Urban low-altitude logistics unmanned aerial vehicle dynamic path planning system

Abstract

This invention relates to a dynamic path planning system for urban low-altitude logistics drones, specifically in the field of intelligent path planning. The system generates a high-fidelity virtual world model in real time, providing a dynamic environmental perception foundation for subsequent planning. It employs a wolf pack algorithm to rapidly generate diverse candidate paths, ensuring the global exploratory nature of the initial solution. In a virtual simulation environment, it simulates multi-round game theory and negotiation among agents, iteratively adjusting path indices based on Nash equilibrium to converge to a Pareto-optimal set of cooperative paths, achieving global optimization at the group level rather than local stress responses. It monitors and predicts potential conflicts in real time, fine-tuning speed and heading angle through a distributed consensus protocol to smoothly resolve conflicts without disrupting the overall cooperative solution. This overcomes the drawbacks of centralized systems, such as high computational latency and limited field of view of single-machine sensors, enabling drone swarms to autonomously coordinate and efficiently avoid obstacles in highly dynamic urban low-altitude environments, significantly improving the safety and operational efficiency of last-mile logistics delivery.

Description

Technical Field

[0001] This invention relates to the field of intelligent path planning, and more specifically, to a dynamic path planning system for urban low-altitude logistics drones. Background Technology

[0002] With the rapid development of the urban low-altitude economy, logistics drones are increasingly widely used in last-mile delivery. In high-density urban areas, such as large central business districts or old residential communities, dozens of logistics drones often perform delivery tasks simultaneously in the same airspace. These drones are constrained by a variety of internal factors during operation, including their remaining battery capacity, payload weight, and differences in flight performance parameters. At the same time, the external environment is also full of dynamic uncertainties: sudden high-rise building construction may extend temporary cranes, medical emergency helicopters may temporarily occupy part of the airspace, strong gusts of wind or sudden rain may force drones to adjust their flight altitude and speed, and even instantaneous dense crowds caused by ground traffic congestion may require drones to actively reduce noise or change routes to avoid disturbing people. In such a complex environment, any change in the path of any drone may trigger a chain reaction due to airspace congestion, leading to chaos in the air traffic order of the entire fleet and even collision accidents.

[0003] Currently, technical solutions for multi-UAV cooperative path planning mainly focus on static path planning for single or small numbers of UAVs, or passive obstacle avoidance using simple rules of maintaining a safe distance. However, in the face of large-scale, highly dynamic urban low-altitude logistics scenarios, existing centralized path planning systems suffer from significant computational latency, making it difficult to complete global path replanning within milliseconds. Traditional algorithms such as A*Sat or Dijkstra's algorithm, while suitable for pathfinding on static maps, cannot effectively handle dynamic obstacle avoidance under multi-target and multi-constraint conditions, especially in achieving cooperative optimization at the group level. Furthermore, the sensors on a single UAV have limited detection range and cannot detect distant objects or blind spots in advance. When encountering sudden obstacles within the area, only reactive, localized evasive maneuvers are possible. Such localized adjustments are often not globally optimal solutions and may even exacerbate congestion and conflict risks in the overall airspace. Therefore, there is an urgent need for a system capable of constructing and continuously updating a virtual world model of the city's low-altitude airspace in real time. This model should not only include static building structures but also cover the positions and movement intentions of all dynamic obstacles and other drones. Based on this, the concept of social simulation is introduced, treating each drone as an intelligent agent with independent decision-making capabilities. By simulating their mutual game, negotiation, and cooperation, a globally optimal group path replanning can be achieved, thereby fundamentally improving the operational efficiency and safety of low-altitude logistics at the city's last mile. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a dynamic path planning system for urban low-altitude logistics drones. The system utilizes a virtual construction module, a path generation module, a collaborative decision-making module, and a conflict resolution module to solve the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically including:

[0006] Virtual building module: Receives real-time data from ground base stations, communication networks and airborne sensors, builds and continuously updates a multi-dimensional state tensor containing static building outlines, dynamic obstacle boundaries and all UAV motion states, and generates a virtual world model;

[0007] Path generation module: Based on the virtual world model, the wolf pack algorithm is used to generate an initial path set containing multiple candidate paths for each drone;

[0008] Collaborative Decision Module: Input the initial path set into the virtual simulation environment constructed by the virtual world model, define each drone as an intelligent agent, execute multi-round game and negotiation between the intelligent agents, iteratively adjust the path index according to the Nash equilibrium, and generate a collaborative path set composed of multiple drones;

[0009] Conflict resolution module: It monitors the deviation of each UAV from the cooperative path set in real time, predicts potential conflict points based on the current position and heading of the UAV, and sends alternative fine-tuning schemes pre-calculated by the cooperative decision module to the relevant UAVs. Through a distributed consensus protocol, it fine-tunes the speed and heading angle to guide the UAVs back to the cooperative path or smoothly resolve the conflict.

[0010] In a preferred embodiment, the virtual building module performs the following operations when building and continuously updating the multidimensional state tensor:

[0011] First, the received real-time data is aligned according to a unified spatiotemporal reference, and a fourth-order multidimensional state tensor is constructed. The four dimensions of the multidimensional state tensor correspond to the sliding window length, the total number of UAVs in the current airspace, the state feature dimension of each UAV, and the number of sampling points of dynamic obstacle boundaries.

[0012] Then, the multidimensional state tensor is decomposed into a tensor ring decomposition, which represents it as a series of cyclic condensations of three-dimensional core tensors. The dimension of each core tensor is determined by its preceding and following ring ranks and the size of that dimension. The ring rank is much smaller than the original size of each dimension.

[0013] The decomposition process employs an incremental alternating least squares algorithm. When a new frame of data arrives, only the core tensor related to the latest time slice and other core tensors adjacent to it in the tensor ring structure are updated, while the remaining core tensors remain unchanged, thus achieving online updates with low storage consumption. Finally, the compressed set of core tensors is used as the virtual world model.

[0014] In a preferred embodiment, during the process of generating the virtual world model, the low-rank structure of the tensor ring is also used to complete the missing data caused by sensor packet loss or occlusion, and to filter out noise.

[0015] Specifically, an observation mask tensor with the same dimension as the multidimensional state tensor is constructed. The observation position elements in the observation mask tensor are all one, and the missing position elements are all zero. Then, the completion and filtering are achieved by solving an optimization problem with nuclear norm regularization. The objective function of the optimization problem includes a fitting error term controlled by the observation mask tensor and a nuclear norm regularization term applied to each core tensor. The coefficient of the nuclear norm regularization term is a preset regularization coefficient.

[0016] The optimization problem converges within milliseconds using a soft thresholding iterative algorithm, thereby completing the missing data while suppressing Gaussian noise and outliers, resulting in a more accurate set of core tensors.

[0017] In a preferred embodiment, the path generation module uses a wolf pack algorithm based on a virtual world model to generate an initial path set for each drone. Specifically, the operation is as follows:

[0018] First, an individual objective function is defined for each drone. The individual objective function comprehensively evaluates the estimated flight time, estimated energy consumption, and safety penalty of a candidate path. The estimated flight time is calculated based on the path length and average cruising speed.

[0019] The estimated energy consumption is calculated by integrating the path length with the wind speed drag function interpolated from the virtual world model;

[0020] Safety penalties are calculated using an exponential function based on the distance from the waypoint to the nearest obstacle.

[0021] Then, using the static obstacle distribution in the virtual world model, an initial reference path is generated for each drone using the fast travel method, and the initial reference path is used as the initial position of the alpha wolf in the wolf pack.

[0022] Next, the remaining drones are randomly divided into two categories: Detective Wolves and Fierce Wolves. Each Detective Wolves generates multiple variant paths around its initial reference path by applying random perturbations that follow a Lévy distribution, thus forming the initial path set.

[0023] In a preferred embodiment, during the process of generating the initial path set using the wolf pack algorithm, the following operations are also performed:

[0024] In each iteration, the scout wolf updates its initial path set based on the current position of the alpha wolf and the path information of neighboring scout wolves. The update rule includes a convergent component learned from the alpha wolf and a collaborative component interacting with neighbors, and is superimposed with a Levy flight exploration step size that decays with the iteration generation.

[0025] The wolves launch a local siege around the alpha wolf's path, and generate multiple closely surrounding candidate paths by spiraling around the alpha wolf's path in a plane perpendicular to the path direction.

[0026] After each generation, the individual objective function is calculated for all paths, and multi-objective sorting is performed according to non-dominated sorting and crowding distance. The top few paths on the Pareto front are selected as the elite path set and directly retained to the next generation. At the same time, the individual with the most non-dominated paths and the largest crowding distance is selected as the new generation alpha wolf.

[0027] After multiple iterations, each drone obtains a set of paths containing multiple candidate paths, each with an estimated energy consumption and time cost.

[0028] In a preferred embodiment, after inputting the initial path set into the virtual simulation environment constructed by the virtual world model and defining each drone as an intelligent agent, the collaborative decision-making module performs the following operations:

[0029] First, a potential game model is constructed for each agent. The strategy is defined as the index of a candidate path selected by the agent from its initial path set. The overall situation function value is defined as the sum of the individual objective function values ​​of all agents plus the weighted sum of the spatiotemporal overlap between any two candidate paths. The payoff function for each agent is defined as the negative marginal contribution of the overall situation function value, such that the payoff function satisfies the potential game conditions.

[0030] Then, in each iteration, each agent calculates the reward for each candidate path based on the strategies of other agents in the previous round, and updates the strategy selection probability using a log-linear learning rule;

[0031] Each agent independently samples according to the updated probability distribution to obtain the new strategy for this round.

[0032] In a preferred embodiment, an inertial momentum update is also introduced in each iteration, and the agent retains the previous policy with a probability that decays with each iteration, and adopts the results of log-linear learning with complementary probabilities.

[0033] After each round, the global situation function value under the current strategy combination is calculated. If the global situation function value does not decrease for several consecutive rounds, a restart mechanism is triggered to reset the strategies of all agents to the strategy combination corresponding to the historical lowest point of the global situation function value.

[0034] The iteration termination condition is that the change in the global situation function value is less than the preset convergence threshold for multiple consecutive rounds;

[0035] After convergence, the set of cooperative paths corresponding to the current policy combination is output. The policy combination output at the end of the iteration is defined as the converged solution. A local Pareto test is performed on the converged solution to check whether there is an agent that achieves Pareto improvement by unilaterally switching candidate paths. If so, the improvement is accepted and convergence is restarted until there are no more Pareto improvement opportunities.

[0036] In a preferred embodiment, the conflict resolution module performs the following operations when monitoring the real-time positions of each UAV and predicting potential conflict points:

[0037] First, Gaussian process regression is used to nonparametrically model the trajectory of each UAV within the future prediction time window. Based on multiple recent historical locations, the mean prediction and covariance matrix are obtained through the squared exponential kernel function, thereby obtaining the probability distribution of the trajectory.

[0038] Then, for any two drones, the cumulative percentage of time during which their distance is less than a preset safe distance threshold within the prediction time window is calculated as the collision probability. When the collision probability exceeds a preset threshold, a conflict warning is triggered, and the predicted conflict time and location are recorded.

[0039] In a preferred embodiment, when sending a negotiation request and alternative fine-tuning scheme to the relevant drone, the following operations are performed:

[0040] For each pair of conflicting UAVs detected, the matching set of fine-tuning parameters is first retrieved from the library of alternative fine-tuning schemes pre-calculated and stored by the collaborative decision-making module;

[0041] If the solution cannot be found, a local quadratic programming problem is solved online. The optimization objective is to minimize the weighted sum of squares of the changes in velocity and heading angle. The constraint is that the distance between the UAV and the other UAV at the moment of conflict should not be less than the safe distance threshold plus the safety margin. The constraint is linearized by first-order Taylor expansion and then solved to obtain a set of optimal changes in velocity and heading angle. Together with the symmetric solution of the other UAV, this constitutes an alternative fine-tuning scheme.

[0042] In a preferred embodiment, when fine-tuning the speed and heading angle through a distributed consensus protocol, the following operations are performed:

[0043] The adjustment vector consisting of the speed change and heading angle change in the alternative fine-tuning scheme is used as the initial value. The relevant UAVs conduct multiple rounds of iterative negotiation through the average consensus algorithm. In each round of iteration, each UAV updates its own adjustment based on the weighted average of the current adjustment of its neighboring UAVs that have conflict with it. After a finite number of rounds of iteration, the adjustment of all relevant UAVs converges to a consistent value.

[0044] Then, the consistent value is allocated to multiple future control cycles, and cubic spline interpolation is used to generate a smooth transition trajectory. At the same time, the current conflicting geometric configuration and the generated fine-tuning scheme are stored in the alternative fine-tuning scheme library for future use.

[0045] The beneficial effects of this invention are as follows: By generating a high-fidelity virtual world model in real time, it provides a dynamic environmental perception basis for subsequent planning; by using the wolf pack algorithm to quickly generate diverse candidate paths, it ensures the global exploratory nature of the initial solution; by simulating multi-round game and negotiation among intelligent agents in a virtual simulation environment, it iteratively adjusts the path index based on Nash equilibrium, converges to the Pareto optimal set of cooperative paths, and achieves global optimization at the group level rather than local stress response; by monitoring and predicting potential conflicts in real time, it fine-tunes speed and heading angle through a distributed consensus protocol, smoothly resolves conflicts without destroying the overall cooperative scheme; and by overcoming the shortcomings of high computational latency and limited field of view of single-machine sensors in centralized systems, it enables UAV swarms to autonomously coordinate and efficiently avoid obstacles in highly dynamic urban low-altitude environments, significantly improving the safety and operational efficiency of last-mile logistics delivery. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention;

[0047] Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0050] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0051] Example 1

[0052] This embodiment provides, for example Figure 1-2 The urban low-altitude logistics drone dynamic path planning system shown includes:

[0053] Virtual building module: Receives real-time data from ground base stations, communication networks and airborne sensors, builds and continuously updates a multi-dimensional state tensor containing static building outlines, dynamic obstacle boundaries and all UAV motion states, and generates a virtual world model;

[0054] Path generation module: Based on the virtual world model, the wolf pack algorithm is used to generate an initial path set containing multiple candidate paths for each drone;

[0055] Collaborative Decision Module: Input the initial path set into the virtual simulation environment constructed by the virtual world model, define each drone as an intelligent agent, execute multi-round game and negotiation between the intelligent agents, iteratively adjust the path index according to the Nash equilibrium, and generate a collaborative path set composed of multiple drones;

[0056] Conflict resolution module: It monitors the deviation of each UAV from the cooperative path set in real time, predicts potential conflict points based on the current position and heading of the UAV, and sends alternative fine-tuning schemes pre-calculated by the cooperative decision module to the relevant UAVs. Through a distributed consensus protocol, it fine-tunes the speed and heading angle to guide the UAVs back to the cooperative path or smoothly resolve the conflict.

[0057] In this embodiment, it is specifically necessary to explain that when the virtual construction module constructs and continuously updates the multidimensional state tensor, it performs the following operations:

[0058] First, the received real-time data is aligned according to a unified spatiotemporal reference, and a fourth-order multidimensional state tensor is constructed. The four dimensions of the multidimensional state tensor correspond to the sliding window length, the total number of UAVs in the current airspace, the state feature dimension of each UAV, and the number of sampling points at the dynamic obstacle boundary. The value of the sliding window length is determined based on the system's real-time requirements and the sensor sampling frequency, with a typical value of thirty to sixty sampling periods, corresponding to a historical data window of three to six seconds. The state feature dimension of each UAV is fixed at six, including three-dimensional position components and three-dimensional velocity components. The number of sampling points at the dynamic obstacle boundary is between twenty and fifty points, adaptively adjusted according to the complexity of the obstacle shape.

[0059] Then, tensor ring decomposition is performed on the multidimensional state tensor, representing it as a series of cyclically condensed three-dimensional core tensors. The dimension of each core tensor is determined by the ring rank before and after it, as well as the size of that dimension. The ring rank is much smaller than the original dimensions of each dimension. The specific calculation process of tensor ring decomposition is as follows: First, the fourth-order multidimensional state tensor is expanded along the first dimension to obtain a two-dimensional matrix. Then, singular value decomposition is performed on this matrix, and the left singular vectors corresponding to the first R0 largest singular values ​​are truncated and retained as the first dimension of the first core tensor. Next, the truncated result is condensed with the original tensor to obtain a new third-order tensor. The above process is repeated to extract all four core tensors in sequence. The ring rank is between five and ten, and the specific value is determined by cross-validation based on the actual data dimensions to ensure a balance between compression ratio and reconstruction accuracy.

[0060] The decomposition process employs an incremental alternating least squares algorithm. When a new frame of data arrives, only the core tensor related to the latest time slice and other core tensors adjacent to it in the tensor ring structure are updated, while the remaining core tensors remain unchanged, thus achieving online updates with low storage consumption. The specific calculation process of the incremental alternating least squares algorithm is as follows: When a new frame of data arrives, the new data is first appended as a time slice to the latest position of the multidimensional state tensor, while the oldest time slice is removed to maintain the sliding window length unchanged. Then, all core tensors except the first and last core tensors are fixed, and alternating least squares updates are performed only on the first and last core tensors. During each update, the core tensor to be updated is treated as a variable, and the remaining core tensors are treated as constants. The updated core tensors are obtained by solving a linear least squares problem. After repeating the alternating updates two to three times, the algorithm converges, completing one online update. Finally, the compressed set of core tensors is used as the virtual world model.

[0061] In the process of generating the virtual world model, the low-rank structure of the tensor ring is also used to complete the missing data caused by sensor packet loss or occlusion, and to filter out noise.

[0062] Specifically, an observation mask tensor with the same dimension as the multidimensional state tensor is constructed. In the observation mask tensor, the elements at the observation positions are set to one, and the elements at the missing positions are set to zero. Then, completion and filtering are achieved by solving an optimization problem with kernel norm regularization. The objective function of the optimization problem includes a fitting error term controlled by the observation mask tensor and a kernel norm regularization term applied to each core tensor. The coefficient of the kernel norm regularization term is a preset regularization coefficient. The value of the regularization coefficient is between 0.01 and 0.1, and the specific value is adaptively adjusted according to the signal-to-noise ratio level. The lower the signal-to-noise ratio, the larger the value of the regularization coefficient is, so as to enhance the suppression effect on noise.

[0063] The optimization problem converges within milliseconds using a soft-thresholding iterative algorithm, thereby suppressing Gaussian noise and outliers while filling in missing data, resulting in a more accurate set of core tensors. The specific computational process of the soft-thresholding iterative algorithm is as follows: First, initialize all core tensors to their values ​​at the end of the previous update; second, reconstruct an approximate multidimensional state tensor using the current set of core tensors through a tensor loop; third, calculate the residual tensor between the approximate tensor and the observed data, and perform a Hadamard product between the residual tensor and the observation mask tensor to obtain a correction term containing only the observation position error; fourth, superimpose the correction term onto the approximate tensor... On the tensor, the updated complete tensor estimate is obtained; in the fifth step, the updated complete tensor estimate is decomposed into a tensor ring again to obtain a new set of core tensors; in the sixth step, a soft thresholding operation is performed on each core tensor, that is, after the core tensor is expanded into a matrix, singular value decomposition is performed, singular values ​​smaller than the regularization coefficient are set to zero, and the remaining singular values ​​are retained after subtracting the regularization coefficient; in the seventh step, the soft-thresholded core tensors are reassembled back into tensor form; in the eighth step, steps two through seven are repeated until the change between the core tensor sets of two adjacent iterations is less than the preset convergence threshold, which is 10 to the power of negative four.

[0064] In this embodiment, it should be specifically explained that the path generation module uses a wolf pack algorithm based on the virtual world model to generate an initial path set for each drone. The specific operation is as follows:

[0065] First, an individual objective function is defined for each drone. This function comprehensively evaluates the estimated flight time, estimated energy consumption, and safety penalty of a candidate path. The estimated flight time is calculated based on the path length and average cruising speed. The specific calculation method for the individual objective function is as follows: multiply the estimated flight time by a first time weighting coefficient, add the estimated energy consumption multiplied by a second time weighting coefficient, and add the safety penalty multiplied by a third time weighting coefficient. The sum of the three weighting coefficients is one, which is dynamically adjusted according to the urgency of the task. For example, in an urgent delivery task, the first time weighting coefficient is 0.6, the second time weighting coefficient is 0.3, and the third time weighting coefficient is 0.1. The estimated flight time is equal to the total path length divided by the average cruising speed. The average cruising speed is obtained by looking up a table based on the drone's remaining battery power and payload weight. The higher the remaining battery power and the lighter the payload, the faster the average cruising speed.

[0066] The estimated energy consumption is calculated based on the path length and the integral of the wind speed drag function interpolated from the virtual world model. The specific calculation method for the estimated energy consumption is as follows: multiply the total path length by the basic energy consumption coefficient, and add the integral value of the wind speed drag along each segment of the path multiplied by the wind drag energy consumption coefficient. The value of the wind speed drag function at any point on the path is obtained by reconstruction interpolation through the tensor loop in the virtual world model. The interpolation process is as follows: first, normalize the three-dimensional coordinates of the point to the grid index range of the tensor loop, then use the loop factor tensor set to calculate the wind speed vector of the point through cyclic shrinking, and then project the wind speed vector onto the tangent direction of the path to obtain the drag component.

[0067] The safety penalty is calculated using an exponential function based on the distance from the waypoint to the nearest obstacle. The specific calculation method for the safety penalty is as follows: For each waypoint on the path, calculate the Euclidean distance from the waypoint to the nearest static or dynamic obstacle in the virtual world model, then divide the negative value of this distance by the safety width coefficient and take the exponent, and then sum the exponent values ​​of all waypoints. The safety width coefficient is set to two meters, so that when the waypoint is less than one meter away from the obstacle, the penalty increases sharply, and when the distance is greater than five meters, the penalty approaches zero.

[0068] Then, using the static obstacle distribution in the virtual world model, an initial reference path is generated for each drone using the fast travel method, and this initial reference path is used as the initial position of the alpha wolf in the wolf pack. The specific calculation process of the fast travel method is as follows: First, the static obstacle areas in the virtual world model are marked as impassable areas, and the remaining areas are marked as passable areas. Second, with the drone's current position as the starting point and the target position as the ending point, a cost field is constructed on the passable area. The value of the cost field is obtained by weighting the distance from the point to the nearest obstacle and the Euclidean distance to the ending point. Third, starting from the starting point, the wavefront is propagated outward step by step using an upwind propagation method to solve the equation of motion. The wavefront is propagated outward step by step until it reaches the ending point. Fourth, the wavefront is traced backward from the ending point along the propagation direction to obtain the shortest time path from the starting point to the ending point, which is the initial reference path.

[0069] Next, the remaining UAVs are randomly divided into two categories: Detective Wolf and Fierce Wolf. Each Detective Wolf generates multiple variant paths around its initial reference path by applying a random perturbation following a Lévy distribution, forming an initial path set. The specific method for generating the Lévy distribution random perturbation is as follows: first, a random number is drawn from a standard normal distribution, and then another random number is drawn from another independent standard normal distribution. The ratio of these two random numbers is calculated and multiplied by a scaling factor to obtain the Lévy step size. This Lévy step size is multiplied by a random unit direction vector to obtain a perturbation vector in three-dimensional space. This perturbation vector is added to each waypoint on the initial reference path to obtain a variant path. Each Detective Wolf repeats the above process five times to generate five variant paths, which, together with the initial reference path, form an initial path set containing six candidate paths.

[0070] During the process of generating the initial path set using the wolf pack algorithm, the following operations are also performed:

[0071] In each iteration, the scout wolf updates its initial path set based on the current alpha wolf position and the path information of neighboring scout wolves. The update rule includes a convergence component learned from the alpha wolf and a cooperative component interacting with neighbors, plus a Levy flight exploration step size that decays with each iteration. The specific calculation process for scout wolf path updates is as follows: For each candidate path of a scout wolf, firstly, the difference vector between the path and the current alpha wolf path is calculated and multiplied by a convergence learning coefficient of 0.5 to obtain the convergence component; then, the sum of the difference vectors between the optimal paths of all neighboring scout wolves within the scout wolf's communication range and the current path is calculated and multiplied by a cooperative learning coefficient of 0.3 to obtain the cooperative component; then, a random perturbation vector following a Levy distribution is generated and multiplied by the current exploration step size to obtain the exploration component; finally, the convergence component, cooperative component, and exploration component are added together and superimposed on the original path to obtain the updated path; the initial value of the exploration step size is 0.1 meters, which decreases by 0.002 meters each generation until it reaches a minimum value of 0.01 meters.

[0072] The wolves execute a local encirclement around the alpha wolf's path, generating multiple closely surrounding candidate paths through a spiral search on a plane perpendicular to the path, centered on the alpha wolf's path. The specific calculation process of the wolf's spiral search is as follows: For each waypoint on the alpha wolf's path, with that waypoint as the center, in a plane perpendicular to the path's tangent, with a radius of five meters, eight directions are taken at 45-degree intervals from zero to 360 degrees. In each direction, a candidate point is obtained by extending radially outward by two meters. These candidate points are then connected in a spiral sequence to form a spiral path encircling the alpha wolf's path. Each wolf generates three such spiral paths, corresponding to different spiral radii: five meters, seven meters, and nine meters.

[0073] At the end of each generation, the individual objective function is calculated for all paths, and multi-objective ranking is performed according to non-dominated ranking and crowding distance. The top few paths on the Pareto front are selected as the elite path set and directly retained to the next generation. At the same time, the individual with the most non-dominated paths and the largest crowding distance is selected as the new generation alpha wolf. The specific calculation process of non-dominated ranking is as follows: For all candidate paths, compare the individual objective function values ​​of any two paths. If the three objective function values ​​of the first path are all no greater than those of the second path and at least one is less than that of the second path, then the first path is said to dominate the second path. All paths not dominated by any other path are classified into the first level. After removing the first level, continue to find the non-dominated paths in the remaining paths and classify them into the second level, and so on until all paths are classified.

[0074] The calculation process of congestion distance is as follows: within the same level, each objective function is sorted from smallest to largest. The congestion distance of the boundary path is set to infinity. The congestion distance of the middle path is equal to the sum of the differences between the two adjacent paths on the objective function divided by the difference between the maximum and minimum values ​​of the objective function. The top eight paths on the Pareto front are selected as the elite path set, that is, the eight paths with the largest congestion distance in the first level.

[0075] The selection method for the new generation alpha wolf is as follows: count the number of non-dominated paths owned by each drone, select the drone with the most non-dominated paths, and if there are multiple drones, further compare the average congestion distance of all non-dominated paths of these drones, and select the drone with the largest average distance as the new generation alpha wolf.

[0076] After multiple iterations, each drone obtains a set of paths containing multiple candidate paths, each with an estimated energy consumption and time cost. The preset number of iterations is forty generations. The iteration terminates when the alpha wolf path no longer changes within forty generations or five consecutive generations.

[0077] In this embodiment, it is specifically noted that after the collaborative decision-making module inputs the initial path set into the virtual simulation environment constructed by the virtual world model and defines each drone as an intelligent agent, it performs the following operations:

[0078] First, a potential game model is constructed for each agent. The strategy is defined as the index of a candidate path chosen by the agent from its initial path set. The overall situation function is defined as the sum of the individual objective function values ​​of all agents plus the weighted sum of the spatiotemporal overlap between any two candidate paths. A payoff function for each agent is defined as the negative marginal contribution of the overall situation function, ensuring that the payoff function satisfies the potential game conditions. The specific calculation method for the overall situation function is as follows: For each drone, based on its currently selected candidate path, the individual objective function value corresponding to that candidate path is read from the path set output by the path generation module. Then, for any two drones, the spatiotemporal overlap of their selected candidate paths is calculated using trajectory extrapolation and spatial occupancy grids in the virtual world model. The overlap value ranges from zero to one. The overall situation function value is obtained by summing all individual objective function values ​​and adding the sum of all pairwise overlaps multiplied by the conflict penalty coefficient. The conflict penalty coefficient is set to ten.

[0079] The reward function of each agent is defined as: the individual objective function value of the agent's current candidate path plus the sum of the spatiotemporal overlap between the agent and the candidate paths selected by all other agents multiplied by the conflict penalty coefficient, and then the negative value is taken.

[0080] In each iteration, each agent calculates the reward for each candidate path based on the strategies of other agents in the previous round, and updates the strategy selection probability using a log-linear learning rule. The rationality coefficient increases with the number of iterations, enabling early exploration and later utilization. The specific calculation process of the log-linear learning rule is as follows: For each candidate path of an agent, first calculate the reward value of that candidate path; then multiply the reward value by the rationality coefficient of the current round, and then take the exponential function to obtain the exponential value of the candidate path; sum the exponential values ​​of all candidate paths to obtain the denominator; divide the exponential value of the candidate path by the denominator to obtain the probability that the candidate path is selected in this round. The rationality coefficient increases as follows: the initial rationality coefficient is 0.1, the maximum rationality coefficient is 10, and the total number of iterations is 100. At the beginning of each iteration, the rationality coefficient is equal to the initial rationality coefficient plus the difference between the maximum rationality coefficient and the initial rationality coefficient multiplied by the ratio of the current round number to the total number of iterations.

[0081] Each agent independently samples according to the updated probability distribution to obtain the new strategy for this round. The specific process of independent sampling is as follows: First, a random number uniformly distributed between zero and one is generated; then, starting from the first candidate path of the agent, the probability of each candidate path being selected is accumulated. When the accumulated sum is greater than or equal to the random number for the first time, the index of the candidate path currently accumulated is selected as the new strategy for this round.

[0082] In each iteration, an inertial momentum update is introduced. The agent retains the previous policy with a probability that decays with each iteration, and uses the result of log-linear learning with complementary probabilities. The specific process of the inertial momentum update is as follows: First, the momentum probability is calculated based on the current round number. The initial value of the momentum probability is 0.3, which is reduced by 0.003 in each round until the minimum value of zero is reached. Then, a random number uniformly distributed between zero and one is generated. If the random number is less than the momentum probability, the agent retains the previous policy unchanged; otherwise, the agent samples the probability distribution calculated by the log-linear learning rule to obtain the new policy for this round.

[0083] After each round, the global situation function value under the current strategy combination is calculated. If the global situation function value does not decrease for several consecutive rounds, a restart mechanism is triggered, resetting the strategies of all agents to the strategy combination corresponding to the historical lowest point of the global situation function value. The criterion for not decreasing for several consecutive rounds is that the global situation function value is not lower than the global situation function value of the previous round for three consecutive rounds. The specific operation of the restart mechanism is as follows: maintain a global variable to record the minimum value of the global situation function value that has appeared in history and its corresponding strategy combination; when a restart is triggered, the strategies of all agents are forcibly set to the strategy combination corresponding to the historical lowest point, and the rationality coefficient is reset to the initial rationality coefficient of 0.1 to start the exploration again.

[0084] The iteration termination condition is that the change amplitude of the global situation function value is less than the preset convergence threshold for multiple consecutive rounds, and the convergence threshold is 10 to the power of negative 4, or the maximum number of iteration rounds is reached; the criterion for judging that the change amplitude is less than the convergence threshold for multiple consecutive rounds is that the absolute change of the global situation function value is less than 10 to the power of negative 4 for ten consecutive rounds; the maximum number of iteration rounds is 100 rounds.

[0085] After convergence, the set of cooperative paths corresponding to the current policy combination is output. The policy combination output at the end of the iteration is defined as the converged solution. A local Pareto test is performed on the converged solution to check whether any agent has achieved Pareto improvement by unilaterally switching candidate paths. If so, the improvement is accepted and convergence is restarted until there are no more Pareto improvement opportunities. The specific process of the local Pareto test is as follows: For each drone in the converged solution, the policy of the drone is switched to each other candidate path in its initial path set in turn, while keeping the policies of all other drones unchanged. For each switch, the individual objective function value and the global situation function value of all drones are recalculated. If the individual objective function value of the drone decreases (i.e., the profit increases) after the switch, and the individual objective function values ​​of all other drones do not increase, then a Pareto improvement is considered to exist. The improvement is accepted, the current policy combination is updated to the switched policy combination, and the iterative convergence process is restarted from the new policy combination until there are no more Pareto improvement opportunities after traversing all candidate paths for all drones.

[0086] The above process is not a random simulation, but rather a real game and negotiation between agents based on defined potential game models, log-linear learning rules, and inertial momentum updates. By iteratively adjusting the path index, a set of cooperative paths composed of multiple UAVs is finally generated. This set satisfies the Nash equilibrium condition and is at the Pareto optimal front. In addition, while generating the set of cooperative paths, the cooperative decision-making module also pre-calculates a set of alternative fine-tuning schemes for each cooperative path, including the allowed speed and heading angle fine-tuning amounts and their corresponding costs, and stores them in the alternative fine-tuning scheme library for the conflict resolution module to call.

[0087] In this embodiment, it should be specifically explained that when the conflict resolution module monitors the location of each UAV in real time and predicts potential conflict points, it performs the following operations:

[0088] First, the deviation of each drone from the cooperative path set is monitored in real time. If the deviation between the actual position of a drone and the expected position at the corresponding time on the cooperative path exceeds a preset threshold (e.g., 0.5 meters), the drone is judged to have veered off course, triggering a conflict warning. Simultaneously, Gaussian process regression is used to nonparametrically model the trajectory of each drone within the future prediction time window. Based on multiple recent historical positions, the mean prediction and covariance matrix are obtained through a squared exponential kernel function, thus obtaining the probability distribution of the trajectory. The specific calculation process of Gaussian process regression is as follows: First, the position coordinates of the drones at the ten most recent historical moments are collected to form a historical trajectory dataset. Second, a squared exponential kernel function is defined, which measures the correlation between any two historical positions. The value of the kernel function is calculated by dividing the square of the Euclidean distance between the two positions by one. The first step is to square the length scale parameter and take its negative exponent. The length scale parameter is set to two meters. The second step is to construct a covariance matrix using the historical trajectory dataset and the kernel function. The element in the i-th row and j-th column of the matrix is ​​the kernel function value between the i-th historical position and the j-th historical position. The third step is to calculate the kernel function value vector between any future prediction time and all historical times. The fourth step is to multiply the kernel function value vector with the inverse of the covariance matrix, and then multiply it by the historical position vector to obtain the mean prediction for that prediction time. The sixth step is to calculate the variance prediction for that prediction time, which is the result of subtracting the kernel function value vector multiplied by the inverse of the covariance matrix and the transpose of the kernel function value vector from the kernel function value value of that prediction time and itself. The mean prediction and the variance prediction together constitute the trajectory probability distribution for that prediction time.

[0089] Then, for any two drones, the cumulative percentage of time during which their distance is less than a preset safe distance threshold within the prediction time window is calculated as the collision probability. The preset safe distance threshold is two meters. When the collision probability exceeds the preset threshold, a conflict warning is triggered, and the predicted conflict time and location are recorded. The specific calculation method for the collision probability is as follows: the prediction time window is divided into one hundred time steps at equal intervals, with each time step lasting 0.03 seconds. For each time step, Monte Carlo sampling is performed based on the trajectory probability distribution of the two drones to generate one thousand possible future position samples. The Euclidean distance between the two drones in each sample is calculated, and the proportion of samples with a distance less than two meters to the total number of samples is counted as the instantaneous collision probability of that time step. The average of the instantaneous collision probabilities of all time steps is used to obtain the collision probability within the entire prediction time window. The preset threshold is 0.05. When the collision probability exceeds 0.05, a conflict warning is triggered, and the time step with the highest instantaneous collision probability is recorded as the conflict time, and the midpoint of the average predicted positions of the two drones at that time is recorded as the conflict location.

[0090] When sending a negotiation request and alternative fine-tuning options to the relevant drone, the following operations are performed:

[0091] For each detected pair of conflicting UAVs, a matching set of fine-tuning parameters is first retrieved from a pool of candidate fine-tuning schemes pre-calculated and stored by the collaborative decision-making module. This pool of candidate fine-tuning schemes is calculated concurrently with the generation of the collaborative path set by the collaborative decision-making module: for each collaborative path, a set of permissible speed and heading angle fine-tuning amounts is pre-calculated to ensure that the adjusted path maintains a safe distance from other UAVs without compromising the stability of the overall situation function. The retrieval method for the candidate fine-tuning scheme pool is as follows: using the relative position vector and relative velocity vector of the two UAVs at the moment of conflict as feature vectors, the Euclidean distance between this feature vector and the feature vectors of all existing schemes in the scheme pool is calculated, and the top three schemes with the smallest distances are selected as candidates. If the smallest distance is less than a preset matching threshold of 0.5 m / s, the set of fine-tuning parameters in the candidate scheme is directly adopted; otherwise, it is determined that no search can be performed.

[0092] If the query cannot be found, a local quadratic programming problem is solved online. The optimization objective is to minimize the weighted sum of squares of the changes in velocity and heading angle. The constraint is that the distance between the adjusted drone and the enemy drone at the moment of conflict should not be less than the safe distance threshold plus a safety margin. The constraints are linearized by a first-order Taylor expansion to obtain a set of optimal changes in velocity and heading angle. These, along with the symmetric solution of the enemy drone, constitute an alternative fine-tuning scheme. The specific solution process of the local quadratic programming problem is as follows: First, define the optimization variables as the changes in velocity and heading angle. The range of the velocity change is from -2 m / s to +2 m / s, and the range of the heading angle change is from -15 degrees to +15 degrees. Second, define the optimization objective as the square of the velocity change multiplied by the weight coefficient 1, plus the square of the heading angle change multiplied by the weight coefficient 0.01, and then multiplied by half. Third, define the constraints. To ensure the distance between the two UAVs at the moment of conflict is no less than the safe distance threshold of two meters plus a safety margin of 0.5 meters, i.e., no less than 2.5 meters; the fourth step is to perform a first-order Taylor expansion of the nonlinear distance function in the constraints at the initial predicted position to obtain a linearized constraint inequality, where the Jacobian matrix is ​​composed of the partial derivatives of the position with respect to velocity and heading angle in the UAV dynamics model; the fifth step is to solve the quadratic programming problem using the Lagrange multiplier method. First, a Lagrange function is constructed, which includes the product of the optimization objective and the linearized constraints. Then, partial derivatives are taken with respect to the optimization variables and the Lagrange multipliers and set to zero to obtain a set of linear equations; the sixth step is to solve this set of linear equations to obtain the optimal velocity change and heading angle change; the seventh step is to assign the optimal change obtained from the solution to the current UAV, and at the same time, invert the change and assign it to the other UAV, forming a pair of symmetrical alternative fine-tuning schemes;

[0093] When fine-tuning speed and heading angle using a distributed consensus protocol, the following operations are performed:

[0094] The adjustment vector, composed of the velocity and heading angle changes in the candidate fine-tuning schemes, is used as the initial value. Related UAVs undergo multiple rounds of iterative negotiation using an average consensus algorithm. In each iteration, each UAV updates its own adjustment based on a weighted average of the current adjustment values ​​of its conflicting neighboring UAVs. After a finite number of iterations, the adjustment values ​​of all related UAVs converge to a consistent value. The specific calculation process of the average consensus algorithm is as follows: First, initialize the adjustment vector of each conflicting UAV to the corresponding value in the candidate fine-tuning scheme; second, determine the conflicting neighbor set of each UAV, i.e., other UAVs that have a direct conflict relationship with that UAV; third, calculate the adjacency weight of each UAV. The adjacency weight is equal to one divided by the larger of the number of conflicting neighbors of the current drone and the number of conflicting neighbors of the opposing drone, plus one; the fourth step is that in each iteration, each drone collects the current adjustment vector of all conflicting neighbors, calculates the difference between the adjustment vector of each neighbor and its own adjustment vector, multiplies it by the corresponding adjacency weight, and then sums them to obtain the correction term; the fifth step is to multiply the correction term by the step size of 0.5 and add it to its own adjustment vector to obtain the updated adjustment vector; the sixth step is to repeat the fourth and fifth steps for a total of five iterations. When the maximum difference between the adjustment vectors of all related drones is less than 0.01 meters per second or 0.1 degree after five iterations, it is considered to have converged, and the average value of the adjustment vectors of all drones is taken as the consensus value;

[0095] Then, the consistency value is allocated to multiple future control cycles, and a transition trajectory is generated using cubic spline interpolation for smooth execution. Simultaneously, the adjusted path is compared with the cooperative path set to ensure that the adjusted path remains within the allowable deviation range of the cooperative path set (e.g., the distance between the adjusted waypoint and the corresponding waypoint on the cooperative path does not exceed 0.8 meters). If the allowable deviation is exceeded, the cooperative decision-making module is automatically triggered to re-enter the system. Starting from the current actual position, a local replanning is performed based on the original initial path set to generate a new cooperative path set, thereby avoiding chain conflicts with other UAVs caused by local adjustments. At the same time, the geometric configuration of the current conflict and the generated fine-tuning scheme are stored in the alternative fine-tuning scheme library for future use. The specific process of generating the transition trajectory using cubic spline interpolation is as follows: First, the velocity change and heading angle change in the consistency value are... The adjustment is uniformly distributed across the next five control cycles, each cycle lasting 0.1 seconds, meaning the adjustment will be completed within 0.5 seconds. The second step involves adding the allocated adjustment amount to the current speed and heading angle at the start of each control cycle to obtain the target speed and heading angle at the end of that cycle. The third step involves constructing five piecewise cubic polynomials, starting with the current speed and heading angle and ending with the target speed and heading angle at the end of the five cycles. Each piecewise polynomial satisfies the boundary conditions of position continuity, velocity continuity, and acceleration continuity. The fourth step involves solving for the coefficients of each piecewise polynomial to obtain a smooth velocity and heading angle curve. The fifth step involves the UAV executing flight commands according to the speed and heading angle calculated by the corresponding piecewise polynomial within each control cycle, thus achieving a smooth transition.

[0096] The specific method for storing the current conflict geometry and the generated fine-tuning scheme into the alternative fine-tuning scheme library is as follows: using the relative position vector and relative velocity vector of the two UAVs at the moment of conflict as the key, and the generated optimal velocity change and heading angle change as the value, the scheme is stored in the scheme library for direct retrieval and calling in future similar conflict scenarios.

[0097] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0098] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0103] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A dynamic path planning system for urban low-altitude logistics drones, characterized in that, Specifically, it includes: Virtual building module: Receives real-time data from ground base stations, communication networks and airborne sensors, builds and continuously updates a multi-dimensional state tensor containing static building outlines, dynamic obstacle boundaries and all UAV motion states, and generates a virtual world model; Path generation module: Based on the virtual world model, the wolf pack algorithm is used to generate an initial path set containing multiple candidate paths for each drone; Collaborative Decision Module: Input the initial path set into the virtual simulation environment constructed by the virtual world model, define each drone as an intelligent agent, execute multi-round game and negotiation between the intelligent agents, iteratively adjust the path index according to the Nash equilibrium, and generate a collaborative path set composed of multiple drones; Conflict resolution module: It monitors the deviation of each UAV from the cooperative path set in real time, predicts potential conflict points based on the current position and heading of the UAV, and sends alternative fine-tuning schemes pre-calculated by the cooperative decision module to the relevant UAVs. Through a distributed consensus protocol, it fine-tunes the speed and heading angle to guide the UAVs back to the cooperative path or smoothly resolve the conflict.

2. The dynamic path planning system for urban low-altitude logistics drones according to claim 1, characterized in that: When the virtual building module builds and continuously updates the multidimensional state tensor, it performs the following operations: First, the received real-time data is aligned according to a unified spatiotemporal reference, and a fourth-order multidimensional state tensor is constructed. The four dimensions of the multidimensional state tensor correspond to the sliding window length, the total number of UAVs in the current airspace, the state feature dimension of each UAV, and the number of sampling points of dynamic obstacle boundaries. Then, the multidimensional state tensor is decomposed into a tensor ring decomposition, which represents it as a series of cyclic condensations of three-dimensional core tensors. The dimension of each core tensor is determined by its preceding and following ring ranks and the size of that dimension. The ring rank is much smaller than the original size of each dimension. The decomposition process employs an incremental alternating least squares algorithm. When a new frame of data arrives, only the core tensor related to the latest time slice and other core tensors adjacent to it in the tensor ring structure are updated, while the remaining core tensors remain unchanged, thus achieving online updates with low storage consumption. Finally, the compressed set of core tensors is used as the virtual world model.

3. The dynamic path planning system for urban low-altitude logistics drones according to claim 2, characterized in that: In the process of generating the virtual world model, the low-rank structure of the tensor ring is also used to complete the missing data caused by sensor packet loss or occlusion, and to filter out noise. Specifically, an observation mask tensor with the same dimension as the multidimensional state tensor is constructed. The observation position elements in the observation mask tensor are all one, and the missing position elements are all zero. Then, the completion and filtering are achieved by solving an optimization problem with nuclear norm regularization. The objective function of the optimization problem includes a fitting error term controlled by the observation mask tensor and a nuclear norm regularization term applied to each core tensor. The coefficient of the nuclear norm regularization term is a preset regularization coefficient. The optimization problem converges within milliseconds using a soft thresholding iterative algorithm, thereby completing the missing data while suppressing Gaussian noise and outliers, resulting in a more accurate set of core tensors.

4. The dynamic path planning system for urban low-altitude logistics drones according to claim 3, characterized in that: The path generation module uses a wolf pack algorithm based on a virtual world model to generate an initial set of paths for each drone. The specific operation is as follows: First, an individual objective function is defined for each drone. The individual objective function comprehensively evaluates the estimated flight time, estimated energy consumption, and safety penalty of a candidate path. The estimated flight time is calculated based on the path length and average cruising speed. The estimated energy consumption is calculated by integrating the path length with the wind speed drag function interpolated from the virtual world model; Safety penalties are calculated using an exponential function based on the distance from the waypoint to the nearest obstacle. Then, using the static obstacle distribution in the virtual world model, an initial reference path is generated for each drone using the fast travel method, and the initial reference path is used as the initial position of the alpha wolf in the wolf pack. Next, the remaining drones are randomly divided into two categories: Detective Wolves and Fierce Wolves. Each Detective Wolves generates multiple variant paths around its initial reference path by applying random perturbations that follow a Lévy distribution, thus forming the initial path set.

5. A dynamic path planning system for urban low-altitude logistics drones according to claim 4, characterized in that: During the process of generating the initial path set using the wolf pack algorithm, the following operations are also performed: In each iteration, the scout wolf updates its initial path set based on the current position of the alpha wolf and the path information of neighboring scout wolves. The update rule includes a convergent component learned from the alpha wolf and a collaborative component interacting with neighbors, and is superimposed with a Levy flight exploration step size that decays with the iteration generation. The wolves launch a local siege around the alpha wolf's path, and generate multiple closely surrounding candidate paths by spiraling around the alpha wolf's path in a plane perpendicular to the path direction. After each generation, the individual objective function is calculated for all paths, and multi-objective sorting is performed according to non-dominated sorting and crowding distance. The top few paths on the Pareto front are selected as the elite path set and directly retained to the next generation. At the same time, the individual with the most non-dominated paths and the largest crowding distance is selected as the new generation alpha wolf. After multiple iterations, each drone obtains a set of paths containing multiple candidate paths, each with an estimated energy consumption and time cost.

6. A dynamic path planning system for urban low-altitude logistics drones according to claim 5, characterized in that: After inputting the initial path set into the virtual simulation environment constructed by the virtual world model and defining each drone as an intelligent agent, the collaborative decision-making module performs the following operations: First, a potential game model is constructed for each agent. The strategy is defined as the index of a candidate path selected by the agent from its initial path set. The overall situation function value is defined as the sum of the individual objective function values ​​of all agents plus the weighted sum of the spatiotemporal overlap between any two candidate paths. The payoff function for each agent is defined as the negative marginal contribution of the overall situation function value, such that the payoff function satisfies the potential game conditions. Then, in each iteration, each agent calculates the reward for each candidate path based on the strategies of other agents in the previous round, and updates the strategy selection probability using a log-linear learning rule; Each agent independently samples according to the updated probability distribution to obtain the new strategy for this round.

7. A dynamic path planning system for urban low-altitude logistics drones according to claim 6, characterized in that: In each iteration, inertial momentum updates are also introduced. The agent retains the previous policy with a probability that decays with each iteration, and adopts the results of log-linear learning with complementary probabilities. After each round, the global situation function value under the current strategy combination is calculated. If the global situation function value does not decrease for several consecutive rounds, a restart mechanism is triggered to reset the strategies of all agents to the strategy combination corresponding to the historical lowest point of the global situation function value. The iteration termination condition is that the change in the global situation function value is less than the preset convergence threshold for multiple consecutive rounds; After convergence, the set of cooperative paths corresponding to the current policy combination is output. The policy combination output at the end of the iteration is defined as the converged solution. A local Pareto test is performed on the converged solution to check whether there is an agent that achieves Pareto improvement by unilaterally switching candidate paths. If so, the improvement is accepted and convergence is restarted until there are no more Pareto improvement opportunities.

8. A dynamic path planning system for urban low-altitude logistics drones according to claim 7, characterized in that: When the conflict resolution module monitors the location of each drone in real time and predicts potential conflict points, it performs the following operations: First, Gaussian process regression is used to nonparametrically model the trajectory of each UAV within the future prediction time window. Based on multiple recent historical locations, the mean prediction and covariance matrix are obtained through the squared exponential kernel function, thereby obtaining the probability distribution of the trajectory. Then, for any two drones, the cumulative percentage of time during which their distance is less than a preset safe distance threshold within the prediction time window is calculated as the collision probability. When the collision probability exceeds a preset threshold, a conflict warning is triggered, and the predicted conflict time and location are recorded.

9. A dynamic path planning system for urban low-altitude logistics drones according to claim 8, characterized in that: When sending a negotiation request and alternative fine-tuning options to the relevant drone, the following operations are performed: For each pair of conflicting UAVs detected, the matching set of fine-tuning parameters is first retrieved from the library of alternative fine-tuning schemes pre-calculated and stored by the collaborative decision-making module; If the solution cannot be found, a local quadratic programming problem is solved online. The optimization objective is to minimize the weighted sum of squares of the changes in velocity and heading angle. The constraint is that the distance between the UAV and the other UAV at the moment of conflict should not be less than the safe distance threshold plus the safety margin. The constraint is linearized by first-order Taylor expansion and then solved to obtain a set of optimal changes in velocity and heading angle. Together with the symmetric solution of the other UAV, this constitutes an alternative fine-tuning scheme.

10. A dynamic path planning system for urban low-altitude logistics drones according to claim 9, characterized in that: When fine-tuning speed and heading angle using a distributed consensus protocol, the following operations are performed: The adjustment vector consisting of the speed change and heading angle change in the alternative fine-tuning scheme is used as the initial value. The relevant UAVs conduct multiple rounds of iterative negotiation through the average consensus algorithm. In each round of iteration, each UAV updates its own adjustment based on the weighted average of the current adjustment of its neighboring UAVs that have conflict with it. After a finite number of rounds of iteration, the adjustment of all relevant UAVs converges to a consistent value. Then, the consistent value is allocated to multiple future control cycles, and cubic spline interpolation is used to generate a smooth transition trajectory. At the same time, the current conflicting geometric configuration and the generated fine-tuning scheme are stored in the alternative fine-tuning scheme library for future use.

Images