Large Language Model-based Heuristic Path Planning Method, Device and Medium for Unmanned Aerial Vehicles

Abstract

The present invention belongs to the technical field of UAV path planning, and discloses a heuristic path planning method, device and medium for UAVs based on large language models. The method includes: modeling the UAV, planning and modeling the flight path based on the flight mission; collecting a data set including flight problem instances and solutions as input parameters for the large language model, training the large language model and then using it for path planning of the flight mission. The heuristic path planning method for UAVs based on large language models provided by the present invention combines the precise path finding ability of traditional path planning algorithms with the global reasoning ability of large language models. The aim is to improve the efficiency of path finding in large-scale scenarios, especially in terms of time and space complexity, while maintaining the validity of the path. By integrating the advantages of the two methods, the limitations of traditional algorithms in terms of computing and memory are solved, and the validity required for path finding is not affected.

Description

Technical Field

[0001] The present invention belongs to the technical field of UAV path planning. Specifically, it relates to a heuristic path planning method, device, and medium for UAVs based on large language models. Background Technique

[0002] As the core of UAV technology, the development of trajectory planning has received extensive attention globally. Early trajectory planning mainly relied on manual settings by technicians, and its accuracy was closely related to the professional level and experience of technicians. However, with the progress of informatization and intelligence, the manually drawn trajectory lines can no longer meet the requirements of the modern battlefield environment. The development of high-tech has brought about the continuous emergence of various reconnaissance means, greatly enhancing the ability to collect environmental information, thus providing strong support for planning high-precision trajectory flight routes.

[0003] During the trajectory planning process, it is necessary to consider the self-condition constraints closely related to the UAV and the constraints related to the mission. More importantly, it is necessary to consider the flight performance constraints, such as different turning radii and different speeds due to different aircraft models. How to quantify these constraint conditions and threat factors, transform them into symbols that the planning system can recognize, and comprehensively consider them is of great significance for the development of trajectory planning technology. Path planning is a basic scientific problem in UAVs and autonomous navigation, which requires the path from the starting point to the end point to be both efficient and able to avoid obstacles. Traditional algorithms, such as A* and its variants, although they can ensure the effectiveness of the path, their computational and memory efficiency decreases significantly when the state space increases. In contrast, large large language models can provide a global insight into the environment through a more extensive environmental analysis by understanding the context. However, they have deficiencies in detailed spatial and temporal reasoning, which may lead to invalid paths. Summary of the Invention

[0004] In view of this, the present invention provides a heuristic path planning method, device, and medium for UAVs based on large language models to solve the above problems.

[0005] To solve the above technical problems, the present invention provides a heuristic path planning method for UAVs based on large language models, including:

[0006] Model the UAV and obtain key flight features and flight objective functions based on the flight mission;

[0007] Collect a data set including flight problem instances and solutions as input parameters for the large language model, train the large language model, and then use it for path planning of the flight mission;

[0008] Among them, the path planning includes:

[0009] Build the infrastructure and data for path planning, calibrate the starting node, and then use the large language model to generate a target list and calculate the priority of nodes;

[0010] Based on the priority of the nodes, update the node ratio in sequence and complete the path planning; and based on the dynamic evaluation of the environmental changes made by the large language model, reconstruct and dynamically update the planned path in real time.

[0011] As an optional method, planning and modeling the flight path based on the flight mission includes:

[0012] Based on the flight mission, obtain the key flight features in the flight mission; perform simultaneous optimization on each key flight feature based on the constraint conditions to generate a flight objective function; the constraint conditions include safety constraint conditions and concealment constraint conditions.

[0013] As an optional method, the key flight features include distance features, average speed features, probability of being monitored by radar, probability of being hit, and probability of fuel consumption.

[0014] As an optional method, modeling the unmanned aerial vehicle includes:

[0015] Based on the physical properties of the target aircraft, construct the dynamic modeling, kinematic modeling, and aerodynamic modeling of the target aircraft;

[0016] The components of the dynamic modeling of the target aircraft include the position of the target aircraft in the inertial coordinate system, the velocity components, angular velocity components, and force components in the body coordinate system, and the mass of the target aircraft;

[0017] The components of the kinematic modeling of the target aircraft include the roll angle, pitch angle, yaw angle, the principal axis moment of inertia and product moment of inertia of the target aircraft in the inertial matrix, and the moment components of the target aircraft in the body coordinate system;

[0018] The components of the aerodynamic modeling of the target aircraft include lift, drag, side force, air density, flight speed, projected area of the wing, angle of attack, and sideslip angle.

[0019] As an optional method, the flight problem instance includes the starting point, ending point, environmental map, and environmental dynamic factors of the flight mission; the environmental dynamic factors are associated with the environmental map.

[0020] As an optional method, training the large language model includes:

[0021] Initialize the flight mission planning conditions, and input the starting point, ending point, heuristic function, neighbor function, environmental map, and environmental dynamic factors;

[0022] Call a large language model to generate a target list, predict environmental changes and / or environmental dynamic factors based on flight problem instances; and perform incremental updates according to the evaluation results to reconstruct the planned path.

[0023] As an alternative, path planning also includes:

[0024] Initialize the path record, add the target node to the list of nodes to be processed, and initialize the best cost and estimated cost of the target node to zero, and set the costs of all other nodes to infinity;

[0025] Calculate the best cost and estimated cost of each node in the list of nodes to be processed, establish the priority of each node for forming the planned path, and form the planned path;

[0026] Input the starting point, target node, and environmental information to the large language model to generate a target list;

[0027] Calculate the priority for each node, and the priority is used for sorting and path selection in path planning;

[0028] When the list of nodes to be processed is not empty or the priority of the starting point is less than the priority of any node in the list, select the node with the lowest priority in the list of nodes to be processed as the starting point for planning. After calculating the priority for each node in turn, call the large language model to predict new obstacles or other dynamic changes according to environmental changes and update the priorities and adjacent relationships of relevant nodes;

[0029] Recalculate the priorities of relevant nodes, add them to the list of nodes to be processed, and then recalculate again to reconstruct the priority order of each node and update the path.

[0030] As an alternative, calculating the priority includes, for each node, calculating its priority based on the best cost, estimated cost, and heuristic function; where

[0031] When calculating the priority for each node in turn:

[0032] If the best cost of this node is greater than its estimated cost, update its best cost to the estimated cost, remove it from the list of nodes to be processed, update the estimated costs of all its neighbor nodes, and add the neighbor nodes not in the list to the list of nodes to be processed;

[0033] If the best cost of this node is not greater than its estimated cost, reset its best cost to infinity, update the estimated costs of itself and all its neighbor nodes at the same time, and add the neighbor nodes not in the list to the list of nodes to be processed.

[0034] On the other hand, the present invention also provides an electronic device, comprising: a memory for storing a computer program; a processor for implementing the steps of any of the above-mentioned large language model-based drone heuristic path planning methods when executing the computer program.

[0035] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned large language model-based drone heuristic path planning method are implemented.

[0036] The beneficial effects of the present invention are as follows:

[0037] The large language model-based drone heuristic path planning method provided by the present invention combines the precise path finding ability of traditional path planning algorithms with the global reasoning ability of large large language models. This hybrid method aims to improve the efficiency of path finding, especially in terms of time and space complexity, while maintaining the validity of the path, especially in large-scale scenarios. By integrating the advantages of the two methods, the path planning algorithm integrating large large language models solves the limitations of traditional algorithms in terms of computing and memory without affecting the validity required for path finding. Description of the Drawings

[0038] Figure 1 It is a schematic flow chart of the large language model-based drone heuristic path planning method provided by the embodiment of the present invention;

[0039] Figure 2 It is a schematic flow chart of the flight path modeling provided by the embodiment of the present invention;

[0040] Figure 3 It is a schematic flow chart of the path planning algorithm provided by the embodiment of the present invention;

[0041] Figure 4 It is a visualization diagram of the simulation result in a certain case provided by the embodiment of the present invention. Detailed Embodiments

[0042] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments.

[0043] Modern path planning technology needs to comprehensively consider the constraints of the UAV itself (such as flight performance, fuel limitation, etc.) and the constraints related to the mission (such as mission time, mission priority, etc.). These external factors need to be effectively quantified and converted into symbols that the planning system can recognize for comprehensive consideration and processing. Traditional path planning algorithms, such as heuristic algorithms and their variants (such as the A* algorithm), have been widely used in the field of path planning. These algorithms usually use heuristic search methods, combined with cost functions and heuristic functions, to find the optimal path from the starting point to the ending point. The D* algorithm is a classic heuristic search algorithm that guides the search process by evaluating the path cost (usually including the actual cost from the starting point to the current node and the estimated cost from the current node to the ending point). The D* algorithm can find the path with the lowest cost in a static environment, but in a dynamic environment, environmental changes will cause the need to recalculate the entire path. The D* algorithm is an extension of the A* algorithm, designed specifically for dynamic environments. It uses an incremental update mechanism to only update the affected part of the path when the environment changes, rather than recalculating the entire path. This makes the path planning algorithm integrated with large language models perform excellently in dynamic environments, but its implementation complexity is relatively high. Large language models, with powerful context understanding and reasoning capabilities, can conduct global analysis and insight into complex environments. Although large language models are insufficient in detailed spatial and temporal reasoning, they can provide a global perspective of the environment, identify potential threats and obstacles. Large language models can process a large amount of environmental data, identify and classify different types of threat factors, and provide corresponding avoidance suggestions. Through the global understanding of the environment, large language models can identify the best path planning strategy to help UAVs effectively avoid obstacles in complex environments.

[0044] Therefore, the present invention proposes a heuristic path planning method for UAVs based on large language models. This method is based on the integration of large language models, combining the accurate path finding ability of traditional path planning algorithms with the global reasoning ability of large language models. This hybrid method aims to improve the efficiency of path finding, especially in terms of time and space complexity, while maintaining the effectiveness of the path, especially in large-scale scenarios. By integrating the advantages of the two methods, an improved heuristic search UAV path planning method based on large language models solves the limitations of traditional algorithms in terms of calculation and memory without affecting the effectiveness required for path finding.

[0045] Please refer to Figures 1 - 3 , as an alternative, in this embodiment, a certain type of fixed-wing UAV is selected to illustrate the implementation process:

[0046] Model fixed-wing UAVs and plan and model flight paths based on flight missions. Collect a dataset containing flight problem examples and solutions as input parameters for a large language model. After training, the large language model is used to plan flight paths. Path planning includes:

[0047] Build the basic structure and data for path planning, calibrate the starting node, use the large language model to generate a list of targets to be processed and calculate the priority of the nodes; based on the priority of the nodes, update the node ratio in sequence and complete the path planning; and reconstruct and dynamically update the path planning in real time based on the dynamic assessment of environmental changes made by the large language model.

[0048] Specifically, this embodiment first performs trajectory planning modeling, converts the description of the above trajectory planning into an optimization objective formula, and regards it as a multi-objective optimization problem. This problem requires optimizing multiple objective functions under the premise of satisfying various constraints. Then, a simulation environment is constructed based on the fixed-wing UAV model. Then, by providing a small number of examples to the large language model, the model can learn patterns and rules from these examples, thereby improving its accuracy and efficiency on new tasks. This method is particularly suitable for scenarios where data is scarce. Finally, based on the analysis of global information by the large language model, the heuristic algorithm is guided to expand the node strategy to improve efficiency. It includes:

[0049] The first step in modeling a flight mission is to capture the key flight characteristics of the mission based on the aircraft's dynamic and aerodynamic equations, establish physical constraints, and design requirements for flight performance. A series of functions are defined to accurately describe these constraints and requirements, ensuring that the model accurately reflects the aircraft's flight characteristics and operational limitations.

[0050] Furthermore, the algorithm design utilizes the D* algorithm, which dynamically updates path and estimated costs to adapt to environmental changes. Furthermore, the path planning algorithm, integrated with a large language model, combines global insights with local search mechanisms, utilizing few-shot fine-tuning and recursive path evaluation techniques to improve the efficiency and optimality of path planning. This integrated approach is better able to cope with dynamic changes in complex environments.

[0051] Finally, after determining the starting and ending points of the planning problem, the obstacle information is first input into the algorithm. Subsequently, the planned path from the starting point to the ending point is generated according to the steps specified in the pseudo code.

[0052] In the above steps, the overall objective function f(x) is established and the distance function is constructed , assuming the total distance is the sum of the Euclidean distances of adjacent points on the path

[0053] Among them, i is an index variable that ranges from 0 to n-1 , where n is the number of points. Therefore, , , and represent the coordinates of the i -th point, while , , and represent the coordinates of the i+1 -th point. This formula calculates the sum of the distances between each point.

[0054]

[0055] Among them, T(x) represents the average time spent, V is the average speed of the drone:

[0056] After that, the constraint function C(x) is constructed:

[0057]

[0058] where RadarRisk() and MissileRisk() represent the probabilities of being detected by the radar and being hit, and are the weight coefficients used to adjust the influence of different risks.

[0059] The fuel consumption constraint function is constructed :

[0060]

[0061] In the formula, FuelRate represents the fuel consumption from xi to xi+1, which depends on the speed and altitude of the aircraft, is the time interval from to , calculated as

[0062] After that, the aircraft is modeled. First, the aircraft dynamics equation is constructed, and its formula is as follows:

[0063]

[0064] x, y, z are the positions of the drone in the inertial coordinate system. u, v, w are the velocity components in the body coordinate system. p, q, r are the angular velocity components in the body coordinate system. m is the mass of the drone. X, Y, Z are the force components in the body coordinate system.

[0065] Construct the kinematic equations, and their formulas are as follows:

[0066]

[0067] , , where \(\phi\), \(\theta\), \(\psi\) are the roll, pitch, and yaw angles, \(I_{xx}\), \(I_{yy}\), \(I_{zz}\) are the moments of inertia about the principal axes of the inertia matrix, \(I_{xy}\), \(I_{yz}\), \(I_{zx}\) are the products of inertia of the inertia matrix, and \(L\), \(M\), \(N\) are the moment components in the body coordinate system.

[0068] Construct the aerodynamic equations of the aircraft, and their formulas are as follows:

[0069]

[0070] where \(L\) is the lift force, i.e., the force perpendicular to the flight direction, which helps the aircraft overcome gravity. \(D\) is the drag force, i.e., the force opposite to the flight direction, which hinders the aircraft's forward movement. \(Y\) is the side force, i.e., the force parallel to the side of the aircraft, which affects the lateral stability of the aircraft. \(\rho\) is the air density, usually in kilograms per cubic meter (kg / m³). \(V\) is the flight speed, usually in meters per second (m / s). \(S\) is the reference area, usually the projected area of the wing, usually in square meters (m²). \(C_{L}\) is the lift coefficient, \(C_{D}\) is the drag coefficient, \(C_{Y}\) is the side force coefficient, and these coefficients depend on the angle of attack and the sideslip angle.

[0071] After the construction is completed, training the large language model includes: initializing the flight mission planning conditions, inputting the starting point, ending point, heuristic function, neighbor function, environmental map, and environmental dynamic factors; calling the large language model to generate a target list, predicting environmental changes and / or environmental dynamic factors based on the flight problem instance; and performing incremental updates according to the evaluation results to reconstruct the planned path.

[0072] At this time, all the preliminary work for path planning is completed, and the UAV path planning is ready to start. In this embodiment, the process of path planning is as follows:

[0073] Initialize the path record, add the target node to the list of nodes to be processed, and initialize the best cost and estimated cost of the target node to zero, and set the costs of all other nodes to infinity;

[0074] Calculate the best cost and estimated cost of each node in the list of nodes to be processed, establish the priority of each node for forming the planned path, and form the planned path;

[0075] Input the starting point, target node, and environmental information to the large language model to generate a target list;

[0076] Calculate the priority for each node, where the priority is used for sorting and path selection in path planning;

[0077] When the list to be processed is not empty or the priority of the starting point is less than the priority of any node in the list, select the node with the lowest priority from the list to be processed to plan the starting point. After calculating the priority for each node in turn, call the large language model to predict new obstacles or other dynamic changes according to the environmental changes and update the priorities and adjacent relationships of the relevant nodes;

[0078] Recalculate the priorities of the relevant nodes, add them to the list to be processed, and then recalculate again, reconstruct the priority order of each node and update the path.

[0079] Among them, calculating the priority includes calculating the priority for each node based on the best cost, estimated cost, and heuristic function; among them,

[0080] When calculating the priority for each node in turn:

[0081] If the best cost of the node is greater than its estimated cost, update its best cost to the estimated cost, remove it from the list to be processed, update the estimated costs of all its neighbor nodes, and add the neighbor nodes not in the list to the list to be processed;

[0082] If the best cost of the node is not greater than its estimated cost, reset its best cost to infinity, update the estimated costs of itself and all its neighbor nodes at the same time, and add the neighbor nodes not in the list to the list to be processed.

[0083] In the above steps, the target list is generated by the large language model, which generates a possible target list to be processed based on environmental information (such as obstacles, starting point, target location, etc.). These target nodes may be some key intermediate states, or special nodes that the algorithm needs to pay attention to during the path planning process. Its role is to provide a high-level target guidance for path planning, so that the method provided in this embodiment can better cope with the changes in the dynamic environment. The target nodes in this list represent important reference points in path planning, but they are not the operation targets when the algorithm directly performs path search.

[0084] The list to be processed stores the nodes that need to be processed during the current path search process. During the path planning process of this embodiment, the nodes in the list to be processed will be continuously sorted and selected according to their costs (i.e., priorities). The nodes in this list are the actual operation objects during the search process. When a node is processed, its priority will be calculated, and path updates (such as cost updates, re-evaluation of neighbor nodes, etc.) will depend on these nodes.

[0085] Based on this, the target list and the list to be processed exist independently, and the nodes in the target list do not directly participate in the priority calculation. The main role of the target list is to provide a high-level goal for path planning, while the list to be processed contains the nodes that the algorithm actually needs to process during the specific search process. When calculating the priority of a node, it is the nodes in the list to be processed that are involved, rather than the nodes in the target list. Each node in the list to be processed calculates its priority based on its current cost and heuristic estimate value to determine its priority.

[0086] In addition, as the drone moves in the environment, the environment may change, such as the appearance or disappearance of obstacles, or a change in the radar range. To cope with these changes, this embodiment performs dynamic evaluation through a large language model. This includes: invoking the large language model to predict changes in the position of obstacles or targets based on the current environmental state. According to the prediction of the model, update the obstacle information or target position in the environment. Update the cost function and adjacency relationship of the affected nodes. Add the affected nodes back to the list to be processed and recalculate their costs. Therefore, once a valid path is found, it is checked whether the estimated cost of the starting point is equal to the optimal cost. If they are equal, it means that the optimal path has been found. Then, backtracking calculation is performed to reconstruct the optimal path from the starting point to the target. Among them, Figure 4 This is a demonstration of a simulation visualization interface based on the solution of this embodiment.

[0087] Based on the above solution, this embodiment combines the D* algorithm with a large language model, and uses the prediction ability of the large language model to anticipate environmental changes, thereby optimizing the path planning of the drone. The D* algorithm itself is suitable for path planning in dynamic environments and can quickly adjust the path, while the large language model provides predictions of environmental changes (such as obstacles, dynamic objects, etc.), enhancing the robustness and intelligence of the algorithm in complex scenarios. This method is suitable for tasks where the drone flies in an unknown or constantly changing environment.

[0088] On the other hand, this embodiment also provides an electronic device, including: a memory for storing a computer program; a processor for implementing the steps of the drone heuristic path planning method based on a large language model as described in any one of the above when executing the computer program.

[0089] On the other hand, this embodiment also provides a computer-readable storage medium, in which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the drone heuristic path planning method based on a large language model as described above.

[0090] The above are only the preferred embodiments of the present invention. It should be noted that the above preferred embodiments should not be construed as limiting the present invention, and the protection scope of the present invention should be defined by the scope of the claims. For those of ordinary skill in the art, without departing from the spirit and scope of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as within the protection scope of the present invention.

Claims

1. A heuristic path planning method for drones based on large language models, characterized in that, Including: Model the UAV, and obtain key flight features and a flight objective function based on the flight mission; Collect a dataset including flight problem instances and solutions as input parameters for the large language model. After training the large language model, use it for path planning of the flight mission; Among them, the training of the large language model includes: Initialize the flight mission planning conditions, and input the starting point, ending point, heuristic function, neighbor function, environmental map, and environmental dynamic factors; Call the large language model to generate a target list, predict environmental changes and / or environmental dynamic factors based on the flight problem instances; and perform incremental updates according to the evaluation results to reconstruct the planned path; The path planning includes: Construct the basic structure and data for path planning, calibrate the starting node, and then use the large language model to generate a target list and calculate the priority of the nodes; Based on the priority of the nodes, update the node ratio in sequence and complete the path planning; and based on the dynamic evaluation of environmental changes made by the large language model, reconstruct and dynamically update the planned path in real time.

2. The heuristic path planning method for drones based on large language models according to claim 1, characterized in that, The key flight features include distance feature, average speed feature, probability of being monitored by radar, probability of being hit, and fuel consumption.

3. A heuristic path planning method for drones based on large language models according to claim 1, characterized in that, Obtaining the flight objective function includes: Based on the constraint conditions, perform simultaneous optimization on each of the key flight features to generate a flight objective function; the constraint conditions include safety constraint conditions and concealment constraint conditions.

4. The heuristic path planning method for drones based on large language models according to claim 1, characterized in that, The modeling of the UAV includes: Based on the physical properties of the target aircraft, construct the dynamic modeling, kinematic modeling, and aerodynamic modeling of the target aircraft; The components of the dynamic modeling of the target aircraft include the position of the target aircraft in the inertial coordinate system, the velocity components, angular velocity components, and force components in the body coordinate system, and the mass of the target aircraft; The components of the kinematic modeling of the target aircraft include roll angle, pitch angle, yaw angle, the principal axis moment of inertia and product moment of inertia of the target aircraft in the inertia matrix, and the moment components of the target aircraft in the body coordinate system; The components of the aerodynamic modeling of the target aircraft include lift, drag, side force, air density, flight speed, projected area of the wing, angle of attack, and sideslip angle.

5. A heuristic path planning method for drones based on large language models according to claim 1, characterized in that, The flight problem instances include the starting point, ending point, environmental map, and environmental dynamic factors of the flight mission; the environmental dynamic factors are associated with the environmental map.

6. A heuristic path planning method for an unmanned aerial vehicle based on a large language model according to claim 1, characterized in that, The path planning further includes: Initialize the path record, add the target node to the list to be processed, and initialize the best cost and estimated cost of the target node to zero, and set the costs of all other nodes to infinity; Calculate the best cost and estimated cost of each node in the list to be processed, establish the priority for each node to form the planned path and form the path planning; Input the starting point, target node, and environmental information to the large language model to generate a target list; Calculate the priority of each node in the target list, and the priority is used for sorting and path selection in path planning; When the list to be processed is not empty or the priority of the starting point is less than the priority of any node in the list, select the node with the lowest priority in the list to plan the starting point. After calculating the priority of each node in turn, call the large language model to predict new obstacles or other dynamic changes according to the environmental changes and update the priorities and adjacent relationships of the relevant nodes; Recalculate the priorities of the relevant nodes, add them to the list to be processed, and then recalculate. Reconstruct the priority order of each node and update the path.

7. A heuristic path planning method for drones based on large language models according to claim 6, characterized in that, Calculating the priority of each node in the target list includes, for each node, calculating its priority based on the best cost, estimated cost, and heuristic function; where When calculating the priority of each node in turn: If the best cost of the node is greater than its estimated cost, update its best cost to the estimated cost, remove it from the list to be processed, update the estimated costs of all its neighbor nodes, and add the neighbor nodes not in the list to the list to be processed; If the best cost of the node is not greater than its estimated cost, reset its best cost to infinity, update the estimated costs of itself and all its neighbor nodes at the same time, and add the neighbor nodes not in the list to the list to be processed.

8. An electronic device, characterized in that, including: A memory for storing computer programs; A processor for implementing the steps of the large language model-based heuristic path planning method for drones according to any one of claims 1 to 7 when executing the computer program.

9. A computer-readable storage medium, characterized in that, A computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the steps of the large language model-based heuristic path planning method for drones according to any one of claims 1 to 7 are implemented.

Images