Intelligent warehouse multi-objective path planning method based on large model

Abstract

The present application relates to the technical field of path planning, and more particularly, to an intelligent warehouse multi-objective path planning method based on a large model, comprising: obtaining a state vector of an intelligent warehouse system, the state vector representing current time pressure, energy consumption pressure and throughput pressure of the system; inputting the state vector into a pre-trained Actor policy network to obtain time weight, energy consumption weight and throughput weight corresponding to the time pressure, energy consumption pressure and throughput pressure. The present application solves the problem of intelligent trade-off of multiple conflicting objectives by constructing a state vector representing the time pressure, energy consumption and throughput pressure of the system in real time, driving the large model to dynamically generate the decision weights of the three, combining the dynamic congestion prediction and cost calculation based on the neural network to plan the globally optimal path with the overall cost for the AGV, and introducing an adaptive re-planning mechanism to cope with environmental changes.

Description

Technical Field

[0001] This invention relates to the field of path planning technology. More specifically, this invention relates to a multi-objective path planning method for intelligent warehousing based on a large model. Background Technology

[0002] Intelligent warehousing is a core component of modern logistics systems, and path planning for autonomous mobile robots (AGVs) is a key technology for improving warehouse automation and operational efficiency. This technology aims to plan optimal travel paths for AGV clusters in the warehouse environment to efficiently and cost-effectively complete the precise handling and sorting of goods. Against the backdrop of rapid e-commerce development, modern intelligent warehouses commonly deploy robot clusters consisting of hundreds or even thousands of AGVs to cope with massive, fragmented, and time-sensitive order demands. The operational efficiency of AGV clusters, especially their path planning and traffic management capabilities, has become a core bottleneck determining the overall warehouse throughput, cost, and fulfillment timeliness.

[0003] However, intelligent warehousing systems are highly dynamic and complex environments. Their real-time status, such as the number of urgent orders, the average remaining power of the AGV cluster, and the overall system load, is constantly changing. Therefore, AGV path planning is not a simple shortest path optimization problem, but a multi-objective optimization challenge that requires real-time trade-offs between conflicting objectives such as order timeliness, equipment energy consumption, and system throughput. How to dynamically adjust decision preferences based on the system's current core contradictions is a major challenge facing existing technologies. For example, during peak order periods, warehouses need to simultaneously dispatch hundreds or thousands of AGV robots to handle massive amounts of scattered orders. Traditional algorithms struggle to simultaneously balance the high timeliness of urgent orders, high overall warehouse throughput, low AGV energy consumption, and avoiding congestion in hotspot areas in such a large-scale and dynamically changing environment, ultimately leading to low system efficiency and high operating costs. Summary of the Invention

[0004] This invention provides a multi-objective path planning method for intelligent warehousing based on a large model. It aims to solve the problem that traditional algorithms in related technologies have difficulty simultaneously considering multiple conflicting objectives in such a large-scale and dynamically changing environment, such as the high timeliness of urgent orders, the high throughput of the overall warehouse, the low energy consumption of AGVs, and the avoidance of congestion in hot spots, which ultimately leads to low system efficiency and high operating costs.

[0005] This invention provides a multi-objective path planning method for intelligent warehousing based on a large model, comprising: acquiring the state vector of the intelligent warehousing system, wherein the state vector represents the system's current timeliness pressure, energy consumption pressure, and throughput pressure; inputting the state vector into a pre-trained Actor policy network to obtain timeliness weights, energy consumption weights, and throughput weights corresponding to the timeliness pressure, energy consumption pressure, and throughput pressure; for each path in the map, performing weighted fusion based on the path's assigned traffic and historical traffic to obtain the optimal predicted traffic value, and calculating the dynamic congestion cost of the path by combining the path type and the throughput pressure; multiplying the timeliness weights, energy consumption weights, and throughput weights by the basic time cost, basic energy consumption cost, and dynamic congestion cost of each path respectively, and summing the results to obtain the global comprehensive cost of each path, using A... The search algorithm plans the optimal path with the lowest overall cost for the AGV. By introducing a pre-trained Actor policy network, it can dynamically generate the optimal weights of the three factors based on the most critical conflict in the current system, and calculate the dynamic congestion cost by combining accurate predictions of future path traffic. Ultimately, it plans a path with the lowest overall cost for the AGV. This method transforms path planning from a static wayfinding problem into an intelligent decision-making process that can dynamically weigh multiple conflicting objectives, thereby significantly improving the overall operating efficiency, robustness, and intelligence level of the system in complex and ever-changing warehousing environments.

[0006] Furthermore, obtaining the state vector includes: obtaining the total number of high-priority orders and orders about to time out, and dividing it by the system's urgent order pressure threshold to obtain an urgency index; obtaining the average battery level of all online AGVs, and normalizing it according to preset low battery and health thresholds to obtain battery health; obtaining the number of orders completed by the system within a preset period, and dividing it by the historical peak throughput to obtain the system load rate; and combining the urgency index, battery health, and system load rate into the state vector. Through a series of explicit and calculable indicators such as the urgency index, battery health, and system load rate, abstract system pressure is transformed into precise numerical values. This enables the large model to make decisions based on a solid data foundation, ensuring the accuracy and sensitivity of its perception of system state, thereby improving the rationality and effectiveness of subsequent dynamic weight adjustments.

[0007] Furthermore, the method also includes: for an AGV currently performing a task, calculating the difference between the new cost of its current remaining path under the current global comprehensive cost and the original cost at the planning time, to obtain the path cost degradation degree; if the path cost degradation degree is greater than a preset adaptive replanning threshold, then repeating the method to replan the path for the AGV. By calculating the cost degradation degree of the current path in real time, the AGV has the ability to evaluate the quality of the path on the way. When environmental changes cause the original path to no longer be optimal, replanning can be triggered in a timely manner, ensuring that the AGV always travels on the optimal path in the current state, greatly enhancing the adaptability and global optimality of the path planning scheme to dynamic environmental changes.

[0008] Furthermore, the adaptive replanning threshold is calculated as follows: The numerator is obtained by adding the base tolerance to the weighted value of energy consumption; the denominator is obtained by adding 1 to the weighted values ​​of timeliness and throughput respectively; and the ratio of the numerator to the denominator is used as the adaptive replanning threshold. The threshold in this scheme dynamically changes according to the current system's decision preferences, i.e., the weights of timeliness, energy consumption, and throughput. For example, when focusing on energy consumption, the system will increase its tolerance for increased costs to avoid the additional energy consumption caused by frequent replanning, making the replanning decision itself more intelligent and consistent with the system's overall strategic goals, avoiding unnecessary computational overhead and operational disturbances.

[0009] Furthermore, the optimal predicted traffic value is calculated as follows: Historical prediction errors, path type, urgency index, and system load rate are input into a pre-trained first neural network to obtain a dynamic confidence score. The product of the dynamic confidence score and the assigned traffic volume is then added to the product of 1 minus the dynamic confidence score and the historical traffic volume to obtain the optimal predicted traffic value. By dynamically fusing these two information sources using a neural network and calculating a dynamic confidence score, it is determined which data should be trusted more under the current operating conditions. This method effectively compensates for the shortcomings of a single data source, obtaining more robust and accurate future traffic prediction results, thus providing a more reliable input for calculating congestion costs.

[0010] Furthermore, the calculation method for the dynamic congestion cost is as follows: The path type and system load rate are input into a pre-trained second neural network to obtain a dynamic congestion index; the optimal predicted flow rate is divided by the difference between the path's maximum capacity and the optimal predicted flow rate to obtain the path saturation; the path saturation value is then raised to the power of the dynamic congestion index to obtain the dynamic congestion cost. By learning the congestion and collapse characteristics of different path types under different system loads through a dedicated neural network, the potential congestion costs that a path may encounter in the future can be more realistically quantified, making the AGV's obstacle avoidance decisions more forward-looking and accurate.

[0011] Furthermore, the training method for the pre-trained Actor policy network includes: adopting an Actor-Critic architecture, inputting the state vector into the Critic value network to obtain the value prediction value, calculating the comprehensive performance score as the reward signal based on the system throughput and the average order timeout rate; calculating the temporal difference error between the reward signal and the value prediction value based on the Bellman equation, and updating the parameters of the Actor policy network and the Critic value network according to the temporal difference error.

[0012] Furthermore, the comprehensive performance score is calculated as follows: the product of system throughput and preset business preference coefficient, minus the product of the difference between 1 and the business preference coefficient and the average order timeout rate, is used as the comprehensive performance score.

[0013] Furthermore, the data sources for obtaining the state vector include: order task list and business priority obtained from the warehouse management system; remaining power, working status and real-time coordinates of each AGV obtained from the AGV cluster management system; and map topology information and historical path traffic data obtained from the motion control system.

[0014] Furthermore, the preset value of the business preference coefficient is 0.7.

[0015] Beneficial effects: By constructing state vectors in real time to represent the system's timeliness, energy consumption, and throughput pressure, the large model is driven to dynamically generate the decision weights of the three. Combined with dynamic congestion prediction and cost calculation based on neural networks, the system plans the optimal path for the AGV's overall cost and introduces an adaptive replanning mechanism to cope with environmental changes. This effectively solves the problem of intelligent trade-offs between multiple conflicting objectives in large-scale, dynamic warehousing scenarios. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating a path planning flowchart according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the dynamic congestion cost variation trend according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating a comparison between the present invention and conventional path planning according to an embodiment of the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] like Figures 1-3 As shown, S101: Data acquisition and state vector construction.

[0019] First, it is necessary to collect and integrate the multi-source heterogeneous data of the intelligent warehousing system in real time, and quantify it into a standardized state vector as the basic input for subsequent decision-making models.

[0020] Specifically, data collection includes the following aspects: Warehouse management system data collection: obtaining a list of all pending and executing order tasks, and reading the business priority of each order in real time, such as VIP expedited orders or regular orders, while obtaining the physical storage location of the goods required for the order, such as shelf area, storage location number, etc.

[0021] AGV cluster data acquisition: Obtain the remaining power percentage, current working status and real-time coordinates of each AGV in the cluster. The working status includes idle, picking up goods and charging.

[0022] Mobility control system data acquisition: By querying the internal traffic management database and dispatch logs, the assigned traffic flow and actual observed traffic flow of each route on the current map are statistically analyzed, and combined with the average traffic flow of the same period in history and historical prediction error data.

[0023] Warehouse environment data acquisition: Obtain high-precision maps of the warehouse, which include topological information such as intersections and paths, as well as the maximum capacity and path type of each path, such as straight lines and intersections.

[0024] Because the raw data collected above varies in size and format (e.g., 50 urgent orders, 45% battery capacity, and 2300 orders per hour), subsequent reinforcement learning models cannot directly process it. Therefore, this invention uses a series of quantization and normalization operations to refine and compress this massive amount of data into a unified, continuous, and standardized warehouse state vector. This vector can clearly describe the core contradictions in the current system operation, such as timeliness pressure, energy consumption pressure, and throughput pressure.

[0025] As a preferred embodiment, the state vector It is a three-dimensional vector in time series, and the specific construction process is as follows.

[0026] Urgency Index This is used to quantify timeliness pressure. The specific quantification process is as follows: The total number of high-priority orders and orders about to expire in the current warehousing system is obtained in real time. In this embodiment, VIP customer orders are defined as high priority, and orders with a delivery time within the next 15 minutes are defined as about to expire. The total number of both is divided by a preset system emergency order pressure threshold, which represents the critical point at which the system begins to experience pressure. In this embodiment, this threshold is preset to 50 orders, resulting in the pressure ratio. Finally, through an upper bound clamping operation, i.e., taking the smaller value between the pressure ratio and 1.0, the ratio is standardized to a continuous value within the range [0,1], which is the urgency index. .when When the value is close to 1, it indicates that the system is under great time pressure.

[0027] Battery health This is used to quantify energy consumption pressure. The specific quantification process is as follows: real-time acquisition of the power percentage of all online AGVs, and calculation of the average power consumption of the entire AGV cluster. Combined with preset low battery thresholds (Preferred percentage in this embodiment is 30%) and health threshold (Preferred percentage in this embodiment is 80%), and the battery health is calculated using the following formula. : It can be seen that when When the value approaches 0, it indicates that the overall power consumption of the AGV cluster is at its lowest, while the system's energy consumption pressure is at its highest. To find the maximum value function, This is a function that takes the minimum value.

[0028] System load rate This is used to quantify throughput pressure. The specific quantification process is as follows: Obtain the number of orders completed by the system in the past period (15 minutes in this embodiment) and convert it into hourly throughput as the system's real-time throughput. Divide this value by the historical peak throughput, i.e., the maximum sustainable throughput obtained from historical data, which is set to 5000 orders / hour in this embodiment, to obtain the load ratio. Similarly, through an upper bound clamping operation, this ratio is standardized to the [0,1] interval to obtain the system load rate. .when When the value is close to 1, it indicates that the throughput pressure is huge and the risk of congestion is extremely high.

[0029] Finally, the above three indicators are combined into a three-dimensional state vector over time. This serves as input for subsequent steps in generating real-time decision-making strategies.

[0030] S102: Generation of continuous policy weights based on reinforcement learning.

[0031] In multi-objective path planning for intelligent warehousing, the system must balance three conflicting objectives in real time: order timeliness, AGV energy consumption, and system throughput. This step employs an "Actor-Critic" reinforcement learning architecture, using the state vector obtained in step S101... As a continuous input, it generates the optimal policy weights for the current state.

[0032] To evaluate the merits of decisions, a global performance indicator that can quantify the overall warehouse operation is needed. This embodiment weighs warehouse throughput and average timeout rate using a business preference coefficient to obtain a comprehensive performance score. This performance score will serve as the reward signal for the reinforcement learning model, whose training objective is to maximize its value. The performance function for this system is based on the premise that an efficient warehousing system should simultaneously possess high throughput and low latency. Its calculation formula is as follows: In the formula, Is the system in The overall performance score obtained at each moment is used to evaluate the final decision; Is the system in throughput at any given moment To the maximum preset throughput, This reflects the load rate; Is the system in The average order timeout rate at any given moment is used to evaluate the average delay of all orders within a period. In this example, a period of 15 minutes is used. This is the business preference coefficient, which is set to 0.7 in this embodiment. It is used to define the system's emphasis on throughput and timeout rate.

[0033] As can be seen from the formula, when throughput Increase or average timeout rate When reduced, overall performance All will improve. Subsequent Actor-Critic models will use... With a single guiding principle, it learns to adapt and adjust online continuously to the current business preferences. Maximize the continuous strategy.

[0034] The state vector obtained in step S1 The input is fed into a pre-trained Actor policy network. This network outputs an activation potential vector at time t through internal multi-layer nonlinear computation. This vector is a set of policy parameters used in subsequent calculations of the final dynamic policy weights. It is the time-dependent activation potential at time t, representing the Actor network's activation potential based on the current state. The urgency index This determines the level of importance attached to ensuring order timeliness. If If the value is larger, the system will be more inclined to choose the path with lower time cost. It represents the energy consumption activation potential at time t, characterizing the Actor network's response based on battery health. This determines the level of importance attached to the goal of saving AGV energy consumption. If If the value is larger, the system will be more inclined to choose the path with lower energy consumption cost. It represents the throughput activation potential at time t, characterizing the Actor network's response to system load. This determines the level of importance attached to the goal of ensuring the total throughput of the system (i.e., avoiding congestion). If If the value is high, the system will prioritize avoiding potentially congested road sections.

[0035] Using the Softmax function to analyze the activation potential vector Normalization is performed to calculate the three objectives—order timeliness, AGV energy consumption, and system throughput—in the current state vector. The specific strategy weights are as follows. Taking the timeliness weight as an example, its calculation formula is: ; This represents the time-sensitivity weight at time t. The potential for activation at time t. This represents the j-th feature in the activation potential vector at time t. This is an exponential function with the natural constant e as its base. Similarly, the energy consumption weight and throughput weight can be calculated using the method described above.

[0036] Then, the state vector The input is fed into a pre-trained Critic value network, which outputs a desired benchmark. This refers to the predicted cumulative return in the current state. When the system completes execution at time [time value missing]... After implementing the strategy, time step t+1 will be reached, yielding the actual overall performance R(t+1) and a new value prediction S(t+1). At this point, the time-series difference error is calculated. To determine the time Whether the strategy and value prediction are accurate, and accordingly adaptively adjust the network parameters. Temporal difference error. The calculation is based on the Bellman equation, which is as follows: In the formula, It is a moment The error score is used to determine whether the strategy adopted by the system in the current state is accurate; Is Under the given conditions, the system's overall performance score; It is a discount factor that ensures the system pays more attention to recent value predictions. In this embodiment, the discount factor is set to 0.95. Is The new value prediction made by the system under the given conditions; Is The value prediction made by the system under the given conditions.

[0037] when When this occurs, it indicates that the current strategy is better than expected and the system is evolving in a positive direction. At this time, the Actor network will adjust its parameters to make the activation potential vector of the next output closer to the current output. When this occurs, it indicates that the Critic network's value predictions are biased. In this case, the Critic network will update its parameters to correct the bias. By approaching zero, the accuracy of predictions is improved. Through this online learning mechanism, the system can continuously optimize its decision-making capabilities.

[0038] S103: Calculate congestion costs.

[0039] In AGV cluster path planning, simply selecting the shortest path or the lowest energy consumption will cause all AGVs to flock to the same optimal path, such as main roads or popular shelving areas, ultimately leading to traffic congestion. Therefore, an adaptive congestion cost mechanism that can accurately predict and quantify future congestion is needed to intelligently guide AGV diversion before congestion occurs, ensuring system throughput and operational efficiency. Specifically, two dedicated learning models are introduced, using the state vector from step S101 as the core input, to predict the future traffic flow of each path and calculate its dynamic congestion cost.

[0040] Through the first small neural network This model aims to learn whether the actual traffic on a path is closer to assigned traffic or historical traffic under different global states. The model is based on the premise that the system's trust in either traffic source should dynamically change under different operating conditions. The calculation formula is as follows: In the formula, For path At any moment The dynamic confidence level in the first small neural network Input historical prediction error Path type Urgency Index and system load rate After that, you can get This is used to subsequently decide whether to place more trust in already assigned traffic or historical traffic. It's important to note that this applies to the path... Path type `<path>` is a static tag used to describe a path. The physical type is obtained from the warehouse environment data collected in step one. This example pre-defines three labels: straight line, corner, and intersection. The default level of confusion is: crossroads > corners > straight lines. The more chaotic the path, the lower the base value of the dynamic reliability of its assigned traffic.

[0041] Combining dynamic confidence For the assigned traffic and historical traffic By performing weighted fusion, the optimal predicted flow value is obtained. The rationale behind this fusion is that dynamic weighting can compensate for the shortcomings of a single data source, resulting in more robust prediction results. The calculation formula is as follows: The optimal predicted flow value obtained through this formula It can proactively anticipate and avoid prediction failures caused by sudden changes in the system's operational status.

[0042] Through a second small neural network This model aims to learn the real-world congestion deterioration characteristics of different road types under varying global loads. The model is based on the premise that different types of roads experience congestion deterioration at different rates under the same load. The calculation formula is as follows: It can be seen that, For path At any moment The dynamic congestion index, in the second small neural network Input path type and system load rate After that, you can get This will be used in subsequent systems to achieve more accurate congestion predictions.

[0043] It is important to note that, The true characteristic of this system is its ability to learn and gain global awareness. For example, at a busy intersection, during peak sales periods, traffic can easily become saturated, leading to rapid congestion and a very fragile congestion collapse characteristic. In this case, the system is extremely sensitive and will strongly discourage AGVs from entering. Conversely, during off-peak hours at night, traffic is lower, the intersection is more stable, and the congestion collapse characteristic is more resilient. In this case, the system is less sensitive and allows AGVs to enter. Ultimately... The model can learn, for the same intersection, the system load rate At very low levels, its congestion and crash characteristics are very low, while... When the value is very high, its congestion collapse characteristic will spike, indicating that the intersection is vulnerable.

[0044] Combined with the optimal predicted flow value and dynamic congestion index Precisely quantify the AGV path selection The potential costs of future congestion. The formula for calculating these costs is: In the formula, It is a path At any moment The cost of congestion; It is a path At any moment The optimal predicted flow value; It is a path Maximum physical capacity, representing the path The maximum physical throughput capacity, for example, a maximum of 3 AGVs can pass through per second, is obtained from the warehouse environment data collected in step S101. It is a very small positive number, such as To prevent the denominator from being 0; It is a path At any moment The dynamic congestion index. For example... Figure 2 It can be seen that the system is very sensitive to congestion during peak hours, and when the path saturation (i.e., traffic flow) exceeds 60%, it begins to increase its growth rate significantly; while during off-peak hours, up to 80%, the system is not sensitive to congestion, and the cost of congestion remains very low.

[0045] S104: Path planning and dynamic replanning.

[0046] This step integrates the aforementioned calculation results to plan the globally optimal path for the AGV and establishes an adaptive replanning mechanism to cope with state changes during operation. Using the strategy weights obtained in step S102, combined with the basic cost of the path and the congestion cost calculated in step S103, a global comprehensive cost is calculated for each path, using the following formula: In the formula, For path exist The cost of congestion at any time Let t be the time-sensitivity weight. Let be the energy consumption weight at time t. The throughput weight at time t It represents the basic time cost, which is the time cost of the AGV traversing the path under ideal conditions without considering congestion. The shortest time, determined by the path The physical length is obtained by dividing the AGV's rated cruising speed; This is the basic energy consumption cost, representing the path the AGV traverses under standard load. The required basic energy is determined by the path The length, slope, and basic power consumption model of the AGV were calculated.

[0047] Optimal path planning: Traditional A The search algorithm is a one-time path planner. When an AGV is executing a planned path, if a global state change occurs, such as a large influx of urgent orders, the AGV will still blindly execute the previously planned old path. However, this old path may no longer be the optimal path, ultimately resulting in extremely high costs. Therefore, the cost degradation degree of the remaining path for each AGV currently executing a task is calculated under the new global state. If the cost of the current path exceeds the replanning threshold due to the global state change, the system re-plans the path.

[0048] To quantify the increase in remaining path cost due to global state changes during AGV execution, a path cost degradation degree is defined. It is calculated by comparing the remaining path cost at the current moment with the expected cost during the original planning. The formula is as follows: In the formula, It is a moment The path cost degradation is used to quantify how much more cost the AGV will have to pay per step in the future compared to expectations. This is the current actual cost, in At that moment, for the remaining old paths The summation of the comprehensive costs yields the result; It is the original expected cost, representing the path calculated before the AGV departs. The overall cost; It is a path Length; It is a very small positive number, such as To prevent the denominator from being zero. Old path. The method of obtaining it is: if the AGV comes from Always set off, at At that moment, located at the node At point (a path consists of several nodes), the remaining old path is... .

[0049] It can be seen that by calculating the AGV's performance... Path cost degradation at any moment This allows us to quantify how much more cost the AGV will incur on average for each step it takes in the future compared to its initial expectations; if Explanation in At any given moment, the cost of AGVs is higher than... Time for remaining path The estimated cost is high.

[0050] The above steps yielded the AGV's performance. Path cost degradation at any moment However, simply based on Judging whether a path needs to be replanned based solely on whether it's greater than 0 is clearly unreasonable. If the cost increase is minimal, the overhead of replanning (such as the energy consumption for AGV emergency stops and turns) is counterproductive. Therefore, a replanning threshold needs to be defined. The system will only replan the path when the value exceeds this threshold. The formula for calculating the replanning threshold is: In the formula, It is a moment The replanning threshold represents the value at time t. The cost tolerance floor; It is the basic tolerance (50), which represents the basic tolerance cost of the system under normal conditions; , and These are order timeliness, AGV energy consumption, and system throughput at time points. The weights; This is the tolerance coefficient, which is set to 1.0 in this embodiment; This is the time sensitivity coefficient, which is set to 2.0 in this embodiment; This is the throughput sensitivity coefficient, which is set to 1.5 in this embodiment.

[0051] As can be seen from the formula, when the weight of time When the value is very high, such as a surge in VIP orders, the denominator increases, causing the threshold to be exceeded. As costs decrease, the system becomes extremely sensitive to cost increases, even if the cost degradation rate decreases. It's very unlikely that a replanning will be triggered. Conversely, when energy consumption weighting... At very high levels, the molecular size increases, and the threshold... The system allows AGVs to tolerate a certain degree of cost variation in order to save energy.

[0052] If the AGV is at time Path cost degradation Exceeding the adaptive replanning threshold The system will then repeat steps S101 to S103, calculate the globally optimal cost based on the latest global state, and call A again. The search algorithm finds a new globally optimal path for the AGV.

[0053] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A multi-objective path planning method for intelligent warehousing based on a large model, characterized in that, include: Obtain the state vector of the intelligent warehousing system, which represents the system's current timeliness pressure, energy consumption pressure, and throughput pressure; The state vector is input into a pre-trained Actor policy network to obtain the timeliness weight, energy consumption weight, and throughput weight corresponding to the timeliness pressure, energy consumption pressure, and throughput pressure. For each path in the map, the assigned traffic and historical traffic of the path are weighted and fused to obtain the optimal predicted traffic value. The dynamic congestion cost of the path is calculated by combining the type of the path and the throughput pressure. The timeliness weight, energy consumption weight, and throughput weight are multiplied by the base time cost, base energy consumption cost, and dynamic congestion cost of each path, and then summed to obtain the global comprehensive cost of each path. A is then used. The search algorithm plans the optimal path with the lowest overall cost for the AGV; The method for calculating the optimal predicted flow value is as follows: The historical prediction error of the path, the path type, the urgency index, and the system load rate are input into the pre-trained first neural network to obtain the dynamic confidence score. The optimal predicted flow value is obtained by multiplying the dynamic confidence level by the assigned flow rate, and adding the product of 1 minus the dynamic confidence level and the historical flow rate.

2. The intelligent warehousing multi-objective path planning method based on a large model according to claim 1, characterized in that, Obtaining the state vector includes: Obtain the total number of high-priority orders and orders that are about to expire, and divide it by the system's emergency order pressure threshold to obtain the urgency index; The average battery level of all online AGVs is obtained and normalized according to preset low battery threshold and health threshold to obtain battery health. Obtain the number of orders completed by the system within a preset period, and divide it by the historical peak throughput to obtain the system load rate; The urgency index, battery health, and system load rate are combined into the state vector.

3. The intelligent warehousing multi-objective path planning method based on a large model according to claim 1, characterized in that, The method further includes: For an AGV currently performing a task, calculate the new cost of its remaining path under the current global comprehensive cost, and the difference between it and the original cost at the planning time to obtain the path cost degradation degree; If the path cost degradation degree is greater than the preset adaptive replanning threshold, the method is repeated to replan the path for the AGV.

4. The intelligent warehousing multi-objective path planning method based on a large model according to claim 3, characterized in that, The adaptive replanning threshold is calculated as follows: The numerator is obtained by adding the base tolerance value to the energy consumption weight. The denominator is obtained by adding 1 to the sum of the weighted values ​​of timeliness weight and throughput weight, respectively. The ratio of the numerator to the denominator is used as the adaptive replanning threshold.

5. The intelligent warehousing multi-objective path planning method based on a large model according to claim 1, characterized in that, The method for calculating the dynamic congestion cost is as follows: The path type and system load rate are input into a pre-trained second neural network to obtain a dynamic congestion index. The path saturation is obtained by dividing the optimal predicted flow by the difference between the maximum capacity of the path and the optimal predicted flow value. The dynamic congestion cost is obtained by raising the value of the path saturation to the power of the dynamic congestion index.

6. The intelligent warehousing multi-objective path planning method based on a large model according to claim 1, characterized in that, The training method for the pre-trained Actor policy network includes: Using an Actor-Critic architecture, the state vector is input into the Critic value network to obtain the value prediction value. The comprehensive performance score is calculated based on the system throughput and the average order timeout rate as a reward signal. The temporal difference error between the reward signal and the value prediction is calculated based on the Bellman equation, and the parameters of the Actor policy network and the Critic value network are updated according to the temporal difference error.

7. The intelligent warehousing multi-objective path planning method based on a large model according to claim 6, characterized in that, The method for calculating the overall performance score is as follows: The comprehensive performance score is calculated by multiplying the system throughput by the preset business preference coefficient, minus the product of 1 and the difference between the business preference coefficient and the average order timeout rate.

8. The intelligent warehousing multi-objective path planning method based on a large model according to claim 1, characterized in that, The data sources for obtaining the state vector include: Order task list and business priorities obtained from the warehouse management system; The remaining power, working status, and real-time coordinates of each AGV are obtained from the AGV cluster management system, as well as map topology information and historical path traffic data obtained from the motion control system.

9. The intelligent warehousing multi-objective path planning method based on a large model according to claim 7, characterized in that, The preset value for the business preference coefficient is 0.7.

Images