Unmanned platform path planning method, device, equipment, medium and program product

Abstract

The application relates to an unmanned platform path planning method, device, equipment, medium and program product. The method comprises the following steps: acquiring static attribute vectors of each processing station in a processing station set, and acquiring dynamic attribute vectors of a current transportation device in the transportation device set of the unmanned platform in the case that the current transportation device travels to one processing station each time; adopting an encoder model in a target path planning model to process the static attribute vectors of each processing station, so as to obtain static feature vectors of each processing station; and adopting a decoder model in the target path planning model to process the static feature vectors of each processing station and the dynamic attribute vectors of the current transportation device, so as to obtain a next processing station to which the current transportation device travels. The method can improve the path planning efficiency of the transportation device, and further improve the transportation efficiency of the transportation device.

Description

Technical Field

[0001] This application relates to the field of path planning technology, and in particular to a path planning method, apparatus, equipment, medium and program product for an unmanned platform. Background Technology

[0002] In unmanned platforms, path planning for transportation devices is crucial, as its rationality directly determines transportation efficiency.

[0003] Path planning in related technologies mainly relies on heuristic optimization algorithms, among which particle swarm optimization (PSO) and genetic algorithms are the most widely used mainstream methods. These algorithms search for near-optimal paths in the solution space by simulating the evolutionary or group behavior mechanisms in nature. For example, PSO draws inspiration from the foraging behavior of bird flocks, finding the optimal solution through iterative updates of individual and group experience; while genetic algorithms are based on the principles of biological evolution, gradually optimizing path schemes through operations such as selection, crossover, and mutation.

[0004] However, in scenarios where transportation demand changes, the relevant technologies require retraining of the model to plan the travel path of the transportation device after the demand changes. This results in low path planning efficiency for the transportation device, which in turn leads to low transportation efficiency. Summary of the Invention

[0005] Based on this, this application provides a path planning method, apparatus, equipment, medium, and program product for unmanned platforms, which can improve the path planning efficiency of transportation devices and thus improve the transportation efficiency of transportation devices.

[0006] Firstly, this application provides a path planning method for an unmanned platform, the method comprising:

[0007] Obtain the static attribute vector of each processing station in the set of processing stations, and obtain the dynamic attribute vector of the current transportation device in the set of transportation devices of the unmanned platform each time it travels to a processing station.

[0008] The encoder model in the target path planning model is used to process the static attribute vectors of each processing station to obtain the static feature vectors of each processing station.

[0009] The decoder model in the target path planning model is used to process the static feature vectors of each processing station and the dynamic attribute vector of the current transportation device to obtain the next processing station to which the current transportation device will travel.

[0010] In some embodiments, the encoder model in the target path planning model is used to process the static attribute vectors of each processing station to obtain the static feature vectors of each processing station, including:

[0011] The context matrix is ​​determined based on the self-attention model in the encoder model and the static attribute vectors of each processing station;

[0012] Based on the context matrix and the expert networks of the hybrid expert model in the encoder model, determine the expert processing matrix;

[0013] The weights of each expert network are determined based on the context matrix and the gating network of the hybrid expert model.

[0014] Based on the processing matrices of each expert and the weights of each expert network, the static feature vectors of each processing station are determined.

[0015] In some embodiments, determining the context matrix based on the self-attention model in the encoder model and the static attribute vectors of each processing station includes:

[0016] The static attribute vectors of each processing station are processed using the first linear projection layer in the encoder model to obtain the linear projection vectors of each processing station.

[0017] The self-attention model in the encoder model is used to process the linear projection vectors of each processing station to obtain the context vectors of each processing station.

[0018] The context matrix is ​​determined based on the linear projection vector of each processing station and the context vector of each processing station.

[0019] In some embodiments, determining the static feature vector of each processing station based on each expert processing matrix and the weights of each expert network includes:

[0020] The weights of each expert network are used to process each expert processing matrix to obtain each weight processing matrix;

[0021] Based on the context matrix and each weight processing matrix, the static feature vector of each processing station is determined.

[0022] In some embodiments, a decoder model from the target path planning model is used to process the static feature vectors of each processing station and the dynamic attribute vector of the current transportation device to obtain the next processing station to which the current transportation device will proceed, including:

[0023] The dynamic attribute vector of the current transportation device is processed by the second linear projection layer in the decoder model to obtain the dynamic feature vector of the current transportation device;

[0024] The dynamic feature vector of the current transportation device and the static feature vector of the current processing station where the transportation device is currently located are concatenated to obtain the concatenated feature vector;

[0025] Based on the attention model in the decoder model, the static feature vectors of each processing station, and the concatenated feature vectors, the next processing station to which the current transport device will travel is determined.

[0026] In some embodiments, determining the next processing station to which the current transport device is headed, based on the attention model in the decoder model, the static feature vectors of each processing station, and the concatenated feature vector, includes:

[0027] The attention model in the decoder model is used to process the static feature vectors and concatenated feature vectors of each processing station to obtain the log probability of the current transportation device heading to each processing station in the set of processing stations.

[0028] The next processing station to which the current transport device will travel is determined based on the logarithmic probability of each processing station in the set of processing stations.

[0029] In some embodiments, the attention model in the decoder model is used to process the static feature vectors and concatenated feature vectors of each processing station to obtain the log probability of the current transport device heading to each processing station in the set of processing stations, including:

[0030] The key linear projection layer and value linear projection layer of the attention model in the decoder model are used to process the static feature vectors of each processing station to obtain the key matrix and value matrix.

[0031] The query linear projection layer of the attention model in the decoder model is used to process the concatenated feature vectors to obtain dynamic query vectors;

[0032] The probabilistic generative network of the attention model in the decoder model is used to process the key matrix, value matrix and dynamic query vector to obtain the log probability of the current transportation device heading to each processing station in the set of processing stations.

[0033] In some embodiments, determining the next processing station to which the current transport device will travel, based on the logarithmic probability of the current transport device traveling to each processing station in the set of processing stations, includes:

[0034] Based on the processing stations that the current transport device has already passed and the log probability of the current transport device heading to each processing station in the set of processing stations, determine the probability of the current transport device heading to each processing station in the set of processing stations.

[0035] The processing station corresponding to the action with the lowest probability is determined as the next processing station to which the current transport device will head.

[0036] In some embodiments, the method further includes:

[0037] Obtain the initial path planning model and scale list; the scale list includes the number of different processing stations.

[0038] The initial path planning model is iteratively trained using training data for each number of processing stations to obtain the target path planning model. The training data for each number of processing stations includes the static attribute vectors of each processing station, multiple starting positions of the transportation device, and the transportation path of the transportation device at each starting position.

[0039] In some embodiments, training data for each number of processing stations is used to iteratively train the initial path planning model to obtain the target path planning model, including:

[0040] The transportation path corresponding to each starting point is determined by the iterative path planning model;

[0041] Determine the total revenue of the transportation route corresponding to each starting point location based on the transportation route corresponding to each starting point location.

[0042] The loss value is determined based on the number of starting points among multiple starting points, the transportation route corresponding to each starting point, and the total revenue of the transportation route corresponding to each starting point.

[0043] If the loss value does not meet the iteration termination condition, continue iterating the path planning model and repeat the above steps until the loss value meets the iteration termination condition, at which point the path planning model of the last iteration is determined as the target path planning model.

[0044] In some embodiments, the total revenue of the transportation route corresponding to each starting point location is determined based on the transportation route corresponding to each starting point location, including:

[0045] Based on the transportation route corresponding to each starting point, determine the total transportation distance and the latest service time for each processing station;

[0046] The total revenue of the transportation route corresponding to each starting point is determined based on the total transportation distance, the latest service time of each processing station, and the service time limit of each processing station.

[0047] Secondly, this application provides an unmanned platform path planning device, the device comprising:

[0048] The acquisition module is used to acquire the static attribute vectors of each processing station in the processing station set, and the dynamic attribute vector of the current transportation device in the unmanned platform's transportation device set each time it travels to a processing station.

[0049] The encoding processing module is used to process the static attribute vectors of each processing station using the encoder model in the target path planning model, so as to obtain the static feature vectors of each processing station.

[0050] The decoding module is used to process the static feature vectors of each processing station and the dynamic attribute vectors of the current transportation device using the decoder model in the target path planning model, so as to obtain the next processing station to which the current transportation device is heading.

[0051] Thirdly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect.

[0052] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method of any one of the first aspects.

[0053] Fifthly, this application provides a computer program product, including a computer program, wherein when the computer program is executed by a processor, it implements the steps of the method in any of the first aspects. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating a path planning method for an unmanned platform provided in some embodiments;

[0056] Figure 2 A schematic diagram illustrating the processing steps corresponding to the unmanned platform path planning method provided in some embodiments;

[0057] Figure 3 A flowchart illustrating a method for determining static feature vectors provided in some embodiments;

[0058] Figure 4 A flowchart illustrating a method for determining the next processing station to which the current transport device is headed, provided in some embodiments;

[0059] Figure 5 A flowchart illustrating a method for determining a target path planning model, provided for some embodiments;

[0060] Figure 6A schematic diagram of the target path planning model provided for some embodiments;

[0061] Figure 7 A schematic diagram of the unmanned platform path planning device provided in some embodiments;

[0062] Figure 8 A schematic diagram of the structure of a computer device provided for some embodiments. Detailed Implementation

[0063] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0065] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, "multiple groups" means two or more, and "each" means each of the multiple, unless otherwise explicitly defined.

[0066] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0067] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0068] Unless otherwise specified, the order of execution steps in the embodiments of this application is not limited. It should also be noted that any step in the embodiments of this application can be executed independently, that is, the execution of any step in the above embodiments can be performed without depending on the execution of other steps.

[0069] In order to at least address the problem of low route planning efficiency of transportation devices in related technologies, which leads to low transportation efficiency of transportation devices, this application provides the following route planning method.

[0070] The path planning method in this application embodiment can be applied to computer devices or processors. A processor can be applied to a computer device. Exemplarily, a computer device may include one or a combination of at least two of the following: a server, a mobile phone, a tablet computer, a computer with transceiver capabilities, a handheld computer, a desktop computer, a personal digital assistant, a portable media player, a smart speaker, a navigation device, a smartwatch, smart glasses, a smart necklace, and other wearable devices, a pedometer, a digital TV, a virtual reality (VR) device, an augmented reality (AR) device, devices in industrial control, devices in self-driving, devices in remote medical surgery, devices in a smart grid, devices in transportation safety, devices in a smart city, devices in a smart home, a vehicle, in-vehicle equipment, in-vehicle modules, etc.

[0071] Figure 1 A flowchart illustrating the path planning method for an unmanned platform provided in some embodiments, such as... Figure 1 As shown, the method includes the following steps:

[0072] S101. Obtain the static attribute vector of each processing station in the processing station set, and obtain the dynamic attribute vector of the current transportation device in the unmanned platform's transportation device set each time it travels to a processing station.

[0073] For example, the processing station set includes multiple processing stations, and the transport device can travel to each processing station in the processing station set. At each processing station, the transport device can perform corresponding business processing. For example, the processing station set includes a warehouse and multiple service stations, and the transport device can transport loads from each service station to the warehouse, or the transport device can transport loads from the warehouse to each service station, or the transport device can transport loads from some service stations to other service stations. In some embodiments, the transport device can go to the warehouse to recharge or replace batteries. For example, the multiple service stations can be included in a service area. In this example, the index for the repository is 0, and the indexes for multiple service sites are... N is an integer greater than or equal to 2.

[0074] The load in this embodiment may include any transportable item, such as goods. Load processing may include loading the load from the processing station onto a transport device, or transferring the load from the transport device to the processing station.

[0075] For example, the transportation device may include a drone. Also for example, the transportation device may include a transport vehicle, delivery vehicle, truck, or other vehicle. This application does not limit the type of transportation device.

[0076] The static attribute vector for each processing station may include a vector composed of multiple static attribute values ​​for each processing station. For example, the static attribute vector may include at least one of the following: a horizontal coordinate (e.g., X-axis coordinate), a vertical coordinate (e.g., Y-axis coordinate), the load to be transported by the transport device, the load processing duration (e.g., the duration from the start of load processing to the completion of load processing by the transport device), the start time of the service time window, and the end time of the service time window. Here, the service time window is the time window during which the transport device is allowed to process the load. For example, the service time window for the i-th processing station is... This indicates that when the transport device is at the i-th processing station, within the service time window... Load handling is performed internally. For example, in transport devices... If a process arrives at the i-th processing station before time i, it may need to wait until time i > i. Only then can load processing be performed. For example, ,in, This indicates the time period for the operational plan.

[0077] The set of transport devices may include at least one transport device. The current transport device can be any transport device in the set of transport devices.

[0078] The dynamic attribute vector of a transportation device can be a vector composed of multiple dynamic attribute values ​​of the transportation device. For example, the dynamic attribute vector can include at least one of the following: current time, remaining load (representing the load capacity or weight that the transportation device can still load), and remaining energy (representing the remaining energy of the transportation device). Remaining energy refers to the amount of energy that can be effectively utilized in the device's current energy storage unit and is a parameter for measuring the device's endurance. In the embodiments of this application, the dynamic attribute vector of the current transportation device is different when the current transportation device arrives at different processing stations.

[0079] S102. Using the encoder model in the target path planning model, the static attribute vectors of each processing station are processed to obtain the static feature vectors of each processing station.

[0080] In some embodiments, the encoder model in the target path planning model may include a first neural network model, which processes the static attribute vectors of each processing station to obtain the static feature vectors of each processing station. The neural network model can be any model capable of vector transformation in related technologies, and this application does not limit the specific implementation of the neural network model.

[0081] In other embodiments, the encoder model in the target path planning model may include a first linear projection layer, which is used to process the static attribute vectors of each processing station to obtain the static feature vectors of each processing station.

[0082] In some embodiments, the dimension of the static feature vector can be greater than the dimension of the static attribute vector. For example, the dimension of the static adjustment vector can range from 32 to 512. For example, the dimension of the static adjustment vector can be 32, 64, 128, 256, or 512, etc.

[0083] S103. Using the decoder model in the target path planning model, the static feature vectors of each processing station and the dynamic attribute vector of the current transportation device are processed to obtain the next processing station to which the current transportation device will travel.

[0084] In some embodiments, the decoder model in the target path planning model may include a second neural network model, which inputs the static feature vectors of each processing station and the dynamic attribute vector of the current transportation device into the second neural network model so that the second neural network model outputs the next processing station to which the current transportation device is headed.

[0085] In the technical solution provided in this application embodiment, a decoder model in the target path planning model is used to process the static feature vectors of each processing station and the dynamic attribute vectors of the current transportation device to obtain the next processing station to which the current transportation device is heading. Thus, in scenarios where transportation demand changes, there is no need to retrain the model. Instead, the path decision is completed based on the dynamic attribute vectors of the current transportation device obtained in real time and the static feature vectors of each processing station obtained in advance, which significantly improves the path planning efficiency of the transportation device and thus improves the transportation efficiency of the transportation device.

[0086] Taking transportation devices including drones and processing station sets including warehouses and multiple logistics distribution points as an example, logistics distribution points are one way to implement service stations. Figure 2 The diagram illustrates the processing steps corresponding to the unmanned platform path planning method provided in some embodiments, such as... Figure 2 As shown, static attribute vectors (i.e., static features) for warehouses and various logistics distribution points can be obtained in advance. These static attribute vectors include spatial topology location, time windows, and user demand. Spatial topology location can include horizontal and vertical coordinates. Time windows can include load processing duration, the start time of the service time window, and the end time of the service time window. User demand can include payloads that need to be transported by drones.

[0087] Each time a drone in the drone ensemble travels to a processing station, its dynamic attribute vector (i.e., dynamic characteristics) can be obtained. This dynamic attribute vector can include the current time, remaining capacity, and remaining energy.

[0088] The static attribute vectors of the warehouse and each logistics distribution point are input into the encoder model (also called the encoder), so that the encoder outputs static feature vectors of the warehouse and each logistics distribution point. The encoder model can include a Mixture of Experts (MoE) based encoder. The static feature vectors of the warehouse and each logistics distribution point are statically embedded (i.e., input) into the decoder model (also called the decoder), so that the decoder model outputs the next processing station to which the current drone should go, based on the current drone's dynamic attribute vector and the static feature vectors of the warehouse and each logistics distribution point, so that the current drone can travel along the planned drone trajectory.

[0089] Figure 3 A flowchart illustrating a method for determining static feature vectors provided in some embodiments, such as... Figure 3 As shown, this method is an explanation of S102, and includes the following steps:

[0090] S1021. Determine the context matrix based on the self-attention model in the encoder model and the static attribute vectors of each processing station.

[0091] Self-attention models are mechanisms that allow a model to assign different attention weights to different parts of the input data. Their core function is to capture the relationships between elements within the input sequence, thereby extracting effective features more accurately. For a description of self-attention models, please refer to the descriptions of related technologies; this application's embodiments will not elaborate further.

[0092] In some embodiments, the self-attention model in the encoder model can be used to process the static attribute vectors of each processing station to obtain the context matrix.

[0093] S1022. Determine the expert processing matrix based on the context matrix and the expert networks of the hybrid expert model in the encoder model.

[0094] A hybrid expert model is a neural network architecture that uses a gating network to dynamically select a small subset of subnetworks, known as experts, for each input, thereby improving model capacity and computational efficiency through sparse activation. For a description of hybrid expert models, please refer to the descriptions of related technologies; this application's embodiments will not elaborate further.

[0095] In some embodiments, the context matrix can be input into each expert network of the hybrid expert model so that each expert network outputs its own expert processing matrix.

[0096] In other embodiments, given the context matrix, it can be normalized to obtain a first normalized matrix; the first normalized matrix is ​​then input into each expert network of the hybrid expert model so that each expert network outputs its own expert processing matrix.

[0097] S1023. Determine the weights of each expert network based on the context matrix and the gating network of the hybrid expert model.

[0098] In some embodiments, the context matrix can be input into a gated network of a hybrid expert model so that the gated network outputs the weights of each expert network.

[0099] In other embodiments, a first normalized matrix is ​​input into a gating network of the hybrid expert model so that the gating network outputs the weights of each expert network.

[0100] S1024. Determine the static feature vector of each processing station based on the processing matrix of each expert and the weight of each expert network.

[0101] In some embodiments, the weights of each expert network can be used to process each expert processing matrix separately (e.g., by corresponding multiplication) to obtain the static feature vector of each processing station.

[0102] In the technical solution provided in this application, a context matrix is ​​determined by the self-attention model in the encoder model and the static attribute vectors of each processing station. This accurately captures the correlation features between the static attributes of each processing station, providing basic data rich in global correlation information for static feature extraction. Based on this context matrix, corresponding expert processing matrices are generated by the expert networks of the hybrid expert model, enabling targeted refinement of the attribute features of different types of stations in the context matrix. At the same time, the weights of each expert network are determined by the gating network in conjunction with the context matrix, which can adaptively match the processing requirements of the attribute features of each processing station and select suitable expert processing capabilities. Finally, the static feature vectors of each processing station are determined by combining each expert processing matrix with the corresponding weights, which can significantly improve the accuracy and relevance of the static feature vector representation. This provides high-quality feature support for the subsequent decoder model to fuse dynamic attribute vectors for path planning, ensuring the accuracy of the decision for the next processing station and thus improving the path planning efficiency of the transportation device.

[0103] In some embodiments, determining the context matrix based on the self-attention model in the encoder model and the static attribute vectors of each processing station includes: processing the static attribute vectors of each processing station using a first linear projection layer in the encoder model to obtain linear projection vectors of each processing station; processing the linear projection vectors of each processing station using the self-attention model in the encoder model to obtain context vectors of each processing station; and determining the context matrix based on the linear projection vectors of each processing station and the context vectors of each processing station.

[0104] The linear projection layer is a fundamental network layer that uses linear transformations to map feature dimensions or transform feature spaces. Its core function is to project the input feature vector onto the target dimension space for subsequent computation by the adaptation model. A description of the linear projection layer can be found in related technical descriptions, and will not be repeated here. For example, the dimension of the linear projection vector is the same as the dimension of the static feature vector.

[0105] For example, determining the context matrix based on the linear projection vector of each processing station and the context vector of each processing station may include: adding the corresponding linear projection vector of each processing station and the context vector of each processing station to obtain the context matrix.

[0106] In the technical solution provided in this application embodiment, the static attribute vectors of each processing station are processed by the first linear projection layer of the encoder model to generate linear projection vectors with uniform dimensions. This can eliminate the interference of the difference in the dimension of attribute vectors of different stations on subsequent calculations and provide adaptable basic features for feature extraction by the self-attention model. Then, by using the self-attention model to process the linear projection vectors, the correlation features between each processing station are accurately captured, generating a context vector containing global correlation information of the stations. Finally, the linear projection vectors and the context vectors are combined to construct a context matrix. This matrix not only retains the original attribute features of the processing stations but also incorporates the correlation features between stations, greatly improving the completeness and accuracy of feature representation. This lays a solid foundation for the subsequent generation of high-quality static feature vectors by the hybrid expert model, thereby improving the decision-making efficiency and accuracy of the transportation device path planning.

[0107] In some embodiments, determining the static feature vector of each processing station based on each expert processing matrix and the weights of each expert network includes: processing each expert processing matrix using the weights of each expert network to obtain each weight processing matrix; and determining the static feature vector of each processing station based on the context matrix and each weight processing matrix.

[0108] For example, the weights of each expert network are multiplied by the corresponding expert processing matrix to obtain each weight processing matrix.

[0109] In some implementations, the context matrix and each weight processing matrix can be added together to obtain the static feature vector of each processing station.

[0110] In other implementations, the context matrix and each weight processing matrix can be added together to obtain the target feature vector of each processing station; the matrix composed of the target feature vectors of each processing station can be normalized to obtain the static feature vector of each processing station.

[0111] In the technical solution provided in this application embodiment, each expert processing matrix is ​​processed by the weights of each expert network to obtain each weight processing matrix. This enables differentiated weighting of the output features of different expert networks, strengthens the effective information of the static attribute features of the adapted processing station, and weakens redundant interference information. Then, by combining the context matrix and each weight processing matrix, the static feature vector of each processing station is determined. This allows the final generated static feature vector to retain both the global correlation features between processing stations in the context matrix and the targeted processing features of the expert model after weight filtering. This significantly improves the accuracy and effectiveness of the static feature vector representation, providing high-quality feature support for the decoder model to fuse the dynamic attribute vectors of the transportation device for path planning. This ensures the accuracy of the decision for the next processing station and improves the efficiency of the transportation device's path planning.

[0112] Figure 4 A flowchart illustrating a method for determining the next processing station to which the current transport device is headed, provided in some embodiments, is an explanation of step S103. The method includes the following steps:

[0113] S1031. The dynamic attribute vector of the current transportation device is processed by the second linear projection layer in the decoder model to obtain the dynamic feature vector of the current transportation device.

[0114] The dynamic feature vector has the same dimension as the static feature vector.

[0115] S1032. The dynamic feature vector of the current transportation device and the static feature vector of the current processing station where the current transportation device is located are concatenated to obtain the concatenated feature vector.

[0116] S1033. Based on the attention model in the decoder model, the static feature vectors of each processing station, and the concatenated feature vectors, determine the next processing station to which the current transport device will travel.

[0117] In the technical solution provided in this application embodiment, the dynamic attribute vector of the current transportation device is processed by the second linear projection layer of the decoder model to generate a dynamic feature vector whose dimension is adapted to subsequent calculations. This eliminates the dimensional difference between the dynamic attribute vector and the static feature vector, improving the effectiveness of feature fusion. By splicing the dynamic feature vector of the current transportation device with the static feature vector of the current processing station, a spliced ​​feature vector is obtained, achieving a preliminary fusion of the real-time state of the transportation device and the inherent attributes of the current processing station. Then, with the help of the attention model of the decoder model, the next processing station is determined by combining the static feature vectors of each processing station with the spliced ​​feature vector. This allows the attention model to accurately focus on the target station features that best match the current transportation state and the current station, improving the pertinence and accuracy of the next station decision. As a result, the model can quickly adapt to changes in transportation state without retraining, significantly improving the path planning efficiency of the transportation device.

[0118] In some embodiments, determining the next processing station to which the current transport device will travel, based on the attention model in the decoder model, the static feature vectors of each processing station, and the concatenated feature vectors, includes: processing the static feature vectors and concatenated feature vectors of each processing station using the attention model in the decoder model to obtain the log probability of the current transport device traveling to each processing station in the set of processing stations; and determining the next processing station to which the current transport device will travel, based on the log probability of the current transport device traveling to each processing station in the set of processing stations.

[0119] In the technical solution provided in this application embodiment, the static feature vectors and concatenated feature vectors of each processing station are processed by the attention model of the decoder model to obtain the log probability of the current transportation device heading to each processing station. The matching relationship between the current state of the transportation device and the inherent attributes of each station can be transformed into a quantitative probability index, realizing an accurate measurement of the adaptability of each candidate station. Then, based on the log probability, the next processing station is determined. The decision screening can be completed by relying on the quantitative probability, avoiding the subjectivity and limitations of human experience decision-making, ensuring that the selected station is highly consistent with the real-time state of the transportation device, significantly improving the accuracy and scientific nature of the path planning decision, and quickly outputting the optimal station decision without retraining the model, effectively improving the path planning efficiency of the transportation device.

[0120] In some embodiments, the attention model in the decoder model is used to process the static feature vectors and concatenated feature vectors of each processing station to obtain the log probability of the current transport device heading to each processing station in the set of processing stations. This includes: processing the static feature vectors of each processing station using the key linear projection layer and value linear projection layer of the attention model in the decoder model to obtain the key matrix and value matrix; processing the concatenated feature vectors using the query linear projection layer of the attention model in the decoder model to obtain the dynamic query vector; and processing the key matrix, value matrix, and dynamic query vector using the probabilistic generation network of the attention model in the decoder model to obtain the log probability of the current transport device heading to each processing station in the set of processing stations.

[0121] The probabilistic generative network is the network remaining in the attention model after removing the key linear projection layer, the value linear projection layer, and the query linear projection layer.

[0122] In the technical solution provided in this application embodiment, the static feature vectors of each processing station are processed by the key linear projection layer and value linear projection layer of the attention model in the decoder model, respectively, to generate dimension-adapted key and value matrices, providing a standardized feature basis for attention calculation. The dynamic query vector is obtained by processing and splicing feature vectors through the query linear projection layer, realizing the dimensionality unification of the real-time status of the transportation device and the attribute features of the current processing station. Then, the key matrix, value matrix and dynamic query vector are processed collaboratively by the probabilistic generation network to output the log probability of the current transportation device heading to each processing station, so as to accurately measure the degree of adaptation of each processing station, thereby greatly improving the accuracy of path planning decisions.

[0123] In some embodiments, determining the next processing station to which the current transport device will travel based on the logarithmic probability of the current transport device traveling to each processing station in the set of processing stations includes: determining the action probability of the current transport device traveling to each processing station in the set of processing stations based on the processing stations the current transport device has already traveled to and the logarithmic probability of the current transport device traveling to each processing station in the set of processing stations; and determining the processing station corresponding to the lowest action probability as the next processing station to which the current transport device will travel.

[0124] For example, the log probability of a processing station already visited is set to the log probability of the current transport device reaching each processing station in the set of processing stations. The target probability of the current transport device heading to each processing station in the set of processing stations is obtained; the formula is used. Determine the probability of the current transport device heading to each processing station in the set of processing stations. Among these, This indicates that the current transportation device selects a processing station at step t. The probability of the action; This represents the actions from step 1 to step t-1, i.e., the selected path in the first t-1 steps. , indicates the processing sites selected from step 1 to step t-1; s represents the problem instance, indicating the optimization problem under the specified number of service sites and the load requirements of each service site; Select a processing station for the current transport device at step t. The target probability; Select a processing station for the current transport device at step t. The target probability; Belongs to the processing site collection In this embodiment, the t-th step decision represents the processing station selected in the t-th step.

[0125] In the technical solution provided in this application, the action probability is determined by combining the processing stations that the current transportation device has already passed with the logarithmic probability of heading to each processing station. This allows for the incorporation of path travel history constraints on the basis of quantitative matching, effectively avoiding invalid decisions that repeatedly plan for stations already passed, and enhancing the guiding value of probability indicators for actual path planning. The processing station corresponding to the lowest action probability is determined as the next processing station. This allows for the direct selection of the target station with the best suitability based on quantitative indicators, eliminating the need for complex multi-round deduction steps, significantly accelerating the path decision-making speed, greatly improving the path planning efficiency of the transportation device, and ensuring the rationality of the planned path, while avoiding transportation delays caused by ambiguity in decision-making.

[0126] In some embodiments, the path planning method may further include a step of determining a target path planning model. Figure 5A flowchart illustrating a method for determining a target path planning model is provided for some embodiments, such as... Figure 5 As shown, the method includes the following steps:

[0127] S501. Obtain the initial path planning model and scale list; the scale list includes the number of different processing stations.

[0128] The structure of the initial path planning model is the same as that of the target path planning model.

[0129] For example, the number of each processing site in the plurality of processing sites may range from 5 to 100. For example, the plurality of processing sites may include at least two of the following: 5, 10, 20, 50, 80, and 100.

[0130] S502. Using the training data for each number of processing stations, iteratively train the initial path planning model to obtain the target path planning model. The training data for each number of processing stations includes the static attribute vector of each processing station, multiple starting positions of the transportation device, and the transportation path of the transportation device at each starting position.

[0131] Iterative training of the initial path planning model involves training the parameters in the initial path planning model to obtain the target path planning model.

[0132] For example, because the multiple starting points of the transport device are different, the dynamic attribute vector of the transport device is different each time it travels to a processing station.

[0133] The technical solution provided in this application adopts a multi-scale hybrid training strategy. By obtaining an initial path planning model and a scale list containing different numbers of processing stations, the target model is obtained by iteratively training the initial model based on the training data under each number of processing stations. This abandons the training mode of single-scale problem instances and avoids the problem that the model is usually trained on a specific scale of problem, and the model performance will significantly decrease once the problem scale (such as the number of processing stations) changes. Therefore, this application embodiment can enhance the adaptability of the model to different numbers of processing stations and improve the reliability of transportation device path planning.

[0134] In some embodiments, training data for each number of processing stations is used to iteratively train an initial path planning model to obtain a target path planning model. This includes: determining the transportation path corresponding to each starting point position based on the iterative path planning model; determining the total revenue of the transportation path corresponding to each starting point position based on the transportation path corresponding to each starting point position; determining a loss value based on the number of starting points among multiple starting points, the transportation path corresponding to each starting point position, and the total revenue of the transportation path corresponding to each starting point position; and continuing to iterate the path planning model and repeating the above steps if the loss value does not meet the iteration termination condition, until the path planning model of the last iteration is determined as the target path planning model if the loss value meets the iteration termination condition.

[0135] In some embodiments, determining the loss value based on the number of starting points among multiple starting points, the transportation path corresponding to each starting point, and the total revenue of the transportation path corresponding to each starting point may include: determining the probability of selecting the transportation path corresponding to each starting point in each iteration using an iterative path planning model based on the transportation path corresponding to each starting point; and determining the loss value based on the number of starting points among multiple starting points, the transportation path corresponding to each starting point, and the total revenue of the transportation path corresponding to each starting point.

[0136] For example, the loss value can be obtained through the following loss function. Determined. Loss function. The method for determining it is as follows: ;in, Indicates the number of starting positions among multiple starting positions; Indicates the k-th starting point; This represents the decision made at step t; Represents the payoff function. , This represents the total revenue of the transportation route corresponding to the starting point k. This represents the probability that, at step t, the iterative path planning model at starting point k selects the transportation path corresponding to starting point k in the training sample.

[0137] In the technical solution provided in this application embodiment, the transportation path corresponding to each starting point position is determined by an iterative path planning model. Then, the corresponding total revenue is calculated based on each path. The loss value is determined by combining the number of starting points, each transportation path, and the total revenue. The model is continuously iterated until the loss value reaches the target to obtain the target path planning model. Thus, the total revenue of the transportation path is the core optimization guide. The model parameters are continuously corrected through multiple rounds of iteration, which effectively improves the rationality of the revenue of the model output path and the planning accuracy. In addition, by performing inference from multiple different starting points of an instance at the same time, the policy gradient is optimized, training convergence is accelerated, and the policy quality is improved.

[0138] In some embodiments, determining the total revenue of the transportation path corresponding to each starting point location based on the transportation path corresponding to each starting point location includes: determining the total transportation distance and the latest service time of each processing station based on the transportation path corresponding to each starting point location; and determining the total revenue of the transportation path corresponding to each starting point location based on the total transportation distance, the latest service time of each processing station, and the service time limit of each processing station.

[0139] For example, the total revenue of the transportation route corresponding to each starting point location. It can be determined using the following formula: Among them, arc segment Distance between, arc segment Indicates processing station i and processing station j; This represents the decision variable, if the transportation device selects the travel arc segment. ,but =1, if the transport device does not select a travel arc segment ,but =0; It is the penalty coefficient; ; This is the actual service time for processing site i. This is the upper limit of the service time for processing site i.

[0140] In the technical solution provided in this application, the total transportation distance and the latest service time of each processing station are determined by the transportation path corresponding to each starting point. Then, the total revenue of the corresponding transportation path is calculated by combining the service time limit of each processing station. The hard time window constraint is relaxed into a soft constraint and included in the objective function as a penalty term. This effectively solves the problems of inaccurate modeling and difficulty in strictly implementing strict delivery time windows in the prior art. When the model faces the problem of excessive load on the logistics network caused by sudden and unpredictable user demand, it can make the optimal trade-off in the unavoidable congestion or delay scenario. It not only ensures the efficiency of the path by optimizing the total transportation distance, but also reduces the impact of service timeout through the soft penalty mechanism, which significantly improves the practicality and robustness of the solution, thereby ensuring the stability and adaptability of the transportation device path planning.

[0141] In some embodiments, multiple service stations belong to a two-dimensional service area. ,in This represents its length and width in meters. The operational planning time domain is represented as... For example, the set of processing sites includes a single repository (vertex 0, i.e., index 0) and One service site, indexed as The vertex set (i.e., the index set) is defined as follows: Each service site Features include demand (The required load capacity or mass of the transport equipment) and hard service time window ,in The duration of on-site service is expressed as follows: If the transport device is If you arrive earlier, you may need to wait. For example, suppose there is an infinite number of isomorphic transport units, each with a payload capacity. and single-route energy budget .

[0142] For example, the energy consumption model for transportation devices considers power consumption during travel and parking (e.g., hovering power consumption for drones) and is linearly related to instantaneous load. Let , Indicates instantaneous load. Indicates maximum load. Driving power. and hovering power They are respectively: , ,in, and These are baseline driving power and baseline hovering power, respectively. and All are incremental power per unit of payload. For example, a warehouse can be set up, and the airborne payload increases along the route due to pickups (e.g., the core task of the transport device in this embodiment is pickup, i.e., departing from the warehouse, going to various service stations to pick up goods, and then returning to the warehouse). Therefore, for example, assuming the transport device only picks up goods at service stations, let... Indicates that the vertex service is complete. The weight delivered afterward, therefore, during the travel arc segment The effective payload during the period is For travel time arc segment , At the driving speed, its driving energy consumption is: Similarly, for arc segments At the vertex The hovering energy consumption is: Therefore, this application can be formalized as a multi-path vehicle picking problem with the objective of minimizing the total driving distance and the number of timeout tasks.

[0143] In this embodiment of the application, when the transportation device is an aircraft such as a drone, driving can be understood as flying.

[0144] For example, the path planning method in the embodiments of this application may use a Markov decision process.

[0145] In the decision-making process, let This indicates the decision-making steps until the task terminates. In the steps... The agent selects an unserved customer or chooses to return to the warehouse early.

[0146] In the state space, the state Defined as ,in Represents the graph structure and provides static site characteristics. , These represent the horizontal and vertical coordinates of processing station i, the load to be transported by the transportation device, the load processing time, the start time of the service time window, and the end time of the service time window, respectively. It is the index t of the current service site; It is the current time of the transportation device's decision at step t; It is the effective load of the transportation device in step t; It is the remaining energy of the transportation device at step t; This is the access indication vector (1 if the site is already served). This state exhibits the Markov property, meaning the state transition depends only on the current state. And the action selected in the current step.

[0147] In action space In this context, the allowed set of actions is defined as: in It is a feasibility indicator function, determined by the energy and capacity constraints of the transportation equipment:

[0148] in, Let t represent the set of all legal actions that the transportation device can choose at step t (the actions are the processing stations to be visited next). N represents the service sites to be served, and N is the set of all service sites with indices 1, 2, ..., N. These are conditions for feasibility assessment. A value of 1 indicates that service site j is currently selectable, while a value of 0 indicates that it is not selectable; This indicates that the action set also includes returning to the warehouse. Station 0 is the warehouse, meaning that the transport device can choose not to go to the new service station, but instead return directly to the warehouse.

[0149] This indicates the load capacity requirement of service station j (e.g., the load capacity / weight that needs to be transported by load devices). This represents the effective load of the transport device at step t, i.e., the capacity or weight of the loaded goods; The maximum effective load of the transport device.

[0150] This represents the remaining energy of the transportation device at step t; This represents the flight energy consumption from index t of the current service station to the next processing station j, which is related to load and distance; This represents the energy consumption of the service at service site j; it is determined based on service duration, waiting time, and load. This represents the flight energy consumption when returning to the warehouse from service station j.

[0151] In real-world scenarios, the sudden and unpredictable nature of user demand often leads to overloaded logistics networks, making service delays for some users unavoidable. To address this challenge, for example, the hard time window constraint is relaxed into a soft constraint and incorporated as a penalty term into the objective function. Therefore, the reward function is set as follows: The reward function enables the path planning model to learn how to avoid service latency.

[0152] The path planning model in this application embodiment is an autoregressive policy network with an encoder-decoder architecture. This network is mainly composed of an encoder based on a hybrid expert network and a self-attention mechanism, and a decoder based on an attention mechanism.

[0153] Figure 6 The diagram below illustrates the structure of a target path planning model provided in some embodiments. This target path planning model can be an autoregressive policy network with an encoder-decoder architecture. The network is mainly composed of an encoder based on a hybrid expert network and a self-attention mechanism, and a decoder based on an attention mechanism.

[0154] Among them, the autoregressive strategy is used for a problem instance. The solution is a site set A sequence (path) on is represented as . All stations on the planned route The term appears exactly once in the middle, while site 0 (warehouse) can appear multiple times. In each step... For example, according to the example and previous actions Calculate the probability of all possible actions. To exclude already visited sites, a masking mechanism is used, for example, to mask previously selected sites (belonging to...). The probability of is set to zero. For example, before decoding, the static embedding of the entire problem is first computed. For each site... Its input representation is a concatenated static feature vector. The initial site embedding is obtained, exemplarily, through a linear projection layer. (i.e., the linear projection vectors of each processing station), where the embedding dimension After processing by the encoder, for example, the global context embedding is obtained: (i.e., the static feature vectors of each processing station).

[0155] The decoder's input includes: the static embedding of the encoder's output. Current processing site Embedded The current environment's dynamic features (i.e., the dynamic attribute vector of the current transportation device) represent the current time, remaining capacity, and remaining energy, respectively. For example, the log probabilities (logits) of each feasible station at step t are calculated in an autoregressive manner using the attention model in the decoder model: , The masking mechanism generates the logarithmic probability after masking, where... Given a problem instance and previous actions Select a site in step t The action probability is calculated using the softmax function. .

[0156] The entire solution trajectory is decomposed into an autoregressive solution form. .in This represents all learnable parameters in the policy network. Wherein, This represents the overall probability that the path planning model generates path π under problem instance s; These are the parameters of the path planning model; This represents the selected path in problem instance s and the previous t-1 steps. Under the given conditions, the processing site is selected in step t. The conditional probability.

[0157] The encoder employs a hybrid expert-enhanced graph attention network. The encoder receives static features (coordinates, demand, time window, service duration) from all stations and embeds them through a multi-layer graph attention mechanism (i.e., a self-attention model within the encoder model). The feedforward network after each graph attention layer is replaced by a hybrid expert module. This module contains multiple expert networks. And a gated network. Embedded for each site The gating network calculates its adaptation weights with each expert. The final output is the weighted sum of the experts. ; This represents the output matrix of the k-th expert network. This allows the model to adaptively extract key topological features for problems of different scales.

[0158] For example, the static attribute vectors (i.e., static features) of each processing station are processed using the first linear projection layer in the encoder model to obtain the linear projection vectors of each processing station. The self-attention model (i.e., multi-head self-attention mechanism) in the encoder model is used to process the linear projection vectors of each processing station to obtain the context vectors of each processing station. When the self-attention model in the encoder model outputs the context vectors of each processing station, the linear projection vectors and context vectors of each processing station are added together to obtain a context matrix. This context matrix is ​​then normalized (i.e., layer normalization, which normalizes the context matrix obtained by each transportation device at each decision step) to obtain a first normalized matrix. This first normalized matrix is ​​input into each expert network (e.g., expert 1-expert 5) of the hybrid expert model so that each expert network outputs its own expert processing matrix. The first normalized matrix is ​​then input into the gating network of the hybrid expert model so that the gating network outputs the weights of each expert network. The weights of each expert network are used to process each expert processing matrix (e.g., corresponding multiplication), and the resulting matrices are normalized (layer normalization) to obtain the static feature vectors (i.e., static embeddings) of each processing station. ).

[0159] The decoder is based on a multi-head attention mechanism. At each step t, the decoder combines the global site embedding (static information) output by the encoder with the current transport device state (dynamic information, such as current time, remaining load, and remaining energy). Through a query-key attention mechanism, it calculates the probability distribution of all unvisited sites and samples or selects the action with the highest probability from it.

[0160] For example, the second linear projection layer in the decoder model is used to process the dynamic attribute vector (i.e., dynamic features) of the current transportation device to obtain the dynamic feature vector (i.e., dynamic features of the current transportation device). ); for the dynamic feature vector of the current transportation device and the static feature vector of the current processing station where the current transportation device is located (i.e. The concatenated feature vectors are obtained by concatenating the key and value linear projection layers of the attention model in the decoder model. The static feature vectors at each processing station are then processed using the key linear projection layer and value linear projection layer of the attention model in the decoder model to obtain the key matrix (K) and value matrix (V). Finally, the concatenated feature vectors are processed using the query linear projection layer of the attention model in the decoder model to obtain the dynamic query vector. The probabilistic generative network of the attention model in the decoder model is used to process the key matrix, value matrix, and dynamic query vector to obtain the log probability of the current transportation device heading to each processing station in the set of processing stations. , (Determined through a multi-head attention network). Based on the processing stations already visited by the current transport device and the logarithmic probability of the current transport device heading to each processing station in the set of processing stations, the action probability of the current transport device heading to each processing station in the set of processing stations is determined (i.e., (Determined through a masking mechanism and a softmax function); the processing station corresponding to the lowest action probability is determined as the next processing station to which the current transport device will proceed.

[0161] For example, a multi-scale hybrid training strategy can be employed to enhance the model's generalization ability. During training, instead of using problem instances of a single scale, a list of scales (e.g., [20, 50, 80], representing the number of service sites) is used, and instances of different scales are randomly sampled for training in each training batch. Simultaneously, policy optimization with multiple optimal solutions is combined. By inferring from multiple different starting points from an instance simultaneously, a baseline with smaller variance is computed, thereby optimizing the policy gradient, accelerating training convergence, and improving policy quality. The loss function is... ,in .

[0162] The embodiments of this application have excellent cross-scale generalization ability. This feature solves the long-standing limitation of learning methods that a model can only adapt to one scale, greatly improving the practicality and flexibility of the model, so that a trained model can be deployed in a variety of scenarios with different business volumes.

[0163] This application's embodiments effectively improve the system's practicality and robustness through an innovative soft time window constraint processing mechanism. By introducing a timeout penalty term into the objective function, the model learns to proactively avoid service delays, significantly reducing the task timeout rate. This enables the system to generate more flexible scheduling schemes that better meet actual operational needs when facing unavoidable congestion and uncertainty in the real world.

[0164] This application proposes a multi-transportation device collaborative logistics delivery path optimization method that comprehensively considers the capacity and energy limitations of transportation devices and the time window constraints of user services. By introducing a hybrid expert structure and multi-scale hybrid training, the model can effectively handle new problem instances with different sizes than the training set without retraining, solving the problem of poor cross-scale adaptability of existing learning methods. By transforming hard time windows into soft penalty terms and incorporating them into the reward function, the model can learn to make optimal trade-offs under unavoidable congestion or delays, improving the practicality and robustness of the solution.

[0165] Based on the same inventive concept, this application also provides a path planning apparatus for implementing the path planning method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more path planning apparatus embodiments provided below can be found in the limitations of the path planning method described above, and will not be repeated here.

[0166] In one exemplary embodiment, Figure 7 This is a structural schematic diagram of an unmanned platform path planning device provided in some embodiments, such as... Figure 7 As shown, the unmanned platform path planning device 700 includes:

[0167] The acquisition module 701 is used to acquire the static attribute vector of each processing station in the processing station set, and the dynamic attribute vector of the current transportation device in the unmanned platform's transportation device set each time it travels to a processing station.

[0168] The encoding processing module 702 is used to process the static attribute vectors of each processing station using the encoder model in the target path planning model to obtain the static feature vectors of each processing station.

[0169] The decoding processing module 703 is used to process the static feature vectors of each processing station and the dynamic attribute vectors of the current transportation device using the decoder model in the target path planning model, so as to obtain the next processing station to which the current transportation device is heading.

[0170] In some embodiments, the encoding processing module 702 includes a context matrix determination unit, an expert processing matrix determination unit, a weight determination unit, and a static feature vector determination unit. The context matrix determination unit is used to determine a context matrix based on the self-attention model in the encoder model and the static attribute vectors of each processing station. The expert processing matrix determination unit is used to determine each expert processing matrix based on the context matrix and each expert network of the hybrid expert model in the encoder model. The weight determination unit is used to determine the weights of each expert network based on the context matrix and the gating network of the hybrid expert model. The static feature vector determination unit is used to determine the static feature vectors of each processing station based on each expert processing matrix and the weights of each expert network.

[0171] In some embodiments, the context matrix determination unit is further configured to process the static attribute vectors of each processing station using the first linear projection layer in the encoder model to obtain the linear projection vectors of each processing station; process the linear projection vectors of each processing station using the self-attention model in the encoder model to obtain the context vectors of each processing station; and determine the context matrix based on the linear projection vectors of each processing station and the context vectors of each processing station.

[0172] In some embodiments, the static feature vector determination unit is further configured to process each expert processing matrix using the weights of each expert network to obtain each weight processing matrix; and to determine the static feature vector of each processing station based on the context matrix and each weight processing matrix.

[0173] In some embodiments, the decoding processing module 703 includes a dynamic feature vector determination unit, a concatenated feature vector determination unit, and a processing station determination unit. The dynamic feature vector determination unit is used to process the dynamic attribute vector of the current transportation device using the second linear projection layer in the decoder model to obtain the dynamic feature vector of the current transportation device. The concatenated feature vector determination unit is used to concatenate the dynamic feature vector of the current transportation device and the static feature vector of the processing station where the current transportation device is currently located to obtain the concatenated feature vector. The processing station determination unit is used to determine the next processing station to which the current transportation device will travel based on the attention model in the decoder model, the static feature vectors of each processing station, and the concatenated feature vector.

[0174] In some embodiments, the processing station determination unit is further configured to use the attention model in the decoder model to process the static feature vector and the concatenated feature vector of each processing station to obtain the log probability of the current transport device heading to each processing station in the set of processing stations; and to determine the next processing station to which the current transport device will head based on the log probability of the current transport device heading to each processing station in the set of processing stations.

[0175] In some embodiments, the processing station determination unit is further configured to process the static feature vectors of each processing station using the key linear projection layer and value linear projection layer of the attention model in the decoder model to obtain the key matrix and value matrix; process the concatenated feature vector using the query linear projection layer of the attention model in the decoder model to obtain the dynamic query vector; and process the key matrix, value matrix, and dynamic query vector using the probabilistic generation network of the attention model in the decoder model to obtain the log probability of the current transport device heading to each processing station in the set of processing stations.

[0176] In some embodiments, the processing station determination unit is further configured to determine the action probability of the current transport device heading to each processing station in the processing station set based on the processing stations that the current transport device has already passed and the logarithmic probability of the current transport device heading to each processing station in the processing station set; and determine the processing station corresponding to the minimum action probability as the next processing station to which the current transport device will head.

[0177] In some embodiments, the route planning device 700 further includes a training module for obtaining an initial route planning model and a scale list; the scale list includes multiple different numbers of processing stations; the initial route planning model is iteratively trained using training data for each number of processing stations to obtain a target route planning model; the training data for each number of processing stations includes the static attribute vector of each processing station, multiple starting positions of the transportation device, and the transportation path of the transportation device at each starting position.

[0178] In some embodiments, the training module is further configured to: determine the transportation path corresponding to each starting position based on the iterative path planning model; determine the total revenue of the transportation path corresponding to each starting position based on the transportation path corresponding to each starting position; determine the loss value based on the number of starting positions among multiple starting positions, the transportation path corresponding to each starting position, and the total revenue of the transportation path corresponding to each starting position; and continue iterating the path planning model if the loss value does not meet the iteration termination condition, and repeat the above steps until the loss value meets the iteration termination condition, at which point the path planning model of the last iteration is determined as the target path planning model.

[0179] In some embodiments, the training module is further configured to determine the total transportation distance and the latest service time of each processing station based on the transportation path corresponding to each starting point location; and to determine the total revenue of the transportation path corresponding to each starting point location based on the total transportation distance, the latest service time of each processing station, and the service time limit of each processing station.

[0180] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0181] Each module in the aforementioned path planning device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0182] In one exemplary embodiment, Figure 8 This is a schematic diagram of the structure of a computer device provided in some embodiments. The computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for wired or wireless communication with external terminals. Wireless communication can be implemented through Wireless Fidelity (WIFI), mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a path planning method. The display unit of the computer device is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0183] Those skilled in the art will understand that Figure 8The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0184] For example, a computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method of any of the above embodiments.

[0185] In one embodiment, a computer-readable storage medium is provided, wherein a computer program, when executed by a processor, implements the steps of the method provided in any of the above embodiments.

[0186] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method provided in any of the above embodiments.

[0187] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the methods described above.

[0188] The processor, functional modules, or functional units in any embodiment of this application may include an integration of one or more of the following: a general-purpose processor, an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a graphics processing unit (GPU), an embedded neural network processing unit (NPU), a controller, a microcontroller, a microprocessor, a programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, a quantum computing-based data processing logic unit, an artificial intelligence (AI) processor, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0189] The memory or computer-readable storage medium in any embodiment of this application may include at least one of non-volatile memory and volatile memory. Non-volatile memory includes integration of one or more of the following: Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Ferromagnetic Random Access Memory (FRAM), Flash Memory, Magnetic Surface Memory, Optical Disc, Compact Disc Read-Only Memory (CD-ROM), Magnetic Tape, Floppy Disk, Flash Memory, Optical Memory, High-Density Embedded Non-Volatile Memory, Resistive Random Access Memory (ReRAM), Magnetoresistive Random Access Memory (MRAM), Ferroelectric Random Access Memory (FRAM), Phase Change Memory (PCM), Graphene Memory, Volatile Memory, etc. Volatile memory includes one or more of the following: Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0190] The acquisition, transmission, storage, use, and processing of data in this application comply with relevant laws and regulations. It should be noted that certain software, components, models, and other existing industry solutions may be mentioned in the embodiments of this application. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0191] 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 application.

[0192] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application's patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A path planning method for an unmanned platform, characterized in that, The method includes: Obtain the static attribute vector of each processing station in the processing station set, and obtain the dynamic attribute vector of the current transportation device in the transportation device set of the unmanned platform each time it travels to a processing station. The context matrix is ​​determined based on the self-attention model in the encoder model of the target path planning model and the static attribute vectors of each processing station; Based on the context matrix and the expert networks of the hybrid expert model in the encoder model, determine each expert processing matrix; The weights of each expert network are determined based on the context matrix and the gating network of the hybrid expert model. Based on the expert processing matrices and the weights of the expert networks, the static feature vectors of each processing station are determined. The decoder model in the target path planning model is used to process the static feature vectors of each processing station and the dynamic attribute vector of the current transportation device to obtain the next processing station to which the current transportation device will travel.

2. The method according to claim 1, characterized in that, The step of determining the context matrix based on the self-attention model in the encoder model and the static attribute vectors of each processing station includes: The static attribute vectors of each processing station are processed using the first linear projection layer in the encoder model to obtain the linear projection vectors of each processing station. The self-attention model in the encoder model is used to process the linear projection vectors of each processing station to obtain the context vectors of each processing station. The context matrix is ​​determined based on the linear projection vector of each processing station and the context vector of each processing station.

3. The method according to claim 1, characterized in that, The step of determining the static feature vector of each processing station based on the expert processing matrices and the weights of each expert network includes: The weights of each expert network are used to process each expert processing matrix to obtain each weight processing matrix; Based on the context matrix and the weight processing matrices, the static feature vectors of each processing station are determined.

4. The method according to claim 1, characterized in that, The set of processing sites includes a warehouse and multiple service sites.

5. The method according to any one of claims 1-4, characterized in that, The step of using the decoder model in the target path planning model to process the static feature vectors of each processing station and the dynamic attribute vector of the current transportation device to obtain the next processing station to which the current transportation device is headed includes: The dynamic attribute vector of the current transportation device is processed by the second linear projection layer in the decoder model to obtain the dynamic feature vector of the current transportation device; The dynamic feature vector of the current transportation device and the static feature vector of the current processing station where the current transportation device is located are concatenated to obtain the concatenated feature vector; Based on the attention model in the decoder model, the static feature vectors of each processing station, and the concatenated feature vector, the next processing station to which the current transport device will travel is determined.

6. The method according to claim 5, characterized in that, The step of determining the next processing station to which the current transport device will travel, based on the attention model in the decoder model, the static feature vectors of each processing station, and the concatenated feature vector, includes: The attention model in the decoder model is used to process the static feature vectors of each processing station and the concatenated feature vectors to obtain the log probability of the current transportation device heading to each processing station in the set of processing stations. The next processing station to which the current transport device will travel is determined based on the logarithmic probability of the current transport device traveling to each processing station in the set of processing stations.

7. The method according to claim 6, characterized in that, The attention model in the decoder model is used to process the static feature vectors of each processing station and the concatenated feature vectors to obtain the log probability that the current transport device will travel to each processing station in the set of processing stations, including: The key linear projection layer and value linear projection layer of the attention model in the decoder model are used respectively to process the static feature vectors of each processing station to obtain the key matrix and value matrix. The concatenated feature vector is processed using the query linear projection layer of the attention model in the decoder model to obtain a dynamic query vector; The probabilistic generative network of the attention model in the decoder model is used to process the key matrix, the value matrix and the dynamic query vector to obtain the log probability of the current transportation device heading to each of the processing stations in the set of processing stations.

8. The method according to claim 6, characterized in that, The step of determining the next processing station to which the current transport device will travel based on the logarithmic probability of the current transport device traveling to each processing station in the set of processing stations includes: Based on the processing stations that the current transport device has already passed and the logarithmic probability of the current transport device heading to each processing station in the set of processing stations, determine the probability of the current transport device heading to each processing station in the set of processing stations. The processing station corresponding to the lowest probability of action is determined as the next processing station to which the current transport device is headed.

9. The method according to any one of claims 1-4, characterized in that, The method further includes: Obtain the initial path planning model and scale list; the scale list includes the number of different processing stations. The initial path planning model is iteratively trained using training data for each number of processing stations to obtain the target path planning model. The training data for each number of processing stations includes the static attribute vector of each processing station, multiple starting positions of the transportation device, and the transportation path of the transportation device at each starting position.

10. The method according to claim 9, characterized in that, The step of iteratively training the initial path planning model using training data at each of the stated number of processing stations to obtain the target path planning model includes: The transportation path corresponding to each starting point is determined by the iterative path planning model; Based on the transportation route corresponding to each starting point location, determine the total revenue of the transportation route corresponding to each starting point location; The loss value is determined based on the number of starting points among the plurality of starting points, the transportation path corresponding to each starting point, and the total revenue of the transportation path corresponding to each starting point. If the loss value does not meet the iteration termination condition, continue iterating the path planning model and repeat the above steps until the loss value meets the iteration termination condition, at which point the path planning model of the last iteration is determined as the target path planning model.

11. The method according to claim 10, characterized in that, The step of determining the total revenue of the transportation route corresponding to each starting point location includes: Based on the transportation path corresponding to each starting point location, determine the total transportation distance and the latest service time of each processing station; The total revenue of the transportation path corresponding to each starting point is determined based on the total transportation distance, the latest service time of each processing station, and the upper limit of service time for each processing station.

12. A path planning device for an unmanned platform, characterized in that, The device includes: The acquisition module is used to acquire the static attribute vector of each processing station in the processing station set, and the dynamic attribute vector of the current transportation device in the transportation device set of the unmanned platform each time it travels to a processing station. The encoding processing module is used to determine the context matrix based on the self-attention model in the encoder model of the target path planning model and the static attribute vectors of each processing station; determine each expert processing matrix based on the context matrix and each expert network of the hybrid expert model in the encoder model; determine the weights of each expert network based on the context matrix and the gating network of the hybrid expert model; and determine the static feature vectors of each processing station based on the expert processing matrices and the weights of each expert network. The decoding processing module is used to process the static feature vectors of each processing station and the dynamic attribute vector of the current transportation device using the decoder model in the target path planning model, so as to obtain the next processing station to which the current transportation device is heading.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

Images