An autonomous motion planning method for swarm drones based on imitation learning

Abstract

The present invention relates to a method for autonomous motion planning of a group of unmanned aerial vehicles based on imitation learning, the method comprising: constructing an expert system: obtaining a data set for imitation learning, including real-time environmental perception, the state of the unmanned aerial vehicle itself, information of other unmanned aerial vehicles in the group, and the unmanned aerial vehicle's own real-time predicted trajectory; constructing a student system: including a neural network reasoning module for predicting the trajectory of the unmanned aerial vehicle through imitation learning, the neural network reasoning module is trained through the data set obtained by the expert system; each unmanned aerial vehicle in the group is respectively configured with a trained student system, and the unmanned aerial vehicle trajectory is predicted through the student system to complete autonomous motion planning. Compared with the prior art, the present invention adopts imitation learning, does not need to design a finite state machine decision process, completely relies on the learning of a neural network for decision-making, integrates autonomous planning and decision-making processes, and saves time.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and in particular to an autonomous motion planning method for a group of unmanned aerial vehicles based on imitation learning. Background Art

[0002] With the development of society and technology, people often encounter drone swarms performing tasks in various scenarios. However, in the current practical application of robots, the performance of some autonomous mobile drones in motion control planning still needs to be improved. In the future, drones will go deeper into more complex and harsh environments to perform tasks that humans cannot do or may require huge costs, such as exploring unknown caves and virgin forests, landfilling nuclear waste, and even going to other planets to perform tasks. At this time, how to enable drone swarms to generate fast, efficient, and mutually avoidant planning paths has become a crucial issue.

[0003] Most of the existing methods for autonomous flight of swarm drones are to execute the four sub-task modules of environmental perception, motion planning, decision control, and then pass the control instructions to the actuator in sequence. When all the sub-modules are executed in sequence, it is inevitable that the total processing time will increase. The entire process must be completed before the drone can avoid obstacles. When the motion planning algorithm is too complex, it will slow down the entire process, causing the drone to be unable to handle the difficulties it faces in real time, resulting in crashes or collisions. Unless the CPU computing power of the airborne end can reach a very strong level, there is no need to consider the complexity of the planning algorithm. In terms of motion planning, two types of methods are considered. One is based on sampling, which can meet the basic collision avoidance conditions, but the solution is not the optimal solution; the other is to solve nonlinear optimization problems containing many constraints. This type of method solves the optimal solution, but it is very difficult and time-consuming to solve. At the decision-making level, most autonomous planning methods consider the use of finite state machine methods. When faced with unknown states, random decisions will be made, or even emergency stops will be made, affecting the overall flight speed. Therefore, how to reduce the time consumed by the entire autonomous planning process, make drones fly safer and faster as much as possible, and make more efficient decisions has become the main technical challenge in the field of swarm drone autonomous flight. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a group of drone autonomous motion planning method based on imitation learning. This method reduces the time consumption problem in the autonomous planning process, allows drones to follow safe trajectories more quickly, and make efficient decisions.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning, the method comprising:

[0007] Build an expert system: Obtain a data set for imitation learning, including real-time environmental perception, the drone’s own status, information about other drones in the group, and the drone’s own real-time predicted trajectory;

[0008] Constructing a student system: including a neural network reasoning module for predicting the trajectory of the drone through imitation learning, wherein the neural network reasoning module is trained by the data set obtained by the expert system, wherein the input of the neural network reasoning module includes real-time environmental perception, the drone's own state, and information about other drones in the group, and the output of the neural network reasoning module includes the real-time predicted trajectory of the drone itself;

[0009] A trained student system is configured for each drone in the group, and the student system is used to predict the drone's trajectory and complete autonomous motion planning.

[0010] Preferably, the expert system comprises:

[0011] Environmental perception module: used to obtain real-time environmental perception, the drone’s own status, and information about other drones in the group;

[0012] Motion planning module: used for global path planning and local trajectory optimization;

[0013] Decision-making control module: Make decisions based on the motion planning results to output the predicted trajectory sequence of the UAV, and generate control instructions based on the predicted trajectory sequence;

[0014] Control instruction execution agency: receives control instructions and executes them.

[0015] Preferably, the student system further comprises:

[0016] Environmental perception module: used to obtain real-time environmental perception, the drone's own status, and information about other drones in the group, and input the acquired information into the neural network reasoning module;

[0017] Control execution module: obtains the real-time predicted trajectory of the UAV itself output by the neural network reasoning module and controls the operation of the UAV.

[0018] Preferably, the real-time environmental perception includes real-time environmental depth image information and the direction of the destination during the flight of the drone.

[0019] Preferably, the drone's own state includes the drone's posture and position information.

[0020] Preferably, the information about other drones in the group includes predicted trajectory information of other drones in the group.

[0021] Preferably, the neural network reasoning module comprises:

[0022] Feature processing branch: Processes real-time environmental perception, the drone’s own status, and information about other drones in the group to obtain feature vectors of three categories;

[0023] Imitation learning network model: Imitation learning is performed based on the input feature vector to obtain the real-time predicted trajectory of the drone itself.

[0024] Preferably, the feature processing branch includes a deep image feature extraction network for extracting features from images.

[0025] Preferably, the deep image feature extraction network includes a MobileNet-V3 network.

[0026] Preferably, the loss function when training the neural network reasoning module includes three parts, namely, the comparison loss between the label and the real-time predicted trajectory of the drone itself output by the neural network reasoning module, the collision loss between the real-time predicted trajectory of the drone itself and obstacles in the environment, and the collision loss between the real-time predicted trajectory of the drone itself and the predicted trajectories of other drones.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] (1) The present invention can enable the entire group of drones to complete an autonomous flight process from a starting point to a preset target point in a complex environment. The process completely relies on autonomous planning, can avoid static obstacles along the way, and ensure that there is no collision between the drones and other drones in the group.

[0029] (2) The present invention adopts imitation learning, so there is no need to design a finite state machine decision-making process. Decisions are made entirely based on the learning of neural networks, integrating autonomous planning and decision-making processes, saving time.

[0030] (3) In the present invention, the autonomous flight speed of the swarm drones of the student system will be faster than that of the expert system, because for the obtained swarm collision-free predicted trajectory sequence, when designing the model predictive controller, larger control speed and acceleration upper and lower limits can be set to enable the robot to track these trajectory points faster.

[0031] (4) The expert system of the present invention is based on nonlinear optimization problems when optimizing trajectories, that is, the predicted trajectories obtained based on mathematical optimization methods can make the data of each successful planning into a data set. Therefore, compared with some methods that obtain predicted trajectory sequences based on sampling, the data utilization rate is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is an overall framework diagram of a method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning according to the present invention;

[0033] Figure 2 A schematic diagram of the training process of the neural network reasoning module in the student system of the present invention;

[0034] Figure 3 The present invention is a schematic diagram of a specific implementation scheme of a method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning. DETAILED DESCRIPTION

[0035] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Note that the following embodiments are merely illustrative in nature, and the present invention is not intended to limit its applicable objects or uses, and the present invention is not limited to the following embodiments.

[0036] Example

[0037] like Figure 1 As shown, this embodiment provides a method for autonomous motion planning of a group of drones based on imitation learning, the method comprising:

[0038] Build an expert system: Obtain a data set for imitation learning, including real-time environmental perception, the drone’s own status, information about other drones in the group, and the drone’s own real-time predicted trajectory;

[0039] Constructing a student system: including a neural network reasoning module for predicting drone trajectories through imitation learning. The neural network reasoning module is trained with a data set obtained by the expert system. The input of the neural network reasoning module includes real-time environmental perception, the drone's own state, and information about other drones in the group. The output of the neural network reasoning module includes the drone's own real-time predicted trajectory.

[0040] A trained student system is configured for each drone in the group, and the student system is used to predict the drone's trajectory and complete autonomous motion planning.

[0041] The following is a specific description of the expert system and the student system:

[0042] 1. Expert system:

[0043] The expert system consists of four sub-modules, namely, environment perception, motion planning, decision control, and control instruction execution mechanism.

[0044] (1) Environmental perception module: used to obtain real-time environmental perception, the drone's own status, and information about other drones in the group. Specifically, it includes the real-time depth image information taken by the onboard depth camera in the current simulation / actual environment during the drone's flight, the direction of the destination, the drone's posture information at the current moment, and the predicted trajectory sequence of other surrounding drones received at the current moment. Only a grid map needs to be established.

[0045] (2) Motion planning module: used for global path planning and local trajectory optimization. The global path planning adopts the hybrid A* algorithm that satisfies the robot dynamics. The local trajectory optimization constructs a nonlinear optimization problem, which includes some constraints, such as the smoothness of the trajectory, the penalty for dynamic infeasibility, the collision penalty between the trajectory and obstacles in the environment, and the group trajectory collision penalty.

[0046] (3) Decision-making control module: The decision-making part of the expert system outputs the predicted trajectory sequence of the drone based on the motion planning results, and generates control instructions based on the predicted trajectory sequence. Specifically, the decision-making part of the expert system adopts a finite state machine. When an individual drone in the group faces any situation, it will select the most appropriate strategy from the limited decision-making, which is simple and efficient. A priority mechanism for drones is added. The priority is determined based on the Euclidean distance between the current drone and the target point. The smaller the Euclidean distance, the higher the priority. After the priority is determined, the finite state machine can operate more effectively. For example, if a narrow area appears in front and only one drone is allowed to pass, the high-priority drone needs to pass through, and the low-priority drone slows down or waits in the air. This can prevent drones with lower priorities from interfering with their own path planning, and finally allow the entire drone group to pass smoothly and avoid the possibility of collision. In addition, the planned trajectory bypasses cylindrical obstacles (such as trees and telephone poles) on both sides. Similarly, the final output is a predicted trajectory sequence that has been optimized and can make reasonable decisions. The control of the expert system adopts a model predictive controller (MPC).

[0047] (4) Control instruction execution agency: Receives control instructions and executes them, thus realizing the entire group autonomous planning process of the expert system.

[0048] 2. Student System

[0049] The student system is divided into three sub-modules, namely, environmental perception, neural network reasoning module, and control execution module.

[0050] (1) Environmental perception module: This module is consistent with the environmental perception module in the expert system and is used to obtain real-time environmental perception, the drone’s own status, and information about other drones in the group. The information is then input into the neural network reasoning module. Specifically, it includes real-time depth image information captured by the onboard depth camera in the current simulation / actual environment during the drone’s flight, the direction of the destination, the drone’s posture information at the current moment, and the predicted trajectory sequence of other surrounding drones received at the current moment.

[0051] (2) Neural network reasoning module: It is trained using the data set obtained through the expert system. The input of the neural network reasoning module includes real-time environmental perception, the drone’s own status, and information about other drones in the group. The output of the neural network reasoning module includes the drone’s own real-time predicted trajectory.

[0052] (3) Control execution module: obtains the real-time predicted trajectory of the UAV output by the neural network reasoning module and controls the operation of the UAV.

[0053] Among them, the neural network reasoning module includes:

[0054] Feature processing branch: Processes real-time environmental perception, the drone’s own status, and information about other drones in the group to obtain feature vectors of three categories. The feature processing branch includes a deep image feature extraction network for extracting features from images, and the deep image feature extraction network includes the MobileNet-V3 network.

[0055] Imitation learning network model: Imitation learning is performed based on the input feature vector to obtain the real-time predicted trajectory of the drone itself.

[0056] The loss function when training the neural network inference module includes three parts: the comparison loss between the label and the drone’s own real-time predicted trajectory output by the neural network inference module, the collision loss between the drone’s own real-time predicted trajectory and obstacles in the environment, and the collision loss between the drone’s own real-time predicted trajectory and the predicted trajectories of other drones.

[0057] The specific implementation process of the present invention is described below with a specific implementation method:

[0058] The expert system needs to run in a simulation environment. The simulation environment used in this method is the Rviz platform, a three-dimensional visualization tool of the robot operating system ROS. A random forest is generated in the visualization scene, and the trees in the forest are replaced by cylinders with a certain width range. The simulation platform also provides a virtual model of the drone, which is only used as the construction of the experimental scene and does not participate in the construction of any drone autonomous flight module.

[0059] The expert system consists of four sub-modules, namely, environmental perception, motion planning, decision control, and control command execution mechanism; the student system consists of three sub-modules, namely, environmental perception, neural network reasoning module trained with the expert system data set, and control execution module; the design reference of each module of the expert system and the student system Figure 1 .

[0060] The environmental perception module of the expert system: contains real-time depth image information of size 640*480 taken by the onboard depth camera in the current simulation / actual environment during the flight of the drone, with the depth in Float64 format, the three-dimensional vector pointing from the current position to the destination as the direction vector of the destination, the attitude information of the drone at the current moment, including three-dimensional position information, speed information, yaw angle, rotation matrix, and the predicted trajectory sequence of other surrounding drones received at the current moment. Only the grid map needs to be established.

[0061] The motion planning module of the expert system is divided into two processes, namely global path planning and local trajectory optimization. The global path planning adopts the hybrid A* algorithm that satisfies the robot dynamics. The local trajectory optimization constructs a nonlinear optimization problem. The penalty terms include some constraints, such as the smoothness of the trajectory, the penalty for dynamic infeasibility, the collision penalty between the trajectory and the obstacles in the environment, and the group trajectory collision penalty. Smoothness is composed of the acceleration during the motion process and the sum of the squares of the derivatives of the acceleration. The penalty for dynamic infeasibility is to ensure that the high-order derivatives of the trajectory in each dimension exist and are continuous, otherwise the trajectory will have a sudden change in the dynamic high-order quantity (such as discontinuity of speed or acceleration), which will eventually cause damage to the actuator. The collision penalty term of the environmental obstacle considers a method of pushing the trajectory away from the local obstacle. It is determined by solving the convex optimization problem of the segment that needs to be improved around the obstacle, selecting a suitable push-off control point, and the gradient direction of the push-off control point is in the vertical direction of the push-off control point. The group trajectory collision penalty term considers the group, ensuring that the predicted trajectory of the drone entering the current drone’s vicinity maintains a certain safe distance from its own trajectory, so as to avoid collisions between drones.

[0062] The decision-making part of the expert system uses a finite state machine. When an individual in the group of drones faces any situation, it will select the most appropriate strategy from the limited decisions, which is simple and efficient. A priority mechanism for drones is added. The priority is determined based on the Euclidean distance between the current drone and the target point. The smaller the Euclidean distance, the higher the priority. After the priority is determined, the finite state machine can operate more effectively. For example, if a narrow area appears in front and only one drone is allowed to pass, the high-priority drone needs to pass through, and the low-priority drone slows down or waits in the air. This can prevent drones with lower priorities from interfering with their own path planning, and ultimately allow the entire drone group to pass smoothly and avoid the possibility of collision. Similarly, the final output is a predicted trajectory sequence that has been optimized and can make reasonable decisions. The control of the expert system uses a model predictive controller (MPC). Finally, the generated control instructions are passed to the executive body of the expert system, and the entire group autonomous planning process of the expert system can be realized.

[0063] The communication between drones is based on ROS. The current position of each drone is exchanged through ROS topics, and the predicted trajectory sequence of surrounding drones is queried according to the distance.

[0064] The difference between the student system and the expert system lies mainly in the use of the neural network module to learn the motion planning module and the finite state machine decision-making part of the expert system.

[0065] Environmental perception module of the student system: consistent with the perception module of the expert system, these are used as inputs to the neural network model.

[0066] Neural network reasoning module of the student system: The neural network is a feature extraction network containing the MobileNet-V3 structure. Figure 2 A schematic diagram of the neural network structure involved in data set training according to the present invention is shown.

[0067] By taking the real-time environmental perception of the expert system, the state of the drone itself, and the predicted trajectories of other drones in the group as training attribute samples, and the real-time predicted trajectories as sample labels, a large amount of data is collected as a data set. This data set is used to train the neural network.

[0068] like Figure 2As shown in the figure, the attributes of the training data include 4 parts, including environmental point cloud information, which is convenient for querying the obstacle point cloud information closest to the current drone through the KD tree. The dimension is (2,), where 2 refers to the size of each batch of data; the drone position and attitude information, the dimension is (2,1,21), the dimension 2 also refers to the size of each batch of data, the dimension 1 refers to taking only the data of the current frame, and the dimension 21 contains the drone's position, angular velocity, linear velocity, rotation matrix coefficient, target direction and other information; the dimension of the depth image after processing becomes (2,1,224,224,3), the first two dimensions are consistent with the previous attribute introduction, 224*224*3 is the size of a single image after processing; the dimension of other drone trajectory information is (2,1,n*60), the explanation of the first two dimensions is the same as above, n represents the number of drones in the group, and 60 represents the predicted three-dimensional coordinates of 20 trajectory points.

[0069] The network consists of two branches, which divide the attributes into two parts, respectively generating latent encodings of deep vision, the attitude and position information of the drone, and extracting the features of the predicted trajectories of other drones in the group, and outputting three predicted trajectories through the forward propagation of the neural network.

[0070] First, for the first depth image branch, the pre-trained MobileNet-V3 architecture is used to effectively extract features from the depth image, and then these features are processed through one-dimensional convolution to generate a multi-feature vector of size 32. Then, for the second branch, the current platform speed and attitude, the required reference direction are concatenated together, and processed by a four-layer perceptron with [64, 32, 32, 32] hidden nodes and LeakyReLU as the activation function, and a one-dimensional convolution is used again to create a 32-dimensional feature vector for this attribute. Then, the two visual and state feature branches are respectively processed by one-dimensional convolution and combined, and then added to the predicted trajectory attribute vector of other drones in the group, and finally processed by another four-layer perceptron with [64, 128, 128] hidden nodes and LeakyReLU activation. The output is three predicted trajectory point sequences with size [2, 3, 60], where dimension 2 represents the size of each batch of data, dimension 3 represents the three trajectories, and 60 represents the length 20 of the predicted trajectory and the position component of the three-dimensional coordinates. The loss function in the network consists of three parts: the comparison loss between the label and the trajectory predicted by the neural network, the collision loss between the trajectory and obstacles in the environment, and the collision loss between the predicted trajectory and other drones. The first loss considers the difference between the label and the output trajectory of the neural network forward propagation, the second loss considers the distance between the output trajectory of the forward propagation and the obstacles in the environmental point cloud at the current position, and the third loss measures the distance between the predicted trajectory information of other drones in the group and the trajectory predicted by the neural network. The Adam optimizer is used for training, the number of each minimum batch is set to 32, and the learning rate is set to 1*10 -3 .

[0071] The control execution module of the student system: The three predicted trajectories with a length of 20 output by the module will be averaged and then fitted with a third-order B-spline curve. The model predictive control (MPC) is also used to follow the trajectory. Finally, the autonomous flight function of the group drone of the student system can be realized.

[0072] It should be noted that: the output of the neural network reasoning module of the student system in this embodiment is 3 prediction trajectories, and in other embodiments, 1 prediction trajectory can be directly output, so that the network scale is reduced and the response speed is faster. Those skilled in the art can adjust the network as needed to obtain 1 or more prediction trajectories. In addition, the length (number) of the prediction trajectory points can also be adjusted as needed, such as adjusted to 5, 10, 15 points, etc.

[0073] Figure 3The method overview diagram in the above-mentioned specific embodiment scheme of the present invention is shown, including the implementation details of the expert system and the student system. The present invention can realize the autonomous flight process of the entire group of drones from the starting point to the preset target point in a complex environment. It completely relies on autonomous planning on the way, can avoid static obstacles along the way, and ensure that there is no collision between drones. Imitation learning is adopted, so there is no need to design a finite state machine decision process, and decisions are made entirely by learning neural networks, integrating autonomous planning and decision-making processes, saving time. The speed of autonomous flight of the group of drones in the student system will be faster than that of the expert system. Because for the obtained group collision-free predicted trajectory sequence, when designing the model predictive controller, a larger control speed and acceleration upper and lower limits can be set to enable the robot to track these trajectory points faster. The expert system is based on a nonlinear optimization problem when optimizing the trajectory, that is, the predicted trajectory obtained based on the mathematical optimization method, and the data of each successful planning can be made into a data set. Therefore, compared with some methods based on sampling to obtain the predicted trajectory sequence, the data utilization rate is higher.

[0074] The above embodiments are merely examples and do not limit the scope of the present invention. These embodiments can also be implemented in various other ways, and various omissions, substitutions, and changes can be made without departing from the technical concept of the present invention.

Claims

1. A method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning, characterized in that: The method includes: Build an expert system: Obtain a data set for imitation learning, including real-time environmental perception, the drone’s own status, information about other drones in the group, and the drone’s own real-time predicted trajectory; Constructing a student system: including a neural network reasoning module for predicting the trajectory of the drone through imitation learning, wherein the neural network reasoning module is trained by the data set obtained by the expert system, wherein the input of the neural network reasoning module includes real-time environmental perception, the drone's own state, and information about other drones in the group, and the output of the neural network reasoning module includes the real-time predicted trajectory of the drone itself; Each drone in the group is equipped with a trained student system, which is used to predict the drone's trajectory and complete autonomous motion planning. The student system also includes: Environmental perception module: used to obtain real-time environmental perception, the drone's own status, and information about other drones in the group, and input the acquired information into the neural network reasoning module; Control execution module: obtains the real-time predicted trajectory of the UAV output by the neural network reasoning module and controls the operation of the UAV; The neural network reasoning module has a two-branch structure, which divides the attributes into two parts, respectively generating potential codes of deep vision, the attitude and position information of the drone, and extracting the features of the predicted trajectories of other drones in the group, and outputting three predicted trajectories after the forward propagation of the neural network; For the first branch structure, the pre-trained MobileNet-V3 architecture is used to extract features from the acquired depth image, and then these features are processed by one-dimensional convolution to generate a multi-feature visual vector of size 32; for the second branch structure, the reference direction required for the current platform's speed and posture is concatenated together and processed by a four-layer perceptron with [64, 32, 32, 32] hidden nodes and LeakyReLU as the activation function, and one-dimensional convolution is used again to create a 32-dimensional state feature vector for this attribute; then, these two visual vectors and the state feature vector are respectively subjected to one-dimensional convolution and combined together, and then added to the predicted trajectory attribute vectors of other drones in the group, and finally processed by another four-layer perceptron with [64, 128, 128] hidden nodes and LeakyReLU activation. The output is three predicted trajectory point sequences with a size of [2, 3, 60], where dimension 2 represents the size of each batch of data, dimension 3 represents three trajectories, and 60 represents the length 20 of the predicted trajectory and the position component of the three-dimensional coordinates; The loss function when training the neural network inference module includes three parts, namely, the comparison loss between the label and the real-time predicted trajectory of the drone itself output by the neural network inference module, the collision loss between the real-time predicted trajectory of the drone itself and obstacles in the environment, and the collision loss between the real-time predicted trajectory of the drone itself and the predicted trajectories of other drones.

2. The method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning according to claim 1 is characterized in that: The expert system comprises: Environmental perception module: used to obtain real-time environmental perception, the drone’s own status, and information about other drones in the group; Motion planning module: used for global path planning and local trajectory optimization; Decision-making control module: Make decisions based on the motion planning results to output the predicted trajectory sequence of the UAV, and generate control instructions based on the predicted trajectory sequence; Control instruction execution agency: receives control instructions and executes them.

3. A method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning according to any one of claims 1 to 2, characterized in that: The real-time environment perception includes real-time environment depth image information and destination direction during the flight of the UAV.

4. A method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning according to any one of claims 1 to 2, characterized in that: The drone's own state includes the drone's posture and position information.

5. The method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning according to any one of claims 1 to 2, characterized in that: The information about other drones in the group includes predicted trajectory information of other drones in the group.

6. The method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning according to claim 1, characterized in that: The neural network reasoning module includes: Feature processing branch: Processes real-time environmental perception, the drone’s own status, and information about other drones in the group to obtain feature vectors of three categories; Imitation learning network model: Imitation learning is performed based on the input feature vector to obtain the real-time predicted trajectory of the drone itself.

7. The method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning according to claim 6 is characterized in that: The feature processing branch includes a deep image feature extraction network for extracting features from images.

8. The method for autonomous motion planning of a swarm of unmanned aerial vehicles based on imitation learning according to claim 7, characterized in that: The deep image feature extraction network includes a MobileNet-V3 network.

Images