Warehousing unmanned distribution autonomous navigation scheduling method and system
By employing a hierarchical collaborative architecture and hybrid heuristic algorithms, combined with a global spatiotemporal reservation map and reinforcement learning, the path conflict and deadlock problems in the unmanned warehouse picking system were solved, enabling efficient multi-robot collaborative operation and dynamic environment adaptation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 中环低碳节能技术(北京)有限公司
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-12
Smart Images
Figure CN122022684A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of intelligent warehousing and logistics automation technology, specifically relating to a method and system for autonomous navigation and scheduling of unmanned warehouse picking. Background Technology
[0002] With the rapid development of e-commerce and intelligent manufacturing, warehousing and logistics automation has become a key link in improving supply chain efficiency. Traditional warehousing operations rely heavily on manual labor, facing challenges such as rising labor costs, unstable operational accuracy, and difficulty in continuously improving operational efficiency. In recent years, unmanned warehousing systems based on mobile robots have gradually become a research hotspot and an industrialization direction. By using autonomously navigated AGVs / AMRs (Automated Guided Vehicles / Autonomous Mobile Robots) to achieve automatic handling and sorting of goods, the automation level of warehousing operations has been significantly improved.
[0003] However, existing unmanned warehouse picking systems still face many technical bottlenecks in practical applications: First, in high-density, dynamically changing warehouse environments, path conflicts and deadlocks are prone to occur when multiple robots work together. Traditional scheduling strategies based on fixed paths or simple avoidance are difficult to guarantee the overall efficiency of the system. Moreover, existing scheduling algorithms mostly adopt a single optimization objective (such as the shortest path) and fail to fully consider multi-dimensional factors such as task timeliness, energy consumption balance and system load balancing, resulting in limited overall operating efficiency.
[0004] In navigation technology, traditional methods often employ extreme architectures of centralized or fully distributed systems. The former carries an excessive computational burden and the risk of single points of failure, while the latter struggles to guarantee global optimization. Furthermore, environmental modeling often relies on static maps, which cannot effectively support dynamic obstacle handling and temporary path adjustments. At the scheduling algorithm level, heuristic algorithms (such as A* and Dijkstra's algorithm) are computationally inefficient in large-scale multi-robot scenarios, while metaheuristic algorithms (such as genetic algorithms and particle swarm optimization) frequently suffer from local optima and slow convergence speeds.
[0005] Therefore, there is an urgent need for a warehouse unmanned picking, navigation and scheduling method and system that combines global optimization and local adaptability, and is capable of autonomous learning and adjustment, in order to cope with the complex and ever-changing needs of modern logistics warehousing. Summary of the Invention
[0006] This application provides a method and system for autonomous navigation and scheduling of unmanned warehouse picking, aiming to solve the problems of low efficiency and limited overall operating efficiency of existing technologies.
[0007] Firstly, a method for autonomous navigation and scheduling of unmanned warehouse order fulfillment, the method comprising:
[0008] S1. Warehouse environment modeling and task parsing: Construct a reconfigurable gridded warehouse map, parse logistics tasks into a triple sequence of start point-operation point-end point, and generate a dynamic task queue.
[0009] S2. Layered collaborative task scheduling: The upper-level central scheduler performs global task allocation and coarse-grained path planning based on a multi-objective optimization model and a hybrid heuristic algorithm. The lower-level distributed autonomous navigation units deployed in each robot receive instructions and are responsible for local perception, planning and feedback.
[0010] S3. Spatiotemporal reservation map construction and conflict pre-resolution: generate spatiotemporal trajectories for the planned path, construct and maintain a global spatiotemporal reservation map to record the time occupancy status of each spatial grid, perform conflict detection during the planning stage and resolve conflicts using a priority-based negotiation adjustment strategy.
[0011] S4. Dynamic execution and real-time monitoring: The robot executes tasks according to the scheduled path and reports the status in real time. The central system monitors the overall progress and responds dynamically to delays, faults and new task events. A rolling optimization window is used to periodically reschedule and optimize the path of unexecuted tasks.
[0012] S5. Multi-objective performance evaluation and adaptive parameter tuning: The system collects multi-dimensional performance indicators during runtime and adaptively adjusts the cost function weights and internal parameters of the scheduling model based on the reinforcement learning mechanism.
[0013] Optionally, in step S1, constructing a reconfigurable rasterized warehouse map specifically includes:
[0014] A global coordinate system is established based on the physical layout of the warehouse, and the warehouse planar area is discretized into a uniform set of grid cells;
[0015] Assign a type attribute to each grid cell. The type attribute includes shelf occupancy cell, two-way passage cell, one-way passage cell, inlet cell, outlet cell, charging station cell, and static obstacle cell.
[0016] The status information of each grid cell is stored, and the status information includes at least the cell type, coordinates, passage cost, and real-time occupancy flag.
[0017] Optionally, in step S2, when the upper-level central scheduler performs global task allocation, it uses the following weighted comprehensive cost function for optimization: in, As a time cost, its calculation is based on the estimated execution time of the task and the estimated completion time of the robot's current task queue; The path length cost is obtained by pre-calculating or estimating the shortest path distance from the starting point to the operation point and from the operation point to the destination. The energy consumption cost is estimated based on path distance, estimated number of turns, and acceleration / deceleration model. Estimate the costs of conflict.
[0018] Optionally, the hybrid heuristic algorithm combines the whale optimization algorithm with a greedy strategy, including:
[0019] During the initialization phase, an initial task sequence containing the tasks to be assigned is randomly generated for each robot as the position vector of the individual whale.
[0020] During the whale optimization phase, the individual position is updated by simulating the encirclement, bubble net attack and random search behavior of humpback whales, and the weighted comprehensive cost is used as the fitness function for global exploration.
[0021] In the local optimization phase of the greedy strategy, the optimal individual and its neighboring individuals are locally refined by traversing the task sequence and trying to swap or insert tasks until the weighted comprehensive cost can no longer be reduced.
[0022] Optionally, in S3, the resolution using a priority-based negotiation adjustment strategy specifically includes: assigning dynamic priorities to the conflicting robot task pairs according to preset rules; maintaining the original path reservation of the high-priority robot unchanged; and adjusting the low-priority robot to attempt local spatial detours or insert waiting time windows.
[0023] If the adjustment fails, it will trigger a rescheduling of the conflicting tasks.
[0024] Optionally, in step S4, the scrolling optimization window mechanism specifically includes:
[0025] Trigger the optimization process using a fixed time period or event period;
[0026] Focus on a subset of all tasks that are in a "waiting" state, and combine the current position and state of all robots as new initial conditions;
[0027] The hybrid heuristic algorithm is used to redistribute and plan paths for the subset of unexecuted tasks, and a new scheme that is superior to the original plan is adopted.
[0028] Optionally, in step S5, the adaptive parameter tuning mechanism based on reinforcement learning includes:
[0029] The parameter optimization process of the scheduling system is modeled as a sequential decision problem, where the state is the performance index vector within the statistical period, the action is the adjustment of the key parameters of the scheduling algorithm, and the reward is the comprehensive feedback value based on multi-objective performance calculation.
[0030] The adaptive parameter tuning agent outputs new parameter configurations based on the system's operating status and is trained using deep reinforcement learning algorithms to achieve dynamic and personalized parameter optimization.
[0031] Secondly, a warehouse unmanned delivery and autonomous navigation scheduling system includes:
[0032] The environmental modeling and management module builds and maintains a reconfigurable gridded warehouse map, a shelving information database, and a central dynamic task queue.
[0033] The central collaborative scheduling module performs global task allocation and coarse-grained path planning based on a multi-objective optimization model and a hybrid heuristic algorithm, and integrates conflict pre-detection functionality.
[0034] The distributed autonomous navigation module is deployed in each mobile robot, receives global commands, performs real-time obstacle avoidance and path replanning based on local perception, and provides feedback on the execution status.
[0035] The spatiotemporal reservation map management module builds and maintains a global spatiotemporal reservation map, recording the time occupancy status of each spatial grid.
[0036] The monitoring and adaptive optimization module monitors the system's operating status in real time, identifies abnormal execution events, and adaptively adjusts the parameters of the scheduling model based on a reinforcement learning mechanism.
[0037] Optionally, the central collaborative scheduling module includes:
[0038] The task allocation unit runs a hybrid heuristic algorithm that combines the whale optimization algorithm and the greedy strategy to perform robot-task matching with the goal of minimizing the weighted comprehensive cost.
[0039] The global path planning unit plans an initial collision-free path for the assigned task and integrates spatiotemporal constraint checks during the planning process.
[0040] The conflict pre-detection unit converts the planned path into a spatiotemporal trajectory and compares it with the global spatiotemporal reservation map to predict potential conflicts.
[0041] Optionally, the monitoring and adaptive optimization module is specifically used for:
[0042] Receive status reports from all robots, and perform progress tracking, deviation calculation, and abnormal event identification and classification.
[0043] Based on the identified event type, a dynamic response mechanism is activated for task delays, robot malfunctions, or new tasks.
[0044] Run an adaptive parameter tuning agent based on reinforcement learning, which takes historical and real-time performance data as input and outputs adjustment actions for scheduling algorithm parameters to achieve self-optimization of long-term running performance.
[0045] Compared with the prior art, this application has at least the following beneficial effects:
[0046] This application adopts a hierarchical collaborative architecture with upper-level central scheduling and lower-level distributed autonomous navigation. The upper layer performs global task allocation and coarse-grained path planning based on a multi-objective optimization model and a hybrid heuristic algorithm that integrates whale optimization and greedy strategies. The lower-level robot is responsible for local perception, real-time obstacle avoidance and fine control, which not only ensures the optimal overall efficiency of the system, but also gives the individual robot the flexibility to cope with dynamic environments, effectively balancing the needs of global optimization and real-time response.
[0047] This application constructs and maintains a global spatiotemporal reservation map, enabling the system to predict potential conflicts during the planning stage and proactively resolve conflicts using a priority-based negotiation adjustment strategy (including local spatial detours and insertion of waiting time windows). This significantly reduces emergency avoidance and parking waiting during operation, and significantly improves the utilization rate of channel resources and system throughput, especially in high-density operation scenarios.
[0048] The rolling optimization window mechanism designed in this application can periodically reschedule and optimize the paths of unexecuted tasks. At the same time, it provides classified dynamic response strategies for abnormal events such as task delays, robot failures, and new tasks, enabling the system to have closed-loop management capabilities, effectively cope with uncertainties in the warehousing environment, and ensure the continuity and reliability of services. Attached Figure Description
[0049] Figure 1 A flowchart illustrating an autonomous navigation and scheduling method for unmanned warehouse picking and distribution, provided as an embodiment of this application;
[0050] Figure 2 This is a schematic diagram of the module connection of an unmanned warehouse picking and dispatching autonomous navigation system provided in one embodiment of this application. Detailed Implementation
[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.
[0052] Combined with appendix Figure 1-2 This application provides a method for autonomous navigation and scheduling of unmanned warehouse picking, comprising the following steps:
[0053] S1: Warehouse environment modeling and task parsing, constructing a reconfigurable gridded warehouse map, parsing logistics tasks into a "start point-operation point-end point" triple sequence, and generating a dynamic task queue;
[0054] Specifically, the warehouse environment modeling process includes: First, establishing a global coordinate system based on a preset warehouse physical layout. Then, discretizing the warehouse planar area into a uniform set of grid cells. Each grid cell has a predefined size and is assigned specific type attributes, including but not limited to: shelf occupancy cells, two-way passage cells, one-way passage cells, inbound cells, outbound cells, charging station cells, and static obstacle cells. The status information of each grid cell is stored in the environmental map database. The status information includes at least cell type, coordinates, passage cost, and real-time occupancy flag. The grid map model supports parametric reconstruction. By adjusting the grid size, cell attribute distribution, and passage rules, it can adapt to warehouse scenarios of different sizes and layouts.
[0055] The task parsing process specifically includes: the system receives a task instruction stream from the upper-level order management system. For each independent logistics task, it performs structured parsing based on its task type: if it is an inbound task, it is parsed into a sequence (robot initial pose, target shelf parking position, specified inbound port); if it is a picking and handling task, it is parsed into a sequence (robot initial pose, current position of the shelf to be picked, specified outbound port); if it is an empty scheduling or shelf return task, it is parsed into a sequence (robot current position, target empty position or shelf empty position). Here, "start point" and "end point" are the starting and ending coordinates of the robot's action, and "operation point" is the specific coordinate of the position where loading or unloading actions must be performed. All parsed tasks are encapsulated into structured task objects, which include a task ID, task type, triplet sequence, priority weight, timestamp, and status identifier.
[0056] The generation and management of the dynamic task queue specifically includes: the system maintains a central task queue, and newly generated structured task objects are inserted into the corresponding positions of the central task queue according to their priority weight and timestamp. The queue supports dynamic insertion and deletion of tasks and provides a list of executable tasks to the scheduling core module in real time. At the same time, it updates the status of each task in the queue, including "waiting", "assigned", "executing" and "completed". This design ensures that the system can continuously respond to the real-time order flow and provide clear and orderly input for subsequent scheduling decisions.
[0057] S2: Hierarchical collaborative task scheduling. The upper-level central scheduler performs global task allocation and coarse-grained path planning based on a multi-objective optimization model and a hybrid heuristic algorithm. The lower-level distributed autonomous navigation units deployed in each robot receive instructions and are responsible for local perception, planning and feedback.
[0058] The hierarchical collaborative task scheduling method is completed collaboratively by the upper-level central scheduler and the lower-level distributed autonomous navigation unit.
[0059] The upper-level central scheduler, acting as the core of global decision-making, has the following workflow: First, it establishes a robot task scheduling model with the objective of minimizing the weighted comprehensive cost. The formula for calculating the weighted comprehensive cost is as follows:
[0060]
[0061] in, As a time cost, its calculation is based on the estimated execution time of the task and the estimated completion time of the robot's current task queue; The path length cost is obtained by pre-calculating or estimating the shortest path distance from the starting point to the operation point and from the operation point to the destination. The energy consumption cost is estimated based on path distance, estimated number of turns, and acceleration / deceleration model. To estimate the cost of conflict, the spatiotemporal trajectory segments of the planned path are pre-matched with the global spatiotemporal reservation map. The number of potential conflict nodes and their severity are statistically analyzed and weighted to obtain the cost. , , , It is an adjustable parameter used to balance the emphasis of different optimization objectives.
[0062] Secondly, the central scheduler employs a hybrid heuristic algorithm that combines Whale Optimization (WOA) with a greedy strategy for robot-task matching and coarse-grained path planning. The specific execution steps of this algorithm are as follows:
[0063] Initialization: The current state of all idle and soon-to-be-idle robots and the queue of tasks to be assigned are taken as input. An initial task sequence containing several tasks to be assigned is randomly generated for each robot as the position vector of the individual whale.
[0064] During the whale optimization phase, within a preset number of iterations, the positions of individual whales (i.e., the task sequence) are updated by simulating humpback whale encirclement, bubble web attacks, and random search behaviors. The fitness function is the aforementioned weighted comprehensive cost. This phase aims to conduct a global exploration to identify areas with potential excellent task allocation schemes.
[0065] A greedy strategy for local optimization involves fine-tuning the optimal individual and its neighborhood individuals generated during the whale optimization phase. Specifically, this involves traversing the task sequence and attempting to swap or insert tasks. If a swap or insertion fails... If the quality decreases, accept the change and iterate this process until no further improvement is possible. This stage aims to rapidly improve the quality of the solution.
[0066] After the algorithm converges, it outputs the task allocation scheme that minimizes the overall cost. For each "start-operation point-end" triplet in the allocation scheme, the scheduler plans a collision-free coarse-grained path based on a simplified environment model (ignoring dynamic obstacles) and the A* algorithm, and binds it to the corresponding task to form an initial path instruction sequence that is sent to the corresponding robot.
[0067] The lower-level distributed autonomous navigation unit is deployed on each mobile robot body. Its workflow includes: First, receiving the task sequence and the corresponding initial path instruction sequence from the central scheduler. Then, when the robot executes along the initial path, its onboard local perception module (including but not limited to lidar, depth camera, and ultrasonic sensor) collects surrounding environmental information in real time and merges it with the neighboring robot broadcast status information received through the wireless network to build a local dynamic environment map. This map is used to detect dynamic obstacles in real time (such as temporarily appearing workers, dropped goods, and other robots that deviate from the predetermined path) and predict potential trajectory conflict risks.
[0068] When a dynamic obstacle is detected obstructing the predetermined path, or when a conflict with a neighboring robot in space and time that cannot be resolved by speed adjustment is predicted, the autonomous navigation unit will initiate local real-time replanning. Local replanning runs on the robot's own computing unit, using the nearest passable location ahead as a temporary target, and quickly plans an alternative sub-path to avoid the obstacle or avoid it in the local dynamic environment map. After completing local obstacle avoidance, the robot will attempt to return to the original global coarse-grained path. At the same time, any local replanning event, significant deviation between the actual execution path and the initial plan, and changes in task status (such as start, delay, completion) will be fed back to the upper-level central scheduler in real time through the communication module. The central scheduler updates its internal task status records and global space-time reservation map based on this feedback information, thereby achieving synchronization and consistency maintenance of the global state.
[0069] S3: Spatiotemporal reservation map construction and conflict pre-resolution, generating spatiotemporal trajectories for planned paths, constructing and maintaining a global spatiotemporal reservation map to record the time occupancy status of each spatial grid, performing conflict detection during the planning phase and resolving conflicts using a priority-based negotiation adjustment strategy;
[0070] The conflict pre-resolution method is based on a global spatiotemporal reservation map, and its specific implementation process includes the following steps:
[0071] First, the spatiotemporal trajectory corridor is generated. After the upper-level central scheduler plans an initial spatial path P (represented by a series of ordered spatial grid coordinate sequences) from the starting point to the ending point for a robot, based on the robot's nominal motion speed model, the estimated precise time point or time period occupied by each path node on path P is calculated. Specifically, combining the robot's start time, the travel time within each grid cell, and the time cost of acceleration, deceleration, and turning, a time attribute is added to path P, thereby generating a trajectory T with spatiotemporal constraints. This trajectory T can be represented as a set of spatiotemporal coordinate pairs: ,in For raster coordinates, The estimated time for the robot to enter or occupy the grid is given. For straight travel segments in the path, this can be simplified to recording the entry time of the starting grid and the exit time of the ending grid to characterize the occupied interval of the path segment on the time axis.
[0072] Secondly, a global spatiotemporal reservation map is constructed and maintained. The system maintains a global data structure within the central scheduler to record the reservation status of all grid resources within a future period. This data structure is essentially a mapping table with grid coordinates (x, y) as keys. Each key corresponds to a time window occupancy table, which records which robot tasks have reserved the grid within multiple consecutive or discrete future time intervals. Whenever a new spatiotemporal trajectory T is generated, the system inserts all spatiotemporal coordinate pairs covered by the trajectory into the reservation table of the corresponding grid in chronological order, marking them as "reserved," and associating them with the corresponding robot ID and task ID.
[0073] Next, conflict detection is performed during the planning phase. When the hybrid algorithm that integrates whale optimization and greedy strategy is evaluated, or when coarse-grained paths are planned for task allocation results, the central scheduler compares the spatiotemporal trajectory T' generated by the path to be evaluated with the global spatiotemporal reservation map in real time. The conflict detection engine traverses each spatiotemporal coordinate pair (x, y, t) in trajectory T' and queries the reservation status of the grid near time t. If it is found that the spatiotemporal coordinate point or time period has been reserved by the trajectory of another robot, and the reservation subjects of the two trajectories are different, it is determined to be a potential spatiotemporal conflict. The system will record the location, time and robot task pairs involved in all conflict points.
[0074] Then, a negotiation-based time window adjustment strategy is implemented to resolve conflicts. Upon detecting a potential conflict, the system initiates a conflict resolution negotiation process. This process assigns priorities to conflicting robot task pairs based on preset dynamic priority rules (which can be calculated based on task urgency, robot task queue length, or path execution progress). The strategy is as follows: the original path reservations of high-priority robots remain unchanged; the system attempts to adjust the spatiotemporal trajectory of low-priority robots. The adjustment methods include: 1) Local spatial detour: performing local path replanning near the conflict point, bypassing the occupied grid, generating a new conflict-free spatiotemporal trajectory segment and updating the reservation; 2) Inserting a waiting time window: at a safe position before the conflict point, inserting a stop waiting time for the low-priority robot, delaying its arrival time at the conflict grid until the grid is released, thereby avoiding overlapping occupation. This negotiation and adjustment process is performed automatically within the scheduler.
[0075] Finally, a rescheduling triggering mechanism is defined. If a feasible conflict-free adjustment solution cannot be found for the low-priority robot through local detours and inserting waiting time windows (e.g., no detour or waiting time exceeds the maximum tolerance threshold), the negotiation is deemed to have failed. At this time, the conflict resolution module will send a signal that the conflict cannot be resolved to the upper-level task allocation module, along with relevant task and robot information. After receiving the signal, the upper-level task allocation module will trigger the rescheduling of the conflicting tasks, which may involve re-evaluating the task allocation scheme or replanning the global path for the relevant robots, thereby fundamentally avoiding irreconcilable resource competition. This mechanism ensures the system's deadlock avoidance and continuous operation capability in complex and high-density operating environments.
[0076] S4: Dynamic execution and real-time monitoring. The robot executes tasks according to the scheduled path and reports the status in real time. The central system monitors the overall progress and responds dynamically to delays, faults and new task events. A rolling optimization window is used to periodically reschedule and optimize the path for unexecuted tasks.
[0077] The dynamic execution and real-time monitoring process aims to ensure the robust execution of the scheduling plan in the physical environment and to manage uncertain events during the execution process in a closed loop.
[0078] First, the robot performing the task moves according to the spatiotemporal scheduled trajectory it receives. During the movement, the robot reports its status information to the central scheduling system in real time through its onboard controller and communication module at a fixed frequency or triggered events (such as reaching a path node or a status change). The reported status information is encapsulated as a structured data message, which includes at least the following: the robot's unique identifier, its precise pose at the current moment (including two-dimensional plane coordinates and orientation), instantaneous speed, the ID of the currently executing task and the current stage of the task (such as "heading to the shelf", "loading", "transporting", "unloading"), the estimated time to reach the next key node, and the codes of local abnormal events detected by the onboard sensors (such as sensing an unforeseen stationary obstacle). The above status data constitutes the main input for the system to perceive the real-time situation of the physical world.
[0079] Secondly, the central scheduling system runs a global execution monitoring module, which continuously receives status reports from all robots and compares them with the planned trajectories in the global spatiotemporal reservation map;
[0080] The monitoring module performs the following core functions: 1) Progress tracking and deviation calculation: Calculate the time and space deviation between the actual position and the planned position of each robot; 2) Status health diagnosis: Determine whether the robot is in normal working condition based on the reported speed, task stage dwell time, etc.; 3) Event identification and classification: Identify execution anomalies based on preset rules. The main event types include: task delay events (actual progress is significantly behind the plan), robot failure / stagnation events (no movement for a long time or reporting of abnormal codes), sudden obstacle events (local obstacle information reported by the robot), and new high-priority task insertion events triggered by external systems.
[0081] The system activates the corresponding dynamic response mechanism based on the identified event type:
[0082] For task delays, the monitoring module first assesses whether the delay will affect the reservation resources of other robots (i.e. whether it will cause new conflicts at subsequent waypoints). If it does not affect the reservation resources, it only updates the estimated arrival time of the robot and moves its subsequent trajectory in the spatiotemporal reservation map accordingly. If it may cause a conflict, it forwards the delay event to the conflict pre-resolution module (S3) to trigger the renegotiation and adjustment of the path of the affected robot.
[0083] In the event of robot failure / stagnation, the system first attempts to put the robot into a safe state and release the future spacetime resources it occupies through instructions. Then, it marks the robot's unfinished tasks as "interrupted" and submits them, along with the affected related tasks, to the head of the scheduler's task queue, triggering an emergency rescheduling of these tasks so that they can be completed by other available robots.
[0084] For new tasks, the system directly inserts the task into the dynamic task queue and, depending on its priority, either processes it in the next optimization cycle or immediately triggers a local rescheduling to quickly allocate resources.
[0085] Finally, the system employs a rolling optimization window mechanism to continuously improve overall efficiency. This mechanism is defined as follows: the system automatically initiates an optimization process at a fixed time period (i.e., the "window" length, such as 30 seconds) or event period (such as after N tasks are completed). In this process, the scheduler does not replan all tasks, but focuses on a subset of tasks that have not yet started execution (i.e., tasks in the "waiting" state). At the same time, it considers the current position and state of all robots as new initial conditions. Based on this updated system state snapshot, the scheduler uses the hybrid heuristic algorithm in S2 to redistribute and plan paths for this subset of unexecuted tasks. The new solution generated by this optimization, which is better than the original plan, will be adopted and updated to the global spatiotemporal reservation map. Subsequent tasks will be executed according to the new plan. This rolling optimization mechanism can periodically correct suboptimal scheduling caused by accumulated errors or incomplete early information, enabling the system to have the ability to continuously self-optimize and adapt to the dynamically changing warehouse operation environment.
[0086] S5: Multi-objective performance evaluation and adaptive parameter tuning. Multi-dimensional performance indicators are collected during system runtime, and the cost function weights and internal parameters of the scheduling model are adaptively adjusted based on the reinforcement learning mechanism to achieve long-term performance optimization.
[0087] The multi-objective performance evaluation and adaptive parameter tuning process aims to enable the system to autonomously learn and optimize its scheduling strategy during long-term operation to adapt to specific job modes and efficiency goals. Its specific implementation includes the following steps:
[0088] First, a multi-dimensional performance indicator collection system is established. During system operation, various functional modules automatically collect and summarize a series of quantitative performance data. Key indicators collected include, but are not limited to: 1) Task completion time: Statistical analysis of the time taken from task assignment to completion, as well as the average completion time and time span of batch tasks; 2) Total travel distance: Accumulated total actual travel distance of all robots within a statistical period; 3) Total estimated energy consumption: Based on the robot's motion model (uniform speed, acceleration, deceleration, turning) and its actual travel path, the total energy consumption within a statistical period is estimated; 4) Number of conflicts: Statistical analysis of the number of conflict events that occur during actual operation and require resolution through stopping, waiting, or emergency avoidance; 5) Equipment utilization rate: Calculation of the ratio of the total effective working time of the robot group to the total available time. All indicators are collected according to preset statistical periods (e.g., per hour, per shift) and stored in a historical performance database.
[0089] Secondly, an adaptive parameter tuning mechanism based on reinforcement learning is designed and implemented. This mechanism models the parameter optimization process of the scheduling system as a sequential decision problem, and its core elements are defined as follows:
[0090] State: Represents the operating status of the system at the end of a statistical period. It is defined as a vector composed of various performance indicators collected during the period, and environmental features such as task density and number of robots can be added.
[0091] Action: Defined as an adjustment to key parameters of the core scheduling algorithm. Specifically, this includes adjusting the combination of weight coefficients in the cost function. , , , ), and adjust specific parameters within the hybrid heuristic algorithm (described in S2), such as the search coefficient of the whale optimization algorithm, the neighborhood search depth of the greedy strategy, etc., with the action space being a set of continuous or discrete parameter values.
[0092] Reward: Calculated based on system performance over a statistical period. The reward function is designed as a comprehensive feedback mechanism for multi-objective performance, for example: .in, to The weights are designed to reflect the management's emphasis on different performance dimensions. The reward values are intended to guide the agent to adjust parameters to maximize long-term cumulative rewards.
[0093] This mechanism is implemented through an adaptive parameter tuning agent, which is trained using deep reinforcement learning algorithms (such as Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO). Its workflow is as follows: at the end of each statistical period, the agent observes the current state and outputs an action (i.e., a new set of parameter values) based on its learned policy network. The scheduling system applies this new set of parameters for task scheduling and path planning in the next statistical period. After the period ends, based on the generated rewards and the new state, a set of experience data is formed and stored in the replay buffer for periodically updating the agent's policy network and value network. Through a large number of online or offline training cycles, the agent eventually learns to automatically output parameter configurations that lead to higher long-term comprehensive performance based on different system operating states (such as peak periods and slow periods).
[0094] Through the aforementioned closed-loop learning process, the system achieves dynamic and personalized parameter optimization. This enables the system to break free from dependence on fixed parameters or manual parameter tuning, automatically adapting to different warehouse layouts, different order flow patterns, and changing operational goals, thereby continuously approaching better overall operating efficiency, energy economy, and system stability in the long term.
[0095] In one embodiment, a warehouse unmanned order fulfillment autonomous navigation and scheduling system is provided, the system comprising:
[0096] The Environment Modeling and Management module is used to build and maintain warehouse maps, shelf information, and task queues. It is responsible for building and maintaining a digital twin warehouse model. Its specific functions include building and storing a reconfigurable rasterized global map based on physical warehouse drawings or scanned data, labeling each grid cell with type attributes (shelf location, two-way / one-way aisle, entrance / exit, charging pile, obstacle zone) and dynamic passage cost. At the same time, the Environment Modeling and Management module maintains a shelf information database, recording the unique ID of each shelf, the information of the stored goods, the current status (idle, occupied, locked) and its precise location in the raster map.
[0097] This module also manages a central dynamic task queue, which is responsible for receiving and parsing external order instructions, converting them into standardized "start point-operation point-end point" task objects, and inserting, sorting and updating the status of the queue according to priority and timestamp, providing clear task input for the scheduling module;
[0098] The central collaborative scheduling module includes the following functional units:
[0099] The task allocation unit receives a dynamic task queue from the environment module and has a built-in multi-objective optimization model. Its core is a weighted comprehensive cost function, which is used to quantify and evaluate the comprehensive cost of different robots performing different tasks. This unit runs a hybrid heuristic algorithm that combines whale optimization (WOA) and greedy strategy. By combining global exploration and local fine-grained search, it finds the robot-task matching scheme with the minimum overall cost under multiple constraints such as time, distance, energy consumption and conflict prediction.
[0100] The global path planning unit plans an initial collision-free path from the start point to the end point for each task triplet in the task assignment results. It uses an improved A* search algorithm as the core path planner and a complex diagonal distance algorithm as the heuristic function to more accurately estimate costs and reduce the number of search nodes. At the same time, it integrates spatiotemporal constraint checks during the search process. By querying the spatiotemporal reservation map, it avoids planning nodes that conflict with other reserved paths in spatiotemporal space, thus incorporating collision avoidance considerations into the planning stage.
[0101] The conflict pre-detection unit works closely with the path planning unit. It receives the planned initial path and expands it into a spatiotemporal trajectory based on the robot's motion model. By comparing this spatiotemporal trajectory with the global resource occupancy table maintained by the spatiotemporal reservation map management module, this unit can predict potential spatiotemporal conflicts (i.e., two or more robots are scheduled to occupy the same spatial grid at the same time) before the task is executed. The detected conflict information will be fed back immediately to trigger negotiation or as a basis for cost evaluation of task allocation.
[0102] Distributed Autonomous Navigation Module: This module is deployed on each mobile robot and is responsible for safe and precise local execution after receiving global commands. Specifically, it includes:
[0103] The local perception unit integrates multiple sensors, such as LiDAR, visual cameras, and inertial measurement units (IMU), to collect raw environmental data around the robot in real time. Through filtering, matching, and fusion algorithms, it constructs and updates a local high-precision environmental map centered on itself. This map is used to identify dynamic obstacles (such as pedestrians, temporary piles of objects, and other robots) that are not marked in the global map, as well as to detect its own positioning deviation.
[0104] Local planning unit: This unit receives global coarse-grained path instructions from the central scheduling module. While traveling along the path, it performs millisecond-level fine-grained motion planning based on real-time environmental information provided by the local perception unit. When dynamic obstacles are detected or trajectory conflicts are predicted in a short time, this unit can plan a smooth sub-path for local obstacle avoidance or collision avoidance in a very short time and guide the robot to rejoin the global path after the obstacle disappears.
[0105] The communication unit is responsible for data exchange between the robot and the central system and other robots. It periodically reports the robot's own status (such as position, speed, task progress, and exception codes) to the central system, while receiving instructions and global information updates from the central system. In some cooperative avoidance scenarios, this unit also supports direct, low-latency status broadcasting and simple negotiation with neighboring robots to achieve faster local collaboration.
[0106] Spatiotemporal Reservation Map Management Module: This module is responsible for maintaining a global, future-oriented spatiotemporal resource reservation system. Its core is a spatiotemporal reservation database, which uses grids as indexes to record the occupied time window and occupant (robot task ID) of each grid on the future timeline. This module provides key atomic operations, including: inserting reservations for planned paths, querying the occupancy status of specified spatiotemporal points, and releasing completed or canceled path reservations. It serves as the basis for conflict pre-detection and supports the execution of conflict negotiation strategies, such as attempting to reserve alternative time windows for low-priority robots or modifying local reservations during the negotiation process.
[0107] The monitoring and adaptive optimization module enables closed-loop optimization and continuous learning of the system. It includes a real-time monitoring dashboard that continuously collects and visualizes overall system performance metrics such as task throughput, robot utilization, average latency, total travel distance, and number of collisions. Embedded within this module is an adaptive parameter tuning agent based on reinforcement learning. This agent uses historical and real-time performance data as state input to adjust the core scheduling algorithm (especially the cost function weights). Using the internal parameters of the hybrid algorithm as actions and the comprehensive performance indicators of the system (such as the reward function) as feedback, the system can autonomously adapt to different work pressures and operational goals through online or offline training, and achieve self-evolution and continuous improvement of long-term operating performance.
[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
Claims
1. A method for autonomous navigation and scheduling of unmanned warehouse order fulfillment, characterized in that, The method includes: S1. Warehouse environment modeling and task parsing: Construct a reconfigurable gridded warehouse map, parse logistics tasks into a triple sequence of start point-operation point-end point, and generate a dynamic task queue. S2. Layered collaborative task scheduling: The upper-level central scheduler performs global task allocation and coarse-grained path planning based on a multi-objective optimization model and a hybrid heuristic algorithm. The lower-level distributed autonomous navigation units deployed in each robot receive instructions and are responsible for local perception, planning and feedback. S3. Spatiotemporal reservation map construction and conflict pre-resolution: generate spatiotemporal trajectories for the planned path, construct and maintain a global spatiotemporal reservation map to record the time occupancy status of each spatial grid, perform conflict detection during the planning stage and resolve conflicts using a priority-based negotiation adjustment strategy. S4. Dynamic execution and real-time monitoring: The robot executes tasks according to the scheduled path and reports the status in real time. The central system monitors the overall progress and responds dynamically to delays, faults and new task events. A rolling optimization window is used to periodically reschedule and optimize the path of unexecuted tasks. S5. Multi-objective performance evaluation and adaptive parameter tuning: The system collects multi-dimensional performance indicators during runtime and adaptively adjusts the cost function weights and internal parameters of the scheduling model based on the reinforcement learning mechanism.
2. The warehouse unmanned picking and dispatching autonomous navigation method and system according to claim 1, characterized in that, In step S1, the construction of the reconfigurable rasterized warehouse map specifically includes: A global coordinate system is established based on the physical layout of the warehouse, and the warehouse planar area is discretized into a uniform set of grid cells; Assign a type attribute to each grid cell. The type attribute includes shelf occupancy cell, two-way passage cell, one-way passage cell, inlet cell, outlet cell, charging station cell, and static obstacle cell. The status information of each grid cell is stored, and the status information includes at least the cell type, coordinates, passage cost, and real-time occupancy flag.
3. The warehouse unmanned picking and dispatching autonomous navigation method and system according to claim 1, characterized in that, In step S2, when the upper-level central scheduler performs global task allocation, it uses the following weighted comprehensive cost function for optimization: in, As a time cost, its calculation is based on the estimated execution time of the task and the estimated completion time of the robot's current task queue; The path length cost is obtained by pre-calculating or estimating the shortest path distance from the starting point to the operation point and from the operation point to the destination. The energy consumption cost is estimated based on path distance, estimated number of turns, and acceleration / deceleration model. Estimate the costs of conflict.
4. The warehouse unmanned picking and dispatching autonomous navigation method and system according to claim 3, characterized in that, The hybrid heuristic algorithm combines the whale optimization algorithm with a greedy strategy, including: During the initialization phase, an initial task sequence containing the tasks to be assigned is randomly generated for each robot as the position vector of the individual whale. During the whale optimization phase, the individual position is updated by simulating the encirclement, bubble net attack and random search behavior of humpback whales, and the weighted comprehensive cost is used as the fitness function for global exploration. In the local optimization phase of the greedy strategy, the optimal individual and its neighboring individuals are locally refined by traversing the task sequence and trying to swap or insert tasks until the weighted comprehensive cost can no longer be reduced.
5. The warehouse unmanned picking and dispatching autonomous navigation method and system according to claim 1, characterized in that, In S3, the resolution using a priority-based negotiation adjustment strategy specifically includes: assigning dynamic priorities to the conflicting robot task pairs according to preset rules; maintaining the original path reservation of the high-priority robot unchanged; and adjusting the low-priority robot to attempt local spatial detours or insert waiting time windows. If the adjustment fails, it will trigger a rescheduling of the conflicting tasks.
6. The warehouse unmanned picking and dispatching autonomous navigation method and system according to claim 1, characterized in that, In S4, the scrolling optimization window mechanism specifically refers to: Trigger the optimization process using a fixed time period or event period; Focus on a subset of all tasks that are in a "waiting" state, and combine the current position and state of all robots as new initial conditions; The hybrid heuristic algorithm is used to redistribute and plan paths for the subset of unexecuted tasks, and a new scheme that is superior to the original plan is adopted.
7. The warehouse unmanned picking and dispatching autonomous navigation method and system according to claim 1, characterized in that, In step S5, the adaptive parameter tuning mechanism based on reinforcement learning includes: The parameter optimization process of the scheduling system is modeled as a sequential decision problem, where the state is the performance index vector within the statistical period, the action is the adjustment of the key parameters of the scheduling algorithm, and the reward is the comprehensive feedback value based on multi-objective performance calculation. The adaptive parameter tuning agent outputs new parameter configurations based on the system's operating status and is trained using deep reinforcement learning algorithms to achieve dynamic and personalized parameter optimization.
8. A warehouse unmanned picking and dispatching autonomous navigation system, characterized in that, include: The environmental modeling and management module builds and maintains a reconfigurable gridded warehouse map, a shelving information database, and a central dynamic task queue. The central collaborative scheduling module performs global task allocation and coarse-grained path planning based on a multi-objective optimization model and a hybrid heuristic algorithm, and integrates conflict pre-detection functionality. The distributed autonomous navigation module is deployed in each mobile robot, receives global commands, performs real-time obstacle avoidance and path replanning based on local perception, and provides feedback on the execution status. The spatiotemporal reservation map management module builds and maintains a global spatiotemporal reservation map, recording the time occupancy status of each spatial grid. The monitoring and adaptive optimization module monitors the system's operating status in real time, identifies abnormal execution events, and adaptively adjusts the parameters of the scheduling model based on a reinforcement learning mechanism.
9. The warehouse unmanned picking and dispatching autonomous navigation system according to claim 8, characterized in that, The central collaborative scheduling module includes: The task allocation unit runs a hybrid heuristic algorithm that combines the whale optimization algorithm and the greedy strategy to perform robot-task matching with the goal of minimizing the weighted comprehensive cost. The global path planning unit plans an initial collision-free path for the assigned task and integrates spatiotemporal constraint checks during the planning process. The conflict pre-detection unit converts the planned path into a spatiotemporal trajectory and compares it with the global spatiotemporal reservation map to predict potential conflicts.
10. The warehouse unmanned picking and dispatching autonomous navigation method and system according to claim 8, characterized in that, The monitoring and adaptive optimization module is specifically used for: Receive status reports from all robots, and perform progress tracking, deviation calculation, and abnormal event identification and classification. Based on the identified event type, a dynamic response mechanism is activated for task delays, robot malfunctions, or new tasks. Run an adaptive parameter tuning agent based on reinforcement learning, which takes historical and real-time performance data as input and outputs adjustment actions for scheduling algorithm parameters to achieve self-optimization of long-term running performance.