Multi-uav task allocation and conflict-free path planning method and system

Abstract

This invention discloses a method and system for multi-UAV task allocation and conflict-free path planning, belonging to the field of industrial IoT UAV technology. For collaborative inspection scenarios involving multiple heterogeneous UAVs, this invention first transforms multi-UAV task allocation into a single-travel salesman problem by using a multi-travel salesman problem transformation method, generating an optimal task sequence that satisfies sensor matching constraints. Then, it constructs a multi-agent reinforcement learning model based on MAPPO, combined with LSTM temporal environment encoding, to train a distributed conflict-free path planning strategy. This invention balances solution efficiency and planning quality, can be scaled to UAV swarms of any size, and effectively solves the problems of existing algorithms struggling to balance efficiency and quality, poor scalability, and insufficient dynamic obstacle avoidance capabilities. It is suitable for industrial inspection scenarios such as hazardous materials warehouses.

Description

Technical Field

[0001] This invention belongs to the field of Industrial Internet of Things (IIoT) and UAV autonomous control technology, specifically relating to a multi-UAV collaborative task allocation and path planning technology, and particularly to a multi-UAV task allocation and conflict-free path planning method and system. Background Technology

[0002] With the rapid development of drone technology, drones have been widely used in scenarios such as hazardous materials warehouse inspection, industrial park environmental monitoring, and power facility inspection in the Industrial Internet of Things (IIoT) field, thanks to their advantages of flexible deployment, high mobility, and ability to carry multiple sensors. In these scenarios, multiple heterogeneous drones equipped with different sensors typically need to work collaboratively to complete two core tasks: first, assigning corresponding inspection tasks to each drone based on its sensor capabilities and the detection requirements of the task points; and second, planning conflict-free flight paths for each drone to ensure that all task points are covered and that there are no collisions between drones or between drones and obstacles.

[0003] Currently, existing technologies for the joint optimization of multi-UAV task allocation and path planning can be mainly divided into two categories: one is based on heuristic algorithms or swarm intelligence optimization algorithms, such as ant colony algorithms, genetic algorithms, and CBSS combined path planning algorithms. These schemes have obvious drawbacks: in order to pursue the optimal solution, multiple iterative searches are required, resulting in extremely low solution efficiency and making it difficult to meet real-time requirements; while in order to improve solution efficiency, the quality of the solution is greatly sacrificed. At the same time, the task allocation and path planning modules are heavily coupled, and the optimal task allocation result is often compromised for the convenience of path planning; in addition, the computational complexity of these schemes increases exponentially with the increase of the number of UAVs and tasks, resulting in extremely poor scalability and making them unsuitable for large-scale UAV swarm scenarios.

[0004] Another type is path planning schemes based on reinforcement learning. Existing schemes mostly use single-agent reinforcement learning, which cannot handle path conflict problems between multiple drones. A few multi-agent reinforcement learning schemes, such as the MADDPG algorithm, are prone to getting trapped in local optima and have unstable training. They also do not fully consider the temporal information of the drone's surrounding environment and have insufficient ability to avoid dynamic obstacles such as other drones in dynamic flight, thus failing to guarantee the success rate of conflict-free path planning.

[0005] In summary, existing technologies cannot simultaneously meet the multiple requirements of task allocation and path planning efficiency, planning quality, scalability, and dynamic obstacle avoidance capabilities in multi-heterogeneous UAV collaborative inspection scenarios. A new technical solution is urgently needed to address these issues. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for multi-UAV task allocation and conflict-free path planning. This method solves the problems of existing algorithms, such as difficulty in balancing solution efficiency and planning quality, poor scalability, and insufficient dynamic obstacle avoidance capabilities. It enables efficient, safe, and scalable collaborative inspection operations of multiple heterogeneous UAVs in industrial IoT scenarios.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0008] A multi-drone task allocation and conflict-free path planning method, applied to a collaborative inspection scenario of multiple heterogeneous drones in the Industrial Internet of Things, includes the following steps:

[0009] S1 Problem Modeling: The joint problem of task allocation and path planning for multiple heterogeneous UAVs is modeled as a multi-agent combined path planning problem, and a finite undirected graph is constructed. vertex set , For the set of drone starting points, For the final assembly of drones, For the target task point set, edge set Define the connectivity between vertices; define the core constraints of the drone: sensor matching constraint (each target task point can only be detected by a drone equipped with the corresponding sensor) and collision safety constraint (the distance between the drone and the obstacle must be greater than the preset safety distance).

[0010] S2 Task Allocation: Based on the Multi-Traveling Salesman Problem (mTSP) transformation method, the multi-UAV task allocation problem is transformed into a single-traveling salesman problem (TSP). Specifically, a transformation graph conforming to sensor matching constraints is constructed, defining four types of edges and their corresponding costs: First type: Edges from the UAV's starting point to its accessible replica nodes, with a cost equal to the shortest path length between the two points in the original undirected graph; Second type: Edges from the starting point of the i-th UAV to the ending point of the (i+1)-th UAV, with a cost of 0, where the edge from the starting point of the last UAV to the ending point of the first UAV has a cost of 0; Third type: Edges between different replica nodes of the same target task point, with a cost of 0; Fourth type: Edges from the predecessor node of a target task point replica node to other replica nodes accessible by that UAV, with a cost equal to the original path length plus a preset penalty value.

[0011] After solving the single TSP problem of the transformation graph using the LKHTSP solver and obtaining the optimal path with the minimum cost, the second type of zero-cost connection edges are removed, and the optimal path is divided into task sequences corresponding to each UAV. The task sequence takes the UAV's starting point as the starting point, the ending point as the ending point, and the matching target task point in between.

[0012] S3 Path Planning Model Construction: A conflict-free path planning model based on multi-agent proximal policy optimization MAPPO is constructed, employing a centralized training and distributed execution framework. The specific design is as follows:

[0013] State space: The UAV’s own state (position, velocity, heading angle), the surrounding environment state (position of dynamic UAV and static obstacles), and the current target point state are integrated. The Long Short-Term Memory (LSTM) network is used to encode the environmental state at continuous time steps, capture the motion trend of dynamic obstacles, and obtain the encoded state features.

[0014] Action space: Defines a continuous action space with yaw angle and acceleration as dimensions, mapping dimensionless action values ​​to the motion parameter range of the actual hardware of the drone, and adapting to different models of heterogeneous drones;

[0015] Reward Function: Construct a total reward function that integrates collision penalty reward and distance reward. The collision penalty reward includes static obstacle collision penalty and dynamic UAV collision penalty, and the distance reward includes target point distance reward and reference trajectory distance reward. The collision penalty is given the highest priority to ensure flight safety, while the distance reward guides the UAV to fly quickly toward the target point, accelerating model convergence.

[0016] S4 Model Offline Training: The MAPPO model is trained offline in a stochastic simulation environment built on OpenAI Gym. A framework of centralized training and distributed execution is adopted. Through experience collection, generalized advantage estimation, and iterative update of network parameters, a convergent conflict-free path planning strategy is obtained.

[0017] S5 Online Path Planning: The trained path planning strategy is deployed to the onboard computer of each UAV. Each UAV receives the task sequence issued by the ground control center, divides the flight process into multiple segmented flight tasks, collects surrounding environmental information in real time through its own onboard sensors, inputs it into the path planning strategy after LSTM encoding, outputs flight action commands, controls the UAV to perform distributed conflict-free path planning between adjacent task points, autonomously avoids static obstacles and other dynamically flying UAVs, completes all detection tasks in sequence and reaches the destination.

[0018] Furthermore, the present invention also provides a multi-UAV task allocation and conflict-free path planning system for implementing the above method, including a task allocation module, a MAPF training module, and a MAPF execution module:

[0019] Task allocation module: A centralized module deployed at the ground control center, including a data receiving unit, a graph transformation unit, and a TSP solving unit. It is used to receive the start point, end point, and sensor type information of the UAV, as well as the location and detection requirements of the target task point. Through the mTSP transformation method and TSP solver, it generates the task sequence of each UAV and sends it to the corresponding UAV.

[0020] MAPF Training Module: An offline training module deployed on a server, including a state preprocessing unit, a policy network unit, a value network unit, and an experience management unit. It is based on the PyTorch framework to build and train the MAPF model, design the state space, action space, and reward function, complete the offline training of the model in a stochastic simulation environment, and output a converged conflict-free path planning policy to the MAPF execution module.

[0021] MAPF Execution Module: A distributed module deployed on the onboard computer of each UAV, including a task parsing unit, an environmental perception unit, an action decision unit, and a flight control unit. It realizes distributed conflict-free path planning and flight control, receives task sequences and path planning strategies, executes distributed conflict-free path planning based on environmental information collected by the UAV's own sensors, and controls the UAV to complete tasks in sequence and avoid obstacles.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] 1. This invention completely decouples task allocation from path planning. By using the mTSP transformation method, the multi-UAV task allocation is transformed into a mature TSP problem, ensuring the optimality of task allocation and the efficiency of the solution. The path planning stage adopts a reinforcement learning scheme, and the inference process has constant time complexity, which greatly shortens the solution time of path planning. At the same time, the reward function guides the UAV to plan the optimal path, achieving a balance between solution efficiency and planning quality.

[0024] 2. This invention adopts the MAPPO framework with centralized training and distributed execution. The path planning strategy trained in a small number of drone environments can be directly extended to large-scale scenarios with any number of drones without retraining. This solves the problem that the computational complexity of existing algorithms increases exponentially with scale, and has extremely strong scalability.

[0025] 3. This invention uses an LSTM network to encode the temporal information of the UAV's surrounding environment, which can effectively capture the movement trends of obstacles such as dynamic UAVs. Combined with high-priority collision penalty rewards, it significantly improves the UAV's ability to avoid static and dynamic obstacles, and the success rate of conflict-free path planning is significantly higher than that of existing algorithms.

[0026] 4. The task allocation process of this invention fully considers the sensor matching constraints of heterogeneous UAVs, and only assigns tasks with detection capabilities to UAVs; the motion space adopts a dimensionless design, which can be flexibly mapped to the hardware motion parameter range of different UAV models, perfectly adapting to the collaborative operation scenario of multiple heterogeneous UAVs.

[0027] 5. The online path planning of this invention adopts a distributed execution mode. Each UAV only needs to make decisions based on the local information collected by its own sensors, without the need for real-time synchronization with other UAVs or ground control centers. This greatly reduces communication overhead, is suitable for communication-constrained environments in industrial scenarios, and has strong engineering practicality and deployability. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall process of the method described in this invention;

[0029] Figure 2 This is a schematic diagram of the overall architecture of the system described in this invention;

[0030] Figure 3 This is a schematic diagram of the training framework of the MAPPO model of the present invention. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to specific embodiments.

[0032] A method and system for multi-UAV task allocation and conflict-free path planning is presented. In this embodiment, a multi-heterogeneous UAV collaborative inspection scenario of a hazardous materials warehouse is used as an example. The warehouse is a two-dimensional planar scene of 20m×20m, and 5 heterogeneous UAVs are deployed. Among them, 2 UAVs are equipped with temperature and humidity sensors, and 3 UAVs are equipped with hazardous materials leakage detection sensors. 10 target task points are set in the warehouse, of which 4 are temperature and humidity detection points and 6 are hazardous materials leakage detection points. Only UAVs equipped with matching sensors can perform detection tasks.

[0033] like Figure 1 As shown, the specific implementation steps of this embodiment are as follows:

[0034] S1. Problem Modeling

[0035] Construct a finite undirected graph to describe the warehouse layout, and define the set of starting points for 5 drones. End point set A set of 10 target task points Define core constraints:

[0036] Sensor matching constraints: Temperature and humidity detection points can only be accessed by drones equipped with temperature and humidity sensors, and hazardous material leak detection points can only be accessed by drones equipped with the corresponding sensors.

[0037] Collision safety constraints: The preset safety distance ε=1m. The Euclidean distance between the drone and static obstacles such as shelves and walls, as well as other dynamically flying drones, must be greater than 1m; otherwise, it is considered a collision.

[0038] Position constraints: The drone is initially located at the corresponding starting point and must reach the corresponding destination after completing the mission;

[0039] S2. Task Assignment

[0040] A transformation graph is constructed based on the mTSP transformation method. For each target task point, a replica node corresponding only to a UAV with detection qualifications is generated. Four types of edges and their corresponding costs are defined, including penalty values. , Given the maximum and shortest path length between vertices in the original undirected graph, the LKH-2.0.9 TSP solver is used to solve the single TSP problem of the transformed graph. After obtaining the optimal path with the minimum cost, the second type of zero-cost edge is removed, and the task sequence corresponding to 5 drones is obtained. Each sequence starts from the drone's starting point, ends at the drone's ending point, and the middle is the matched target task point.

[0041] S3. Path Planning Model Construction

[0042] The MAPPO model is built based on the PyTorch framework, and the specific design is as follows:

[0043] State space: The state of the UAV itself (x / y coordinates, x / y direction velocity, heading angle cosθ / sinθ), the state of 5 surrounding dynamic UAVs, the position of 20 static obstacles, and the position of the current target point are integrated; the environmental state of the first 2 time steps is encoded using an LSTM network to obtain the encoded state features;

[0044] The encoded state-space expression is:

[0045]

[0046] in, Let t represent the position, velocity, and heading angle of the UAV itself. Let t be the current position state of the target point. For the encoding function of the LSTM network, The surrounding obstacle status includes the positions of dynamic drones and static obstacles, and p is the historical time step used by the LSTM network.

[0047] Action space: Defines a dimensionless continuous action space. Mapped to the actual parameter range: yaw angle acceleration ;

[0048] Reward function: The total reward function is When a collision occurs and Output a penalty value of -100; output 0 for no collision. and All values ​​are The closer the distance to the target point and the closer it is to the reference trajectory, the higher the reward value, i.e., the closer it is to 1. The reference trajectory is pre-generated by the JSP fast path search algorithm.

[0049] S4. Offline Model Training

[0050] A stochastic simulation environment was built using OpenAI Gym, with the following training hyperparameters set: 100 parallel environments, minibatch size = 8, training steps = 100M, MLP hidden layer size [128, 128], activation function ReLU, optimizer Adam, PPO pruning factor 0.2, discount factor γ = 0.99, and learning rate 0.005. During iterative training, state-action-reward data was collected and stored in a buffer pool, randomly shuffled, and the advantage and target values ​​were calculated. Figure 3 As shown, the policy network and value network are updated until the average reward of the model converges and the number of drone collisions tends to stabilize, thus obtaining the trained path planning policy.

[0051] S5. Online Path Planning

[0052] The trained path planning strategy was deployed to the onboard computers of five UAVs. Each UAV received the task sequence from the ground control center, parsed and generated segmented flight targets, collected surrounding environmental information in real time through lidar and depth cameras, and input it into the policy network after LSTM encoding. The flight control unit then mapped the commands to motor control parameters to control the UAV to fly in segments. After reaching the task point, the UAV hovered for 2 seconds to complete the detection. It autonomously avoided obstacles throughout the entire process without path conflicts.

[0053] This embodiment also provides a multi-UAV task allocation and conflict-free path planning system for implementing the above method, such as... Figure 2 As shown, it includes a task allocation module, a MAPF training module, and a MAPF execution module:

[0054] Task allocation module: A centralized module deployed at the ground control center, including a data receiving unit, a graph transformation unit, and a TSP solving unit. It is used to receive the start point, end point, and sensor type information of the UAV, as well as the location and detection requirements of the target task point. Through the mTSP transformation method and TSP solver, it generates the task sequence of each UAV and sends it to the corresponding UAV.

[0055] MAPF Training Module: An offline training module deployed on a server, including a state preprocessing unit, a policy network unit, a value network unit, and an experience management unit. It is based on the PyTorch framework to build and train the MAPF model, design the state space, action space, and reward function, complete the offline training of the model in a stochastic simulation environment, and output a converged conflict-free path planning policy to the MAPF execution module.

[0056] The MAPF training module includes a state preprocessing unit, a policy network unit, a value network unit, and an experience management unit.

[0057] The state preprocessing unit is used to perform temporal encoding on the environmental state collected by the UAV using an LSTM network, and output the encoded features to the policy network unit.

[0058] The policy network unit is an actor network shared by multiple agents, used to output flight action decisions for the UAV based on encoded state features;

[0059] The value network unit is a centralized commentator network used to calculate the global value based on the global state and the actions of all drones, guiding the parameter updates of the policy network.

[0060] The experience management unit is used to collect, store, and shuffle state, action, and reward data during the training process, providing data support for network parameter updates.

[0061] MAPF Execution Module: A distributed module deployed on the onboard computer of each UAV, including a task parsing unit, an environmental perception unit, an action decision unit, and a flight control unit. It realizes distributed conflict-free path planning and flight control, receives task sequences and path planning strategies, executes distributed conflict-free path planning based on environmental information collected by the UAV's own sensors, and controls the UAV to complete tasks in sequence and avoid obstacles.

[0062] The MAPF execution module includes a mission parsing unit, an environmental perception unit, an action decision-making unit, and a flight control unit.

[0063] The task parsing unit is used to parse the received task sequence and generate segmented flight targets for adjacent task points;

[0064] The environmental perception unit is used to collect environmental information about surrounding static obstacles and dynamic drones through the lidar and depth camera sensors carried by the drone.

[0065] The action decision unit is used to output action commands of yaw angle and acceleration based on environmental perception information and the trained path planning strategy.

[0066] The flight control unit is used to map action commands into hardware control parameters of the UAV, and control the UAV to complete segmented flight and reach the target point.

[0067] Tests show that in this embodiment, task allocation takes less than 1 second, path planning is solved at a speed that is more than 80% faster than the traditional CBSS algorithm, the maximum flight time of all UAVs is reduced by more than 30% compared to the CBSS algorithm, the path planning success rate without conflict is more than 98%, and the trained strategy can be directly extended to a cluster of up to 20 UAVs without retraining, demonstrating excellent scalability.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-UAV task allocation and conflict-free path planning, characterized in that, Includes the following steps: S1. Problem Modeling: The joint problem of task allocation and path planning for multiple heterogeneous UAVs is modeled as a multi-agent combined path planning problem. A finite undirected graph containing the set of UAV starting points, the set of target task points, and the set of ending points is constructed, and collision safety constraints and sensor matching constraints for UAVs are defined. S2. Task Allocation: Based on the multi-traveling salesman transformation method (mTSP), the multi-UAV task allocation problem is transformed into a single-traveling salesman transformation problem (TSP). After obtaining the optimal solution through the TSP solver, the task sequence corresponding to each UAV is restored. S3. Path Planning Model Construction: Construct a conflict-free path planning model based on MAPPO (Multi-Agent Proximal Policy Optimization), specifically including: State Space Design: Integrating the UAV's own state, surrounding dynamic UAV states, static obstacle states, and current target point state, using a Long Short-Term Memory (LSTM) network to encode the temporal information of the surrounding environment to obtain the encoded state features; Action Space Design: Defining a continuous action space with yaw angle and acceleration as dimensions, mapping dimensionless action values ​​to the motion parameter range of the UAV's actual hardware; Reward Function Design: Constructing a total reward function that integrates collision penalty rewards and distance rewards. The collision penalty rewards include static obstacle collision penalties and dynamic UAV collision penalties, and the distance rewards include target point distance rewards and reference trajectory distance rewards. S4. Offline Model Training: The MAPPO model constructed in step S3 is trained offline in a stochastic simulation environment. A framework of centralized training and distributed execution is adopted. Through experience collection, generalized advantage estimation, and iterative update of network parameters, a convergent conflict-free path planning strategy is obtained. S5. Online Path Planning: Each UAV receives the task sequence obtained in step S2, and based on the path planning strategy trained in step S4, performs distributed conflict-free path planning between adjacent task points in the task sequence, autonomously avoids static and dynamic obstacles, completes all tasks in sequence, and reaches the destination.

2. The method for multi-UAV task allocation and conflict-free path planning according to claim 1, characterized in that, The specific process in step S2 is as follows: S21. Based on the finite undirected graph and sensor matching constraints in step S1, construct a transformation graph: generate a copy node for each target task point that corresponds only to the UAV with detection qualification, and define the vertex set of the transformation graph as the set of the UAV starting point, the copy nodes accessible to each UAV, and the UAV ending point. S22. Define four types of edges in the transformed graph and their corresponding costs: Type 1: Edges from the starting point of a drone to its accessible replica nodes, with a cost equal to the shortest path length between the two points in the original undirected graph; Type 2: Edges from the starting point of the i-th drone to the ending point of the (i+1)-th drone, with a cost of 0, where the edge from the starting point of the last drone to the ending point of the first drone has a cost of 0; Type 3: Edges between different replica nodes of the same target task point, with a cost of 0; Type 4: Edges from the predecessor node of a target task point replica node to other replica nodes accessible by the drone, with a cost equal to the original path length plus a preset penalty value. S23. The LKHTSP solver is used to solve the single traveling salesman problem of the transformed graph. After obtaining the optimal path with the minimum cost, the second type of zero-cost edge is removed, and the optimal path is divided into task sequences corresponding to each UAV. The task sequence takes the UAV's starting point as the starting point and the ending point as the ending point, with the matching target task point in between.

3. The method for multi-UAV task allocation and conflict-free path planning according to claim 1, characterized in that, In step S3, the encoded state feature expression is: ； in, Let t represent the position, velocity, and heading angle of the UAV itself. Let t be the current position state of the target point. For the encoding function of the LSTM network, The surrounding obstacle status includes the positions of dynamic drones and static obstacles, and p is the historical time step used by the LSTM network.

4. The method for multi-UAV task allocation and conflict-free path planning according to claim 1, characterized in that, In step S3, the expression for the total reward function is: ； Wherein, the weight coefficients satisfy ; This is a dynamic drone collision penalty reward; it is 0 when there is no collision and outputs a negative penalty value when a collision occurs. This is the penalty reward for collisions with static obstacles; it is 0 if there is no collision, and outputs a negative penalty value when a collision occurs. The reward is based on the distance to the target point; the closer the target point is to the target point, the higher the reward value. The reference trajectory is pre-generated by a fast path search algorithm, and the closer the UAV is to the reference trajectory, the higher the reward value.

5. The multi-UAV task allocation and conflict-free path planning method according to claim 1, characterized in that, In step S1, the collision safety constraint is achieved through a preset safety distance ε. A collision is determined when the Euclidean distance between the drone and the obstacle is less than ε. The expression for the collision safety state is: ； in, Let be the position of the i-th drone at time t. The set of obstacles includes static obstacles and dynamic drones. Let be the Euclidean distance between the i-th drone and obstacle o.

6. A multi-UAV task allocation and conflict-free path planning system, used to implement the method described in any one of claims 1-5, characterized in that, It includes a task allocation module, a MAPF training module, and a MAPF execution module, which work together: The task allocation module is a centralized module deployed at the ground control center. It is used to receive the start point, end point, and sensor type information of the UAV, as well as the location and detection requirements of the target task point. Through the mTSP transformation method and TSP solver, it generates the task sequence of each UAV and sends it to the corresponding UAV. The MAPF training module is an offline training module deployed on a server. It is used to build the MAPF path planning model, design the state space, action space and reward function, complete the offline training of the model in a stochastic simulation environment, and output a converged conflict-free path planning strategy to the MAPF execution module. The MAPF execution module is a distributed module deployed on the onboard computer of each UAV. It is used to receive task sequences and path planning strategies, perform distributed conflict-free path planning based on environmental information collected by the UAV's own sensors, and control the UAV to complete tasks in sequence and avoid obstacles.

7. A multi-UAV task allocation and conflict-free path planning system according to claim 6, characterized in that, The MAPF training module includes a state preprocessing unit, a policy network unit, a value network unit, and an experience management unit. The state preprocessing unit is used to perform temporal encoding on the environmental state collected by the UAV using an LSTM network, and output the encoded features to the policy network unit. The policy network unit is an actor network shared by multiple agents, used to output flight action decisions for the UAV based on encoded state features; The value network unit is a centralized commentator network used to calculate the global value based on the global state and the actions of all drones, guiding the parameter updates of the policy network. The experience management unit is used to collect, store, and shuffle state, action, and reward data during the training process, providing data support for network parameter updates.

8. A multi-UAV task allocation and conflict-free path planning system according to claim 6, characterized in that, The MAPF execution module includes a mission parsing unit, an environmental perception unit, an action decision-making unit, and a flight control unit. The task parsing unit is used to parse the received task sequence and generate segmented flight targets for adjacent task points; The environmental perception unit is used to collect environmental information about surrounding static obstacles and dynamic drones through the lidar and depth camera sensors carried by the drone. The action decision unit is used to output action commands of yaw angle and acceleration based on environmental perception information and the trained path planning strategy. The flight control unit is used to map action commands into hardware control parameters of the UAV, and control the UAV to complete segmented flight and reach the target point.

Images