An unmanned aerial vehicle end-to-end flight method, system, terminal and storage medium facing strong wind disturbance and thrust saturation constraint

Abstract

The application belongs to the field of data processing, and discloses an unmanned aerial vehicle end-to-end flight method, system, terminal and storage medium facing strong wind disturbance and thrust saturation constraints, the method comprising: calculating the relative wind speed vector of the unmanned aerial vehicle in the target simulation environment; calculating the air disturbance vector according to the three-dimensional direction unit vector and the relative wind speed vector; calculating the physical load according to the air disturbance vector, carrying out orthogonal decomposition processing to obtain parallel components and vertical components, and calculating the limit remaining available thrust margin; obtaining the discrete depth map and motion state data, and carrying out feature coding processing to obtain the feature coding result; carrying out alignment fusion processing on the limit remaining available thrust margin and the feature coding result to obtain the comprehensive feature vector; and generating the acceleration instruction vector according to the comprehensive feature vector to control the unmanned aerial vehicle to complete the end-to-end flight task. The application significantly improves the safety and stability of the unmanned aerial vehicle flight in the extreme strong wind environment.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an end-to-end flight method, system, terminal, and computer-readable storage medium for unmanned aerial vehicles (UAVs) under strong wind disturbance and thrust saturation constraints. Background Technology

[0002] In highly dynamic application scenarios such as complex low-altitude operations, urban canyon inspections, forest navigation, and high-speed autonomous obstacle avoidance, rotary-wing UAVs need to maintain efficient and safe autonomous flight under conditions of limited line-of-sight, complex obstacle topology, and strong sudden wind disturbances. Traditional UAV navigation architectures typically employ a hierarchical, cascaded design of "path planning - trajectory generation - trajectory tracking control." In this traditional approach, the upper-level path and trajectory planners usually perform geometric optimizations based on a nominal physical model of relatively static air, completely ignoring the severe impact of unsteady aerodynamic disturbances such as gusts, shear winds, and terrain-induced wake turbulence on the actual performance of the UAV. This leads to the trajectory generated by the upper-level planner often exceeding the maximum thrust limit of the physical motor in strong wind environments, resulting in severe trajectory divergence or even attitude instability during lower-level control and tracking.

[0003] To address these issues, end-to-end visual flight methods based on deep reinforcement learning have gradually become a research hotspot in recent years. However, existing visual end-to-end flight planning methods cannot adapt to sudden changes in relative wind speed under strong random winds, which leads to the inability to guarantee the safety of UAV flight and the failure to successfully complete end-to-end flight missions.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide an end-to-end flight method, system, terminal, and computer-readable storage medium for unmanned aerial vehicles (UAVs) under strong wind disturbances and thrust saturation constraints. This invention aims to solve the problem that existing visual end-to-end flight planning methods cannot adapt to sudden changes in relative wind speed under strong random winds, which leads to compromised flight safety and unsuccessful completion of end-to-end flight missions.

[0006] To achieve the above objectives, the present invention provides an end-to-end flight method for unmanned aerial vehicles (UAVs) oriented towards strong wind disturbances and thrust saturation constraints. The end-to-end flight method for UAVs oriented towards strong wind disturbances and thrust saturation constraints includes the following steps: The low-frequency steady-state reference wind speed vector, random gust wind speed vector, and velocity vector of the UAV in the target simulation environment are obtained. The low-frequency steady-state reference wind speed vector and the random gust wind speed vector are superimposed to obtain the total wind speed vector of the three-dimensional spatial composite wind field. The relative wind speed vector is calculated based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector. Obtain a three-dimensional directional unit vector, calculate a three-dimensional projection component based on the three-dimensional directional unit vector and the relative wind speed vector, and calculate an air disturbance vector based on the three-dimensional projection component and the three-dimensional directional unit vector; The physical load is calculated based on the air disturbance vector, and the physical load is orthogonally decomposed to obtain parallel and vertical components. The remaining usable thrust margin is then calculated based on the parallel and vertical components. A discrete depth map and motion state data are acquired, and feature encoding processing is performed on the discrete depth map and the motion state data to obtain high-dimensional visual space obstacle avoidance feature encoding and motion feature encoding. The air disturbance vector and the limit remaining available thrust margin are initially fused to obtain the physical perception code. The physical perception code, the high-dimensional visual space obstacle avoidance feature code, and the motion feature code are then aligned and fused to obtain the comprehensive feature vector. The temporal hidden state vector is obtained based on the comprehensive feature vector, and the expected nominal control acceleration command vector of the UAV is generated based on the temporal hidden state vector, so as to control the UAV to complete the end-to-end flight mission according to the expected nominal control acceleration command vector of the UAV.

[0007] Optionally, the described end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints includes the following steps: acquiring the low-frequency steady-state reference wind speed vector, random gust wind speed vector, and velocity vector of the UAV in the target simulation environment; superimposing the low-frequency steady-state reference wind speed vector and the random gust wind speed vector to obtain the total wind speed vector of the three-dimensional spatial composite wind field; and calculating the relative wind speed vector based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector. Specifically, this includes: A target simulation environment is constructed. The low-frequency steady-state reference wind speed vector corresponding to the target simulation environment is generated by the direction vector normalization method, and the random gust wind speed vector corresponding to the target simulation environment is generated by the first-order discrete shaping filter. The low-frequency steady-state reference wind speed vector and the random gust wind speed vector are superimposed to obtain the total wind speed vector of the three-dimensional spatial composite wind field. Obtain the current velocity vector of the UAV and calculate the difference between the velocity vector and the total wind speed vector of the three-dimensional spatial composite wind field to obtain the relative wind speed vector.

[0008] Optionally, the end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints, wherein the step of obtaining a three-dimensional direction unit vector, calculating a three-dimensional projection component based on the three-dimensional direction unit vector and the relative wind speed vector, and calculating an air disturbance vector based on the three-dimensional projection component and the three-dimensional direction unit vector specifically includes: Obtain the three-dimensional orientation unit vector of the UAV, and calculate the three-dimensional projection component of the relative wind speed vector on the three-dimensional orientation unit vector; Wherein, the three-dimensional direction unit vector is the three-dimensional direction unit vector of the forward axis, left axis and up axis in the body coordinate system of the UAV in the world coordinate system; The pure quadratic air resistance three-dimensional vector value is calculated based on the three-dimensional directional unit vector and the three-dimensional projection component, and the air disturbance vector experienced by the UAV in a windy environment is calculated based on the pure quadratic air resistance three-dimensional vector value.

[0009] Optionally, the UAV end-to-end flight method for strong wind disturbance and thrust saturation constraint includes the following steps: calculating the physical load based on the air disturbance vector, performing orthogonal decomposition on the physical load to obtain parallel and vertical components, and calculating the ultimate remaining available thrust margin based on the parallel and vertical components. The physical load borne by the UAV at its current position is calculated based on the air disturbance vector; Determine the current target planning direction, and perform orthogonal decomposition on the physical load according to the current target planning direction to obtain parallel and vertical components; Determine the maximum available thrust limit of the UAV, and calculate the limit remaining available thrust margin of the UAV in the current target planning direction based on the maximum available thrust limit, the parallel component, and the vertical component.

[0010] Optionally, the end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints, wherein acquiring discrete depth maps and motion state data, and performing feature encoding processing on the discrete depth maps and motion state data to obtain high-dimensional visual space obstacle avoidance feature encoding and motion feature encoding, specifically includes: A discrete depth map is obtained, and a feature encoding is extracted from the discrete depth map using a preset convolutional neural network to obtain a high-dimensional visual space obstacle avoidance feature encoding. Acquire the motion state data of the UAV, wherein the motion state data includes the current speed, target speed, and current attitude data; The motion state data is projected using a multilayer perceptron to obtain motion feature codes.

[0011] Optionally, the end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints, wherein the initial fusion processing of the air disturbance vector and the limit remaining available thrust margin to obtain a physical perception code, and the alignment and fusion processing of the physical perception code, the high-dimensional visual space obstacle avoidance feature code, and the motion feature code to obtain a comprehensive feature vector, specifically includes: The air disturbance vector and the ultimate remaining available thrust margin are initially fused to obtain the physical feature vector. The physical feature vector is projected through a preset MLP network layer to obtain the physical perception code; The physical perception encoding, the high-dimensional visual space obstacle avoidance feature encoding, and the motion feature encoding are aligned and fused to obtain a comprehensive feature vector.

[0012] Optionally, the UAV end-to-end flight method for strong wind disturbance and thrust saturation constraints, wherein obtaining the temporal hidden state vector based on the comprehensive feature vector and generating the UAV's desired nominal control acceleration command vector based on the temporal hidden state vector specifically includes: Determine the gated recurrent unit network, input the comprehensive feature vector into the gated recurrent unit network, and output the temporal hidden state vector; Determine the fully connected layer, input the temporal hidden state vector into the fully connected layer, and output the expected nominal control acceleration command vector of the UAV.

[0013] Furthermore, to achieve the above objectives, the present invention also provides an end-to-end flight system for unmanned aerial vehicles (UAVs) oriented towards strong wind disturbances and thrust saturation constraints, wherein the end-to-end flight system for UAVs oriented towards strong wind disturbances and thrust saturation constraints includes: The relative wind speed vector calculation module is used to obtain the low-frequency steady-state reference wind speed vector, random gust wind speed vector, and velocity vector of the UAV in the target simulation environment. The low-frequency steady-state reference wind speed vector and the random gust wind speed vector are superimposed to obtain the total wind speed vector of the three-dimensional spatial composite wind field. The relative wind speed vector is calculated based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector. An air disturbance vector calculation module is used to obtain a three-dimensional directional unit vector, calculate a three-dimensional projection component based on the three-dimensional directional unit vector and the relative wind speed vector, and calculate an air disturbance vector based on the three-dimensional projection component and the three-dimensional directional unit vector. The remaining available thrust margin calculation module is used to calculate the physical load based on the air disturbance vector, perform orthogonal decomposition on the physical load to obtain parallel and vertical components, and calculate the ultimate remaining available thrust margin based on the parallel and vertical components. The feature encoding processing module is used to acquire discrete depth maps and motion state data, and to perform feature encoding processing on the discrete depth maps and motion state data to obtain high-dimensional visual space obstacle avoidance feature codes and motion feature codes. The alignment and fusion processing module is used to perform initial fusion processing on the air disturbance vector and the limit remaining available thrust margin to obtain the physical perception code, and to perform alignment and fusion processing on the physical perception code, the high-dimensional visual space obstacle avoidance feature code and the motion feature code to obtain a comprehensive feature vector. An acceleration command vector generation module is used to obtain a temporal hidden state vector based on the comprehensive feature vector, and to generate a UAV expected nominal control acceleration command vector based on the temporal hidden state vector, so as to control the UAV to complete an end-to-end flight mission according to the UAV expected nominal control acceleration command vector.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an end-to-end flight program for UAVs facing strong wind disturbances and thrust saturation constraints stored in the memory and executable on the processor, wherein when the end-to-end flight program for UAVs facing strong wind disturbances and thrust saturation constraints is executed by the processor, the steps of the end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints as described above are implemented.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an end-to-end flight program for a UAV oriented to strong wind disturbance and thrust saturation constraints, and when the end-to-end flight program for a UAV oriented to strong wind disturbance and thrust saturation constraints is executed by a processor, it implements the steps of the end-to-end flight method for a UAV oriented to strong wind disturbance and thrust saturation constraints as described above.

[0016] In this invention, the low-frequency steady-state reference wind speed vector, random gust wind speed vector, and velocity vector of the UAV in the target simulation environment are obtained. The low-frequency steady-state reference wind speed vector and the random gust wind speed vector are superimposed to obtain the total wind speed vector of the three-dimensional spatial composite wind field. The relative wind speed vector is calculated based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector. A three-dimensional directional unit vector is obtained. A three-dimensional projection component is calculated based on the three-dimensional directional unit vector and the relative wind speed vector. An air disturbance vector is calculated based on the three-dimensional projection component and the three-dimensional directional unit vector. The physical load is calculated based on the air disturbance vector. The physical load is orthogonally decomposed to obtain parallel and vertical components. The invention calculates the remaining usable thrust margin; acquires discrete depth maps and motion state data, and performs feature encoding processing on the discrete depth maps and motion state data to obtain high-dimensional visual space obstacle avoidance feature encoding and motion feature encoding; performs initial fusion processing on the air disturbance vector and the remaining usable thrust margin to obtain physical perception encoding, and performs alignment and fusion processing on the physical perception encoding, the high-dimensional visual space obstacle avoidance feature encoding, and the motion feature encoding to obtain a comprehensive feature vector; obtains a temporal hidden state vector based on the comprehensive feature vector, and generates a UAV expected nominal control acceleration command vector based on the temporal hidden state vector to control the UAV to complete the end-to-end flight mission according to the UAV expected nominal control acceleration command vector. This invention introduces the wind field into a differentiable physical simulation environment and inputs air resistance disturbances as explicit physical features into the policy network, endowing the policy network with wind disturbance perception awareness when planning actions. This significantly improves the network convergence stability, obstacle avoidance safety, and physical executability of the UAV end-to-end flight strategy in extreme strong wind environments, ensuring the smooth execution of the end-to-end flight mission and the safety and stability of UAV flight. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the UAV end-to-end flight method for strong wind disturbance and thrust saturation constraint according to the present invention; Figure 2 This is a schematic diagram of the overall process of a preferred embodiment of the UAV end-to-end flight method for strong wind disturbance and thrust saturation constraint of the present invention. Figure 3 This is a structural diagram of a preferred embodiment of the UAV end-to-end flight system for strong wind disturbance and thrust saturation constraints of the present invention; Figure 4 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] In highly dynamic application scenarios such as complex low-altitude operations, urban canyon inspections, forest navigation, and high-speed autonomous obstacle avoidance, rotary-wing UAVs need to maintain efficient and safe autonomous flight under conditions of limited line-of-sight, complex obstacle topology, and strong sudden wind disturbances. Traditional UAV navigation architectures typically employ a hierarchical cascaded design of (path planning - trajectory generation - trajectory tracking control). In this traditional approach, the upper-level path and trajectory planners usually perform geometric optimizations based on a nominal physical model of relatively static air, completely ignoring the severe impact of unsteady aerodynamic disturbances such as gusts, shear winds, and terrain-induced wake turbulence on the actual performance of the UAV. This leads to severe trajectory divergence or even attitude instability during lower-level control and tracking in strong wind environments, as the trajectory generated by the upper-level planner often exceeds the maximum thrust limit of the physical motor.

[0020] In recent years, end-to-end vision-based flight methods based on deep reinforcement learning have gradually become a research hotspot. This method directly maps input visual sensor data (such as depth maps) and motion states to subsequent motion planning commands by constructing a policy network. However, existing end-to-end network flight planning schemes exhibit the following significant drawbacks when facing strong random wind disturbances and inherent thrust saturation constraints in hardware: 1. The drag model is too idealistic and lacks wind disturbance perception capability: Existing end-to-end methods typically assume that air drag is a linear or simple quadratic function with fixed parameters during training, which cannot adapt to sudden changes in relative wind speed under strong random winds. Because the network lacks explicit physical disturbance perception features as input, the strategy can only passively learn indirectly from trajectory errors, resulting in a serious lack of robustness in the actions planned by the network when wind conditions change rapidly.

[0021] 2. Completely ignoring the physical limits of the actuator, resulting in unexecutable planned actions: Most end-to-end vision policy networks treat the network output layer as an unbounded ideal physical quantity during simulation optimization, failing to incorporate the maximum thrust saturation constraint of the actuator (i.e., the upper limit of the motor's physical output) into the computational graph of the network planning. When encountering strong headwinds or crosswinds, in order to complete the obstacle avoidance task, the network is very likely to plan unrealizable acceleration commands that exceed the physical limits of the motor, causing the underlying mechanical collapse of the system.

[0022] 3. Mismatch between traditional disturbance compensation techniques and end-to-end planning: Existing disturbance observers or feedforward compensation control strategies mostly focus on error correction in the lower-level control loop, failing to simultaneously optimize "wind disturbance characteristics" and "thrust margin boundaries" as network decision constraints during the upper-level network training phase. This separation between planning and control results in the policy network lacking awareness of physical boundaries during end-to-end training, leading to slower convergence speed of the policy in strong wind scenarios.

[0023] To address the problems of unstable network training, physically unexecutable planned actions, and insufficient anti-disturbance capability in existing vision-based end-to-end flight planning methods under strong random wind disturbances and actuator thrust saturation constraints, this invention proposes an end-to-end flight method for UAVs under strong wind disturbances and thrust saturation constraints. The core idea of ​​this invention is to introduce the wind field into a differentiable physics simulation environment and innovatively implement a joint mechanism of disturbance feature perception and force constraint boundary in the network's planning decision-making process.

[0024] First, this invention inputs air resistance disturbances as explicit physical features into the policy network, endowing the network with "wind disturbance awareness" when planning actions. Simultaneously, this resistance is used to calculate the "available thrust margin" of the actuator in the target direction under the current wind resistance; this thrust margin is also input as a feature supplement to the network, serving as its "boundary awareness" when planning actions. Second, this invention directly embeds a continuously differentiable force saturation soft barrier penalty function into the computational graph loss function used for network training. This function, by continuously penalizing unrealizable commands output by the policy network that exceed the physical execution capabilities of the motor, forces the network policy to spontaneously learn the optimal action planning for hazard avoidance and speed tracking within the physical thrust boundary during end-to-end optimization.

[0025] The solution of this invention does not require external real wind speed sensor hardware. It relies purely on differentiable physics simulation to construct a loss boundary with clear physical meaning during the training period, which significantly improves the network convergence stability, obstacle avoidance safety and physical feasibility of the end-to-end flight strategy in extreme strong wind environments.

[0026] The preferred embodiment of the present invention describes an end-to-end flight method for unmanned aerial vehicles (UAVs) oriented towards strong wind disturbances and thrust saturation constraints, such as... Figure 1 and Figure 2 As shown, the end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints includes the following steps: Step S10: Obtain the low-frequency steady-state reference wind speed vector, random gust wind speed vector, and velocity vector of the UAV in the target simulation environment. Superimpose the low-frequency steady-state reference wind speed vector and the random gust wind speed vector to obtain the total wind speed vector of the three-dimensional spatial composite wind field. Calculate the relative wind speed vector based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector.

[0027] like Figure 2 As shown, the end-to-end network flight planning method proposed in this invention is generally divided into three major stages: environment setting, constraint optimization, and network planning.

[0028] Specifically, a target simulation environment is constructed, and a low-frequency steady-state reference wind speed vector corresponding to the target simulation environment is generated by the direction vector normalization method, and a random gust wind speed vector corresponding to the target simulation environment is generated by the first-order discrete shaping filter; the low-frequency steady-state reference wind speed vector and the random gust wind speed vector are superimposed to obtain the total wind speed vector of the three-dimensional spatial composite wind field.

[0029] The process of calculating wind field environment settings and air resistance disturbances in this invention is as follows: In the aviation field, the DrydenWind Turbulence Model is commonly used to describe atmospheric turbulence. This model obtains a colored gust sequence with spatial scale, intensity, and correlation time by shaping and filtering random white noise. To simulate real low-altitude operational scenarios, this invention introduces this wind field environment into a differentiable dynamics training simulator.

[0030] The total wind speed (i.e., the total wind speed vector of the three-dimensional spatial composite wind field in this invention) is composed of the superposition of the low-frequency steady-state reference wind and the high-frequency random gust, and its expression is as follows: ; in, : indicates the first At any given moment, the total wind speed vector of the three-dimensional spatial composite wind field in the simulation environment (i.e., the total wind speed vector of the three-dimensional spatial composite wind field in this invention). : Represents the low-frequency steady-state reference wind speed vector generated by normalizing the direction vector and scaling it (i.e., performing vector modulus calculation); : indicates the first The constant reference is to the Dryden project, which generates a time-domain correlated random gust wind speed vector from a first-order discrete shaping filter.

[0031] Obtain the current velocity vector of the UAV and calculate the difference between the velocity vector and the total wind speed vector of the three-dimensional spatial composite wind field to obtain the relative wind speed vector.

[0032] Suppose that the rotary-wing UAV is in the The position vector, velocity vector, and current attitude rotation matrix at each moment are as follows: , , The expression for the relative wind speed vector generated by the relative motion between the drone and the wind field is: ; in, : indicates the first The relative wind speed vector of the drone with respect to the surrounding airflow at any given time.

[0033] Step S20: Obtain the three-dimensional direction unit vector, calculate the three-dimensional projection component based on the three-dimensional direction unit vector and the relative wind speed vector, and calculate the air disturbance vector based on the three-dimensional projection component and the three-dimensional direction unit vector.

[0034] Specifically, the three-dimensional directional unit vector of the UAV is obtained, and the three-dimensional projection component of the relative wind speed vector on the three-dimensional directional unit vector is calculated; wherein, the three-dimensional directional unit vector is the three-dimensional directional unit vector of the forward axis, left axis and up axis in the body coordinate system of the UAV in the world coordinate system; the pure quadratic air resistance three-dimensional vector value is calculated based on the three-dimensional directional unit vector and the three-dimensional projection component, and the air disturbance vector experienced by the UAV in the windy field is calculated based on the pure quadratic air resistance three-dimensional vector value.

[0035] Understandably, in order to accurately simulate nonlinear aerodynamic drag under high speed and strong winds, this invention abandons the ideal linear drag assumption and projects the relative wind speed onto the three axes of the UAV's body to construct a pure quadratic drag model, as shown in the following expression: ; ; in, : These represent the projection components of the relative wind speed vector on the forward, left, and upward axes of the aircraft (i.e., the three-dimensional projection components in this invention). : These represent the three-dimensional direction unit vectors of the front, left, and top axes (i.e., the front axis, left axis, and top axis of the UAV body coordinate system) in the world coordinate system, respectively; : Represents the corresponding three-dimensional vector value of pure secondary air resistance; : Indicates the preset secondary air resistance coefficient.

[0036] Furthermore, based on the aforementioned three-dimensional vector value of pure secondary air resistance, the air disturbance vector experienced by the current UAV in a windy environment is set as... The feature used for subsequent calculation of the maximum thrust margin boundary and injection into the network as a planning feedforward sensing feature is expressed as: .

[0037] Step S30: Calculate the physical load based on the air disturbance vector, perform orthogonal decomposition on the physical load to obtain the parallel component and the vertical component, and calculate the ultimate remaining available thrust margin based on the parallel component and the vertical component.

[0038] Specifically, the physical load borne by the UAV at its current location is calculated based on the air disturbance vector.

[0039] To ensure that the strategy network can know the maximum physical boundary of the current actuator (i.e., the UAV) in real time during planning, this invention derives the thrust margin and saturation loss in the differentiable dynamics diagram. The specific implementation process is as follows: Let the maximum usable thrust allowed by the UAV hardware system be... First, calculate the drone's position at the current point to counteract gravitational acceleration. And the complete environmental drag disturbance (i.e., the air disturbance vector). The physical load that must be borne Its expression is: ; Where m is the weight of the drone.

[0040] The current target planning direction is determined, and the physical load is orthogonally decomposed according to the current target planning direction to obtain parallel and vertical components; the maximum available thrust limit of the UAV is determined, and the limit of the remaining available thrust margin of the UAV in the current target planning direction is calculated according to the maximum available thrust limit, the parallel component and the vertical component.

[0041] It is understood that this invention incorporates this comprehensive environmental load (i.e., physical load). Along the current target planning direction Perform orthogonal decomposition, dividing into parallel components. With vertical component Its expression is: ; ; in, : Represents the projection value of the environmental load onto the flight direction of the UAV target, i.e., the parallel component; : This represents the vertical lateral force load vector, or vertical component, that the actuators of the UAV must cancel out in the lateral direction in order to maintain flight in the target direction.

[0042] After deducting the lateral loads that must be offset, the actuator's maximum remaining available thrust margin along the target planning direction under the current wind field. The expression is: ; Furthermore, during the end-to-end training phase, if the policy network outputs excessive control acceleration in order to forcibly avoid obstacles... This leads to an estimated value of the actual thrust required by the motor. When exceeding the safe physical boundary, this invention utilizes continuously differentiable... The barrier function directly incorporates this constraint into the loss function for gradient backpropagation optimization, as shown in the following expression: ; ; in, : This represents the force saturation constraint loss function term, which has a very small gradient when it is not saturated, and an exponentially exponential gradient when it approaches or exceeds the boundary, thereby constraining the network actions; : Indicates the preset safety margin for saturation protection; : Indicates the total number of samples in a single network training iteration; : Represents a standard continuously differentiable activation barrier function. It is an exponential function.

[0043] Step S40: Obtain discrete depth map and motion state data, and perform feature encoding processing on the discrete depth map and motion state data to obtain high-dimensional visual space obstacle avoidance feature encoding and motion feature encoding.

[0044] Specifically, a discrete depth map is acquired, and feature encoding is extracted from the discrete depth map using a preset convolutional neural network to obtain a high-dimensional visual space obstacle avoidance feature code; motion state data of the UAV is acquired, wherein the motion state data includes the current speed, target speed, and current attitude data; the motion state data is projected using a multilayer perceptron to obtain a motion feature code.

[0045] The end-to-end policy network constructed in this invention aims to explicitly integrate environmental physical features into the action planning flow. The policy network in the first... It continuously receives heterogeneous information from multiple sources and performs parallel feature extraction, including the following information: 1. Input a discrete depth map rendered in real time by the camera. High-dimensional visual space obstacle avoidance feature encoding is extracted using a pre-defined convolutional neural network (CNN). Input kinematic information such as current velocity, target velocity, and current attitude data, and project it onto a multi-layer perceptron (MLP) to encode motion features. .

[0046] Step S50: Initially fuse the air disturbance vector with the remaining available thrust margin to obtain the physical perception code, and then perform alignment and fusion processing on the physical perception code, the high-dimensional visual space obstacle avoidance feature code, and the motion feature code to obtain the comprehensive feature vector.

[0047] Specifically, the air disturbance vector and the remaining available thrust margin are initially fused to obtain a physical feature vector; the physical feature vector is then projected through a preset MLP network layer to obtain a physical sensing code. This invention integrates the calculated wind extra disturbance (i.e., air disturbance vector) and available thrust margin (i.e., ultimate remaining available thrust margin) into a physical characteristic vector. Physical sensing encoding is achieved through independent MLP network layer projection. .

[0048] The physical perception encoding, the high-dimensional visual space obstacle avoidance feature encoding, and the motion feature encoding are aligned and fused to obtain a comprehensive feature vector.

[0049] Multi-source heterogeneous features (i.e., physical perception encoding, high-dimensional visual spatial obstacle avoidance feature encoding, and motion feature encoding) are spatially aligned and fused through addition or concatenation, and then input into a gated recurrent unit (GRU) network to enhance long-term motion memory and output changes based on wind field disturbance trends. The expression is as follows: ; in, : indicates the first The comprehensive feature vector after multi-source feature alignment and fusion at any time.

[0050] Step S60: Obtain the temporal hidden state vector based on the comprehensive feature vector, and generate the UAV expected nominal control acceleration command vector based on the temporal hidden state vector, so as to control the UAV to complete the end-to-end flight mission according to the UAV expected nominal control acceleration command vector.

[0051] Specifically, a gated recurrent unit network is determined, and the comprehensive feature vector is input into the gated recurrent unit network to output a temporal hidden state vector; a fully connected layer is determined, and the temporal hidden state vector is input into the fully connected layer to output the expected nominal control acceleration command vector of the UAV.

[0052] The expression is as follows: ; ; in, : These represent the temporal hidden state vectors of the gated recurrent unit (GRU) at the current time and the previous time, respectively; : Represents the nonlinear temporal evolution function of a standard gated cyclic unit; : This is a fully connected layer, representing a fully connected output layer that maps hidden states to a 3D action space.

[0053] In the training forward propagation and actual online operation of the network, this invention combines the expected instructions of the network planning with feedforward compensation and sends them to the dynamic simulation or controller for execution.

[0054] The key innovations of this invention include: 1. Strategy network planning architecture for joint perception of wind disturbance and thrust boundary: The strategy network in this invention not only receives visual depth maps and motion states, but also proposes to explicitly inject physical features (extra wind disturbance and available thrust margin in direction) calculated by the secondary drag model. This enables the black-box strategy network to perceive the current wind disturbance direction, amplitude and remaining actuator capacity before making obstacle avoidance and path planning actions, thus solving the problem of blindness in action planning in strong wind environments.

[0055] 2. A configurable barrier constraint mechanism for the maximum physical thrust limit: In the computational graph of end-to-end network policy optimization, this invention transforms the absolute physical limitations of the hardware motor into continuously differentiable mathematical penalties by embedding a saturation barrier function, guiding the network to actively plan a physically executable, non-overshooting, smooth optimal flight trajectory.

[0056] In summary, this invention discloses an end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints. It includes an end-to-end visual policy network architecture that jointly injects physical disturbance features and hardware thrust boundary features; a calculation process for calculating available thrust margin by orthogonally projecting the complete drag load onto the target planning direction; and an end-to-end loss function optimization method that uses the softplus function to constrain the network planning output during network gradient training to prevent motor thrust saturation.

[0057] Alternatives or variations of the present invention include: 1. Replacement of disturbance observation and calculation sources: In the simulation stage, the wind-induced disturbance of the present invention is obtained by the true value analysis of the differentiable environment. When the actual aircraft is ported to the airborne system, the wind-induced disturbance characteristics can be obtained by the simulation dynamic residual calculation, or they can be replaced by the real-time wind resistance physical quantity estimated by the airborne high-frequency IMU (Inertial Measurement Unit), electronic speed feedback, micro Pitot tube array, or external disturbance observer (such as high-order Kalman filter) based on multi-sensor fusion.

[0058] 2. Equivalent replacement of network backbone structure: The current neural network backbone used for long-term feature fusion is the CNN-GRU architecture (i.e., a deep learning model that combines convolutional neural networks and gated recurrent units). Depending on the computing power available on a specific platform, its recurrent layers can be equivalently replaced by Temporal Convolutional Networks (TCN), lightweight self-attention mechanism models (Transformer), state-space models (SSM, such as the Mamba architecture), or ordinary recurrent networks with attention mechanisms, etc.

[0059] 3. Replacement of the mathematical expression of the force saturation barrier constraint: The force saturation constraint in the loss function of this invention adopts the softplus function. In actual engineering evolution, this constraint form can be replaced by the hinge loss function, a higher-order quadratic penalty term, or an explicit differentiable model predictive control hard constraint solver layer attached to the end of the differentiable dynamics, all of which can achieve similar gradient penalty effects.

[0060] The design options to avoid this include: 1. Planning avoidance scheme based on direct measurement of hardware array: Competitors may abandon the analytical derivation of air disturbance terms through a pure quadratic drag model in dynamics, and instead attach a miniature ultrasonic anemometer array or multi-directional Pitot tube around the fuselage of the UAV to directly measure the relative wind speed through hardware and substitute it into empirical formulas to calculate the thrust margin. In this way, they may attempt to avoid the technical features of this invention, which rely on the characteristic flow of disturbances that are purely dependent on differentiable dynamic calculation graphs.

[0061] 2. A layered avoidance scheme decoupling planning from underlying strong adaptive control: Competitors may completely omit wind disturbance characteristics and thrust margin characteristics in the end-to-end planning network, allowing the upper-layer network to still output a nominal planning trajectory without boundaries; however, in the underlying control loop, they may employ an extremely aggressive underlying strong disturbance rejection control algorithm to forcibly absorb thrust saturation and wind disturbances. This scheme, which separates physical boundary awareness from the planning layer, is a typical avoidance design.

[0062] In addition to the above, the present invention can be supplemented with the following technical contents based on the solution: 1. Seamless Collaboration Potential for High-Dimensional Semantic Planning Modules: This invention represents a typical low-dimensional physical end-to-end flight planning scheme addressing strong winds and mechanical constraints. In future complex autonomous special operations of robots, this invention can serve as an efficient, low-level safe flight planner, seamlessly integrating with upper-level high-dimensional semantic planning layers such as Visual Language Models (VLM) or Large Language Models (LLM). Due to their large number of parameters, upper-level semantic models typically have high inference latency, only needing to provide a coarse target direction unit vector. In contrast, the end-to-end network flight method at the bottom layer of this invention is responsible for perfectly avoiding strong wind resistance and motor saturation hazards with extremely low latency, thus securing a valuable macroscopic inference time window for the upper-level large model.

[0063] 2. Combination of Cross-Platform Migration and Online System Identification of Drag Coefficient: To enable this invention to possess excellent cross-platform migration and self-evolving capabilities that improve stability over time, a lightweight online physical system identification module can be integrated into the framework in the future. During actual flight, the secondary drag coefficient of a UAV may experience slight drift due to random loads, battery wear, or airframe deformation. By dynamically correcting these physical parameters through online identification, the physical characteristics input to the network can maintain absolute kinematic consistency.

[0064] Furthermore, such as Figure 3 As shown, based on the above-mentioned end-to-end flight method for UAVs accommodating strong wind disturbances and thrust saturation constraints, the present invention also provides a corresponding end-to-end flight system for UAVs accommodating strong wind disturbances and thrust saturation constraints, wherein the end-to-end flight system for UAVs accommodating strong wind disturbances and thrust saturation constraints includes: The relative wind speed vector calculation module 51 is used to obtain the low-frequency steady-state reference wind speed vector, random gust wind speed vector and velocity vector of the UAV in the target simulation environment, and to superimpose the low-frequency steady-state reference wind speed vector and the random gust wind speed vector to obtain the total wind speed vector of the three-dimensional spatial composite wind field, and to calculate the relative wind speed vector based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector. The air disturbance vector calculation module 52 is used to obtain a three-dimensional direction unit vector, calculate a three-dimensional projection component based on the three-dimensional direction unit vector and the relative wind speed vector, and calculate an air disturbance vector based on the three-dimensional projection component and the three-dimensional direction unit vector. The remaining available thrust margin calculation module 53 is used to calculate the physical load based on the air disturbance vector, perform orthogonal decomposition on the physical load to obtain the parallel component and the vertical component, and calculate the ultimate remaining available thrust margin based on the parallel component and the vertical component. The feature encoding processing module 54 is used to acquire discrete depth maps and motion state data, and to perform feature encoding processing on the discrete depth maps and motion state data to obtain high-dimensional visual space obstacle avoidance feature codes and motion feature codes. The alignment and fusion processing module 55 is used to perform initial fusion processing on the air disturbance vector and the limit remaining available thrust margin to obtain the physical perception code, and to perform alignment and fusion processing on the physical perception code, the high-dimensional visual space obstacle avoidance feature code and the motion feature code to obtain a comprehensive feature vector. The acceleration command vector generation module 56 is used to obtain a temporal hidden state vector based on the comprehensive feature vector, and generate a UAV expected nominal control acceleration command vector based on the temporal hidden state vector, so as to control the UAV to complete the end-to-end flight mission according to the UAV expected nominal control acceleration command vector.

[0065] Furthermore, such as Figure 4 As shown, based on the above-mentioned end-to-end flight method and system for UAVs facing strong wind disturbances and thrust saturation constraints, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 4 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0066] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores an end-to-end flight program 40 for UAVs facing strong wind disturbances and thrust saturation constraints. This end-to-end flight program 40 can be executed by the processor 10 to implement the end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints described in this application.

[0067] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the end-to-end flight method of the UAV facing strong wind disturbance and thrust saturation constraints.

[0068] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface.

[0069] In one embodiment, when the processor 10 executes the UAV end-to-end flight program 40 for strong wind disturbance and thrust saturation constraints stored in the memory 20, it implements the steps of the UAV end-to-end flight method for strong wind disturbance and thrust saturation constraints as described above.

[0070] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an end-to-end flight program for a UAV oriented to strong wind disturbance and thrust saturation constraints, and the end-to-end flight program for a UAV oriented to strong wind disturbance and thrust saturation constraints, when executed by a processor, implements the steps of the end-to-end flight method for a UAV oriented to strong wind disturbance and thrust saturation constraints as described above.

[0071] In summary, this invention provides an end-to-end flight method, system, terminal, and storage medium for UAVs facing strong wind disturbances and thrust saturation constraints. The method includes: acquiring a low-frequency steady-state reference wind speed vector, a random gust wind speed vector, and a velocity vector of the UAV in a target simulation environment; superimposing the low-frequency steady-state reference wind speed vector and the random gust wind speed vector to obtain a total wind speed vector of a three-dimensional spatial composite wind field; and calculating a relative wind speed vector based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector; acquiring a three-dimensional directional unit vector; calculating a three-dimensional projection component based on the three-dimensional directional unit vector and the relative wind speed vector; and calculating an air disturbance vector based on the three-dimensional projection component and the three-dimensional directional unit vector; calculating a physical load based on the air disturbance vector; and performing orthogonal decomposition processing on the physical load to obtain a flat... The system calculates the maximum remaining available thrust margin based on the parallel and vertical components; it acquires discrete depth maps and motion state data, and performs feature encoding processing on the discrete depth maps and motion state data to obtain high-dimensional visual space obstacle avoidance feature codes and motion feature codes; it performs initial fusion processing on the air disturbance vector and the maximum remaining available thrust margin to obtain physical perception codes, and performs alignment fusion processing on the physical perception codes, the high-dimensional visual space obstacle avoidance feature codes, and the motion feature codes to obtain a comprehensive feature vector; it obtains a temporal hidden state vector based on the comprehensive feature vector, and generates a UAV expected nominal control acceleration command vector based on the temporal hidden state vector to control the UAV to complete the end-to-end flight mission according to the UAV expected nominal control acceleration command vector. This invention introduces the wind field into a differentiable physical simulation environment and inputs air resistance disturbances as explicit physical features into the strategy network, giving the strategy network wind disturbance perception awareness when planning actions. This significantly improves the network convergence stability, obstacle avoidance safety, and physical feasibility of the UAV end-to-end flight strategy in extreme strong wind environments, ensuring the smooth execution of end-to-end flight missions and the safety and stability of UAV flight.

[0072] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0073] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0074] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An end-to-end flight method for unmanned aerial vehicles (UAVs) accommodating strong wind disturbances and thrust saturation constraints, characterized in that, The end-to-end flight method for UAVs facing strong wind disturbances and thrust saturation constraints includes: The low-frequency steady-state reference wind speed vector, random gust wind speed vector, and velocity vector of the UAV in the target simulation environment are obtained. The low-frequency steady-state reference wind speed vector and the random gust wind speed vector are superimposed to obtain the total wind speed vector of the three-dimensional spatial composite wind field. The relative wind speed vector is calculated based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector. Obtain a three-dimensional unit vector, calculate a three-dimensional projection component based on the three-dimensional unit vector and the relative wind speed vector, and calculate an air disturbance vector based on the three-dimensional projection component and the three-dimensional unit vector, specifically including: Obtain the three-dimensional orientation unit vector of the UAV, and calculate the three-dimensional projection component of the relative wind speed vector on the three-dimensional orientation unit vector; Wherein, the three-dimensional direction unit vector is the three-dimensional direction unit vector of the forward axis, left axis and up axis in the body coordinate system of the UAV in the world coordinate system; The pure quadratic air resistance three-dimensional vector value is calculated based on the three-dimensional directional unit vector and the three-dimensional projection component, and the air disturbance vector experienced by the UAV in the windy area is calculated based on the pure quadratic air resistance three-dimensional vector value. The physical load is calculated based on the air disturbance vector, and the physical load is orthogonally decomposed to obtain parallel and vertical components. The ultimate remaining available thrust margin is calculated based on the parallel and vertical components, specifically including: The physical load borne by the UAV at its current position is calculated based on the air disturbance vector; Determine the current target planning direction, and perform orthogonal decomposition on the physical load according to the current target planning direction to obtain parallel and vertical components; Determine the maximum available thrust limit of the UAV, and calculate the limit remaining available thrust margin of the UAV in the current target planning direction based on the maximum available thrust limit, the parallel component and the vertical component; A discrete depth map and motion state data are acquired, and feature encoding processing is performed on the discrete depth map and the motion state data to obtain high-dimensional visual space obstacle avoidance feature encoding and motion feature encoding. The air disturbance vector and the limit remaining available thrust margin are initially fused to obtain the physical perception code. The physical perception code, the high-dimensional visual space obstacle avoidance feature code, and the motion feature code are then aligned and fused to obtain the comprehensive feature vector. The temporal hidden state vector is obtained based on the comprehensive feature vector, and the expected nominal control acceleration command vector of the UAV is generated based on the temporal hidden state vector, so as to control the UAV to complete the end-to-end flight mission according to the expected nominal control acceleration command vector of the UAV.

2. The end-to-end flight method for UAVs under strong wind disturbance and thrust saturation constraints according to claim 1, characterized in that, The process of acquiring the low-frequency steady-state reference wind speed vector, random gust wind speed vector, and velocity vector of the UAV in the target simulation environment, superimposing the low-frequency steady-state reference wind speed vector and the random gust wind speed vector to obtain the total wind speed vector of the three-dimensional spatial composite wind field, and calculating the relative wind speed vector based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector, specifically includes: A target simulation environment is constructed. The low-frequency steady-state reference wind speed vector corresponding to the target simulation environment is generated by the direction vector normalization method, and the random gust wind speed vector corresponding to the target simulation environment is generated by the first-order discrete shaping filter. The low-frequency steady-state reference wind speed vector and the random gust wind speed vector are superimposed to obtain the total wind speed vector of the three-dimensional spatial composite wind field. Obtain the current velocity vector of the UAV and calculate the difference between the velocity vector and the total wind speed vector of the three-dimensional spatial composite wind field to obtain the relative wind speed vector.

3. The end-to-end flight method for UAVs under strong wind disturbance and thrust saturation constraints according to claim 1, characterized in that, The process of acquiring discrete depth maps and motion state data, and performing feature encoding processing on the discrete depth maps and motion state data to obtain high-dimensional visual space obstacle avoidance feature encoding and motion feature encoding specifically includes: A discrete depth map is obtained, and a feature encoding is extracted from the discrete depth map using a preset convolutional neural network to obtain a high-dimensional visual space obstacle avoidance feature encoding. Acquire the motion state data of the UAV, wherein the motion state data includes the current speed, target speed, and current attitude data; The motion state data is projected using a multilayer perceptron to obtain motion feature codes.

4. The end-to-end flight method for UAVs under strong wind disturbance and thrust saturation constraints according to claim 1, characterized in that, The initial fusion process of the air disturbance vector and the ultimate remaining available thrust margin to obtain a physical perception code, and the alignment and fusion process of the physical perception code, the high-dimensional visual space obstacle avoidance feature code, and the motion feature code to obtain a comprehensive feature vector, specifically includes: The air disturbance vector and the ultimate remaining available thrust margin are initially fused to obtain the physical feature vector. The physical feature vector is projected through a preset MLP network layer to obtain the physical perception code; The physical perception encoding, the high-dimensional visual space obstacle avoidance feature encoding, and the motion feature encoding are aligned and fused to obtain a comprehensive feature vector.

5. The end-to-end flight method for UAVs under strong wind disturbance and thrust saturation constraints according to claim 1, characterized in that, The step of obtaining the temporal hidden state vector based on the comprehensive feature vector, and generating the expected nominal control acceleration command vector for the UAV based on the temporal hidden state vector, specifically includes: Determine the gated recurrent unit network, input the comprehensive feature vector into the gated recurrent unit network, and output the temporal hidden state vector; Determine the fully connected layer, input the temporal hidden state vector into the fully connected layer, and output the expected nominal control acceleration command vector of the UAV.

6. An end-to-end flight system for unmanned aerial vehicles (UAVs) accommodating strong wind disturbances and thrust saturation constraints, characterized in that, The UAV end-to-end flight system oriented towards strong wind disturbance and thrust saturation constraints is used to implement the UAV end-to-end flight method oriented towards strong wind disturbance and thrust saturation constraints as described in any one of claims 1-5, wherein the UAV end-to-end flight system oriented towards strong wind disturbance and thrust saturation constraints includes: The relative wind speed vector calculation module is used to obtain the low-frequency steady-state reference wind speed vector, random gust wind speed vector, and velocity vector of the UAV in the target simulation environment. The low-frequency steady-state reference wind speed vector and the random gust wind speed vector are superimposed to obtain the total wind speed vector of the three-dimensional spatial composite wind field. The relative wind speed vector is calculated based on the total wind speed vector of the three-dimensional spatial composite wind field and the velocity vector. An air disturbance vector calculation module is used to obtain a three-dimensional directional unit vector, calculate a three-dimensional projection component based on the three-dimensional directional unit vector and the relative wind speed vector, and calculate an air disturbance vector based on the three-dimensional projection component and the three-dimensional directional unit vector. The remaining available thrust margin calculation module is used to calculate the physical load based on the air disturbance vector, perform orthogonal decomposition on the physical load to obtain parallel and vertical components, and calculate the ultimate remaining available thrust margin based on the parallel and vertical components. The feature encoding processing module is used to acquire discrete depth maps and motion state data, and to perform feature encoding processing on the discrete depth maps and motion state data to obtain high-dimensional visual space obstacle avoidance feature codes and motion feature codes. The alignment and fusion processing module is used to perform initial fusion processing on the air disturbance vector and the limit remaining available thrust margin to obtain the physical perception code, and to perform alignment and fusion processing on the physical perception code, the high-dimensional visual space obstacle avoidance feature code and the motion feature code to obtain a comprehensive feature vector. An acceleration command vector generation module is used to obtain a temporal hidden state vector based on the comprehensive feature vector, and to generate a UAV expected nominal control acceleration command vector based on the temporal hidden state vector, so as to control the UAV to complete an end-to-end flight mission according to the UAV expected nominal control acceleration command vector.

7. A terminal, characterized in that, The terminal includes: a memory, a processor, and an end-to-end flight program for UAVs oriented to strong wind disturbances and thrust saturation constraints, stored in the memory and executable on the processor. When the end-to-end flight program for UAVs oriented to strong wind disturbances and thrust saturation constraints is executed by the processor, it implements the steps of the end-to-end flight method for UAVs oriented to strong wind disturbances and thrust saturation constraints as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an end-to-end flight program for a drone oriented to strong wind disturbances and thrust saturation constraints, which, when executed by a processor, implements the steps of the end-to-end flight method for a drone oriented to strong wind disturbances and thrust saturation constraints as described in any one of claims 1-5.

Images