Large-model-driven unmanned aerial vehicle relay mode vehicle-vehicle-electricity cooperative distribution method

Abstract

The invention, which relates to the low-altitude technology and artificial intelligence field, discloses a large-model-driven unmanned aerial vehicle relay mode vehicle-vehicle-electricity cooperative distribution method comprising the following steps: analyzing order demand information through a multi-mode large model, performing distribution mode decision making in combination with an environment state and constraints, and outputting an optimal distribution mode. The orders are aggregated and combined according to the characteristics of time, semantic similarity and the like. And based on the real-time state and risk assessment, a deep value learning model optimizes a distribution path. Unmanned aerial vehicle tasks are executed according to distance sorting and feasibility matching, the state is updated in real time during execution, and the optimization model is fed back. According to the method, global collaborative optimization of vehicles, machines and electricity is realized under dynamic order and energy constraints, and the service rate and the total income are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of low-altitude technology and artificial intelligence, in particular to a large model driven vehicle-aircraft-electricity collaborative delivery method in a UAV relay mode. BACKGROUND

[0002] With the rapid growth of delivery and instant delivery demand, traditional manual riders or single vehicle delivery methods are difficult to balance efficiency, cost and service quality in high-density urban environments. Unmanned aerial vehicles (UAVs) have flexible and fast end delivery capabilities, but are limited by endurance, load capacity and regulatory restrictions, and cannot independently complete meal pickup and long-distance transportation. Autonomous vehicles have strong load and endurance capabilities, but have low end delivery efficiency in complex urban road networks.

[0003] Existing research has proposed the idea of vehicle and UAV collaborative delivery, but most methods rely on rule setting or mathematical programming models, and often consider vehicle scheduling, UAV scheduling and charging energy problems independently in the optimization process, lacking joint optimization mechanisms across modalities. This separate modeling approach is difficult to achieve global optimal scheduling results under the influence of dynamic order random generation, limited vehicle power and uncertain environmental factors. How to achieve efficient allocation and scheduling of multi-customer point orders while ensuring the joint constraints of UAVs and vehicles still lacks effective solutions.

[0004] To address the above problems, the present method proposes to use a large model to automatically capture order demand information, including text, voice, picture perception and analysis, and output order semantic vectors to jointly schedule the vehicle-aircraft-electricity collaborative delivery modeling model. Introduce the hub as a key node in the collaborative delivery system. On the one hand, the hub connects electric vehicles and UAVs, allowing UAVs to only be responsible for short-radius end delivery, thereby breaking through the endurance and load capacity bottleneck and reducing energy consumption pressure; on the other hand, the hub has charging and energy supply functions, providing real-time energy support for electric vehicles and UAVs, thereby extending the operating time and reducing scheduling interruptions caused by insufficient power. In addition, the hub serves as a task aggregation and decomposition platform, supporting order consolidation and spatial grouping allocation, allowing tasks with similar geographic locations to be dispatched centrally, effectively reducing repeated paths and overall delivery costs. By introducing the hub, the vehicle and UAV scheduling process is decoupled, making it easier to select the delivery mode and schedule resources under a hierarchical reinforcement learning framework, thereby significantly improving the system's adaptive ability and service quality in dynamic environments. SUMMARY

[0005] The present application is directed to the cooperative optimization problem of automatic driving trucks and unmanned aerial vehicles in instant delivery of take-out food, and proposes a truck-aircraft-electricity cooperative delivery method in the large model driven unmanned aerial vehicle relay mode, which can realize the coordinated optimization of delivery and scheduling under the conditions of dynamic orders and energy constraints. The specific technical solutions are as follows: S1: Obtain order demand information, including pickup point location, customer delivery point location and user personalized constraint information (such as delivery time limit, temperature control requirement, fragile marking, etc.), analyze order text, voice, picture, etc. through the demand perception module of the multi-modal large model, model the analysis results and structured information of the order, output the order semantic vector and uncertainty, and trigger the rule-based safety fallback when the uncertainty exceeds the threshold or there is external knowledge conflict; S2: The system makes delivery mode decision based on the constraints and environment state parsed out by the first layer near-end learning strategy model, combines the mode prior given by the large model and the feasibility mask, and outputs “automatic driving truck direct delivery” or “truck delivery to transfer station – unmanned aerial vehicle secondary delivery”; The customer points determined for unmanned aerial vehicle delivery are assigned to the corresponding transfer stations to form delivery tasks; S3: After completing the mode selection, combine the orders of the same time step and the same mode, and aggregate and combine the orders with the geometric / time window features and the semantic similarity generated by the large model; S4: According to the order and vehicle state, make real-time decision by the second layer deep value learning model under the first layer mode constraint, accept the candidate path and risk summary generated by the large model, and apply feasibility mask and soft weight to the expanded Markov decision process (MDP) action set, select the vehicle-action pair according to the weighted value greed, and preferentially select the action related to the delivery order; S5: When the transfer station receives multiple unmanned aerial vehicle delivery tasks, first sort the to-be-delivered points according to the distance from the transfer station, and divide them into several task groups according to the preset maximum grouping number Introduce the task score generated by the large model and the unmanned aerial vehicle-task feasibility for one-to-one matching, and each unmanned aerial vehicle executes a group of delivery tasks; In the execution process, the unmanned aerial vehicle completes the delivery according to the “transfer station – delivery point – return station” path, and removes the allocated tasks from the to-be-delivered task table after completing the task; S6: During the cooperative execution of the vehicle and the unmanned aerial vehicle, the system state such as the vehicle position, the electric quantity, the unmanned aerial vehicle state and the transfer station queue is updated in real time, the execution log and the sensing data are submitted to the large model for summary and abnormal attribution, and the two-layer model is iteratively optimized online / offline according to the summary;

[0006] The above-mentioned method is adopted for the cooperative optimization of automatic driving trucks and unmanned aerial vehicles, considering multiple transfer stations and several unmanned aerial vehicles, and the unmanned aerial vehicles only go back and forth between the transfer stations; Further, the demand perception module is specifically as follows: (1) Text processing submodule: for natural language processing (NLP) of text notes in user orders, through word segmentation, entity recognition and semantic analysis, key information is extracted; (2) Speech recognition submodule: for converting user voice input into text and combining with the text processing submodule for semantic analysis to generate corresponding delivery demand labels; (3) Image recognition submodule: for processing user uploaded pictures or videos, through convolutional neural network (CNN) to identify item categories and packaging features, output temperature control requirements or fragility labels; The analysis results of text, voice and image are combined with the structured information of the order (pick-up point , customer point , weight , time window ) to input into a multi-modal large model , which includes a text encoder, an acoustic encoder, a visual encoder and a structured field encoder, and a cross-modal fusion layer thereon, for generating an order semantic vector and its uncertainty measure : The is mapped to a structured constraint parameter set by a constrained mapper : and a corresponding weight vector , for subsequent scheduling and path decision; when exceeds a set threshold, a safety fallback process of rule engine / artificial confirmation is triggered.

[0007] Further, the first layer proximal learning strategy model refers to establishing a mode selection network based on deep reinforcement learning, constructing a first layer model through the spatiotemporal distribution of historical orders, transfer station information and environmental information, and the core task is to decide the delivery mode of customer points. The objective function is as follows: Where is the discount factor, is the parameterized policy, is the reward function. In the first layer proximal learning strategy, the state space is constructed for the orders generated at each time step, considering order features, transfer stations and environmental information. Specifically, the state space design consists of the following: ​(1) The multi-modal information of the order is analyzed by the demand perception module to obtain a delivery constraint representing the user's individualization, which includes at least one or more of the following: temperature control demand level, fragility identification, posture retention requirement, or order priority. (2) Manhattan distance from the pickup point to the customer point, to represent the driving cost of the vehicle to directly complete the order; (3) Euclidean distance from the transfer station to the customer point, to represent the flight cost of the UAV to execute the order; (4) Time window of customer demand , to describe the timeliness constraint and urgency of the order; (5) Order weight , to represent the weight limit and energy consumption difference when the UAV executes the task; (6) Real-time queue length of the transfer station , to depict the congestion of the transfer station task; (7) Idle rate of the delivery vehicle fleet , to depict the available vehicles for scheduling vehicles; (8) Retrieval enhancement vector by multi-modal large model combined with external knowledge sources (such as weather, airspace restrictions, road construction, building access control, availability of charging piles), to supplement semantic constraint information that pure geometric and queue features cannot cover, which can be obtained by online or offline retrieval; (9) Natural language rules and temporary notifications are mapped to formal constraint indicators by the multi-modal large model, including but not limited to "prohibition of take-off and landing / flight prohibition radius / time period control / special placement", etc., to explicitly expose the feasible region boundary and risk level in the state. All the above features are adaptively normalized and weighted encoded, and the weights can be output by the multi-modal large model or estimated by the system online to enhance the transferability in different business scenarios. In the design of the action space, the delivery mode is abstracted as a binary decision, i.e. the vehicle direct delivery mode and the vehicle delivery to transfer station - UAV secondary delivery mode. Specifically, the action set is defined as , wherein represents that the automatic driving truck directly completes the pickup and delivery tasks, represents that the automatic driving truck completes the pickup and delivery to the transfer station, and then the UAV completes the secondary delivery. In the design of the reward function, the application adopts the relative advantage principle: for each order generated at each time step, the estimated cost of the vehicle direct delivery mode and the UAV delivery mode is calculated respectively, and the cost difference of the selected mode relative to the optimal mode is taken as the basic reward; wherein the vehicle mode cost is composed of the driving distance and energy consumption from the pickup point to the delivery point, and the UAV mode cost includes the driving distance from the vehicle pickup point to the transfer station, the round-trip flight distance of the UAV and the take-off and landing energy consumption. On this basis, a soft constraint term is introduced, including the penalty for exceeding the customer time window, the congestion and waiting penalty caused by the queue length of the transfer station, and the penalty for exceeding the flight range or load of the UAV; at the same time, the reward is dynamically adjusted in combination with the idle rate of the vehicle fleet, when the idle rate of the vehicle fleet is low, the reward of the vehicle mode is reduced and the reward of the UAV mode is increased, and vice versa when the idle rate of the vehicle fleet is high.

[0008] Further, the combined processing of multiple orders generated at the same time step means that for the vehicle direct delivery mode order, the orders of the same pickup point or close geographical position are combined into one delivery path by using the order splitting strategy, so as to reduce the vehicle driving mileage and improve the carrying efficiency; for the order of the UAV delivery mode, the vehicle delivers the order to the transfer station task queue, and uses the spatial clustering method to group the delivery points, so that the task points in the same group are close in geographical position, thereby reducing the flight path length and energy consumption of the UAV.

[0009] The transfer station task queue is a set of delivery tasks to be executed by the transfer station UAV group. During task execution, the UAV is subject to power and flight radius constraints: the remaining power must satisfy , wherein represents the minimum power to ensure safe return; and the flight radius must satisfy , wherein represents the maximum flight radius. When the power of the UAV is insufficient or after completing the task and returning, the UAV can be charged at the transfer station or directly returned to standby, thereby ensuring the sustainability of the UAV in long-term operation and the stability of task execution.

[0010] Further, the second layer deep value learning model means that according to the obtained optimal delivery mode information and the state of the vehicle, the agent related data is obtained, and an MDP model is constructed, which is represented as (S, A, T, R, γ), wherein S is the state space, A is the action space, T is the state transition model, R is the reward model, and γ is the discount factor;

[0011] The vehicle state value function is used to measure the future value of the vehicle space-time state , which represents the vehicle The estimated cumulative reward from the current state to the end of operation time, wherein The vehicle Perform an action The state of the last arrival, the optimization target under the consideration of future value , wherein is a discount factor, indicating the importance of the model to the future value, and the objective function is as follows:

[0012] A neural network-based vehicle state value prediction model is established, and the state S is fitted through a neural network to obtain a vehicle state value function Then, the algorithm converges through a double neural network and an experience pool;

[0013] The architecture of the neural network includes an input layer, a first convolutional layer, a second convolutional layer, a max-pooling layer, a first fully connected layer, a second fully connected layer, and an output layer. The state S is input through the neural network, and the parameterized state value evaluation function is defined as A TD(0) method is used to estimate The vehicle state value function is obtained as follows: , wherein represents the parameters of the target network.

[0014] The input variable of the model is the state vector S in the model construction, which is represented as , wherein is the time required when the vehicle state changes to "dispatchable", is the percentage of the remaining power when the vehicle state changes to "dispatchable", is the road network position when the vehicle state changes to "dispatchable", is the distance from the current position of the vehicle to the nearest charging station, is the time step per day, is the date time step; In the design of the action space, the action space represents all feasible action sets (A ) at time step t, including staying (S ), transferring (T ), charging (C ), discharging (D ), and matching (M ); The "dispatchable" autonomous driving truck selects the executable actions of the agent, specifically including: a. If under the dispatching instruction, the autonomous driving truck selects to stay ( ), the vehicle will stay in place to wait for future travel demand; b. If the autonomous delivery vehicle chooses to relocate ( ) under dispatch instruction, the dispatch instruction guides the vehicle to relocate to another location. The vehicle continues to move within the dispatch time window , and the reachable new location is the place within the time distance of from the original location. The system decides whether the vehicle relocates to the different relocation place according to the space-time value of the place. Through the instruction, the vehicle can relocate to a place where the probability of orders appearing is greater or go to a charging station; c. If the autonomous delivery vehicle chooses to charge ( ) under dispatch instruction, if the vehicle is at the charging station, the vehicle receiving the instruction continues to charge or immediately starts charging. When the vehicle is already in a full battery state, the instruction is invalid; d. If the autonomous delivery vehicle chooses to discharge ( ) under dispatch instruction, if the vehicle is at the charging station, the vehicle receiving the instruction continues to discharge or immediately starts discharging. When the vehicle is already in a no battery state, the instruction is invalid; e. If the autonomous delivery vehicle chooses to match ( ) under dispatch instruction, the instruction assigns an order to the vehicle, i.e., the action related to the order set . Each "match" action is bound to a delivery order, and when the vehicle executes a specified "match" instruction, it means that the vehicle needs to go to the merchant to pick up the meal, and then deliver the order to the destination. The vehicle is in an "unschedulable" state before it arrives at the customer's destination. Under the premise of keeping the action set {a1,…,a5} unchanged, the action feasibility mask and soft preference coefficient generated by the multimodal large model are introduced: when the retrieval enhanced information or semantic rule indicates that a certain action is not feasible under the current constraints, the action is shielded; when multiple actions are feasible, the strategy network output is temperature scaled and calibrated according to the prior and uncertainty to reduce the risk of misjudgment due to incomplete information. For relocation (a2), the large model generates a candidate destination set and an arrival feasible interval based on demand heat prediction and road risk, and the strategy network selects a relocation location with higher expected value from the candidate set. For charging / discharging (a3 / a4), the large model maps the power grid announcement, time-of-use electricity price, and station congestion prediction into the allowed time window and penalty coefficient, which are used to restrict or encourage energy management behavior in specific time periods. In the design of the reward function, according to the system dispatch instruction, if the electric delivery vehicle chooses the action stay ( ), relocate ( ), the operator's revenue is given by , Since this instruction does not generate direct cost or revenue, both of them are 0. Since the model can utilize the peak-valley feature of time-of-use electricity price and Vehicle to Grid (V2G) technology for charging and discharging, if the electric truck chooses to act as charging ), discharging ), the operator's revenue is given by , where is a piecewise function with respect to time, representing the time-of-use electricity price at time in a day, : the duration of charging and discharging action, : representing the average charging power and average discharging power (kW), : representing the efficiency coefficient of charging and discharging (0 ≤ 1), : representing the battery degradation cost per unit energy of charging and discharging. : representing the ancillary service reward for discharging during grid peak hours, : an indicator function, taking value 1 during grid peak hours (i.e. electricity price peak hours) and 0 otherwise. represents the cost of charging, represents the revenue of discharging; when the system assigns a "charging" instruction to the vehicle, the charging cost will be generated, thus the negative revenue will be obtained. On the contrary, when discharging, the operator will obtain positive revenue. Under the dispatching instruction, if the electric truck chooses to match ), each "match" action is associated with an order , for any order in the set, its key time parameters are defined as follows: : the start time of order ; : the estimated pickup time of order ; : the estimated delivery time of order ; : the maximum acceptable waiting time of order ; : order The shortest travel time from the origin to the destination. : order The required delivery time limit. Wherein: is the takeout waiting time constraint, and the pickup time is not later than the latest pickup time. Under the condition of meeting the waiting time constraint, the reward of the action is represented as follows: Wherein represents the income obtained by the vehicle per unit time. is the order representation quantity. When (single order), this item is 0; when (group order), this item is positive. Due to the different positions of the vehicle, the waiting time of the customer is different when the system dispatches different vehicles to deliver the specified order, and is used to represent the compensation for the customer's point waiting time. When the system assigns different vehicles to execute the same order, the operator has different degrees of income due to the different initial positions of the vehicle. represents the group order reward coefficient. On this basis, according to the analysis results of the multi-modal large model on the constraints such as "time window, temperature control, fragile, upright, priority", etc. Set the reward shaping weight, and adaptively re-calibrate the items such as waiting, default, and deviation; When the search enhanced information indicates that it is forbidden to fly, bad weather, temporary restrictions, peak electricity prices or station congestion abnormalities, risk penalties are superimposed or conservative fallback is triggered; Optionally, the interpretable output of the large model on the action preference is used as a distillation regularity for offline training to shorten the cold start and improve sample efficiency.

[0015] Further, the vehicle combines the current state and the predicted future value to make real-time path planning and scheduling decisions.

[0016] By greedily selecting the vehicle-action pair with the largest weight, the action related to carrying passengers is preferentially selected. Finally, the greedy matching result is used as the initial solution of the integer programming solver to further solve it to achieve maximization.

[0017] Beneficial effects: The large model driven unmanned aerial vehicle relay mode vehicle-aircraft-electricity collaborative distribution method provided by the application can automatically complete the selection and scheduling of the distribution mode without manual fixed rules. Through high-level reinforcement learning to determine the distribution mode of the order, and combining the low-level value network and global matching optimization, the order can be reasonably distributed between the autonomous driving truck and the unmanned aerial vehicle, thereby reducing the waste of vehicle and unmanned aerial vehicle resources, improving the comprehensive utilization rate of the two types of transportation tools, and improving the overall service efficiency of the system. In addition, the method considers the transfer station constraint and unmanned aerial vehicle energy management in the scheduling process, which can effectively avoid the task interruption problem caused by the return or insufficient power of the unmanned aerial vehicle, and improve the stability and safety of the unmanned aerial vehicle in executing tasks.

[0018] At the same time, by introducing the time-of-use electricity price, order time window, weight constraint and transfer station queuing condition and other extended features, the application can realize more refined scheduling decisions, reduce the energy consumption and operating cost of the system while ensuring the order timeliness rate. Therefore, the application not only improves the collaborative ability of the autonomous driving truck and the unmanned aerial vehicle, but also realizes efficient resource allocation and green and low-carbon operation in the dynamic and complex urban distribution scene, and has significant application value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments of the application. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation to the scope, and for those skilled in the art, other related drawings can be obtained without creative labor on the basis of these drawings.

[0020] Figure 1 The flow chart of the large model driven unmanned aerial vehicle relay mode vehicle-aircraft-electricity collaborative distribution method provided by the embodiments of the application;

[0021] Figure 2 The demand perception module schematic diagram of the large model driven unmanned aerial vehicle relay mode vehicle-aircraft-electricity collaborative distribution method provided by the embodiments of the application;

[0022] Figure 3 The collaborative distribution framework diagram of the large model driven unmanned aerial vehicle relay mode vehicle-aircraft-electricity collaborative distribution method provided by the embodiments of the application. DETAILED DESCRIPTION

[0023] The application will be further described in detail below in combination with the drawings and specific embodiments. However, it should not be understood that the scope of the above-mentioned subject matter of the application is limited to the following embodiments, and any technology realized on the basis of the content of the application belongs to the scope of the application. As Figure 1As shown, the large model driven unmanned aerial vehicle relay mode of the vehicle-aircraft-electricity collaborative distribution method considers multiple transfer stations and several unmanned aerial vehicles, and the specific steps are as follows:

[0024] The order demand information is obtained, including the pickup point location, the customer distribution point location, and the user personalized constraint information (such as delivery time limit, temperature control requirement, fragile mark, etc.), the demand perception module of the multi-modal large model is used to analyze the order text, voice, picture, etc., the analysis result is combined with the order structured information to model, and the order semantic vector and uncertainty are output, when the uncertainty exceeds the threshold or the external knowledge conflicts, the rule-based safety fallback is triggered; In this step, a delivery order is randomly generated, which includes a pickup point location and a distribution point location, and an automatic driving car is used to pick up food at a takeout restaurant, and an unmanned aerial vehicle can only take off from a transfer station and return to the transfer station, and does not participate in picking up food; the order demand information is obtained, including the pickup point location, the customer distribution point location, and the user personalized constraint information (such as delivery time limit, temperature control requirement, fragile mark, etc.), and the analysis result of the demand perception module on the order text, voice, image is combined with the order structured information (pickup point , customer point , weight , time window ) is input into the multi-modal large model , which includes a text encoder, an acoustic encoder, a visual encoder, and a structured field encoder, and a cross-modal fusion layer thereon, for generating an order semantic vector and its uncertainty : The is mapped to a structured constraint parameter set by a constraint mapper : and the corresponding weight vector , for subsequent scheduling and path decision; when exceeds the set threshold, the rule engine / artificial confirmation safety fallback process is triggered.

[0025] Step 2: The system determines the distribution mode based on the parsed constraints and environmental state by the first layer near-end learning strategy model, combines the mode prior given by the large model and the feasibility mask, and outputs "automatic driving car direct delivery" or "car delivery to transfer station-unmanned aerial vehicle secondary distribution"; the customer points determined for unmanned aerial vehicle distribution are allocated to corresponding transfer stations to form delivery tasks: Step 2.1: State space design In the first layer of reinforcement learning, the state space is constructed for each order generated at each time step, comprehensively considering order features, transit stations, and environmental information. Specifically, the state vector consists of the following: (1) The multimodal information of the order is parsed by the demand perception module to represent the user's personalized delivery constraints, which include at least one or more of the following: temperature control requirement level, fragility label, posture maintenance requirement, or order priority level. (2) The Manhattan distance between the pickup point and the customer's location is used to represent the driving cost of the vehicle directly completing the order; (3) The Euclidean distance from the transit station to the customer point is used to characterize the flight cost of the drone executing the order; (4) Time window of customer demand This is used to describe the time constraints and urgency of an order; (5) Order weight This is used to characterize the payload limitations and energy consumption differences of drones when performing missions; (6) Real-time queue length of the transfer station It is used to depict the congestion situation of transfer station tasks; (7) Idle rate of delivery fleet This is used to depict the availability of vehicles for dispatching vehicles; (8) A retrieval enhancement vector composed of a multimodal large model and external knowledge sources (such as weather, airspace restrictions, road construction, building access control, and charging pile availability) is used to supplement semantic constraint information that is difficult to cover by pure geometry and queue features. The external knowledge can be obtained through online or offline retrieval. (9) The natural language rules and temporary notifications are mapped into formal constraint indicators by the multimodal large model, including but not limited to “no take-off and landing / no-fly radius / time period control / special placement”, which are used to explicitly expose the feasible domain boundary and risk level in the state. All of the above features are adaptively normalized and weighted. The weights can be output by a multimodal large model or estimated online by the system to enhance portability in different business scenarios. Step 2.2: Motion Space Design In the design of the action space, this invention abstracts the delivery mode into a binary decision, namely, the vehicle direct delivery mode and the vehicle-to-transfer-station-drone secondary delivery mode. Specifically, the action set is defined as... ,in This indicates that autonomous trucks will directly handle the food pickup and delivery tasks. This means that autonomous trucks pick up the food and transport it to a transfer station, where drones then complete the secondary delivery. At each time step The system will receive new order requests and update the status of vehicles and drones. Decision variables. Used to determine customers In time Delivery methods (vehicles or drones). If the customer is served by a drone, further processing is available through... The variable determines its corresponding transfer station allocation. The constraints are as follows: in, Indicates at time step UAV from transit station Serving customers The feasibility of this. Step 2.3: Reward Function Design In designing the reward function, this invention adopts the principle of relative advantage: for each order generated at each time step, the estimated costs of the vehicle delivery mode and the drone delivery mode are calculated separately, and the cost difference between the selected mode and the optimal mode is used as the basic reward. The vehicle mode cost consists of the travel distance and energy consumption from the pickup point to the delivery point, while the drone mode cost includes the travel distance from the pickup point to the transfer station, the round-trip flight distance of the drone, and take-off and landing energy consumption. Based on this, soft constraints are introduced, including penalties for customer time window expiration, congestion and waiting penalties caused by transfer station queue length, and penalties for drone range or load exceeding limits. Simultaneously, the reward is dynamically adjusted based on the fleet idle rate. When the fleet idle rate is low, the reward for the vehicle mode is reduced and the reward for the drone mode is increased; conversely, when the fleet idle rate is high, the reward is reduced. Specifically: Relative advantage reward: in, The cost of the optimal delivery method for the current order. The cost of the delivery mode selected for the order; if the vehicle mode is selected, then... If you select drone mode, then , It should be a small positive number to prevent the denominator from being zero. Soft constraint penalties: in: This refers to the actual delivery time. To request a delivery time, This is the timeout amount; This refers to the queue length at the transfer station; Waiting time at the transfer station; and is an indicator function, which is 1 when the UAV range or payload is out of limit, otherwise 0; is a penalty coefficient. Dynamic adjustment of reward: wherein: is the fleet idle rate (between 0 and 1); is the action selected by the agent (0 represents the vehicle mode, and 1 represents the UAV mode); is the adjustment intensity coefficient. Final reward:

[0026] Step 3: After the first layer of proximal learning strategy model selects the delivery mode for each order, multiple orders of the same delivery mode at the same time step are combined. For orders in the vehicle direct delivery mode, orders with similar geographical locations or the same pickup point are combined into one delivery path using the order-packing strategy to reduce vehicle travel mileage and improve carrying efficiency. For orders in the UAV delivery mode, the vehicle delivers the orders to the task queue in the transfer station, and uses the spatial clustering method to group the delivery points so that the task points in the same group are close to each other in geographical location, thereby reducing the flight path length and energy consumption of the UAV. Define the combined order set as For any order in the set, its key time parameters are defined as follows: is the start time of order ; is the estimated pickup time of order ; is the estimated delivery time of order ; is the maximum acceptable waiting time of order ; is the shortest travel time of order from the starting point to the ending point; is the required delivery deadline of order . To ensure the timeliness and feasibility of the delivery service, the order combination must meet the following two core constraints The estimated delivery time of each order in the set must not be later than its required delivery time, with a reasonable tolerance allowed. wherein, is the delivery time tolerance of an order . The total travel time of serving the set of orders must be guaranteed to complete the delivery before the most urgent order deadline. The total travel time of serving the set of orders must be guaranteed to complete the delivery before the most urgent order deadline. wherein, is the total travel time of the planned path serving the set of orders .

[0027] Step 4: According to the order and vehicle state, make real-time decisions under the first layer mode constraint by the second layer deep value learning model, accept the candidate path and risk summary generated by the large model, and impose feasibility masks and soft weights on the expanded Markov Decision Process (MDP) action set, select the vehicle-action pair according to the weighted value greedy, and prefer the actions related to the delivery order. The specific implementation is as follows: Construct the MDP model (S, A, T, R, γ) of state space S, action space A, state transition model T, reward model R, and discount factor γ to describe the autonomous driving truck operation decision-making process. After the system combines the order delivery mode decided by the first layer proximal learning strategy model, a neural network architecture including convolutional layers, pooling layers, and fully connected layers is designed to input the vehicle state S into the network. The TD(0) method is used to estimate the state value function, providing accurate guidance for subsequent decision-making.

[0028] Define actions including staying, transferring, charging, discharging, and matching, and dynamically select the optimal action to respond to operational demand and passenger orders according to the current situation. Action space represents the set of all feasible actions ( ) at time step , including staying ( ), transferring ( ), charging ( ), discharging ( ), and matching ( );

[0029] Under the dispatching instruction, the electric truck selects the action to stay ( ), transfer ( ), and the operator's revenue is respectively as formula The system assigns the "charge" instruction to the vehicle, and the vehicle will be charged. The cost of charging is shown in the figure. Since the instruction does not generate direct cost or benefit, both the cost and the benefit are 0.

[0030] The model can use the peak-valley characteristics of the time-of-use electricity price and the vehicle-to-grid (V2G) technology to charge and discharge. The benefits of the actions "charging" and "discharging" are calculated using the formulas wherein is a piecewise function with respect to time, indicating the time-of-use electricity price at time of day : the duration of the charging and discharging action, : representing the average charging power and the average discharging power (kW), : representing the efficiency coefficient of charging and discharging (0 ≤ 1), : representing the battery degradation cost per unit of energy for charging and discharging. : representing the auxiliary service reward for discharging during the grid peak period, : an indicator function, taking the value 1 during the grid peak period (i.e., the peak period of the electricity price), and 0 otherwise. When the system assigns the "charging" instruction to the vehicle, the charging cost will be generated, and thus a negative benefit will be obtained. In contrast, when discharging, the operator will obtain a positive benefit.

[0031] Under the dispatching instruction, the electric truck selects a match ), and each "match" action is associated with an order : is the waiting time constraint of the takeout, and the pickup time is not later than the latest pickup time;

[0032] Under the satisfaction of the waiting time constraint, the return of the action is calculated by , wherein represents the benefit obtained by the vehicle per unit time. is the order representation quantity. When (single order), this item is 0; when (group order), this item is positive. represents the group order reward coefficient. Due to the different locations of the vehicles, the waiting time of the order is different when the system dispatches different vehicles to deliver the specified order. The use of represents the compensation to the customer per unit of waiting time. When the system dispatches different vehicles to deliver the same order, the operator has different degrees of benefit due to the different initial positions of the vehicles.

[0033] ​For each "dispatchable" vehicle, the model enumerates all actions, calculates the value of the vehicle-action pair wherein represents the current state of the vehicle, represents the estimated future cumulative return of the vehicle performing the action in the current state:

[0034] Step 5: When the transfer station receives multiple UAV delivery tasks, first sort the to-be-delivered points by distance from the transfer station, and divide them into several task groups with a preset maximum number of groups, introduce the task score generated by the large model, and one-to-one match with UAV-task feasibility, each UAV executes a group of delivery tasks; during execution, the UAV completes delivery according to the "transfer station-delivery point-return station" path, and removes the allocated tasks from the to-be-delivered task table after completing the task; During task execution, the UAV is subject to power constraints and flight radius constraints: its remaining power must satisfy wherein represents the minimum power to ensure safe return; its flight radius must satisfy wherein represents the maximum flight radius. At the same time, the weight of the UAV should satisfy: wherein, represents the weight of the order , and represents the maximum load capacity of the UAV; When the UAV's power is insufficient or after completing the task and returning, it can be charged at the transfer station or directly returned to standby, thereby ensuring the sustainability of the UAV in long-term operation and the stability of task execution.

[0035] Step 6: During the joint execution of vehicles and UAVs, real-time update of system state information, including vehicle location, power, UAV state, transfer station task queue, etc., submission of execution logs and sensor data to the large model for summary and abnormal attribution, and optimization of decision-making strategy of the double-layer reinforcement learning model based on state feedback. Use historical order data and time-of-use electricity price information for offline pre-training, analyze multi-modal information through the large model, build the first layer of near-end learning strategy model and the second layer of deep value learning network, and establish the vehicle state value prediction model. Through experience pool and double neural network technology, optimize the convergence and generalization ability of the algorithm.

[0036] In the implementation method, the training process of the neural network model is as follows: S1. Initialize the neural networks of the first layer and the second layer; S2. Initialize the experience replay buffer; each experience includes the current state, the action taken, the reward obtained, the next state, and whether it is terminated; S3. The system obtains real-time data from the simulated evolving road network environment in order to explore more feasibility. Specifically as follows: a Initialize the environment and obtain the initial state. When a new order arrives, parse the multi-modal information through the large model; b Based on the parsed constraints and state, the system uses the first layer of proximal learning strategy to make delivery mode decision for each time step of the generated order; c Assign orders to the corresponding queue (truck order queue or drone order queue) according to the mode; d Use the second layer model to make scheduling decisions; S4. Select an action; the agent selects an action according to the current state and the Q value function, using algorithm to balance the exploration and utilization process, to probability of randomly selecting a feasible action for the vehicle fleet, to probability of global decision-making after action evaluation and then assigning the globally matched results to the vehicle fleet; S5. Execute the action and observe the result; the agent executes the selected action and observes the next state and the reward obtained; S6. Store the experience; store the first layer transition to the first layer experience buffer, and store the second layer current state, the action taken, the reward obtained, the next state, and whether it is terminated to the experience replay buffer; S7. Randomly sample a batch of experiences from the experience replay buffer; S8. The first layer uses the collected experiences to update the policy network and the value network, and the second layer uses a target network to estimate in

[0013] according to the classical DQN idea in reinforcement learning, and uses the evaluation network to periodically update the target network; S9. Calculate the target Q value; for each sampled experience, calculate the target Q value: wherein, is the reward obtained, γ is the discount factor, is the vehicle state value function; S10. Calculate the loss function; use the mean square error between the target Q value and the Q value estimated by the neural network as the loss function to update the parameters of the neural network: wherein represents the parameters of the evaluation network; S11. Update the neural network; update the parameters of the neural network using the backpropagation algorithm to minimize the loss function; S12. Repeat steps S3-S9 until a preset number of training iterations is reached or convergence is achieved. In combination with time-of-use pricing and V2G technology, charging and discharging strategies are developed to maximize the use of grid resources and reduce operating costs.

[0037] In a large-scale scheduling system, global matching can quickly calculate to obtain the optimal solution. The specific steps are as follows: Input: Set of vehicle-action pair weight vectors Output: Optimal matching set Initialize empty set Weight vectors Sort from large to small "Match" is preferred to other actions under the same weight for vehicle-action pair , in do if ( not in ) and (matching action not in ) then if is matching action) then ( ) Integer programming solver (initial solution = ) return

[0038] In combination with the current state and the predicted future value, first greedily select the vehicle-action pair with the largest weight, and second, prefer actions related to delivery orders. Finally, the greedy matching result is used as the initial solution of the integer programming solver for further solving to maximize the formula.

Claims

1. A vehicle-machine-electronics collaborative delivery method in a large-scale model-driven UAV relay mode, characterized in that, Includes the following steps: S1: Obtain order demand information, including the location of the pickup point, the location of the customer's delivery point, and user-personalized constraints (such as delivery time limit, temperature control requirements, fragile markings, etc.). The demand perception module of the multimodal big model parses the order text, voice, images, etc., and combines the parsing results with the order structured information to jointly model and output the order semantic vector and uncertainty. When the uncertainty exceeds the threshold or there is an external knowledge conflict, a rule-based safety fallback is triggered. S2: Based on the resolved constraints and environmental state, the system makes delivery mode decisions using the first-layer near-end learning strategy model. Combining the mode priors and feasibility masks given by the large model, it outputs "direct delivery by autonomous truck" or "delivered by truck to transfer station - secondary delivery by drone". Customers who are determined to be delivered by drone are assigned to the corresponding transfer station to form delivery tasks. S3: After completing the pattern selection, orders with the same pattern at the same time step are combined and aggregated and grouped together using geometric / time window features and semantic similarity generated by the large model. S4: Based on the order and vehicle status, the second-layer deep value learning model makes real-time decisions under the first-layer pattern constraints. It accepts the candidate paths and risk summaries generated by the large model and applies feasibility masks and soft weights to the extended Markov Decision Process (MDP) action set. It greedily selects vehicle-action pairs according to weighted value and prioritizes actions related to the delivery order. S5: When the transfer station receives multiple drone delivery tasks, it first sorts the delivery points according to their distance from the transfer station and groups them into a preset maximum number of groups. The task is divided into several task groups, and the task score generated by the large model is matched one by one with the drone-task feasibility. Each drone performs a group of delivery tasks. During the execution, the drone completes the delivery according to the path of "transfer station - delivery point - return station" and removes the assigned task from the pending task list after the task is completed. S6: During the collaborative execution of vehicles and drones, the system status such as vehicle location, battery level, drone status, and transfer station queue is updated in real time. The execution logs and sensor data are submitted to the large model for summarization and anomaly attribution, and the two-layer model is optimized and iterated online / offline accordingly. The system also includes: 1) A fleet of autonomous trucks, serving as mobile delivery and energy storage units, connected to a charging infrastructure network that includes V2G functionality; 2) Drone swarms, serving as last-mile delivery units originating from transit stations; 3) Distributed transfer station network, serving as a collaborative coupling node between trucks and drones.

2. The vehicle-machine-electronics collaborative delivery method in the large-model-driven UAV relay mode according to claim 1, characterized in that, The demand perception module in step S1 includes a text processing submodule, a speech recognition submodule, and an image recognition submodule, which integrates the parsing results of text, speech, and images with the order structured information (pickup point). Customer Point ,weight Time window Joint input to multimodal large model The It includes a text encoder, an acoustic encoder, a visual encoder, and a structured field encoder, along with a cross-modal fusion layer, for generating order semantic vectors. and its uncertainty measure : Will via constraint mapper Mapped to a set of structured constraint parameters: and corresponding weight vector This is used for subsequent scheduling and path decision-making; when When the set threshold is exceeded, a safety rollback process involving rule engine / manual confirmation is triggered.

3. The vehicle-machine-electronics collaborative delivery method in the large-model-driven UAV relay mode according to claim 1, characterized in that, The first-layer proximal learning strategy model in step S2 refers to establishing a pattern selection network based on deep reinforcement learning. The first-layer proximal learning strategy is implemented through the spatiotemporal distribution of historical orders, transit station information, and environmental information. Its core task is to decide the delivery mode of the customer point. The pattern selection network receives the "pattern prior distribution" and "uncertainty" output by the multimodal large model in parallel, as well as the external knowledge features (weather, airspace / no-fly zone, road construction, access control rules, etc.) obtained by retrieval and reinforcement, which are used to calibrate the strategy input and limit the feasible domain. In the first-layer proximal learning policy model, the state space is constructed for each order generated at each time step, comprehensively considering order features, transit stations, and environmental information; the specific design of the state space, action space, and reward function is as follows: State space design: (1) Based on the multimodal information of the order, the demand perception module analyzes and uses it to represent the user's personalized delivery constraints, which include at least one or more of the following: temperature control requirement level, fragility label, posture maintenance requirement, or order priority level; (2) The Euclidean distance from the transit station to the customer point is used to characterize the flight cost of the UAV executing the order; (3) Time window of customer demand This is used to describe the time constraints and urgency of an order; (4) Order weight This is used to characterize the payload limitations and energy consumption differences of drones when performing missions; (5) Real-time queue length at the transfer station It is used to depict congestion at transfer stations; (6) Idle rate of delivery fleet This is used to depict the availability of vehicles for dispatching vehicles; (7) The retrieval enhancement vector, which is a combination of a multimodal large model and external knowledge sources (such as weather, airspace restrictions, road construction, building access control, and charging pile availability), is used to supplement the semantic constraint information that is difficult to cover by pure geometry and queue features. The external knowledge can be obtained through online or offline retrieval. (8) The natural language rules and temporary notifications are mapped into formal constraint indicators by the multimodal large model, including but not limited to "no take-off and landing / no-fly radius / time period control / special placement", which are used to explicitly expose the feasible domain boundary and risk level in the state; All of the above features are adaptively normalized and weighted. The weights can be output by a multimodal large model or estimated online by the system to enhance portability in different business scenarios. Decision space design: This invention abstracts the delivery model into a binary decision, namely, the direct vehicle delivery model and the vehicle delivery to the transfer station model. The drone secondary delivery mode is specifically defined as {0,1}, where 0 indicates that the autonomous truck directly completes the food pickup and delivery tasks, and 1 indicates that the autonomous truck picks up the food and transports it to the transfer station, where the drone then completes the secondary delivery. When the external knowledge of the large model indicates the existence of no-fly zones, extreme weather, or access restrictions, a mask is applied to action "1". This binary action structure ensures the simplicity and scalability of the decision-making process. Reward function design: This invention employs the principle of relative advantage: for each order generated at each time step, the estimated costs of vehicle-to-delivery mode and drone delivery mode are calculated separately, and the cost difference between the selected mode and the optimal mode is used as the basic reward. The vehicle mode cost consists of the travel distance from the pickup point to the delivery point and energy consumption, while the drone mode cost includes the travel distance from the pickup point to the transfer station, the round-trip flight distance of the drone, and take-off and landing energy consumption. Based on this, the analysis results of constraints such as time window, temperature control, fragility, upright position, and priority from a multimodal large model generate shaping weights for timeliness breach, temperature control deviation, fragility / attitude violation, etc., in the reward, used for adaptive recalibration. Simultaneously, soft constraints are introduced, including penalties for customer time window timeouts, congestion and waiting penalties caused by transfer station queue length, and penalties for drone range or load exceeding limits. Furthermore, the reward is dynamically adjusted based on the fleet idle rate: when the fleet idle rate is low, the reward for the vehicle mode is reduced and the reward for the drone mode is increased; conversely, when the fleet idle rate is high, the reward is increased.

4. The vehicle-machine-electronics collaborative delivery method in the large-model-driven UAV relay mode according to claim 1, characterized in that, In step S3, after the first-layer near-end learning strategy model selects a delivery mode for each order, multiple orders with the same delivery mode at the same time step are combined. For orders in the vehicle direct delivery mode, a group-buying strategy is used to merge orders with similar geographical locations or the same pick-up point into a single delivery route. Semantic similarity and uncertainty generated by a multimodal large model (covering at least temperature control level, fragility / uprightness, time window relaxation, and merchant / pick-up point consistency) are introduced to calibrate the weight and feasibility of group-buying. If necessary, the group-buying strategy is rolled back to pure geographical group-buying based on uncertainty gating. For orders in the drone delivery mode, the vehicle delivers the order to the transfer station task queue. Spatial clustering is used to group the delivery points, and the grouping and queue priority are corrected by combining the task score and feasibility mask output by the multimodal large model (based on retrieval-enhanced weather / airspace restrictions, load and range constraints, etc.) to reduce the drone flight path length and energy consumption. When there is external knowledge conflict or high uncertainty, the strategy is rolled back to a combination strategy based on hard constraints.

5. The vehicle-machine-electronics collaborative delivery method in the large-model-driven UAV relay mode according to claim 1, characterized in that, The second-layer deep value learning model in step S4 refers to acquiring relevant data about the agent based on the obtained optimal delivery mode information and vehicle status, constructing an MDP model, establishing a vehicle state value prediction model based on a neural network, and receiving candidate paths and risk summaries, pattern priors and uncertainties output by the multimodal large model in parallel, as well as retrieving enhanced external knowledge (weather / airspace restrictions, road events, time-of-use electricity pricing, charging station congestion prediction, etc.) to limit the feasible region and calibrate the value estimation. The specific state space, action space, and reward function design of the model are as follows: State space design: The input variable is the state vector S in model construction, denoted as: ,in It is the time required for the vehicle status to become "schedulable". It represents the remaining battery percentage when the vehicle status changes to "schedulable". It refers to the road network location where a vehicle is when its status changes to "schedulable". The vehicle is from its current location Distance to the nearest charging station It's the daily time step. It is a date and time step, supplemented with value priors and uncertainties provided by a multimodal large model, as well as retrieval-enhancing external knowledge features (including weather / airspace restrictions, road events, time-of-use pricing, and charging station congestion predictions), which are then normalized / weighted and incorporated into the model. ; Motion space design: Action space This indicates that the agent is at time step All possible action sets ( ), including agent dwell time ( ), transfer ( ),Charge( ), discharge ( ),match( If under a scheduling instruction, the specific actions are as follows: 1) Stay ( The vehicles will remain stationary, awaiting future travel demand. 2) Transfer ( The dispatch instruction guides the vehicle to another location; the vehicle is within the dispatch time window. If you continue moving within the area, the new location you can reach is a distance from your original location in time. Within the specified locations; the system decides whether to transfer the vehicle to the location based on the spatiotemporal value of different transfer locations; a multimodal large model generates a set of candidate destinations and reachability ranges based on demand heat prediction and road risk, and the policy network selects the best option within the candidates to pass the instruction, so that the vehicle can transfer to a location with a higher probability of ordering or go to a charging station; 3) Charging ( If the vehicle is at a charging station, the vehicle that receives this instruction will continue charging or start charging immediately; the instruction is invalid if the vehicle is already fully charged. 4) Discharge ( If the vehicle is at a charging station, the vehicle that receives this instruction will continue to discharge or begin discharging immediately; if the vehicle is already out of power, the instruction is invalid. 5) Matching ( This instruction assigns a trip order to a vehicle, which is related to the order set. Related actions; In maintaining the action set Under the premise of no change, an action feasibility mask and soft preference coefficient generated by a multimodal large model are introduced: when the retrieved enhanced information or semantic rules indicate that an action is not feasible under the current constraints, the action is masked; when multiple actions are feasible, the output of the policy network is temperature scaled and calibrated based on prior and uncertainty to reduce the risk of misjudgment due to incomplete information. For transfer ( The large model generates a set of candidate destinations and feasible reach intervals based on demand heat prediction and road risk. The policy network selects transfer locations with high expected value within the candidate set. For charging / discharging ( The large model maps grid announcements, time-of-use pricing, and site congestion predictions into allowable time windows and penalty coefficients, which are used to restrict or encourage energy management behavior during specific periods. Reward function design: According to the system dispatch instructions, if the electric truck selects to pause ( ), transfer ( The rewards received by operators are distributed using the following methods: This indicates that since the instruction does not generate any direct costs or benefits, both benefits are zero. If the electric truck chooses to charge ( ), discharge ( The revenue obtained by the operators is calculated using the formula... It means that among them A piecewise function of time, representing the time-of-use electricity price at different times of the day; : Duration of charging and discharging action (hours); : Indicates average charging power and average discharging power; , : Represents the efficiency coefficient for charging and discharging. ; This represents the cost of battery degradation per unit of energy during charging and discharging. This refers to ancillary service rewards for discharging electricity during peak grid hours. : Indicator function, takes a value of 1 during peak grid hours (i.e. peak electricity price hours), otherwise takes a value of 0; This indicates the cost incurred during charging. Indicates the benefits generated by the discharge; Under the dispatch instruction, if the electric truck selects to match ( Each "match" action is associated with an order. Association for any order in the set Its key time parameters are defined as follows: :Order The initiation time; Order Estimated pickup time; Order Estimated delivery time; :Order Maximum acceptable waiting time; :Order The shortest travel time from the origin to the destination; :Order Required delivery time limit; in: There are restrictions on the waiting time for food delivery; the pickup time must not be later than the latest pickup time. Under the condition of satisfying the waiting time constraint, the reward for this action is represented as follows: in This represents the revenue a vehicle generates per unit of time. It represents the quantity of the order. (For a single order) this item is 0, when This item will be positive when (group buying); This indicates compensation for the customer's waiting time. This indicates the group-buying reward coefficient; Based on this, reward shaping weights are set according to the analysis results of constraints such as "time window, temperature control, fragility, uprightness, and priority" in the multimodal large model, and adaptive recalibration is performed on items such as waiting / default / deviation; when the retrieval enhancement information indicates no-fly zone, severe weather, temporary control, peak electricity price, or abnormal site congestion, risk penalties are superimposed or conservative rollback is triggered. Optionally, the interpretable output of the large model on action preferences can be used as a distillation regularization for offline training to shorten the cold start and improve sample efficiency.

6. The vehicle-machine-electronics collaborative delivery method in the large-model-driven UAV relay mode according to claim 1, characterized in that, Step S5 includes: When a transit station receives multiple drone delivery tasks, it first sorts the delivery points according to their distance from the transit station, and then calibrates the sorting results using task scores generated by a multimodal large model (which at least reflect time window relaxation, temperature control / fragility / uprightness constraint costs, weather and airspace risks, and customer priority). If the model uncertainty exceeds a threshold or there is an external knowledge conflict, it reverts to distance-based sorting. Based on this, it groups the drones into groups using a preset maximum number of groups. The task is divided into several task groups, and the grouping is corrected according to the payload / range / weather and airspace feasibility mask output by the large model, so that the task points in the same group are geographically close to each other and satisfy the feasible domain constraint. Then, the idle UAV is matched with the task group one by one. During the matching process, the "UAV-task group" matching prior and infeasibility mask generated by the large model are introduced to reduce the solution cost and mismatch risk. During execution, the battery status of each drone is monitored, and peak-shifting charging and charging queue optimization are performed for drones returning to the transfer station based on enhanced time-of-use pricing and charging station congestion prediction. Once all delivery tasks in any task group have been completed, the task is removed from the list of tasks to be dispatched. Each of the drones is configured to: execute the assigned task group and fly along a round-trip route of "transfer station - delivery point - return station"; replan or backtrack when the large model risk summary indicates that the route is not feasible or the risk increases; and autonomously move into the charging facility for energy replenishment if its battery level is lower than the preset threshold after returning to the transfer station and landing.

Images