Drone route planning method, device and equipment applied to target accelerator card

Abstract

Embodiments of the present application provide a method, device and equipment for unmanned aerial vehicle route planning applied to a target accelerator card. The method comprises: obtaining first position information of the unmanned aerial vehicle, second position information of a target point, third position information of an obstacle and a route planning target; determining state information of the unmanned aerial vehicle according to the first position information, the second position information and the third position information; constructing a potential field model according to the state information, the first position information, the second position information and the third position information, the potential field model comprising an attractive field and a repulsive field, the attractive field being used to guide the unmanned aerial vehicle to converge to the target point in accordance with a potential energy gradient descent direction, and the repulsive field being used to generate a repulsion effect to make the unmanned aerial vehicle avoid the obstacle; generating a target potential field reward according to the potential field model; and generating a control strategy of the unmanned aerial vehicle according to the state information, the potential field reward and the route planning target. The problem of poor real-time performance of the unmanned aerial vehicle route planning is solved.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) route planning technology, and in particular to a UAV route planning method, apparatus and equipment applied to a target acceleration card. Background Technology

[0002] Unmanned aerial vehicle (UAV) route planning is a core component of autonomous flight control systems. Its main objective is to generate a safe, smooth, and efficient feasible path from the starting point to the target point for the UAV in complex environments.

[0003] In related technologies, methods such as heuristic search, grid mapping, or artificial potential fields can be used. However, while these methods can obtain relatively high-quality paths in static scenarios, in dynamic environments, due to the continuous changes in the motion state of obstacles, the algorithm needs to frequently recalculate the cost function and update the path, causing the computational load to increase sharply, resulting in poor real-time performance of UAV path planning. Summary of the Invention

[0004] This application provides a method, apparatus, and device for UAV route planning applied to a target acceleration card, in order to solve the problem of poor real-time performance of UAV path planning.

[0005] In a first aspect, embodiments of this application provide a method for unmanned aerial vehicle (UAV) route planning applied to a target acceleration card, including:

[0006] Acquire the first location information of the drone, the second location information of the target point, the third location information of the obstacles, and the route planning target;

[0007] The status information of the UAV is determined based on the first location information, the second location information, and the third location information;

[0008] Based on the state information, first position information, second position information, and third position information, a potential field model is constructed. The potential field model includes a gravitational field and a repulsive field. The gravitational field is used to guide the UAV to converge toward the target point according to the direction of potential energy gradient descent, and the repulsive field is used to generate a repulsive effect so that the UAV can avoid obstacles.

[0009] Based on the potential field model, a target potential field reward is generated, which is the reward coefficient obtained from the potential energy value of the potential field model.

[0010] Based on the state information, target potential field reward, and route planning objectives, a control strategy for the UAV is generated.

[0011] This embodiment completes the UAV path planning calculation through target acceleration card hardware. By integrating multi-dimensional position information of UAV, target point, and obstacles, an artificial potential field model containing gravitational and repulsive fields is constructed. The potential energy of the model is converted into a dedicated potential field reward and combined with the flight status and planned target to output control strategies. It can stably guide the UAV to approach the target by relying on gravity and reliably avoid various obstacles by relying on repulsion. It also relies on the target acceleration card to improve the overall calculation speed and real-time performance of path planning, effectively optimize UAV flight control decisions, and improve the real-time performance of UAV path planning.

[0012] In one possible implementation, a potential field model is constructed based on state information, first position information, second position information, and third position information, including:

[0013] Construct a gravitational field based on the first and second position information;

[0014] Obtain the preset safety threshold between the drone and the obstacle;

[0015] A repulsive field is constructed based on the first location information, the third location information, and the preset safety threshold.

[0016] Determine the first weight corresponding to the gravitational field and the second weight corresponding to the repulsive field;

[0017] A potential field model is constructed based on the gravitational field, repulsive field, first weight, and second weight.

[0018] In this embodiment, the potential field model continuously points towards the target point through a gravitational field and exerts a repulsive effect on the neighborhood of obstacles through a repulsive field. The weighting is adjusted to ensure that target convergence and obstacle avoidance are uniformly expressed within the same model. This allows the UAV to output control data with directional consistency and safety constraints in complex environments, and reduces trajectory fluctuations when target approach and obstacle approach occur simultaneously.

[0019] In one possible implementation, a potential field model is constructed based on the gravitational field, the repulsive field, the first weight, and the second weight, including:

[0020] Acquire historical status information of the drone and historical location information of obstacles within the sliding window;

[0021] Based on historical status information and historical location information, predict the fourth location information of the drone in the first moment of the future;

[0022] When the difference between the first position indicated by the second position information and the fourth position information is greater than the first threshold, and the difference between the second position indicated by the third position information and the fourth position information is greater than the second threshold, the first weight and the second weight are updated according to the first strategy.

[0023] When the difference between the third position information and the fourth position information is less than the third threshold, the first weight and the second weight are updated according to the second strategy.

[0024] A potential field model is constructed based on the gravitational field, repulsive field, updated first weight, and updated second weight.

[0025] This embodiment achieves weight updates by jointly predicting historical state and location information, allowing the weights to be adjusted based on future location trends rather than solely on the current instantaneous position. By employing different weighting strategies under varying positional differences, the potential field model can adaptively adjust to the approach of the target and the proximity of obstacles. Consequently, the guidance direction and obstacle avoidance intensity of the potential field model remain matched, thereby improving the stability and real-time performance of route planning in dynamic environments.

[0026] In one possible implementation, the target potential field reward is generated based on the potential field model, including:

[0027] Obtain the first potential field value of the UAV in the potential field model, and the historical potential field value of the UAV in the potential field model at the previous historical moment;

[0028] Generate the target potential field reward based on the first potential field value and the historical potential field values.

[0029] This embodiment introduces a comparison of potential field values ​​at adjacent time points, mapping the continuous potential energy changes in the potential field model to a reward output. This ensures that the reward remains consistent with the dynamic relationship between the UAV's current position, the target point's position, and the obstacle's position, thereby enabling the control strategy to form continuous feedback between target convergence and obstacle avoidance. By adopting this target potential field reward, the reward signal does not rely on a single static potential value but can describe the trend changes in the flight state, thus improving the stability of control strategy updates and the consistency of response to environmental changes.

[0030] In one possible implementation, a target potential field reward is generated based on a first potential field value and historical potential field values, including:

[0031] Obtain the mission completion status of the drone. The mission completion status is used to indicate whether the drone has collided with an obstacle, whether the drone's mission duration has exceeded the preset duration, and / or whether the drone has successfully reached the target point.

[0032] The first potential field reward is generated based on the first potential field value and the historical potential field values.

[0033] The second potential field reward is determined based on the drone's mission completion status;

[0034] The target potential field reward is generated based on the first potential field reward and the second potential field reward.

[0035] In this embodiment, the potential field change is incorporated into the reward generation process along with the results of the task, such as collision and timeout. The target potential field reward can simultaneously represent the local approach trend and the global task outcome, enabling the control strategy to maintain stable convergence when approaching the target and to obtain the final feedback immediately when a collision or timeout occurs. This makes the reward signal for UAV route planning more continuous and has clear constraints.

[0036] In one possible implementation, the control strategy for the UAV is generated based on state information, target potential field reward, and route planning objective, and further includes:

[0037] The first probability of the control strategy, the second probability of the historical control strategy before the update, and the dominance parameter of the control strategy are determined. The first probability is used to indicate the probability that the UAV will choose the control strategy in the same state, the second probability is used to indicate the probability that the UAV will choose the historical control strategy in the same state, and the dominance parameter is used to indicate the reward of the control strategy compared with the other executable strategies of the UAV.

[0038] Based on the first probability, the second probability, and the dominance parameter, the pruning strategy loss of the control strategy is determined. The pruning strategy loss is used to limit the range of change between the control strategy and the historical control strategy.

[0039] Determine the mean square error between the predicted state value and the target return of the value network for the control strategy, as well as the entropy regularization term;

[0040] The target optimization strategy is determined based on the pruning strategy loss, mean square error, and entropy regularization term.

[0041] The control strategy is optimized based on the target optimization strategy.

[0042] This embodiment ensures that policy updates remain consistent with historical policies and incorporates potential field rewards into the reward learning process. By constraining the magnitude of policy changes and introducing error and entropy terms, the control policy can achieve stable optimization while maintaining its exploratory capabilities, thereby improving control continuity and path tracking accuracy in complex spatial domains.

[0043] In one possible implementation, prior to obtaining the drone's route planning target, the method further includes:

[0044] The data includes the drone's flight distance, mission duration, threat coefficient, and the third weight of the drone's flight distance, the fourth weight of the mission duration, and the fifth weight of the threat coefficient, which indicates the distance between the drone and obstacles.

[0045] Based on the UAV's flight distance, mission duration, threat level, third weight, fourth weight, and fifth weight, a route planning objective is determined. This objective is used to control the weighted sum and minimum of flight distance, mission duration, and threat level.

[0046] The objective construction method in this embodiment introduces joint constraints from three types of indicators: distance, time, and threat. This allows the planning results to move beyond relying solely on a single shortest path or obstacle avoidance indicator, achieving comprehensive optimization through weight adjustment. The resulting route planning objective provides a unified optimization criterion for subsequent control strategies, enabling the UAV to maintain mission continuity in complex obstacle environments and keeping flight risks within acceptable limits.

[0047] In one possible implementation, the route planning objective is determined based on the UAV's flight distance, mission duration, threat level, third weight, fourth weight, and fifth weight, and further includes:

[0048] Determine the first position range and power overload conditions corresponding to the drone. The first position range represents the position range in which the drone is not allowed to fly. The power overload conditions include the drone's maximum lateral offset angle and maximum longitudinal offset angle.

[0049] The first position information of the control drone is not within the first position range, and the turning radius of the drone is less than the power overload condition.

[0050] This embodiment incorporates no-fly zones, dynamic boundaries, and maximum mission runtime into the route planning objective, ensuring that the target calculation results simultaneously satisfy spatial, dynamic, and temporal constraints. Furthermore, it synchronously verifies the first position information and turning radius during the control output phase, thereby ensuring the executability and stability of the generated path.

[0051] Secondly, embodiments of this application provide a drone route planning device applied to a target acceleration card, comprising:

[0052] The acquisition module is used to acquire the first position information of the UAV, the second position information of the target point, the third position information of the obstacles, and the route planning target;

[0053] The first determining module is used to determine the state information of the UAV based on the first position information, the second position information, and the third position information;

[0054] The construction module is used to construct a potential field model based on the state information, first position information, second position information and third position information. The potential field model includes a gravitational field and a repulsive field. The gravitational field is used to guide the UAV to converge toward the target point according to the direction of potential energy gradient descent, and the repulsive field is used to generate a repulsive effect so that the UAV can avoid obstacles.

[0055] The first generation module is used to generate a target potential field reward based on the potential field model. The target potential field reward is a reward coefficient obtained from the potential energy value of the potential field model.

[0056] The second generation module is used to generate the control strategy for the UAV based on the state information, potential field reward, and route planning objectives.

[0057] In one possible implementation, the building module is specifically used for,

[0058] Construct a gravitational field based on the first and second position information;

[0059] Obtain the preset safety threshold between the drone and the obstacle;

[0060] A repulsive field is constructed based on the first location information, the third location information, and the preset safety threshold.

[0061] Determine the first weight corresponding to the gravitational field and the second weight corresponding to the repulsive field;

[0062] A potential field model is constructed based on the gravitational field, repulsive field, first weight, and second weight.

[0063] In one possible implementation, the building module is specifically used for,

[0064] Acquire historical status information of the drone and historical location information of obstacles within the sliding window;

[0065] Based on historical status information and historical location information, predict the fourth location information of the drone in the first moment of the future;

[0066] When the difference between the first position indicated by the second position information and the fourth position information is greater than the first threshold, and the difference between the second position indicated by the third position information and the fourth position information is greater than the second threshold, the first weight and the second weight are updated according to the first strategy.

[0067] When the difference between the third position information and the fourth position information is less than the third threshold, the first weight and the second weight are updated according to the second strategy.

[0068] A potential field model is constructed based on the gravitational field, repulsive field, updated first weight, and updated second weight.

[0069] In one possible implementation, the first generation module is specifically used for,

[0070] Obtain the first potential field value of the UAV in the potential field model, and the historical potential field value of the UAV in the potential field model at the previous historical moment;

[0071] Generate the target potential field reward based on the first potential field value and the historical potential field values.

[0072] In one possible implementation, the first generation module is specifically used for,

[0073] Obtain the mission completion status of the drone. The mission completion status is used to indicate whether the drone has collided with an obstacle, whether the drone's mission duration has exceeded the preset duration, and / or whether the drone has successfully reached the target point.

[0074] The first potential field reward is generated based on the first potential field value and the historical potential field values.

[0075] The second potential field reward is determined based on the drone's mission completion status;

[0076] The target potential field reward is generated based on the first potential field reward and the second potential field reward.

[0077] In one possible implementation, the second generation module is specifically used for,

[0078] The first probability of the control strategy, the second probability of the historical control strategy before the update, and the dominance parameter of the control strategy are determined. The first probability is used to indicate the probability that the UAV will choose the control strategy in the same state, the second probability is used to indicate the probability that the UAV will choose the historical control strategy in the same state, and the dominance parameter is used to indicate the reward of the control strategy compared with the other executable strategies of the UAV.

[0079] Based on the first probability, the second probability, and the dominance parameter, the pruning strategy loss of the control strategy is determined. The pruning strategy loss is used to limit the range of change between the control strategy and the historical control strategy.

[0080] Determine the mean square error between the predicted state value and the target return of the value network for the control strategy, as well as the entropy regularization term;

[0081] The target optimization strategy is determined based on the pruning strategy loss, mean square error, and entropy regularization term.

[0082] The control strategy is optimized based on the target optimization strategy.

[0083] In one possible implementation, a second determining module is also included.

[0084] The second determining module is used to obtain the drone's flight distance, mission duration, threat coefficient, and the third weight of the drone's flight distance, the fourth weight of the mission duration, and the fifth weight of the threat coefficient, the threat coefficient being used to indicate the distance between the drone and obstacles;

[0085] Based on the UAV's flight distance, mission duration, threat level, third weight, fourth weight, and fifth weight, a route planning objective is determined. This objective is used to control the weighted sum and minimum of flight distance, mission duration, and threat level.

[0086] In one possible implementation, the second determining module is further configured to,

[0087] Determine the first position range and power overload conditions corresponding to the drone. The first position range represents the position range in which the drone is not allowed to fly. The power overload conditions include the drone's maximum lateral offset angle and maximum longitudinal offset angle.

[0088] The first position information of the control drone is not within the first position range, and the turning radius of the drone is less than the power overload condition.

[0089] Thirdly, embodiments of this application provide an electronic device, including: a processor and a memory communicatively connected to the processor;

[0090] The memory stores the instructions that the computer executes;

[0091] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0092] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0093] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0094] The UAV route planning method, apparatus, and device applied to a target acceleration card provided in this application embodiment acquire first position information of the UAV, second position information of the target point, third position information of obstacles, and a route planning target; determine the state information of the UAV based on the first, second, and third position information; construct a potential field model based on the state information, first, second, and third position information, the potential field model including a gravitational field and a repulsive field, the gravitational field being used to guide the UAV to converge toward the target point according to the potential energy gradient descent direction, and the repulsive field being used to generate a repulsive effect to enable the UAV to avoid obstacles; generate a target potential field reward based on the potential field model, which is a reward coefficient obtained from the potential energy value of the potential field model; and generate a control strategy for the UAV based on the state information, the potential field reward, and the route planning target. In this way, the path planning calculation of the UAV can be completed through the target acceleration card hardware. By integrating the multi-dimensional position information of the UAV, the target point, and the obstacles, an artificial potential field model containing gravitational and repulsive fields is constructed. The potential energy of the model is converted into a dedicated potential field reward and combined with the flight status and the planned target to output control strategies. It can not only rely on gravity to stably guide the UAV to approach the target and rely on repulsion to reliably avoid various obstacles, but also rely on the target acceleration card to improve the overall calculation speed and real-time performance of path planning, effectively optimize the UAV flight control decision, and improve the real-time performance of UAV path planning. Attached Figure Description

[0095] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0096] Figure 1 A schematic diagram of the system architecture provided for this application;

[0097] Figure 2 A flowchart illustrating the UAV route planning method applied to a target acceleration card provided in this application;

[0098] Figure 3 A schematic diagram of a potential field model provided in an embodiment of this application;

[0099] Figure 4 A schematic diagram illustrating the process of constructing a potential field model provided in an embodiment of this application;

[0100] Figure 5 A schematic diagram of the process for generating the target potential field reward provided in an embodiment of this application;

[0101] Figure 6 This application provides a complete flowchart of UAV route planning applied to a target acceleration card, as illustrated in the embodiments of this application.

[0102] Figure 7A schematic diagram of the UAV route planning device applied to a target acceleration card provided in this application;

[0103] Figure 8 A schematic diagram of another UAV route planning device applied to a target acceleration card provided in this application;

[0104] Figure 9 A schematic diagram of the structure of the electronic device provided in this application.

[0105] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0106] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0107] First, let me explain the terms used in this application:

[0108] Drones: Drones, short for unmanned aerial vehicles, are aircraft that fly autonomously without the need for onboard personnel, using ground-based remote control equipment or pre-programmed procedures. They carry a power system, flight control module, and payloads such as cameras and sensors. Their size ranges from small aerial photography drones to large industrial and military models. They are widely used in civilian fields such as aerial surveying, agricultural and forestry plant protection, power line inspection, logistics distribution, and emergency rescue. They also play a role in reconnaissance and patrol scenarios, relying on radio and satellite positioning to achieve stable flight and operations.

[0109] Drone route planning: Drone route planning refers to the process by which operators or flight control systems automatically or manually plan an optimal flight path from the takeoff point to the target point based on operational needs, airspace restrictions, obstacle distribution, and endurance constraints. The system will avoid obstacles such as buildings, trees, and signal towers in real time, and rationally plan turning, hovering, and return-to-home nodes to ensure safe and stable flight of the drone throughout the entire process. It can be adapted to various operational scenarios such as surveying and inspection, plant protection spraying, and logistics delivery. It can also dynamically adjust the flight path based on the real-time environment to improve operational efficiency and reduce the risk of collisions and loss of contact.

[0110] Drone route planning is mainly used in scenarios such as urban low-altitude inspection, disaster relief, and power line inspection. In these scenarios, drones typically need to combine their own position information, target point position information, and obstacle position information to continuously generate flight control strategies in order to complete tasks such as reaching the target, avoiding obstacles, and maintaining a feasible flight path in complex airspace.

[0111] Existing UAV route planning schemes typically employ search methods, sampling methods, numerical optimization methods, or reinforcement learning methods to establish spatial relationships between the UAV and target points and obstacles based on environmental perception results, and output flight paths or control commands accordingly.

[0112] Specifically, search-based planning methods can perform global searches using a discretized environment grid, making them suitable for static scenarios. However, they have high computational complexity and are difficult to meet the real-time requirements of dynamic obstacle scenarios.

[0113] Sampling-based planning methods can generate paths by randomly sampling in a continuous space. Although they have high-dimensional space scalability, the path smoothness is poor and replanning is frequent, resulting in insufficient stability in dynamic environments.

[0114] Numerical optimization-based planning methods, such as model predictive control, can dynamically adjust trajectories, but they are highly dependent on the accuracy of environmental modeling and have high computational costs.

[0115] Reinforcement learning-based planning methods can learn optimal policies through interaction with the environment and have adaptive capabilities. However, the training process is slow to converge, the policy fluctuates significantly, and it is prone to local optima in high-dimensional action spaces.

[0116] The above scheme can complete the basic planning task when the environment changes slowly or the obstacle distribution is relatively stable. However, in scenarios where dynamic obstacles change frequently, it often requires repeated replanning or relies on a large number of iterative calculations to update the path.

[0117] However, when factors such as moving vehicles, temporary obstacles, and sudden interference coexist, existing solutions are prone to problems such as increased planning delays, local decision instability, and trajectory oscillations. Specifically, solutions based on traditional rules or optimization are insufficiently responsive to sudden environmental changes, making it difficult to balance target convergence and real-time obstacle avoidance; solutions relying solely on reinforcement learning are prone to unstable reward feedback due to non-stationary environmental states, thus affecting the quality of policy convergence. These shortcomings not only reduce the efficiency of UAVs in completing tasks in complex environments but also increase the risk of collisions when target approach and obstacle approach coexist, and further lead to insufficient computing power utilization after migrating to a target acceleration card.

[0118] To address the aforementioned technical issues, this application provides a UAV route planning method applied to a target acceleration card. This method utilizes the target acceleration card hardware to perform UAV path planning calculations. By fusing multi-dimensional positional information from the UAV, target point, and obstacles, it constructs an artificial potential field model containing both gravitational and repulsive fields. The model's potential energy is converted into a dedicated potential field reward, which, combined with flight status and the planned target, outputs a control strategy. This approach not only stably guides the UAV towards the target using gravity but also reliably avoids various obstacles using repulsive forces. Furthermore, it leverages the target acceleration card to improve the overall computational speed and real-time performance of path planning, effectively optimizing UAV flight control decisions and enhancing the real-time performance of UAV path planning.

[0119] Figure 1 A schematic diagram of the system architecture provided for this application, such as Figure 1 As shown, the system architecture includes a drone. The drone's flight objective is to fly from the starting point to the target point, without touching any obstacles during the flight, and to minimize the flight time or flight path.

[0120] The drone can be equipped with a target acceleration card, on which an algorithm model for drone path planning can be loaded and run.

[0121] The target accelerator card can be a computing hardware board deployed inside a drone for drone path planning. The target accelerator card has a different software stack than existing accelerator cards, and it is different from the main control chip on the drone. The target accelerator card has stronger parallel tensor computing capabilities and can be dedicated to carrying path planning algorithm models that integrate potential field models and deep reinforcement learning. It can independently complete intensive computing tasks such as obstacle and target position measurement information processing, real-time potential field reward solving, and control strategy reasoning, thereby reducing the computing pressure on the main control and improving the real-time computing performance of trajectory planning and online obstacle avoidance in dynamic environments.

[0122] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0123] Figure 2This is a flowchart illustrating the UAV route planning method applied to a target accelerator card provided in this application. This method can be executed by a UAV route planning device applied to the target accelerator card. This device can be implemented via a computer program; it can also be implemented via a medium storing the relevant computer program, such as a USB flash drive and / or optical disc; or it can be implemented via a physical device integrating or installing the relevant computer program, such as an electronic device. The following explanation uses a target accelerator card as an example. Figure 2 As shown, the method includes the following steps S201-S205.

[0124] S201. Obtain the first position information of the UAV, the second position information of the target point, the third position information of the obstacles, and the route planning target.

[0125] The first location information can be used to represent the current location of the drone, serving as the basic input for subsequent state construction and path planning.

[0126] In one possible implementation, the drone can be equipped with onboard GPS and gyroscopes, which can be used to obtain the drone's initial location information.

[0127] The first location information may include the real-time spatial position, attitude, and velocity information of the UAV in the global or local coordinate system, such as position coordinates, flight heading angle, flight angular velocity, flight speed, and other airborne motion state data.

[0128] The second position information can be used to represent the location of the target point. The target point can be the spatial coordinates of the target point that the UAV needs to reach, track, or hover in this flight mission, and can be a static fixed target or a dynamic moving target. The second position information can be used to establish the target pointing direction for the UAV to converge towards the target point.

[0129] Third-party location information can be used to represent the location of obstacles, which is used to establish environmental constraints and obstacle avoidance basis.

[0130] In one possible implementation, the third location information can be preset location information, that is, the location of the obstacle is fixed information, such as fixed static obstacles (such as buildings, trees, mountains).

[0131] In another possible implementation, the drone can be equipped with detection devices such as cameras and lidar, which can be used to detect the real-time spatial position, boundary range, and distance information of static and dynamic obstacles (such as other aircraft and mobile devices).

[0132] Route planning objectives can be target conditions used to constrain and optimize the flight path of UAVs.

[0133] In one possible implementation, the route planning objective can be determined as follows: obtain the UAV's flight distance, mission duration, threat coefficient, and the third weight of the UAV's flight distance, the fourth weight of the mission duration, and the fifth weight of the threat coefficient; determine the route planning objective based on the UAV's flight distance, mission duration, threat coefficient, third weight, fourth weight, and fifth weight, and the route planning objective is used to control the minimum weighted sum of flight distance, mission duration, and threat coefficient.

[0134] The threat coefficient can be used to indicate the distance between the drone and obstacles. The real-time distance between the drone and obstacles can be quantified as the threat coefficient, which is inversely proportional to the distance; the closer the distance, the higher the threat coefficient.

[0135] The strength of the repulsive field is dynamically adjusted according to the threat level to ensure a stronger repulsive effect when an obstacle approaches, thereby improving obstacle avoidance safety.

[0136] Understandably, the threat coefficient quantifies the distance between the drone and obstacles, while the weighted minimum objective balances path planning between safety and efficiency by adjusting the weights of flight distance, mission duration, and threat coefficient. For example, in high-threat scenarios, priority is given to reducing the threat value, while in low-threat scenarios, the focus is on shortening the flight distance, thereby achieving dynamic optimization of multiple objectives.

[0137] In weighted and minimum objectives, threat coefficients and dynamic constraints (such as turning radius) need to be treated in a hierarchical manner: threat coefficients have higher priority than dynamic constraints to ensure safe obstacle avoidance; dynamic constraints, as secondary constraints, ensure path feasibility by limiting lateral / longitudinal offset angles. For example, when the threat coefficient is too high, the path is adjusted first to reduce the threat value, and then the turning radius is optimized while still meeting the threat coefficient requirements.

[0138] In one possible implementation, the route planning objective can be expressed as:

[0139]

[0140] in, This represents the optimal route strategy, and T represents the mission runtime of the UAV. This indicates the position of the drone at time t. Indicates the flight distance of the drone. This represents the threat factor of the drone, which is the shortest distance between the drone's location and the obstacle at time t. This indicates the third weight, namely, the importance of the drone's flight distance in the route planning objective. This indicates the fourth weight, namely the importance of the drone's mission runtime in the route planning objective. This indicates the fifth weight, namely the importance of the threat level of drones in the route planning objectives.

[0141] Understandably, the goal of the above route planning is to find an optimal route strategy. This ensures that the drone can not only safely perform small maneuvers to avoid dynamic obstacles, but also that the drone complies with aircraft dynamics constraints.

[0142] In one possible implementation, the route planning objective also includes constraints, including determining the first position range corresponding to the UAV, power overload conditions, controlling the first position information of the UAV to not be within the first position range, and the UAV's turning radius being smaller than the power overload conditions.

[0143] The first position range represents the area where the drone is not allowed to fly. The dynamic overload condition includes the drone's maximum lateral and longitudinal offset angles. By limiting the drone's lateral and longitudinal offset angles, the dynamic overload condition ensures that the turning radius conforms to the drone's dynamic constraints. For example, the maximum offset angle limit prevents the drone from losing control due to sharp turns, thereby ensuring that the generated path is physically feasible and avoiding flight risks caused by sudden trajectory changes.

[0144] The constraints can be expressed as:

[0145]

[0146] in, Indicates the range of the first position. Indicates the position of the UAV at time t. Perform the second derivative, This indicates the power overload conditions of the drone, which include the offset angle in the x-direction. and the offset angle in the y direction That is, the dynamic overload condition can be expressed as . Indicates the third weight. Indicates the fourth weight. This indicates the fifth weight.

[0147] In one possible implementation, the constraint may also include that the drone's mission runtime is less than or equal to a preset duration, which can be used to indicate the maximum mission runtime limit of the drone.

[0148] In practical implementation, the UAV can continuously acquire its own position data (first position information) during flight. The second position information of the target point and the third position information of obstacles can be provided by the onboard perception system or an external mission system. The first, second, and third position information, along with the route planning target, can be timestamped and their coordinate systems unified. Data from different data interfaces can be mapped to the same navigation reference frame, forming a standard input data packet for the current planning cycle. This input data packet is then written to the input buffer of the target acceleration card to trigger subsequent state information determination steps. Based on this processing method, the spatial information between the UAV, the target point, and obstacles can have a consistent data benchmark within the same calculation time, providing directly callable input for subsequent state construction, potential field calculation, and strategy reasoning.

[0149] In one possible implementation, the dynamic differential model of the UAV in continuous time can be represented by the following formula:

[0150]

[0151] in, and This indicates the position information of the drone on the flight plane, i.e. This represents the coordinates of the UAV in the x-direction at time t. This represents the coordinates of the UAV in the y-direction at time t. Indicates the speed of the drone. Indicates the drone's heading. This indicates the angular velocity of the drone. This indicates the drone's acceleration. Understandable. This represents the velocity of the drone in the x-direction at time t. This represents the velocity of the drone in the y-direction at time t. This represents the acceleration of the drone at time t. Let t represent the angular velocity of the UAV at time t.

[0152] S202. Determine the status information of the UAV based on the first location information, the second location information, and the third location information.

[0153] The state information of a drone can be a comprehensive flight state set that integrates the drone's own position, the relative position of the target point, and the relative position of the obstacle. It can fully reflect the current flight environment and the aircraft's operating condition and can be used as the core input dimension of reinforcement learning algorithms.

[0154] Understandably, the UAV route planning problem can be abstracted into a Restricted Markov Decision Process (CMDP) model. In this model, the UAV's flight state (such as position, velocity, heading angle, etc.) and its spatial relationship with the target point and obstacles together constitute the state space; the UAV's executable velocity commands and yaw rate constitute the action space; the environmental state evolves with the actions, forming state transitions. By designing a reward function, the UAV can be guided to gradually approach the target in a dynamic environment while maintaining the smoothness of its flight trajectory; by setting constraint functions, collisions between the UAV and obstacles can be judged as violations, and the success or failure of the mission can be clearly distinguished, thereby ensuring that path planning achieves a balance between safety and mission completion rate.

[0155] That is, the state information of the UAV can be composed of the UAV's flight state and its spatial relationship with the target point and obstacles, which together constitute the state space.

[0156] In one possible implementation, the drone's state information It can be represented as:

[0157]

[0158] in, Let t represent the coordinates of the UAV in the x-direction at time t. Let t represent the coordinates of the UAV in the y-direction at time t. This represents the flight speed of the drone at time t. This represents the turning angle of the drone at time t. This indicates the drone's measurement information of obstacles. This indicates the measurement information of the target point by the drone.

[0159] Understandable. It can be determined based on the first location information. It can be determined based on the first and third location information. It can be determined based on the first location information and the second location information.

[0160] In practice, after obtaining the first, second, and third position information processed by a unified coordinate system, the relative positional relationship between the UAV and the target point, as well as the relative positional relationship between the UAV and each obstacle, can be calculated first. When the third position information contains multiple obstacles, obstacle information participating in state construction can be selected according to a predetermined method, or the positional relationships of multiple obstacles can be combined and represented to control the state dimension and adapt to the input requirements of the target acceleration card.

[0161] During the state information determination process, to ensure consistent input scale, each state component can be normalized. If invalid obstacle entries exist, fixed placeholder values ​​are filled and valid marker bits are added, ensuring that the inference model on the target accelerator card maintains a constant tensor shape during batch processing. After the above processing, the state information can form a state vector suitable for subsequent calculations. Based on the above analysis, this step, by converting the raw position data into a compact and computable state representation, enables the subsequent potential field model to be directly calculated based on spatial relationships, and makes the control strategy generation process independent of repeated parsing of the original sensor data, thereby shortening the single-cycle planning link.

[0162] S203. Construct a potential field model based on the state information, first position information, second position information, and third position information.

[0163] The potential field model can be a mathematical model used to describe the potential energy relationship between the drone, the target point and the obstacle. It can unify the target convergence requirement and the obstacle avoidance requirement into a calculable potential energy relationship.

[0164] The potential field model can include a gravitational field and a repulsive field. The gravitational field can be used to provide an attractive force for the UAV to move toward the target point. The gravitational field can guide the UAV to converge toward the target point in the direction of the potential energy gradient descent. The repulsive field can be used to exert a repulsive force on obstacles, that is, it can generate a repulsive force to enable the UAV to avoid obstacles.

[0165] Understandably, the gravitational field can be constructed using the target point as the point of minimum potential energy, forming a virtual gravitational force along the negative gradient of potential energy, driving the drone to continuously converge toward the target. The farther away from the target, the higher the potential energy; the closer to the target, the lower the potential energy, guiding the drone to complete its mission and approach the target.

[0166] A repulsive field can be constructed using an obstacle as the area of ​​maximum potential energy. The closer to the obstacle, the higher the potential energy becomes, generating a virtual repulsive force that deviates from the obstacle, forcing the drone to avoid dangerous areas. Once the drone is beyond a safe distance, the repulsive force approaches zero and does not affect normal flight.

[0167] In one possible implementation, Figure 3 For a schematic diagram of a potential field model provided in an embodiment of this application, please refer to [link / reference]. Figure 3 The sinking side can be considered as a gravitational field, and the lowest sinking point can be considered as the point of minimum potential energy of the target point; the convex side can be considered as a repulsive field, and the highest convex point can be considered as the point of maximum potential energy of the obstacle.

[0168] In practice, the current cycle state vector and corresponding position data can be read from the target acceleration card. A gravitational field is constructed based on the relative positional relationship between the UAV and the target point, and a repulsive field is constructed based on the relative positional relationship between the UAV and obstacles. These gravitational and repulsive fields are then combined to form a potential field model. For cases with multiple obstacles, the repulsive field can be constructed by integrating the effects of multiple obstacles on the UAV. After the potential field model is constructed, the potential energy information corresponding to the model can be written into an intermediate cache as direct input for generating the target potential field reward.

[0169] In this way, by uniformly constructing gravitational and repulsive fields, the two types of control requirements of UAV convergence to the target point and obstacle avoidance are transformed into relationships in the same potential energy space. This allows subsequent strategy generation to not only utilize state information but also directly use the potential energy distribution to express the priority direction of the current action, thereby enhancing the decision-making stability in dynamic environments.

[0170] S204. Generate the target potential field reward based on the potential field model.

[0171] The target potential field reward is the reward coefficient obtained based on the potential energy value of the potential field model. It is a standardized reward value calculated by combining gravitational potential energy and repulsive potential energy, and is the core evaluation index of reinforcement learning algorithms.

[0172] Understandably, the potential field model reflects the attraction of a target and the repulsion of an obstacle in a spatial geometric sense, while the process of generating control strategies requires quantifiable reward indicators, thus necessitating the conversion of potential energy states into reward coefficients.

[0173] In one possible implementation, the target potential field reward can refer to the first potential field reward obtained based on the potential energy value of the potential field model. The target potential field reward can also include the sparse reward (i.e., the second potential field reward) shaped by whether the UAV's current flight mission is completed. In other words, the target potential field reward can be obtained by combining the first potential field reward and the second potential field reward.

[0174] In this way, by converting the potential energy value in the potential field model into a directional reward coefficient, the target convergence trend and obstacle avoidance trend are introduced into the control strategy generation process in numerical form, so that the control strategy can respond to the current geometric position relationship and its corresponding potential field distribution.

[0175] S205. Generate the control strategy for the UAV based on the status information, target potential field reward, and route planning target.

[0176] The control strategy can be used to control the UAV to perform route planning and obstacle avoidance actions. It determines the UAV's behavioral output in the current control cycle based on state information, target potential field reward, and route planning objectives.

[0177] By constructing a potential field model that includes both gravitational and repulsive fields, the approach requirements of the UAV to the target point and the obstacle avoidance requirements are uniformly mapped to potential energy gradients, enabling the control strategy to dynamically adjust the flight path based on changes in potential field energy. Potential field rewards, through comparison of potential field values ​​at adjacent time points, transform the potential energy trend into a continuous feedback signal, allowing the control strategy to form a closed-loop response between target convergence and obstacle avoidance, thereby improving real-time performance and path smoothness in dynamic environments.

[0178] In one possible implementation, the generated control strategy may include the drone's acceleration and angular velocity.

[0179] Status information can provide the current flight status and relative spatial relationship, target potential field reward provides the benefit evaluation of the current status relative to the potential energy target, and route planning target provides the constraints and target conditions at the mission level. The three are input together to form an executable control decision.

[0180] In one possible implementation, a control policy generation model, which can be a Restricted Markov Decision Process (CMDP) model, can be deployed on the target acceleration card. In the current cycle, state information, target potential field reward, and route planning objective are assembled into a unified input, which is then calculated by the target acceleration card to output the control policy. The control policy can be represented as one or more combinations of the UAV's speed control, heading control, attitude control, and altitude control variables.

[0181] In this way, by jointly processing state information, target potential field reward, and route planning objective, a control output oriented to the current moment is formed, enabling the UAV to maintain obstacle avoidance response while converging towards the target point. Furthermore, since the strategy generation-related calculations are deployed on the target acceleration card, the computation link within a single cycle is shorter, which can meet the real-time planning requirements in dynamic and complex environments.

[0182] In one possible implementation, after generating the control strategy for the UAV, the process further includes: determining a first probability of the control strategy, a second probability of the historical control strategy before the update, and the dominance parameter of the control strategy; determining a pruning strategy loss of the control strategy based on the first probability, the second probability, and the dominance parameter, the pruning strategy loss being used to limit the variation range between the control strategy and the historical control strategy; determining the mean square error between the target reward and the actual reward of the control strategy, and the entropy regularization term; determining a target optimization strategy based on the pruning strategy loss, the mean square error, and the entropy regularization term; and optimizing the control strategy based on the target optimization strategy.

[0183] The first probability can be used to indicate the probability that the UAV will choose a control strategy in the same state, the second probability can be used to indicate the probability that the UAV will choose a historical control strategy in the same state, and the dominance parameter can be used to indicate the reward of the control strategy compared to the other executable strategies of the UAV.

[0184] In one possible implementation, the first probability can be expressed as The second probability can be expressed as ,in, Represents the policy function. This represents the weight parameters of the new policy network to be updated. This represents the weight parameters of the old policy network (i.e., the weight parameters of the historical control policy before the update). This represents the specific actions of the drone at time t, i.e., the action space output. This represents the state parameters of the UAV at time t, i.e., the state space input. The weight parameters of the new policy network to be updated and the weight parameters of the old policy network can be preset values ​​by the user.

[0185] In one possible implementation, the pruning strategy loss It can be represented as:

[0186]

[0187] in, This represents the empirical mean of a finite batch of samples. This represents the ratio of the first probability to the second probability, i.e., , This represents the weight parameters of the new policy network to be updated. Indicates the dominant parameter. This represents the pruning strategy parameters, used to control the maximum update magnitude of the strategy in a single iteration. The weight parameters, dominance parameters, and pruning strategy parameters of the new policy network to be updated can be preset values ​​by the user.

[0188] In one possible implementation, the mean square error between the predicted state value and the target return in the value network of the control policy is... It can be represented as:

[0189]

[0190] in, This represents the empirical mean of a finite batch of samples. This represents the value network's prediction of the state value. This indicates the target reward. The target reward can be a value preset by the user.

[0191] In one possible implementation, the entropy regularization term It can be represented as:

[0192]

[0193] in, This represents the empirical mean of a finite batch of samples. This represents the first probability.

[0194] In one possible implementation, the objective optimization strategy can be expressed as:

[0195]

[0196] in, The weighting coefficient for the mean squared error. The weighting coefficients for the entropy regularization term, the mean square error, and the entropy regularization term can be preset by the user.

[0197] In this way, during the training process, the control strategy can generate multiple drone flight trajectories, collect status, action, reward and cost information, and then iteratively update the strategy based on the objective optimization until the strategy converges. This ensures the smoothness of the path and the success rate of the task, while limiting the update step size to ensure the stability and convergence speed of the training.

[0198] Understandably, the pruning strategy loss limits the probability ratio between the old and new strategies, preventing excessive policy updates from causing trajectory oscillations; the mean squared error measures the deviation between the policy output and the target reward, ensuring the policy converges towards the optimal solution; and the entropy regularization term encourages action diversity, preventing the policy from getting trapped in local optima. The joint optimization of these three factors enables the control strategy to maintain a balance between exploration and exploitation in complex spatial domains, improving the smoothness and safety of path tracking.

[0199] In one possible implementation, after constructing the potential field model and the target potential field reward, the model can be applied to the target accelerator card. However, the target accelerator card differs from existing accelerator card devices in terms of device drivers, compilation and runtime interfaces, basic operator libraries, collection communication libraries, and memory management mechanisms, thus requiring adaptation for the target accelerator card.

[0200] The target accelerator card can accelerate the construction of potential field models and policy reasoning through parallel tensor computing capabilities. For example, the matrix operation unit of the target accelerator card can be used to batch process potential field gradient calculations, thereby reducing path planning latency and meeting the real-time requirements in dynamic environments.

[0201] The specific adaptation process may include hardware environment adaptation, data computation optimization, operator acceleration, and mixed precision.

[0202] Hardware environment adaptation includes building a dedicated container image for the target accelerator card (such as a deep computing unit), configuring compatibility for the operating system, drivers, and deep learning framework, and ensuring that the operator libraries on which the algorithm depends can be correctly called on the deep computing unit.

[0203] Data computation optimization includes adjusting the batch processing methods for state sampling, trajectory storage, and reward calculation to match the memory management and parallel characteristics of the deep computing unit, thereby avoiding memory access bottlenecks.

[0204] Operator acceleration and mixed precision involve replacing key computational steps (including forward or backward propagation of the policy network and value network, gradient accumulation) with acceleration library operators from the target accelerator card, combined with mixed precision and quantization strategies, to ensure that the training process improves throughput efficiency while maintaining convergence.

[0205] The trained model can be deployed on the target accelerator card to achieve low-latency real-time control of the drone.

[0206] One possible implementation could include quantizing the policy network parameters to reduce storage and computational overhead, while pruning redundant layer structures to ensure that the inference phase meets the real-time requirements under the dynamic constraints of the UAV; utilizing the matrix operation acceleration capabilities provided by the target acceleration card to perform batch calculations on the input data to output action distributions and quickly generate control policies; and based on hardware-optimized forward inference, enabling the UAV to update heading and speed commands in dynamic environments with millisecond-level latency, thereby achieving smooth path planning and real-time obstacle avoidance.

[0207] Another possible implementation could include using 16-bit half-precision floating-point numbers (i.e., FP16 precision quantization) for key computational layers (such as attention mechanisms and convolutional layers) of the policy network and value network to reduce memory usage and improve computational efficiency; and optimizing memory scheduling of domestic accelerator cards by adjusting the batch processing method of state sampling and trajectory storage to reduce memory access latency and improve inference speed.

[0208] In this embodiment, when route planning for a drone is required, the following can be obtained: first position information of the drone, second position information of the target point, third position information of the obstacles, and a route planning target. The first position information can represent the current position of the drone, the second position information can represent the position of the target point, and the third position information can represent the position of the obstacles. The route planning target can be a target condition used to constrain and optimize the drone's flight path. Based on the first, second, and third position information, the drone's state information is determined. The drone's state information can be a comprehensive flight state set that integrates the drone's own position, the relative position of the target point, and the relative position of the obstacles. Based on the state information, the first, second, and third position information, a potential field model is constructed. The potential field model can be a mathematical model used to describe the potential energy relationship between the drone, the target point, and the obstacles. The potential field model can include gravitational and repulsive fields. Based on the potential field model, a target potential field reward is generated. Based on the state information, the target potential field reward, and the route planning target, a control strategy for the drone is generated. The control strategy can be used to control the drone to perform route planning and obstacle avoidance actions. Thus, by constructing state information from the first, second, and third position information using the above method, and further mapping it uniformly into a potential field model including gravitational and repulsive fields, the potential energy value is converted into a target potential field reward and input into the control strategy generation process along with the route planning target. This couples the target point convergence requirement, obstacle avoidance requirement, and task constraints within the same computational link. Consequently, the control strategy reflects both the real-time spatial relationship between the UAV and the target point and obstacles in each planning cycle, as well as the direction of action reward corresponding to the potential field distribution. This allows for the continuous output of executable flight commands in scenarios with dynamic obstacle changes, thereby improving the real-time performance of route planning.

[0209] Below, in conjunction with Figure 4 The process of constructing the potential field model is explained.

[0210] Figure 4 This is a schematic flowchart illustrating the process of constructing a potential field model according to an embodiment of this application. Based on the above embodiments, see [link to relevant documentation]. Figure 4 The method includes:

[0211] S401. Construct a gravitational field based on the first position information and the second position information.

[0212] The gravitational field can be constructed based on the relative position of the UAV and the target point. That is, the position coordinates of the UAV can be determined based on the first position information, and the position coordinates of the target point can be determined based on the second position information, thereby constructing the gravitational field.

[0213] In one possible implementation, the gravitational field can be constructed based on the Euclidean distance between the UAV and the target point. The gravitational field can be represented as a potential function that monotonically increases as the distance between the UAV and the target point decreases, with the gradient of the potential function pointing towards the target point.

[0214] In one possible implementation, the potential value of the gravitational field... It can be represented as:

[0215]

[0216] Where x represents the coordinates of the drone in the x-direction, and y represents the coordinates of the drone in the y-direction. This represents the coordinates of the target point in the x-direction. This represents the coordinates of the target point in the y-direction. This represents the gravitational coefficient, which can be used to adjust the magnitude of gravity. The gravitational coefficient can be a preset value and can be a positive number less than 1.

[0217] S402. Obtain the preset safety threshold between the drone and the obstacle, and construct a repulsive field based on the first position information, the third position information and the preset safety threshold.

[0218] The preset safety threshold can be a pre-set value, which can be set according to the size of the drone, its flight speed, and the perception error.

[0219] A preset safety threshold can be used to define the danger range of obstacles. When the distance between the drone and the obstacle is less than this threshold, the repulsive field strength increases sharply to ensure that the drone stays away from the obstacle; the gravitational field guides the drone to approach the target through the potential energy gradient at the target point. By dynamically adjusting the weights of the gravitational and repulsive fields, the potential field model maintains a balance between target approach and obstacle avoidance, avoiding trajectory oscillations or local extrema problems.

[0220] In one possible implementation, the repulsive field can be calculated based on the relative distance between the first and third position information and a preset safety threshold. When the distance is less than the preset safety threshold, an increasing repulsive potential energy is generated, and when the distance is greater than the preset safety threshold, the repulsive term decays to zero or close to zero.

[0221] In one possible implementation, the repulsive field can be determined based on the Lyapunov function. Based on the stability of the Lyapunov function, it can be guaranteed that the UAV trajectory will not produce oscillations or local minima, and the UAV's flight trajectory will converge to the set hovering area, i.e., outside the modeled obstacle model.

[0222] In one possible implementation, the potential value of the repulsive field is... It can be represented as:

[0223]

[0224] Where r represents the straight-line distance between the drone and the obstacle. This indicates a preset safety threshold.

[0225] The straight-line distance r between the drone and the obstacle can be expressed as:

[0226]

[0227] Where x represents the coordinates of the drone in the x-direction, and y represents the coordinates of the drone in the y-direction. This represents the coordinates of the obstacle in the x-direction. This represents the coordinates of the obstacle in the y-direction.

[0228] S403. Determine the first weight corresponding to the gravitational field and the second weight corresponding to the repulsive field. Based on the gravitational field, the repulsive field, the first weight, and the second weight, construct a potential field model.

[0229] The first weight can be used to indicate the proportion of the gravitational field's contribution to the potential field model, while the second weight can be used to indicate the proportion of the repulsive field's contribution to the potential field model.

[0230] In one possible implementation, the first weight and the second weight can be preset values.

[0231] In one possible implementation, the potential field model can be expressed as:

[0232]

[0233] in, Indicates the first weight. This indicates the second weight.

[0234] In one possible implementation, the potential field model can be constructed as follows: Historical state information of the UAV and historical position information of obstacles within a sliding window are obtained; based on the historical state information and historical position information, the fourth position information of the UAV at a future first moment is predicted; when the difference between the second position information and the fourth position information indicates a first position greater than a first threshold, and the difference between the third position information and the fourth position information indicates a second position greater than a second threshold, the first weight and the second weight are updated according to a first strategy; when the difference between the third position information and the fourth position information indicates a third position less than a third threshold, the first weight and the second weight are updated according to a second strategy; and a potential field model is constructed based on the gravitational field, repulsive field, updated first weight, and updated second weight.

[0235] The first strategy could be to increase the first weight and adaptively decrease the second weight. The second strategy could be to increase the second weight and adaptively decrease the first weight.

[0236] Understandably, when the target point is moving away rapidly and there are no dangerous obstacles in the vicinity, the first weight can be increased and the second weight can be decreased simultaneously, thereby widening the target attraction range and accelerating the convergence of the drone to the target point; when the obstacle is moving rapidly towards the drone, the second weight can be increased and the first weight can be decreased simultaneously, thereby increasing the priority of the repulsive field, reducing the distance between the drone and the obstacle, and avoiding the risk of collision in advance.

[0237] Sliding window prediction forecasts future position trends by analyzing the historical states of the drone and obstacles. When the target point is far away and obstacles are sparse, the gravitational field weights are increased to accelerate convergence; when obstacles are approaching, the repulsive field weights are increased to prioritize obstacle avoidance. This dynamic weight adjustment mechanism enables the potential field model to adapt to environmental changes and improves path stability in complex scenarios.

[0238] In this way, by jointly predicting historical state and historical position information, the weight updates are no longer based solely on the current instantaneous position, but are adjusted in conjunction with future position trends. By employing different weight strategies under different positional differences, the potential field model can adaptively adjust according to the approach of the target and the degree of obstacle approach. As a result, the guidance direction and obstacle avoidance intensity of the potential field model can remain matched, thereby improving the stability and real-time performance of route planning in dynamic environments.

[0239] exist Figure 4 In the illustrated embodiment, when a potential field model needs to be constructed, a gravitational field can be constructed based on the first and second position information. The gravitational field can be constructed based on the relative position of the UAV and the target point. A preset safety threshold between the UAV and the obstacle is obtained. A repulsive field is constructed based on the first and third position information and the preset safety threshold. The repulsive field can be determined based on the Lyapunov function. A first weight corresponding to the gravitational field and a second weight corresponding to the repulsive field are determined. Based on the gravitational field, the repulsive field, the first weight, and the second weight, a potential field model is constructed. The first weight can be used to indicate the contribution ratio of the gravitational field to the potential field model, and the second weight can be used to indicate the contribution ratio of the repulsive field to the potential field model. In this way, through the above method, the potential field model can continuously point towards the target point through the gravitational field and exert a repulsive effect on the neighborhood of the obstacle through the repulsive field. The weights work together to make target convergence and obstacle avoidance uniformly expressed within the same model. This allows the UAV to output control basis with directional consistency and safety constraints in complex environments and reduces trajectory fluctuations when target approach and obstacle approach occur simultaneously.

[0240] Below, in conjunction with Figure 5The process of generating the target potential field reward is explained.

[0241] Figure 5 This is a schematic diagram illustrating the process of generating a target potential field reward as provided in an embodiment of this application. Based on the above embodiments, see also... Figure 5 The method includes:

[0242] S501, Obtain the first potential field value of the UAV in the potential field model, and the historical potential field value of the UAV in the potential field model at the previous historical moment.

[0243] Historical potential values ​​can be the potential values ​​cached from the previous historical moment.

[0244] S502. Generate the target potential field reward based on the first potential field value and the historical potential field values.

[0245] The target potential field reward can be the first potential field reward obtained based on the first potential field value and the historical potential field value, or it can be a combination of the first potential field reward and the second potential field reward based on the task completion status.

[0246] Understandably, by comparing the potential field values ​​at the current moment with those at historical moments, the potential field reward can reflect the trend changes of the UAV in the potential field space. For example, a decrease in potential energy indicates that the target is approaching, while an increase in potential energy indicates that the obstacle avoidance requirement is increasing. This trend reward signal enables the control strategy to continuously adjust its action output based on dynamic environmental changes, rather than relying on a single static potential value, thereby improving the stability and consistency of the strategy response.

[0247] After obtaining the first potential field value at the current moment, the historical potential field value cached from the previous historical moment can be read synchronously, and the difference or ratio between the two can be calculated to form the potential field change. When the change indicates that the UAV's potential energy is converging towards the target, a positive target potential field reward is generated; when the change indicates that the potential energy is moving away from the target or increasing towards the obstacle, a negative or weakened target potential field reward is generated. The magnitude of the reward value can be determined by the direction, magnitude, and rate of change, and is output to the subsequent control strategy generation module as control feedback.

[0248] In one possible implementation, the target potential field reward can be generated as follows: obtain the mission completion status of the UAV; generate a first potential field reward based on the first potential field value and historical potential field values; determine a second potential field reward based on the mission completion status of the UAV; and generate the target potential field reward based on the first and second potential field rewards.

[0249] The first potential field reward can refer to the potential field reward obtained based on the potential energy value of the potential field model, while the second potential field reward can refer to the sparse reward shaped based on whether the UAV's current flight mission is completed.

[0250] In one possible implementation, the first potential field reward It can be represented as:

[0251]

[0252] in, Represents the historical potential field value. This represents the first potential field value. This represents the weighting coefficient of the first power field reward.

[0253] In one possible implementation, the second potential field reward It can be represented as:

[0254]

[0255] in, This indicates whether the task was successful. The feedback value can be a number greater than 0.

[0256] The mission completion status can be used to indicate whether the drone has collided with an obstacle, whether the drone's mission duration has exceeded the preset time, and / or whether the drone has successfully reached the target point. It is understood that when the drone collides with an obstacle, the drone's mission duration exceeds the preset time, and / or the drone fails to reach the target point, the current mission can be considered unsuccessful; when the drone does not collide with an obstacle, the drone's mission duration does not exceed the preset time, and the drone successfully reaches the target point, the current mission can be considered successful.

[0257] In one possible implementation, the target potential field reward r can be expressed as:

[0258]

[0259] Understandably, the second potential field reward provides sparse but explicit endgame feedback through task completion status (such as collision or timeout). For example, a negative feedback value is assigned upon collision, and a negative feedback value is assigned upon timeout. This endgame feedback, together with the first potential field reward, constitutes a complete reward system, enabling the control strategy to maintain a balance between local convergence and the global task objective, and avoiding task failure due to misleading local rewards.

[0260] In this way, by incorporating the potential field change along with the task results such as collision and timeout into the reward generation process, the target potential field reward can simultaneously represent the local approach trend and the global task outcome, enabling the control strategy to maintain stable convergence when approaching the target and to receive immediate final feedback when a collision or timeout occurs. This makes the reward signal for UAV route planning more continuous and has clear constraints.

[0261] exist Figure 5In the illustrated embodiment, when a target potential field reward needs to be generated, the first potential field value of the UAV in the potential field model and the historical potential field value of the UAV in the potential field model at the previous historical moment can be obtained. The historical potential field value can be the potential field value cached at the previous historical moment. Based on the first potential field value and the historical potential field value, the target potential field reward is generated. In this way, in the above method, by introducing the comparison of potential field values ​​at adjacent moments, the continuous potential energy changes in the potential field model are mapped to the reward output, so that the reward is consistent with the dynamic relationship between the current position of the UAV, the position of the target point, and the position of the obstacle, thereby enabling the control strategy to form continuous feedback between target convergence and obstacle avoidance. After adopting this target potential field reward, the reward signal does not depend on a single static potential value, but can describe the trend changes of the flight state, thereby improving the stability of the control strategy update and the consistency of response to environmental changes.

[0262] Below, in conjunction with Figure 6 This paper explains the complete process of the UAV route planning method applied to the target acceleration card.

[0263] Figure 6 This is a complete flowchart illustrating the drone route planning applied to a target acceleration card, as provided in an embodiment of this application. Please refer to... Figure 6 ,include:

[0264] S601: Obtain the first location information of the UAV, the second location information of the target point, the third location information of the obstacles, and the route planning target.

[0265] S602. Determine the status information of the UAV based on the first location information, the second location information, and the third location information.

[0266] S603. Construct a gravitational field based on the first position information and the second position information.

[0267] S604. Obtain the preset safety threshold between the drone and the obstacle, and construct a repulsive field based on the first position information, the third position information and the preset safety threshold.

[0268] S605. Determine the first weight corresponding to the gravitational field and the second weight corresponding to the repulsive field. Based on the gravitational field, the repulsive field, the first weight, and the second weight, construct a potential field model.

[0269] S606, Obtain the first potential field value of the UAV in the potential field model, and the historical potential field value of the UAV in the potential field model at the previous historical moment.

[0270] S607. Generate the target potential field reward based on the first potential field value and the historical potential field values.

[0271] S608 generates a control strategy for the UAV based on the status information, target potential field reward, and route planning target.

[0272] S609, Determine the first probability of the control strategy, the second probability of the historical control strategy before the update, and the dominance parameters of the control strategy.

[0273] S610. Determine the pruning strategy loss of the control strategy based on the first probability, the second probability, and the dominance parameter.

[0274] S611. Determine the mean square error between the predicted state value and the target return of the value network for the control strategy, as well as the entropy regularization term. Based on the pruning strategy loss, mean square error, and entropy regularization term, determine the target optimization strategy.

[0275] S612. Optimize the control strategy according to the target optimization strategy.

[0276] This application also provides a drone route planning device for use with a target acceleration card, wherein the resource recovery device is integrated into an electronic device.

[0277] Figure 7 This is a schematic diagram of the UAV route planning device applied to the target acceleration card provided in this application, as shown in the figure. Figure 7 As shown, the UAV route planning device 70 for target acceleration cards provided in this embodiment includes:

[0278] The acquisition module 71 is used to acquire the first position information of the UAV, the second position information of the target point, the third position information of the obstacles, and the route planning target;

[0279] The first determining module 72 is used to determine the state information of the UAV based on the first position information, the second position information and the third position information;

[0280] Module 73 is used to construct a potential field model based on state information, first position information, second position information and third position information. The potential field model includes a gravitational field and a repulsive field. The gravitational field is used to guide the UAV to converge toward the target point according to the direction of potential energy gradient descent, and the repulsive field is used to generate a repulsive effect so that the UAV can avoid obstacles.

[0281] The first generation module 74 is used to generate a target potential field reward based on the potential field model. The target potential field reward is used to convert the potential energy value of the potential field model into a reward coefficient.

[0282] The second generation module 75 is used to generate the control strategy of the UAV based on the state information, potential field reward and route planning target.

[0283] The UAV route planning device for target acceleration card provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0284] In one possible implementation, the building module 73 is specifically used for,

[0285] Construct a gravitational field based on the first and second position information;

[0286] Obtain the preset safety threshold between the drone and the obstacle;

[0287] A repulsive field is constructed based on the first location information, the third location information, and the preset safety threshold.

[0288] Determine the first weight corresponding to the gravitational field and the second weight corresponding to the repulsive field;

[0289] A potential field model is constructed based on the gravitational field, repulsive field, first weight, and second weight.

[0290] In one possible implementation, the building module 73 is specifically used for,

[0291] Acquire historical status information of the drone and historical location information of obstacles within the sliding window;

[0292] Based on historical status information and historical location information, predict the fourth location information of the drone in the first moment of the future;

[0293] When the difference between the first position indicated by the second position information and the fourth position information is greater than the first threshold, and the difference between the second position indicated by the third position information and the fourth position information is greater than the second threshold, the first weight and the second weight are updated according to the first strategy.

[0294] When the difference between the third position information and the fourth position information is less than the third threshold, the first weight and the second weight are updated according to the second strategy.

[0295] A potential field model is constructed based on the gravitational field, repulsive field, updated first weight, and updated second weight.

[0296] In one possible implementation, the first generation module 74 is specifically used for,

[0297] Obtain the first potential field value of the UAV in the potential field model, and the historical potential field value of the UAV in the potential field model at the previous historical moment;

[0298] Generate the target potential field reward based on the first potential field value and the historical potential field values.

[0299] In one possible implementation, the first generation module 74 is specifically used for,

[0300] Obtain the mission completion status of the drone. The mission completion status is used to indicate whether the drone has collided with an obstacle or whether the drone's mission duration has exceeded the preset time.

[0301] The first potential field reward is generated based on the first potential field value and the historical potential field values.

[0302] The second potential field reward is determined based on the drone's mission completion status;

[0303] The target potential field reward is generated based on the first potential field reward and the second potential field reward.

[0304] In one possible implementation, the second generation module 75 is specifically used for,

[0305] The first probability of the control strategy, the second probability of the historical control strategy at the previous historical moment, and the dominance parameter of the control strategy are determined. The first probability is used to indicate the probability that the UAV will choose the control strategy in the same state, the second probability is used to indicate the probability that the UAV will choose the historical control strategy in the same state, and the dominance parameter is used to indicate the reward of the control strategy compared with the other executable strategies of the UAV.

[0306] Based on the first probability, the second probability, and the dominance parameter, the pruning strategy loss of the control strategy is determined. The pruning strategy loss is used to limit the range of change between the control strategy and the historical control strategy.

[0307] Determine the mean square error between the target return and the actual return of the control strategy, as well as the entropy regularization term;

[0308] The target optimization strategy is determined based on the pruning strategy loss, mean squared error, and entropy regularization term.

[0309] The control strategy is optimized based on the target optimization strategy.

[0310] The UAV route planning device for target acceleration card provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0311] Figure 8 A schematic diagram of another UAV route planning device applied to a target acceleration card provided in this application is shown. Figure 7 On the basis of, such as Figure 8 As shown, the UAV route planning device 70 for target acceleration cards provided in this embodiment further includes: a second determination module 76.

[0312] The second determining module 76 is used to obtain the drone's flight distance, mission duration, threat coefficient, and the third weight of the drone's flight distance, the fourth weight of the mission duration, and the fifth weight of the threat coefficient, wherein the threat coefficient is used to indicate the distance between the drone and obstacles.

[0313] Based on the UAV's flight distance, mission duration, threat level, third weight, fourth weight, and fifth weight, a route planning objective is determined. This objective is used to control the weighted sum and minimum of flight distance, mission duration, and threat level.

[0314] In one possible implementation, the second determining module 76 is further configured to,

[0315] Determine the first position range and power overload conditions corresponding to the drone. The first position range represents the position range in which the drone is not allowed to fly. The power overload conditions include the drone's maximum lateral offset angle and maximum longitudinal offset angle.

[0316] The first position information of the control drone is not within the first position range, and the turning radius of the drone is less than the power overload condition.

[0317] The UAV route planning device for target acceleration card provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0318] Figure 9 A schematic diagram of the structure of the electronic device provided in this application. Figure 9 As shown, the electronic device 90 provided in this embodiment includes: a processor 901 and a memory 902 communicatively connected to the processor 901.

[0319] Optionally, the electronic device 90 also includes a communication component 903. The processor 901, memory 902, and communication component 903 are connected via a bus.

[0320] In the specific implementation process, the processor 901 executes the computer execution instructions stored in the memory 902, causing the processor 901 to perform the above-mentioned method.

[0321] The specific implementation process of processor 901 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0322] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0323] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0324] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0325] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0326] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0327] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as read-only memory (ROM), static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0328] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in application-specific integrated circuits (ASICs). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0329] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0330] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0331] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0332] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a server to execute all or part of the steps of the methods of the various embodiments of the present invention.

[0333] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for UAV route planning applied to a target acceleration card, characterized in that, include: Acquire the first location information of the drone, the second location information of the target point, the third location information of the obstacles, and the route planning target; The status information of the drone is determined based on the first location information, the second location information, and the third location information; Based on the state information, the first position information, the second position information, and the third position information, a potential field model is constructed. The potential field model includes a gravitational field and a repulsive field. The gravitational field is used to guide the UAV to converge toward the target point according to the direction of potential energy gradient descent, and the repulsive field is used to generate a repulsive effect so that the UAV can avoid the obstacle. Based on the potential field model, a target potential field reward is generated, wherein the target potential field reward is a reward coefficient obtained based on the potential energy value of the potential field model; The control strategy for the UAV is generated based on the state information, the target potential field reward, and the route planning target.

2. The method according to claim 1, characterized in that, Based on the state information, the first position information, the second position information, and the third position information, a potential field model is constructed, including: The gravitational field is constructed based on the first location information and the second location information; Obtain the preset safety threshold between the drone and the obstacle; The repulsive field is constructed based on the first location information, the third location information, and the preset safety threshold. Determine the first weight corresponding to the gravitational field and the second weight corresponding to the repulsive field; The potential field model is constructed based on the gravitational field, the repulsive field, the first weight, and the second weight.

3. The method according to claim 2, characterized in that, Based on the gravitational field, the repulsive field, the first weight, and the second weight, the potential field model is constructed, including: Obtain the historical state information of the UAV and the historical position information of the obstacle within the sliding window; Based on the historical state information and the historical location information, predict the fourth location information of the UAV at the first future moment; When the difference between the second location information and the first location information indicated by the fourth location information is greater than the first threshold, and the difference between the third location information and the second location information indicated by the fourth location information is greater than the second threshold, the first weight and the second weight are updated according to the first strategy. When the difference between the third position information and the third position information indicated by the fourth position information is less than the third threshold, the first weight and the second weight are updated according to the second strategy. The potential field model is constructed based on the gravitational field, the repulsive field, the updated first weight, and the updated second weight.

4. The method according to any one of claims 1-3, characterized in that, Based on the potential field model, a target potential field reward is generated, including: Obtain the first potential field value of the UAV in the potential field model, and the historical potential field value of the UAV in the potential field model at the previous historical moment; Based on the first potential field value and the historical potential field value, a target potential field reward is generated.

5. The method according to claim 4, characterized in that, Based on the first potential field value and the historical potential field value, a target potential field reward is generated, including: The task completion status of the drone is obtained, which is used to indicate whether the drone has collided with the obstacle, whether the drone's task duration exceeds a preset duration, and / or whether the drone has successfully reached the target point. A first potential field reward is generated based on the first potential field value and the historical potential field value; The second potential field reward is determined based on the mission completion status of the drone. The target potential field reward is generated based on the first potential field reward and the second potential field reward.

6. The method according to any one of claims 1-3, characterized in that, Based on the state information, the target potential field reward, and the route planning target, a control strategy for the UAV is generated, which further includes: A first probability of the control strategy, a second probability of the historical control strategy before the update, and an advantage parameter of the control strategy are determined. The first probability is used to indicate the probability that the UAV will select the control strategy in the same state. The second probability is used to indicate the probability that the UAV will select the historical control strategy in the same state. The advantage parameter is used to indicate the reward benefit of the control strategy compared with the other executable strategies of the UAV. Based on the first probability, the second probability, and the dominance parameter, the pruning strategy loss of the control strategy is determined, and the pruning strategy loss is used to limit the range of change between the control strategy and the historical control strategy. Determine the mean square error between the predicted state value and the target return of the value network for the control strategy, as well as the entropy regularization term; The target optimization strategy is determined based on the pruning strategy loss, mean square error, and entropy regularization term. The control strategy is optimized according to the target optimization strategy.

7. The method according to any one of claims 1-3, characterized in that, Before obtaining the drone's route planning target, the following steps are also included: The flight distance, mission duration, and threat coefficient of the UAV are obtained, as well as the third weight of the flight distance, the fourth weight of the mission duration, and the fifth weight of the threat coefficient, wherein the threat coefficient is used to indicate the distance between the UAV and the obstacle; Based on the UAV's flight distance, mission duration, threat coefficient, third weight, fourth weight, and fifth weight, a route planning objective is determined. This route planning objective is used to control the weighted sum of the flight distance, mission duration, and threat coefficient to be minimized.

8. The method according to claim 7, characterized in that, Based on the UAV's flight distance, mission duration, threat level, third weight, fourth weight, and fifth weight, the route planning objective is determined, which also includes: Determine the first position range and power overload conditions corresponding to the drone. The first position range represents the position range in which the drone is not allowed to fly. The power overload conditions include the maximum lateral offset angle and the maximum longitudinal offset angle of the drone. The first position information of the control drone is not located within the first position range, and the turning radius of the control drone is less than the power overload condition.

9. A UAV route planning device applied to a target acceleration card, characterized in that, include: The acquisition module is used to acquire the first position information of the UAV, the second position information of the target point, the third position information of the obstacles, and the route planning target; The first determining module is used to determine the status information of the UAV based on the first location information, the second location information, and the third location information; A construction module is used to construct a potential field model based on the state information, the first position information, the second position information, and the third position information. The potential field model includes a gravitational field and a repulsive field. The gravitational field is used to guide the UAV to converge toward the target point according to the direction of potential energy gradient descent, and the repulsive field is used to generate a repulsive effect so that the UAV avoids the obstacle. The first generation module is used to generate a target potential field reward based on the potential field model, wherein the target potential field reward is a reward coefficient obtained based on the potential energy value of the potential field model. The second generation module is used to generate the control strategy of the UAV based on the state information, the target potential field reward, and the route planning target.

10. An electronic device, characterized in that, include: A processor and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-8.

12. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-8.

Images