Unmanned aerial platform flexible obstacle active crossing method and related device

Abstract

The invention discloses an unmanned flight platform flexible obstacle active crossing method and a related device, and relates to the technical field of unmanned aerial vehicles, and the method comprises the steps: building a flexible obstacle interactive simulation environment based on a physical engine; in a flexible obstacle interactive simulation environment, taking a visual data flow and a state data flow of the unmanned flight platform as input, taking a motor accelerator instruction as output, and training a cascade deep neural network model based on a near-end strategy optimization algorithm; driving a sensor interface through an ROS system to obtain real-time vision and state data, inputting the real-time vision and state data to the trained cascade deep neural network model, and generating a motor control instruction; and determining an active crossing strategy of the unmanned flight platform according to the motor control instruction. The problems of extra load and pneumatic interference caused by a traditional protection device can be avoided.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a method and related apparatus for the active obstacle crossing of an unmanned aerial platform. Background Technology

[0002] Traditional unmanned aerial platforms are primarily designed for operations in open spaces, with their environmental perception systems focusing on conventional obstacle avoidance, utilizing sensors such as LiDAR and cameras for path planning and obstacle avoidance. However, with the increasing demand for complex indoor operations in fields like emergency rescue and facility inspection, the application scenarios of unmanned aerial platforms are expanding to confined environments such as the interiors of high-rise buildings and enclosed spaces. These indoor environments often contain obstacles with special properties, such as doors, windows, nets, and curtains, or semi-dynamic obstacles, posing new challenges that existing obstacle avoidance systems struggle to handle. Unlike rigid obstacles like doors and windows, flexible obstacles such as nets and curtains are flexible and lightweight, requiring no strong interaction with rigid obstacles, and are more susceptible to airflow disturbances generated by the flight platform. Compared to the second-order dynamics exhibited in interactions with rigid obstacles like doors and windows, the interaction between the flight platform and flexible obstacles is far more complex, requiring not only intricate contact interaction modeling but also consideration of the risk of flexible obstacles entanglement in the flight platform's propellers.

[0003] However, existing solutions for addressing the entanglement of flexible obstacles on flight platform propellers suffer from the following problems: 1) Existing anti-entanglement configurations rely on physical barriers to prevent flexible obstacles from contacting the propellers. While improving safety, this leads to increased load resulting in power loss, aerodynamic layout interference, and sensor performance degradation, leading to efficiency losses. 2) Existing end-to-end methods are mainly used for indoor obstacle avoidance, optimizing real-time obstacle avoidance, but they have limitations in flexible obstacle scenarios. 3) There is insufficient consideration of the interaction dynamics of flexible obstacles. Anti-entanglement configurations only provide passive protection and do not address the dynamics of active interaction between the UAV and flexible obstacles. End-to-end methods also lack modeling of the airflow-flexible obstacle coupling effect, which can easily increase the risk of entanglement or system instability. Summary of the Invention

[0004] The purpose of this application is to provide a method and related device for active obstacle crossing of unmanned aerial platforms, which can effectively solve the problem of active obstacle crossing of unmanned aerial platforms when facing flexible obstacles, avoid the problems of additional load and aerodynamic interference caused by traditional protective devices, and improve the operating efficiency and safety in complex indoor environments.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] Firstly, this application provides a method for an unmanned aerial platform to actively traverse flexible obstacles, including:

[0007] S1. Based on a physics engine, construct a flexible obstacle interaction simulation environment;

[0008] S2. In the flexible obstacle interaction simulation environment, using the visual data stream and state data stream of the unmanned aerial platform as input and the motor throttle command as output, a cascaded deep neural network model is trained based on the near-end strategy optimization algorithm; the reward function of the near-end strategy optimization algorithm includes at least the crossing progress reward, high speed reward, propeller stop reward, collision and crash penalty, and entanglement risk penalty; the near-end strategy optimization algorithm is used to collaboratively optimize a complete behavioral strategy including pre-acceleration, active propeller stop crossing, and safe recovery;

[0009] S3. Real-time visual and status data are acquired through the ROS system's sensor interface and input into the trained cascaded deep neural network model to generate motor control commands; the trained cascaded deep neural network model is deployed in the embedded inference engine of the unmanned flight platform.

[0010] S4. Determine the active crossing strategy of the unmanned aerial platform according to the motor control command.

[0011] Optionally, performing step S1 specifically includes:

[0012] Based on the physics engine, the physical properties of flexible obstacles are parametrically modeled, and a set of discrete parameters for flexible obstacles and a dynamic model of material flexibility are defined.

[0013] Based on the physics engine, scene conditions including obstacle spatial configuration, visual interference, and physical disturbance are modeled;

[0014] Based on a physics engine, physical modeling and sensor model adaptation are performed on unmanned aerial platforms and their external devices.

[0015] Optionally, the set of discrete parameters for the flexible obstacle includes the curtain width, mesh diameter, and the coefficient of variation of the mesh spacing; the dynamic model of material flexibility is determined by Young's modulus, Poisson's ratio, and damping coefficient.

[0016] Optionally, the cascaded deep neural network model includes a visual encoding layer and a policy network; the visual encoding layer uses VisionTransformer to extract features from the input RGBD image and outputs a visual feature vector; the policy network uses a multilayer perceptron to fuse the visual feature vector with the state vector of the unmanned aerial platform and outputs a normalized motor throttle command.

[0017] Optionally, the formula for the reward function of the near-end policy optimization algorithm is:

[0018] r t =λ1rnavig +λ2r vel +λ3r stop -λ4r crash -λ5r entangle ;

[0019] Where, r navig =Δd is the progress reward, Δd is the amount of time the platform and the target point are shortened within a single time step; r vel For high-speed rewards; r stop As a reward for stopping the paddle; r crash Penalties for collisions and crashes; entangle λ1 is the entanglement risk penalty; λ2 is the crossing progress reward coefficient; λ3 is the high speed reward coefficient; λ4 is the collision and crash penalty coefficient; λ5 is the entanglement risk penalty coefficient.

[0020] Optionally, the formula for the high-speed reward is:

[0021]

[0022] Among them, I near Let v be an indicator function indicating that the platform approaches a flexible obstacle smaller than the fuselage length, and v is the current flight speed scalar. max denoted as the maximum level flight speed scalar, α as the crossing attenuation coefficient, and κ as the speed sensitivity coefficient.

[0023] Optionally, the formula for the stop-paddle bonus is:

[0024]

[0025] Where β is the stop strength coefficient, ω i This is the normalized throttle command for the motor.

[0026] Optionally, the objective function of the near-end policy optimization algorithm is:

[0027]

[0028] Wherein, KL(θ,θ) old ) represents the KL divergence value between dissimilar strategies, and β1 represents the adaptive weights. Let the initial objective function be... s t ,a t Let θ be the state variables and policy output actions of the Markov process at time t, and θ be the current neural network policy parameters. old For the old policy parameters that have not been updated, J(θ), These are the current loss function and the loss function under the old policy, respectively, π. θ (a t |st ), These represent the new and old policies of the neural network, A(s) t ,a t ) represents the advantage function estimated based on sampling using the old strategy. This represents the expected cumulative return gradient under the old strategy.

[0029] Secondly, this application provides an unmanned aerial platform, comprising:

[0030] External attachments are bolted to the front of the unmanned aerial platform's fuselage.

[0031] The depth camera is mounted on the front of the unmanned aerial platform and located below the external attachments.

[0032] An embedded computing platform equipped with the TensorRT inference engine for running deployed cascaded deep neural network models;

[0033] A flight controller is used to receive motor commands generated by the cascaded deep neural network model and control the unmanned flight platform.

[0034] Thirdly, this application provides an external attachment device for an unmanned aerial platform. The external attachment device adopts an integrated carbon fiber composite material shell. The external attachment device has a streamlined warhead profile when viewed from above, with the front end tapering into a pointed cone-shaped guide section and the rear end gradually expanding to match the nose section of the flight platform. The external attachment device has an asymmetrical airfoil profile when viewed from the side, with the front section being a progressively upward-curving lift guide surface and the rear section transitioning into a horizontally extended platform.

[0035] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0036] This application provides a method and related device for active obstacle crossing of flexible obstacles by an unmanned aerial platform. An interactive simulation environment built using a physics engine accurately simulates the characteristics of flexible obstacles, providing high-fidelity training data for cascaded deep neural network models. The near-end strategy optimization algorithm, through a multi-objective reward function design, significantly improves crossing efficiency while ensuring flight safety. The weighted combination of crossing progress rewards and high-speed rewards enables the platform to dynamically adjust its crossing strategy. The propeller stop reward mechanism effectively reduces the risk of entanglement with flexible obstacles by monitoring the propeller status in real time. The dual constraints of collision and crash penalties and entanglement risk penalties ensure the robustness of the algorithm in complex scenarios. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 A flowchart illustrating an embodiment of this application provides a method for actively traversing flexible obstacles on an unmanned aerial platform.

[0039] Figure 2 This is a schematic diagram of a cascaded deep neural network model structure provided in an embodiment of this application;

[0040] Figure 3 A device structure diagram provided in an embodiment of this application;

[0041] Figure 4 A software package architecture diagram provided for an embodiment of this application;

[0042] Figure 5 A flowchart illustrating the execution of an edge deployment algorithm provided in an embodiment of this application;

[0043] Figure 6 A schematic diagram illustrating the process of an unmanned aerial platform actively traversing flexible obstacles, as provided in an embodiment of this application.

[0044] Figure 7 This is an overall configuration diagram of an unmanned aerial platform provided in an embodiment of this application;

[0045] Figure 8 This is a front view of an unmanned aerial platform provided in an embodiment of this application.

[0046] 1-External attachments to the aircraft, 2-Unmanned flying platform, 3-Depth camera. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] Unlike static obstacles or moving dynamic obstacles in traditional obstacle avoidance scenarios, semi-dynamic obstacles remain static under normal conditions but can be displaced by specific external forces. In typical application scenarios such as emergency rescue, flight platforms often face the need to traverse semi-dynamic obstacles, such as passing through partially open doors and windows or entering confined spaces during post-disaster search and rescue operations. This active traversal requirement differs fundamentally from conventional obstacle avoidance behavior, demanding that the platform possess obstacle passability assessment and interactive traversal capabilities.

[0049] Unlike rigid obstacles such as doors and windows, flexible obstacles such as nets and curtains possess characteristics such as flexibility and lightweight, do not require strong force interactions with corresponding rigid obstacles, and are more susceptible to airflow disturbances generated by the flight platform. Compared to the second-order dynamics exhibited in the interaction process with rigid obstacles such as doors and windows, the interaction process between the flight platform and flexible obstacles is more complex, not only in terms of complex contact interaction modeling, but also in terms of the risk of flexible obstacles becoming entangled in the flight platform's propeller blades.

[0050] To address the problem of flexible obstacles entangled in propeller blades, relevant anti-entanglement configuration designs have been proposed. Utility model patent application CN202223183453.1 proposes a drone with an anti-entanglement structure, which effectively blocks obstacles from entangled by covering the rotor area with anti-entanglement rings and protective nets, ensuring flight safety and preventing damage to the aircraft. Utility model patent application CN202323498520.3 proposes a collision-resistant drone frame, which is installed outside the motor with a protective shell and uses protective nets and threaded mesh covers to form a mesh barrier, preventing leaves and other debris from entangled in the propellers and avoiding drone loss of control. Meanwhile, to address the problem of operating in complex indoor environments on flight platforms, relevant learning-based end-to-end methods have been proposed. The invention patent with application number CN202410129508.5 proposes a high-speed obstacle avoidance system and method for UAVs based on an end-to-end optoelectronic network. It generates depth maps in real time through jointly trained optical hardware and decoders, and plans obstacle avoidance paths with a lightweight decision generator, achieving high-speed autonomous navigation with millisecond-level latency. The invention patent with application number CN202510065720.4 proposes an end-to-end UAV autonomous navigation method based on differential theory. It models the optimal trajectory hybrid distribution through neural networks and extracts safety constraints. It plans the spatiotemporal optimal trajectory by combining differentiable numerical optimization, achieving real-time navigation without mapping.

[0051] However, the above concept has the following problems: (1) Existing anti-entanglement configurations (such as the aforementioned protective net, anti-entanglement ring and protective shell) block flexible obstacles from contacting the blades through physical barriers. Although this improves safety, it introduces multiple efficiency losses such as increased load leading to power loss, aerodynamic layout interference, and sensor performance degradation. (2) Existing end-to-end methods are mainly aimed at indoor obstacle avoidance tasks. Although they optimize the real-time performance of obstacle avoidance, they have fundamental limitations in flexible obstacle scenarios, such as not considering the traversability of flexible obstacles, lack of adaptation to active traversal strategies in action space design, and simulation-real gap. (3) There is a lack of consideration for the interaction dynamics of flexible obstacles. The anti-entanglement configuration only provides passive protection, but does not solve the dynamics problem when the UAV interacts actively with the flexible obstacle. The end-to-end method also lacks modeling of the airflow-flexible obstacle coupling effect, which may lead to increased entanglement risk or system instability.

[0052] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] Example 1

[0054] like Figure 1 As shown, this embodiment provides a method for active obstacle crossing of an unmanned aerial platform, including:

[0055] S1. Based on a physics engine, construct a flexible obstacle interaction simulation environment;

[0056] S2. In the flexible obstacle interaction simulation environment, using the visual data stream and state data stream of the unmanned aerial platform 2 as input and the motor throttle command as output, a cascaded deep neural network model is trained based on the near-end strategy optimization algorithm; the reward function of the near-end strategy optimization algorithm includes at least the crossing progress reward, high speed reward, propeller stop reward, collision and crash penalty, and entanglement risk penalty; the near-end strategy optimization algorithm is used to collaboratively optimize a complete behavioral strategy including pre-acceleration, active propeller stop crossing, and safe recovery;

[0057] S3. Real-time visual and status data are acquired through the ROS system drive sensor interface and input into the trained cascaded deep neural network model to generate motor control commands; the trained cascaded deep neural network model is deployed in the embedded inference engine of the unmanned flight platform 2.

[0058] S4. Determine the active crossing strategy of the unmanned flight platform 2 according to the motor control command.

[0059] In some embodiments, when performing step S1, the specific steps may be as follows:

[0060] Based on the physics engine, the physical properties of flexible obstacles are parametrically modeled, and a set of discrete parameters for flexible obstacles and a dynamic model of material flexibility are defined.

[0061] Based on the physics engine, scene conditions including obstacle spatial configuration, visual interference, and physical disturbance are modeled;

[0062] Based on the physics engine, physical modeling and sensor model adaptation are performed on the unmanned aerial platform 2 and its external devices.

[0063] Specifically, to support the training and validation of end-to-end deep reinforcement learning algorithms for the unmanned aerial platform, a high-fidelity flexible obstacle interaction simulation environment was constructed. The simulation is based on the GarmentLab physics engine framework, integrates multi-scale deformable body dynamics models, and builds a dynamic scene library covering differentiated net curtain structures.

[0064] First, the physical properties of the flexible obstacle are modeled, and the discrete parameter set of the mesh curtain is defined, including the curtain width (e.g., 5-40cm), the mesh diameter (e.g., 1-20cm), and the coefficient of variation of the hole spacing (e.g., 0.1-0.8). Then, a dynamic model of material flexibility is introduced, and the deformation response characteristics of different materials (e.g., nylon / polyester / gauze, etc.) are quantified by Young's modulus (e.g., 1-100MPa), Poisson's ratio (e.g., 0.2-0.45), and damping coefficient (e.g., 0.05-0.3).

[0065] Then, complex scene conditions are modeled, where obstacle spatial configurations can include various semi-dynamic structures such as single / multiple curtains and mesh obstacles. Visual disturbances can introduce parameters such as different colors, textures, and lighting conditions, while physical disturbances can introduce environmental wind fields, mechanical vibrations, and other disturbances. To accelerate the policy training and optimization process, a large-scale parallel training architecture is designed. Based on the PhysX physics backend used in GarmentLab, CUDA cores can be used for acceleration, compressing the computational latency of flexible obstacle dynamics. At the same time, distributed instance deployment is carried out based on the Isaac Sim benchmark platform used in GarmentLab, generating differentiated scenes in parallel and training.

[0066] Similarly, the unmanned aerial platform 2 and its external devices require refined physical modeling and simulation environment adaptation. Custom adjustments are made based on the UAV dynamics model and sensors provided by the OmniDrones framework: After importing the 3D model into the simulation environment, the physical parameters of the flight platform are adapted, adjusting mass distribution, moment of inertia, structural stiffness, and motor dynamics models according to the actual aircraft configuration; a depth camera model is placed on the lower front side of the flight platform fuselage, and parameters such as the camera's field of view and resolution are adjusted to address the sensor occlusion issue caused by the external devices. To achieve the migration from simulation to physical deployment, a multi-stage domain adaptation framework is constructed. Dynamic domain randomization is implemented in the simulation environment, allowing for the introduction of perturbations in the physical parameters of flexible obstacles, such as injecting ±20% mass fluctuations and ±15% Young's modulus variations in the later stages of training; environmental noise is also injected, such as adding Gaussian noise to RGB images and depth map fluctuation errors.

[0067] The established simulation environment was used in conjunction with the subsequently constructed network for training, following a simulation-training-transfer closed loop. During the training phase, the simulation environment generated real-time visual and state data streams as network input through a depth camera model and a dynamics model. It also received throttle commands from the network output to control the constructed flight platform model for active interaction and traversal. Simultaneously, the GarmentLab physics engine provided real-time feedback on collision detection, entanglement risk, and motion state, calculating a reward function to drive the algorithm to update network weights. The physical parameter perturbations and sensor noise injected in the later stages of training enabled the model to learn generalization features across simulation-real-world differences, enhancing its robustness to real-world uncertainties.

[0068] In some embodiments, when performing step S2, the specific steps may be as follows:

[0069] To further achieve an end-to-end mapping closed loop from visual perception to motor commands, a system is constructed as follows: Figure 2 The cascaded deep neural network model shown.

[0070] The model input mainly includes two data streams: one is the visual data stream, which is the normalized, dimension-reduced RGBD image I∈R. H×W×C The resolution H×W is 320×240 pixels, and the number of channels C is 4.

[0071] Another type is the state data stream, which is a 33-dimensional state vector s composed of the state data of the unmanned aerial platform 2 itself. t =[p,p d ,v,q,u t-1:k ],in For the body position, For the desired position, For linear velocity, For attitude quaternions, This refers to the historical k-step motor instructions (k=5).

[0072] In the cascaded deep neural network model, the visual encoding layer uses the Vision Transformer (ViT-S / 16 model) as the visual backbone network. The preprocessing layer segments the image in the visual data stream into 16×16 pixel blocks, resulting in 300×1024 dimensional block data. y∈{0,16,...,224}, through linear projection.

[0073] E i =P i W e +b e ;

[0074] Among them, weight Bias weights After adding positional encoding, the resulting embedding vector is generated as follows:

[0075]

[0076] Where e [CLS] For the [CLS] token vector, pos i Encode the vector for the corresponding position.

[0077] Specifically, the encoding layer contains 12 Transformer modules, each with 6 self-attention heads, and the output is the [CLS] token vector of the last Transformer module. Finally, after passing through a fully connected layer, 512-dimensional visual features are output. The weight matrix Encoding matrix Then, a multilayer perceptron (MLP) is used as the policy network to achieve feature fusion and action decision-making. The feature fusion layer will combine visual features f v With state vector s t Concatenate them into a 545-dimensional joint feature vector f f The policy backbone network is a three-layer fully connected network.

[0078]

[0079] Where i = 1-3 corresponds to hidden layers with 512-256-64 neurons respectively, each layer uses the LeakyReLU activation function, and the final output is a 4-dimensional action vector a. t =[ω1,ω2,ω3,ω4] represents the normalized throttle command mapped to the four motors.

[0080] Then, the network parameters are trained based on the Proximal Policy Optimization (PPO) algorithm. The key design of the training framework includes the reward function r. t=λ1r navig +λ2r vel +λ3r stop -λ4r crash -λ5r entangle .

[0081] Where, r navig =Δd is the progress reward, and Δd is the amount of time the platform and the target point are shortened within a single time step; As a reward for high speed, Let v be an indicator function indicating that the platform approaches a flexible obstacle smaller than the fuselage length, and v is the current flight speed scalar. max α is the maximum level flight speed scalar, α is the crossing attenuation coefficient, which defines the ideal curtain-crossing speed threshold, and κ is the speed sensitivity coefficient, which adjusts the speed sensitivity, i.e., the larger the parameter, the steeper the penalty. For the deceleration bonus, β is the deceleration intensity coefficient, which controls the degree of deceleration; that is, the larger the parameter, the higher the low-throttle bonus. crash For collision and crash penalties, contact detection of non-flexible obstacles in the physics simulation engine is used for judgment; entangle To mitigate the risk of entanglement, the contact time between the flexible obstacle and the fuselage is calculated using a physics simulation engine. A set of coefficients can be set as follows: α = 0.8, κ = 0.6, β = 0.5, λ1 = 0.4, λ2 = 0.5, λ3 = 0.5, λ4 = 1.0, λ5 = 1.0. Through this collaborative optimization, the following traversal strategy can be obtained:

[0082] Specifically, the PPO algorithm training process is as follows:

[0083] The proximal policy optimization algorithm used in training is as follows, based on importance sampling. Modify the gradient of the loss function for the same strategy as before:

[0084]

[0085] Where, π θ (a t |s t ), These are the new and old strategies for the neural network, respectively. For strategy π θ The expected cumulative return gradient under the given conditions, A(s) t ,a t ) is the advantage function estimated based on the old strategy sampling.

[0086] Then by formula The objective function can be deduced from the gradient as follows:

[0087]

[0088] Where st ,a t Let θ be the state variables and policy output actions of the Markov process at time t, and θ be the current neural network policy parameters. old For old policy parameters that have not been updated, This represents the current loss function and the loss function under the old policy. To limit excessive differences in distribution between different policies, a constraint needs to be added to the objective function, namely the KL divergence:

[0089]

[0090] Where KL(θ,θ) old Let θ be the KL divergence value between dissimilar strategies, and β1 be the adaptive weight. Set upper and lower bounds for the KL divergence value; if KL(θ, θ) = 0, then... k KL max The upper limit then increases by β1; if KL(θ,θ) k ) <KL min The lower limit is then reduced by β1. After completing multiple rounds of training policy updates, when the training error is less than the training convergence threshold ∈ conv When the training result accuracy meets the requirements, training is complete; when the training error exceeds the training termination threshold, training is complete. term If the training results diverge, training is terminated and restarted; if the number of training iterations exceeds the specified number of training terminations N, training is terminated again. term If the training results do not meet the requirements, the initial parameters of the tracking control spiking neural network are readjusted and the above process is repeated until training is completed, and finally the policy network parameters are saved.

[0091] In some embodiments, when performing steps S3-S4, the specific steps may be as follows:

[0092] Once the unmanned aerial platform 2 identifies a flexible obstacle ahead using its visual sensors, its active obstacle-crossing strategy involves first entering a pre-acceleration phase, maintaining a speed bonus r. vel Guide the drone to ascend to the preset curtain-penetrating speed to ensure sufficient kinetic energy reserves to cope with subsequent lift loss;

[0093] When the critical distance is reached, the policy network determines the stop reward function r. stop Triggering a motor stop command brings the propellers to a complete stop, mimicking the folding behavior of birds, while also allowing for pre-adjustment of the fuselage attitude to avoid entanglement or collision.

[0094] During the journey, the journey progress reward is r. navig The platform is continuously incentivized to maintain a straight trajectory and efficiently traverse obstacle areas. After successfully navigating the obstacle, the strategy-driven motor quickly restarts and restores stable lift.

[0095] The entire process includes collision and crash penalty items.crash Throughout the entire flight, the system maintains control over flight altitude and attitude safety boundaries to ensure controllable flight even during rotor shutdown and lift loss. The strategy utilizes a phase-adaptive mechanism based on the reward function, enabling the UAV to autonomously complete the entire behavioral chain of acceleration through the obstacle course, active rotor shutdown, and safe recovery.

[0096] Then, perform edge-side model deployment and adaptation, such as... Figure 5 As shown, the trained model is quantized to FP16 precision, the number of parameters is further compressed, and it is adapted to the TensorRT inference engine; (The following is a separate, unrelated sentence: "Writing...") Figure 4 The ROS software package shown acquires the visual and state data streams required for model inference through the ROS platform's sensor driver interface and MAVROS topics. It then inputs the motor commands output by the model into the MAVROS underlying control interface, achieving an end-to-end mapping closed loop from visual perception to motor commands. For heterogeneous embedded platforms on the edge, the container construction pipeline uses Docker container technology to unify and encapsulate runtime environments such as ROS, Python, and TensorRT, achieving environment standardization and rapid deployment, reducing configuration and compatibility issues, improving the overall system efficiency and stability, and supporting flexible migration and expansion across different hardware platforms. Figure 3 As shown, the end-to-end deep reinforcement learning algorithm of the unmanned flight platform 2 is deployed on the flight platform, carrying an Nvidia Orin NX 16GB high-performance embedded platform, including an Arm Cortex-A78AE microprocessor (CPU) and an Ampere graphics processor (GPU), as well as an Intel RealSense D435i depth camera and a Raex Nora+ flight controller.

[0097] This application also provides a scenario of the unmanned aerial platform 2 actively traversing flexible obstacles, such as... Figure 6 As shown, in this scenario, when the platform approaches the flexible obstacle, the depth camera 3 collects environmental information in real time and transmits the visual and state data streams to the trained cascaded deep neural network model via the ROS software package. The model quickly processes this data and outputs precise motor commands to guide the drone in adjusting its flight attitude and speed. During the crossing process, the policy network dynamically optimizes flight behavior based on feedback from the reward function, ensuring that the drone passes through the obstacle area on the optimal trajectory while avoiding collisions and entanglement. Ultimately, the drone successfully crosses the flexible obstacle, verifying the efficiency and robustness of the proposed unmanned aerial platform 2 in complex environments.

[0098] Example 2

[0099] like Figure 7 and Figure 8 As shown, this application provides an unmanned aerial platform, including:

[0100] External attachment 1 is bolted to the front of the fuselage of the unmanned aerial platform 2;

[0101] Depth camera 3 is mounted on the front side of the fuselage of the unmanned aerial platform 2 and located below the external attachment device 1.

[0102] An embedded computing platform equipped with the TensorRT inference engine for running deployed cascaded deep neural network models;

[0103] A flight controller is used to receive motor commands generated by the cascaded deep neural network model and control the unmanned flight platform 2.

[0104] Example 3

[0105] This embodiment provides an external attachment device 1 for an unmanned aerial platform 2. The external attachment device 1 adopts an integrated carbon fiber composite material shell. The overall configuration is designed with hollowing out for weight reduction and aerodynamic optimization, which combines structural strength and low wind resistance.

[0106] The external attachment 1 of the fuselage has a streamlined warhead outline when viewed from above. The front end tapers into a pointed cone-shaped guide section, and the rear end gradually expands to match the cross section of the nose of the flight platform.

[0107] The external attachment 1 of the fuselage has an asymmetrical airfoil profile in side view, with the front section being a gradually upward-curving guide surface and the rear section transitioning into a horizontally extended platform.

[0108] This configuration can adapt to the pitch attitude of the flight platform during forward flight, effectively guiding flexible obstacles (such as net curtains) to slide away from the core area along the curved surface.

[0109] In summary, this application has the following technical effects:

[0110] First, by translating the biological characteristics of birds folding their wings into an engineering-feasible sequence of actions, the system achieves coordinated control of biomimetic behavior and mechanical systems. Combined with the aerodynamically optimized configuration of the lightweight external attachment device 1, it physically guides the flexible obstacle away from the core area, fundamentally avoiding the additional load and aerodynamic interference problems caused by traditional protective devices, significantly improving crossing efficiency and reducing energy consumption. Second, a real-time risk assessment and response mechanism is established, comprehensively optimizing multiple dimensions such as speed control, entanglement avoidance, and flight stability within a millisecond timescale. Relying on a high-fidelity flexible obstacle interaction simulation system and an end-to-end deep reinforcement learning algorithm, a closed-loop optimization from visual perception to action decision-making is achieved, ensuring the accuracy and timeliness of the active crossing strategy. Finally, an environmental adaptive capability is formed. By sensing physical characteristics online and combining the dynamic randomization perturbation of a multi-stage domain adaptation framework, the gap between simulation and reality is effectively bridged, enhancing the model's robustness and adaptability in changing real environments, and comprehensively improving operational safety and crossing success rate. These advantages address the limitations of existing technologies, such as inefficient protective devices, insufficient end-to-end interaction capabilities, and lack of dynamic modeling, providing an efficient and reliable traversal solution for the unmanned aerial platform 2 in complex indoor environments containing flexible obstacles.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for active obstacle crossing of an unmanned aerial platform, characterized in that, include: S1. Based on a physics engine, construct a flexible obstacle interaction simulation environment; S2. In the flexible obstacle interaction simulation environment, using the visual data stream and state data stream of the unmanned aerial platform as input and the motor throttle command as output, a cascaded deep neural network model is trained based on the near-end strategy optimization algorithm; the reward function of the near-end strategy optimization algorithm includes at least the crossing progress reward, high speed reward, propeller stop reward, collision and crash penalty, and entanglement risk penalty; the near-end strategy optimization algorithm is used to collaboratively optimize a complete behavioral strategy including pre-acceleration, active propeller stop crossing, and safe recovery; S3. Real-time visual and status data are acquired through the ROS system's sensor interface and input into the trained cascaded deep neural network model to generate motor control commands; the trained cascaded deep neural network model is deployed in the embedded inference engine of the unmanned flight platform. S4. Determine the active crossing strategy of the unmanned aerial platform according to the motor control command.

2. The method for active obstacle crossing of an unmanned aerial platform according to claim 1, characterized in that, When performing step S1, the specific steps include: Based on the physics engine, the physical properties of flexible obstacles are parametrically modeled, and a set of discrete parameters for flexible obstacles and a dynamic model of material flexibility are defined. Based on the physics engine, scene conditions including obstacle spatial configuration, visual interference, and physical disturbance are modeled; Based on a physics engine, physical modeling and sensor model adaptation are performed on unmanned aerial platforms and their external devices.

3. The method for active obstacle crossing of an unmanned aerial platform according to claim 2, characterized in that, The set of discrete parameters for the flexible obstacle includes the curtain width, mesh diameter, and the coefficient of variation of the mesh spacing; the dynamic model of material flexibility is determined by Young's modulus, Poisson's ratio, and damping coefficient.

4. The method for active obstacle crossing of an unmanned aerial platform according to claim 3, characterized in that, The cascaded deep neural network model includes a visual encoding layer and a policy network. The visual encoding layer uses VisionTransformer to extract features from the input RGBD image and outputs a visual feature vector. The policy network uses a multilayer perceptron to fuse the visual feature vector with the state vector of the unmanned aerial platform and outputs a normalized motor throttle command.

5. The method for active obstacle crossing of an unmanned aerial platform according to claim 4, characterized in that, The formula for the reward function of the near-end policy optimization algorithm is as follows: r t =λ1r navig +λ2r vel +λ3r stop -λ4r crash -λ5r entangle ; Where, r navig =Δd is the progress reward, Δd is the amount of time the platform and the target point are shortened within a single time step; r vel For high-speed rewards; r stop As a reward for stopping the paddle; r crash Penalties for collisions and crashes; entangle λ1 is the entanglement risk penalty; λ2 is the crossing progress reward coefficient; λ3 is the high speed reward coefficient; λ4 is the collision and crash penalty coefficient; λ5 is the entanglement risk penalty coefficient.

6. The method for active obstacle crossing of an unmanned aerial platform according to claim 5, characterized in that, The formula for the high-speed reward is as follows: Among them, I near Let v be an indicator function indicating that the platform approaches a flexible obstacle smaller than the fuselage length, and v is the current flight speed scalar. max denoted as the maximum level flight speed scalar, α as the crossing attenuation coefficient, and κ as the speed sensitivity coefficient.

7. The method for active obstacle crossing of an unmanned aerial platform according to claim 6, characterized in that, The formula for the stop-row bonus is: Where β is the stop strength coefficient, ω i This is the normalized throttle command for the motor.

8. The method for active obstacle crossing of an unmanned aerial platform according to claim 7, characterized in that, The objective function of the near-end strategy optimization algorithm is: Wherein, KL(θ,θ) old ) represents the KL divergence value between dissimilar strategies, and β1 represents the adaptive weights. Let the initial objective function be... s t ,a t Let θ be the state variables and policy output actions of the Markov process at time t, and θ be the current neural network policy parameters. old For old policy parameters that have not been updated, These are the current loss function and the loss function under the old policy, respectively, π. θ (a t |s t ), These represent the new and old policies of the neural network, A(s) t ,a t ) represents the advantage function estimated based on sampling using the old strategy. This represents the expected cumulative return gradient under the old strategy.

9. An unmanned aerial platform, characterized in that, include: External attachments are bolted to the front of the unmanned aerial platform's fuselage. The depth camera is mounted on the front of the unmanned aerial platform and located below the external attachments. An embedded computing platform equipped with the TensorRT inference engine for running deployed cascaded deep neural network models; A flight controller is used to receive motor commands generated by the cascaded deep neural network model and control the unmanned flight platform.

10. An external attachment device for an unmanned aerial platform, characterized in that, The external attachments are made of a single piece of carbon fiber composite material. The external attachments, when viewed from above, have a streamlined warhead profile, tapering at the front end to a pointed cone-shaped guide section and gradually widening at the rear end to match the nose section of the flight platform. The external attachments, when viewed from the side, have an asymmetrical airfoil profile, with the front section being a progressively upward-curving lift guide surface and the rear section transitioning into a horizontally extended platform.

Images