quickly allocate contiguous two-dimensional array memory space in the heap area. Its total scale is No more than Under the constraints, it always remains within the hardware's carrying capacity. Simultaneously, the background process of the state space connectivity relaxation and aggregation unit 102.2 starts running. Based on the block layout scale parameters faced by the current system, it automatically triggers the state space relaxation and aggregation algorithm, establishes a bidirectional hash key-value mapping table between the underlying physical nodes and the corresponding abstract relaxation supernodes, and constructs the macroscopic transfer cost matrix between relaxation supernodes based on the minimum value strategy using the full-source shortest path distance matrix.

[0051] During the real-time acquisition and feedback perturbation phase of the spatiotemporal cost road network, this phase is not a one-off action, but rather a persistent daemon process that resides in the system background, operating asynchronously and in parallel with the main dynamic programming thread and possessing the highest system preemption priority. The congestion perception and dynamic weight adjustment module 103, triggered by a timer, scans the entire frequency band of all online agents using a fixed and extremely high-frequency system heartbeat to obtain the currently uploaded physical location coordinates and instantaneous velocity vectors. Direction. The real-time spatiotemporal flow density monitoring unit 103.1 uses computational geometry algorithms to project discrete point data onto a topological graph network, continuously refreshing each directed edge in the network graph. Real-time dynamic traffic values ​​carried internally The dynamic congestion weighting coefficient calculation unit 103.2 extracts the theoretical physical saturation capacity limit data of each lane and calculates the current congestion index. The penalty amplification factor is calculated using a continuous exponential penalty function. At the end of each clock cycle, the cost function reconstruction and hot update feedback unit 103.3 is based on the latest dynamic edge weights. The shortest path cost between affected node pairs is incrementally recalculated, and the latest cost matrix, which integrates the underlying physical distance and dynamic congestion penalties, is written into the backup page of the double buffer. Then, an atomic pointer swap is used to switch it to the active page, allowing the core dynamic programming state transition execution unit 102.3 to safely read it during subsequent state transition addressing. Furthermore, when the multi-agent deadlock prevention and cooperative control module 104 submits an edge locking or unlocking request through the event-driven injection interface, the cost function reconstruction and hot update feedback unit 103.3 will perform an additional instant buffer write and atomic switch outside of the regular timing period, ensuring that the locking penalty information takes effect in the active cost matrix with zero latency.

[0052] In the dynamic programming optimization stage based on the relaxed bitmask model, when the target intelligent vehicle is in a standby state at the sorting and distribution station, or has just completed operations in a certain area and is about to cross the boundary into the next sorting zone, the core dynamic programming state transition execution unit 102.3 receives an interrupt request and is triggered to wake up. The system mobilizes the multi-core computing power pool and performs transition and convergence derivation of all state sets by executing binary logic bit operations. After the system program enters the loop, it uses bit operation statements... To determine the target picking point index Whether it has not been accessed yet; after the condition is met, the system core thread retrieves the most recently switched page from the currently active page of the double buffer, which was reconstructed by the cost function and hot update feedback unit 103.3. Numerical values; then, strictly following the state transition equation, updates are performed. The minimum cumulative cost value. Throughout the solution process, since the core set comparison and lookup operations are converted into binary shifts and bitwise logic operations performed at the cache register level, and the relaxation strategy of the state space connectivity relaxation and aggregation unit 102.2 shields invalid cross-block path expansion operations, even under extreme test cases, the state-relaxed bitmask dynamic programming engine 102 can still ensure that the logical derivation of the full-dimensionality reduction dynamic programming array is completed within a few milliseconds. The connectivity restoration and path repair unit 102.4 then performs connectivity gap scanning and path repair verification on the obtained optimal solution set. After confirming that the physical continuity is correct, it formally encapsulates and outputs it as an optimal picking task access sequence based on the current global spatiotemporal congestion state. If the current agent's complete wave task contains multiple sub-task groups, after each sub-task group has been solved, the kinematic perception and order wave analysis unit 101.2 performs inter-group connection sorting, splicing the optimal sequences within each group into a complete access path covering all picking points.

[0053] During the phase of driving and real-time fine-tuning correction by the underlying intelligent agent, the planned picking access sequence and the set of physical movement trajectory instructions mapped to it are sent by the system's network scheduling layer to the underlying control terminal of the corresponding intelligent agent through an encrypted channel based on a reliable transmission protocol. After receiving the instruction message, the intelligent agent's motor and steering device start smooth movement according to the given sequence and trajectory coordinates. During the life cycle of the intelligent agent's movement, the multi-agent deadlock prevention and cooperative control module 104 continuously scans and monitors the unobstructed status of the physical channel within a certain range ahead of its travel path. Assuming that a sudden and unpredictable intersection event occurs during operation, such as a manually driven forklift not directly controlled by the system mistakenly entering a channel set for one-way travel and causing a blockage, the narrow channel mutex lock scheduling and deadlock prevention unit 104.1 will detect the physical conflict of the corresponding mutex lock and immediately report a traffic interruption alarm to the main gateway. At the same logical moment, it will write the maximum penalty of the locked segment into the activity cost matrix through the event-driven injection interface. Upon receiving an alarm, the system does not need to perform a global reset operation to clear the agent's original complete task queue. Instead, while preserving the original task framework, the local trajectory replanning and order reversal unit 104.2 extracts the mask state of the agent's remaining unfinished tasks and, with the highest priority, re-triggers the core dynamic programming state transition execution unit 102.3 to solve the problem on the latest cost matrix, which now includes a maximum value penalty. Because the state compression mechanism of the state-relaxed bitmask dynamic programming engine 102 gives the system extremely low computational latency, even intelligent vehicles on the verge of stopping can successfully receive the new alternative solutions recalculated by the system using the updated cost matrix within a few hundred milliseconds before their heavy-load braking action has fully stabilized. At the physical execution level, this solution may manifest as the intelligent vehicle initiating a turning logic on the spot, turning around to abandon the congested passage ahead and autonomously finding a way to an alternative detour passage; in more complex scenarios, it may manifest as the intelligent vehicle temporarily suspending the picking action of the target product on the nearest shelf, and instead, based on the new access order output by the replanning, prioritizing the next target point in an open and unobstructed area.

[0054] The aforementioned coordinated process of action perception, penalty feedback, instantaneous recalculation, instruction issuance, and underlying error correction execution occurs continuously and cyclically throughout the entire storage space at a frequency of tens of times per second, until the underlying variables held by the agent are reached. All binary status bits were logically set to 1 by the system, physically corresponding to the completion of all picking tasks it carried. The intelligent vehicle was then guided back to the distribution and shipping waiting area, marking the completion of a closed-loop operation for this entity.

[0055] Based on the above embodiments, those skilled in the art will understand that the specific implementation of the present invention is not limited to the above content. The system deployment can flexibly adopt various architecture modes such as fully cloud-based centralized computing, cloud-edge collaborative computing, or fully edge-distributed computing. The bit width of the bitmask can be selected as 32-bit or 64-bit integer depending on the actual subtask group size. The single-enumerable threshold... The aggregation granularity of the relaxed supernodes can be adjusted according to the hardware configuration of the deployment environment. The aggregation granularity of the relaxed supernodes can be adaptively configured based on the block topology of the repository. The strategy for determining the macroscopic transfer cost between relaxed supernodes is not limited to a minimum value strategy; a weighted average or other reasonable representative metric can also be used. The mathematical form of the congestion penalty function is not limited to an exponential function; a polynomial function, a piecewise linear function, or other monotonically increasing nonlinear mapping function that remains continuous at a threshold can also be used. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A warehouse management system based on dynamic programming picking paths, characterized in that, include: The data acquisition and digital twin preprocessing module is used to map the physical warehousing environment into a topological directed graph model, receive order data streams, perform order clustering and kinematic parameter binding to generate a picking point set, and establish the bijective mapping relationship between topological graph nodes and bitmask indices. A state-relaxed bitmask dynamic programming engine is connected to the data acquisition and digital twin preprocessing module. It is used to compress the dynamic picking subset into a bitmask state and perform connectivity state relaxation in the search domain. Multiple physical in and out nodes are merged into relaxed supernodes. Then, the state transition equation is executed using low-level bit operations. Under the constraint that the number of picking points is limited by a preset threshold, the globally optimal macroscopic picking sequence is derived with controllable computational overhead. The congestion perception and dynamic weight adjustment module is connected to the state relaxation-based bitmask dynamic programming engine. It is used to collect the coordinates and velocity vectors of the working entity in real time to calculate the spatiotemporal flow density, generate a dynamic congestion index based on the capacity limit, and reconstruct the dynamic edge weight matrix that integrates real-time traffic density through a continuous nonlinear mapping function. The dynamic edge weight matrix is ​​then injected into the state transition cost of the dynamic programming engine at high frequency. The multi-agent deadlock prevention and collaborative control module is connected to the state-relaxation-based bitmask dynamic programming engine and the congestion perception and dynamic weight adjustment module, respectively. It is used to apply mutex lock control commands in real time on restricted road sections, synchronously trigger the cost hot update mechanism to inject maximum value penalties, and induce subsequent agents to perform local trajectory replanning and order reversal while retaining the original unfinished task framework.

2. The warehouse management system based on dynamic programming picking paths according to claim 1, characterized in that: The data acquisition and digital twin preprocessing module includes a multi-dimensional road network topology directed graph construction unit and a kinematics perception and order wave analysis unit. The multi-dimensional road network topology directed graph construction unit is used to construct a static directed topology graph model and assign physical distance length, geometric width attributes, and theoretical capacity limits to each directed edge. The kinematics perception and order wave analysis unit is used to establish a bijective mapping function. The binary bit sequence number in the bitmask is matched one-to-one with the corresponding physical node identifier in the topology diagram, and the hardware kinematic parameters of the device undertaking the task are bound together.

3. A warehouse management system based on dynamic programming picking paths according to claim 2, characterized in that: The kinematic sensing and order wave analysis unit is also configured to: when the number of picking points in a single wave... Exceeding the system's preset single-time enumerable threshold When this occurs, a hierarchical task decomposition strategy is triggered, dividing the original set of picking points into several groups based on spatial proximity and block affiliation, with each group containing no more than [number of elements]. The subtask groups are divided into subtask groups. Each subtask group independently solves the optimal access sequence and then completes the splicing through lightweight inter-group connection sorting.

4. A warehouse management system based on dynamic programming picking paths according to claim 1, characterized in that: The bitmask dynamic programming engine based on state relaxation includes: a bitmask state dimensionality reduction definition unit, a state space connectivity relaxation and aggregation unit, a core dynamic programming state transition execution unit, and a connectivity restoration and path repair unit.

5. A warehouse management system based on dynamic programming picking paths according to claim 4, characterized in that: The state space connectivity relaxation and aggregation unit introduces a many-to-one dimensionality reduction mapping function. This will connect multiple physical nodes located at the edge of the same logical block that are equivalent in terms of macroscopic path direction decision-making. At the logical level, they are merged and mapped into abstract relaxed supernodes. The unique identifier is stored as a hash key in the dynamic programming dictionary to prevent the exponential explosion of node dimensions; the transfer cost between the relaxed supernodes is the minimum of the shortest path costs between the merged physical nodes as the macroscopic transfer cost between the supernodes.

6. A warehouse management system based on dynamic programming picking paths according to claim 4, characterized in that: The core dynamic programming state transition execution unit employs a bottom-up triple loop mechanism for optimal substructure iterative updates, using bitwise operations to filter effective pick points and unvisited targets to be explored. The core state transition equation it triggers is: in, This is the current mask. To ensure the current picking point index is valid, Index unvisited points for the target to be explored. and This is the identifier of the corresponding physical node in the topology graph. This refers to the dynamic integrated path cost.

7. A warehouse management system based on dynamic programming picking paths according to claim 4, characterized in that: The connectivity restoration and path repair unit is configured to perform a constrained shortest path search in the local neighborhood space along the generated preliminary macroscopic path array, physically connect and complete the connectivity gaps caused by relaxation aggregation, and recalculate dynamic programming to eliminate redundant back-visits in the local area.

8. A warehouse management system based on dynamic programming picking paths according to claim 1, characterized in that: The congestion perception and dynamic weight adjustment module includes a real-time spatiotemporal flow density monitoring unit, a dynamic congestion weight coefficient calculation unit, and a cost function reconstruction and hot update feedback unit. The dynamic congestion weight coefficient calculation unit calculates the congestion index coefficient based on the number of real-time active operation entities on the directed edge and the theoretical capacity. The penalty weight coefficient is calculated using a continuous nonlinear mapping function. Its mapping logic is: when Greater than or equal to the warning threshold At that time, activate the exponential penalty: in This is a hyperparameter for congestion sensitivity.

9. A warehouse management system based on dynamic programming picking paths according to claim 8, characterized in that: The cost function reconstruction and hot update feedback unit sets the dynamic edge weights as follows: Furthermore, through a double buffering mechanism and atomic pointer swapping operation, the shortest path cost that incorporates dynamic edge weights and the expected queuing waiting loss are written into the activity cost matrix buffer for dynamic programming addressing and reading, thus eliminating the risk of read-write concurrency tearing.

10. A warehouse management system based on dynamic programming picking paths according to claim 1, characterized in that: The multi-agent deadlock prevention and cooperative control module includes a narrow channel mutex scheduling and deadlock prevention unit and a local trajectory replanning and order reversal unit. When the narrow channel mutex scheduling and deadlock prevention unit applies a locking control command, it synchronously calls the event-driven injection interface to force the passage cost of the locked segment to a maximum value and triggers an atomic switch. The local trajectory replanning and order reversal unit extracts the mask state of the unfinished picking points of subsequent agents, re-triggers optimization calculation on the activity cost matrix containing the maximum value penalty, and outputs a new route to avoid the locked channel or adjusts the access sequence of unvisited picking targets.

Images