An air trimming method and device based on quadratic programming end control

Abstract

The application provides an air cutting method and device based on quadratic programming end control. The method provided by the application comprises the following steps: constructing a bidirectional cognitive task planning architecture, wherein the bidirectional cognitive task planning architecture comprises a slow thinking layer and a fast thinking layer; extracting an unmanned aerial vehicle air cutting skill through the slow thinking layer and generating a task path point; converting the task path point into a Bezier curve flight trajectory containing a convex polyhedron safety channel through the fast thinking layer, controlling the unmanned aerial vehicle to move along the Bezier curve flight trajectory to above a cutting target; based on a cutting target image fed back by an unmanned aerial vehicle on-board camera in real time, issuing a mechanical arm end speed instruction through a force feedback hand controller, integrating the speed instruction to obtain a position instruction, combining a quadratic programming model containing an unmanned aerial vehicle rotation matrix and angular velocity compensation to solve control parameters, driving a mechanical arm to carry an on-board scissors tool to a cutting point, and controlling the on-board scissors tool to open and close to perform air cutting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aerial pruning, and in particular to an aerial pruning method and device based on quadratic programming end control. BACKGROUND

[0002] Aerial pruning operations have vital application requirements in many modern industries. For example, in garden maintenance, the branches of tall trees need to be pruned; in power inspection, the creeping plants that affect the safety of power transmission lines need to be cleaned up; in emergency rescue, it may be necessary to quickly cut off obstacles to open up rescue routes. The common feature of these operation scenarios is that the target position is in the high altitude or dangerous and complex environment that personnel cannot directly reach. Using manual methods for operation not only has low efficiency, but also poses a great threat to the personal safety of the operation personnel. Therefore, it is of clear practical necessity and broad application prospect to develop automated and intelligent aerial pruning technology to replace manual labor with machines to perform such high-risk tasks.

[0003] Currently, the mainstream technical solution for aerial pruning is to use a drone (an aerial operation robot platform) as a carrying platform, install a mechanical arm at the bottom of the drone, and install a pair of scissors driven by a motor at the end of the mechanical arm. This way liberates the operator from the dangerous high-altitude operation environment, realizes remote operation, and is a common practice under current technical conditions. However, the above method has several significant defects. First, its task planning capability is insufficient, and it relies heavily on the real-time judgment and operation of the operator, with low intelligence level, and in the face of complex tasks and environments, the operator is required to be high and prone to errors. Second, in terms of control, the traditional control method is difficult to overcome the instability of the drone itself and the influence of external disturbances (such as wind and airflow) when the drone and the mechanical arm work together, resulting in poor positioning accuracy of the end of the mechanical arm, and the mechanical arm cannot maintain accurate and stable relative position with the pruning target in a dynamic environment, directly affecting the pruning effect and operation safety.

[0004] Therefore, there is an urgent need for a method that, through intelligent task planning and quadratic programming end collaborative control, realizes precise, stable, and autonomous automated pruning operations of high-altitude targets by the drone in complex environments. SUMMARY

[0005] Therefore, the present application provides an aerial pruning method and device based on quadratic programming end control, which realizes precise, stable, and autonomous automated pruning operations of high-altitude targets by the drone in complex environments through intelligent task planning and quadratic programming end collaborative control.

[0006] Specifically, the present application is realized by the following technical solutions:

[0007] The first aspect of the application provides an aerial cutting method based on quadratic programming end control, the method comprising:

[0008] A bidirectional cognitive task planning architecture is constructed, which comprises a slow thinking layer and a fast thinking layer;

[0009] The slow thinking layer extracts the aerial cutting skills of the UAV and generates task path points;

[0010] The fast thinking layer converts the task path points into a Bezier curve flight trajectory containing a convex polyhedron safety channel, controls the UAV to move along the Bezier curve flight trajectory to directly above the cutting target;

[0011] Based on the real-time feedback of the cutting target image of the UAV on-board camera, the mechanical arm end speed command is issued through the force feedback hand controller, the position command is obtained by integrating the speed command, and the control parameters are solved by combining the quadratic programming model containing the UAV rotation matrix and angular velocity compensation, to drive the mechanical arm to carry the on-board scissors tool to the cutting point, and control the on-board scissors tool to open and close for aerial cutting.

[0012] The second aspect of the application provides an aerial cutting device based on quadratic programming end control, the device comprising a construction module, a generation module and a control module;

[0013] The construction module is used to construct a bidirectional cognitive task planning architecture, which comprises a slow thinking layer and a fast thinking layer;

[0014] The generation module is used to extract the aerial cutting skills of the UAV through the slow thinking layer and generate task path points;

[0015] The control module is used to convert the task path points into a Bezier curve flight trajectory containing a convex polyhedron safety channel through the fast thinking layer, and control the UAV to move along the Bezier curve flight trajectory to directly above the cutting target;

[0016] The control module is also used to control the UAV to move along the Bezier curve flight trajectory to directly above the cutting target based on the real-time feedback of the cutting target image of the UAV on-board camera, issue the mechanical arm end speed command through the force feedback hand controller, obtain the position command by integrating the speed command, and solve the control parameters by combining the quadratic programming model containing the UAV rotation matrix and angular velocity compensation, to drive the mechanical arm to carry the on-board scissors tool to the cutting point, and control the on-board scissors tool to open and close for aerial cutting.

[0017] The air cutting method and device based on quadratic programming end control provided in the application, through the division and cooperation of the slow thinking layer and the fast thinking layer of the bidirectional cognitive task planning architecture, the slow thinking layer first extracts an air cutting skill and generates a task path point based on global planning to ensure that a path covers key nodes in the whole process from taking off to cutting and then evacuating, thereby providing a macro framework for subsequent operations; then the fast thinking layer converts discrete path points into a Bezier curve flight trajectory containing a convex polyhedron safety channel, which not only guarantees a safe distance between the unmanned aerial vehicle and obstacles during flight through the convex polyhedron safety channel, but also matches the dynamics of the unmanned aerial vehicle by means of the smoothness of the Bezier curve, thereby realizing stable movement; finally, the quadratic programming model fuses real-time states such as the rotation matrix and angular velocity compensation of the unmanned aerial vehicle, takes the position command obtained by integrating the end speed command as a target, and solves optimal control parameters to drive the mechanical arm to accurately position, and the human-computer interaction of the force feedback hand controller ensures the flexibility and precision of the cutting operation. This closely linked step design forms a closed loop from macro planning to micro control, the global planning of the slow thinking layer guarantees the integrity and safety of the task, the trajectory generation of the fast thinking layer realizes the smoothness and environmental adaptability of the path, and the real-time solving of the quadratic programming model solves the accurate control problem of the mechanical arm under the dynamic attitude of the unmanned aerial vehicle, and the three cooperate to not only improve the automation level of the air cutting, but also reduce the planning and control difficulty in a complex environment through layered processing, and finally realize efficient, safe and accurate air cutting operation. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of the air cutting method based on quadratic programming end control provided for Embodiment One of the application;

[0019] Figure 2 A structural schematic diagram of the air cutting device based on quadratic programming end control provided for Embodiment Two of the application. DETAILED DESCRIPTION

[0020] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, unless otherwise indicated, like numbers in the attached drawings refer to the same or similar elements. The embodiments described in the following exemplary embodiments are not meant to represent all embodiments in accordance with the application.

[0021] The terms used in the present application are merely for the purpose of describing particular embodiments and are not intended to limit the present application. As used in the present application, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "and / or," as used herein, refer to and encompass any or all possible combinations of one or more of the associated listed items.

[0022] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0023] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0024] Figure 1 The flowchart illustrates the aerial clipping method based on quadratic programming end-point control provided in Embodiment 1 of this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0025] S101. Construct a bidirectional cognitive task planning architecture.

[0026] The bidirectional cognitive task planning architecture includes a slow thinking layer and a fast thinking layer.

[0027] Specifically, the bidirectional cognitive task planning architecture is a layered collaborative task planning system adapted to UAV aerial trimming scenarios, consisting of a slow thinking layer and a fast thinking layer. "Bidirectional" refers to the closed-loop linkage between the two layers, formed by top-down task decomposition and bottom-up constraint feedback. Specifically, the slow thinking layer, based on the overall task requirements of aerial trimming (such as trimming targets and environmental information), completes macro-level skill extraction, task decomposition, and pathpoint generation, and passes the planning results (pathpoints) as input to the fast thinking layer, guiding it to construct micro-trajectories. During the process of converting pathpoints into flight trajectories, the fast thinking layer verifies the planning results of the slow thinking layer in conjunction with real-time environmental constraints (such as obstacle positions and UAV dynamics limitations). If a safety risk is found in a pathpoint (such as obstacles nearby or failure to meet trajectory fitting conditions), feedback is sent back to the slow thinking layer, triggering it to adjust and optimize the pathpoint, forming a bidirectional closed loop of planning-verification-optimization.

[0028] It should be noted that, due to the significant characteristics of the aerial pruning scene, such as complex environment, high precision requirement and coherent task flow, traditional single planning layer (only macro planning or only micro execution) cannot meet the demand. In traditional aerial operation planning, macro path planning (such as only generating target points) and micro trajectory execution (such as only fitting paths) are separated: if only relying on the slow thinking layer to do macro planning, the generated path points may ignore the unmanned aerial vehicle dynamics limitations (such as flight speed, mechanical arm movement range) or real-time environmental obstacles, resulting in subsequent trajectory unable to be executed; if only relying on the fast thinking layer to do micro trajectory fitting, lacking of macro task logic guidance (such as pruning priority, operation safety specification), may appear “trajectory feasible but pruning task fails” (such as missing the optimal pruning point, colliding with non-target branches). The bidirectional architecture ensures that trajectory fitting always revolves around the pruning task target through the forward transfer of “slow thinking layer to determine direction, fast thinking layer to do execution”, and ensures that macro planning conforms to actual execution conditions through the reverse linkage of “fast thinking layer to feedback constraints, slow thinking layer to adjust planning”, avoiding disconnection between the two layers.

[0029] S102, extracting an unmanned aerial vehicle aerial pruning skill through the slow thinking layer and generating a task path point.

[0030] Specifically, the unmanned aerial vehicle aerial pruning skill refers to a series of scene-specific operation capabilities and rules necessary for unmanned aerial vehicles to achieve precise, safe and efficient pruning in aerial pruning operations, which is an abstract refinement of aerial pruning operation experience, operation specification and equipment characteristics. Specifically, it includes: motion control skills: such as hovering accuracy control in pruning scenes (need to maintain millimeter-level position stability to avoid cutting deviation due to shaking), minimum flight distance control within the operation radius of the on-board scissors (to ensure the safe distance of the unmanned aerial vehicle from the pruning target and obstacles); environmental adaptation skills: such as emergency obstacle avoidance skills in pruning areas (rapid response strategies for complex obstacles such as tree branches and power lines), attitude stability skills under air flow disturbance (to resist the influence of rotor air flow on pruning accuracy); tool coordination skills: such as linkage control skills of the mechanical arm and the on-board scissors (to ensure precise matching of the opening and closing time of the scissors and the position of the unmanned aerial vehicle), shear force feedback adjustment skills (to adjust the output force of the mechanical arm according to the thickness of the branches to avoid tool damage or incomplete cutting).

[0031] Further, the task path point refers to a series of discrete key spatial coordinates generated by the slow thinking layer based on the aerial pruning task requirements, which is a key node set for the unmanned aerial vehicle to fly from the take-off point to the pruning target area and then to the evacuation point after the operation is completed. Each task path point contains three-dimensional spatial coordinates (X, Y, Z) and corresponding time stamps, specifically including: start and evacuation points (safe coordinates for the unmanned aerial vehicle to take off and return after the operation is completed); transition points (transit coordinates to avoid obstacles during flight); pruning target points (operation coordinates accurately corresponding to the pruning objects).

[0032] In a specific implementation, an aerial cutting task description file is constructed; the description file is semantically parsed to extract a UAV aerial cutting skill; the UAV aerial cutting skill is feature abstracted to form a task constraint set and a skill set, an application program interface for adapting to an aerial cutting task is constructed based on the skill set; a point cloud map is read through the application program interface, the point cloud map is converted into structured data, a cutting scene task planning test set is constructed based on the structured data and cutting task requirements, the test set contains multiple sets of cutting task input conditions and corresponding expert planning results; the test set is input to a task planning operation unit, planning logic of an iterative training optimization unit is optimized, unit internal parameters are adjusted using feedback data of the expert planning results to update planning reasoning capability; a cutting task description is input to the optimized task planning operation unit, combined with real-time map structured data transmitted by the application program interface, a task execution step containing cutting target coordinate information and a corresponding flight path point are output through unit internal planning logic operation.

[0033] Specifically, according to the aerial pruning task and the existing unmanned aerial vehicle operation manual, a description file for the aerial pruning task is written, which contains key information such as operation flow, safety regulations, and pruning tool instructions during task execution. The description file is input into the large language model, and the large language model is guided to extract aerial pruning skills through prompt, such as: "Identify the core operation capabilities required for the unmanned aerial vehicle from the task description, including position control accuracy, obstacle avoidance response speed, and tool coordination logic, etc.", and the large language model outputs quantitative skill rules (such as "hovering error ≤ 3 mm" and "adjusting path within 1 second when encountering obstacles"). The large language model abstracts the extracted skills to generate a task constraint set (such as "distance from obstacles ≥ 0.5 meters") and a skill set (such as "positioning → adjusting attitude → starting pruning" flow); based on the skill set, the large language model generates application program interfaces (APIs) (such as call format, input and output parameters) to adapt to the function call requirements of the aerial pruning task. After reading the point cloud map data through the API, the large language model combines the output of the point cloud analysis tool (such as PCL) to convert the three-dimensional point cloud into structured data (such as "obstacle: power line, coordinates (x, y, z), length 5 meters" and "target branch: coordinates (a, b, c), diameter 5 cm"), forming textual environmental information that the large language model can understand. Input diverse pruning scene descriptions (such as "dense forest environment multi-branch pruning" and "single branch pruning under strong wind") into the large language model to guide the model to generate multiple input conditions; at the same time, input expert planning cases to let the large language model mimic expert logic to generate corresponding path point results, and combine them to form a "input condition-expert result" test set. Input the test set into the large language model, and use the fine-tuning ability of the large language model to generate the planning logic of the unit with the expert results as supervision data, and iteratively optimize the planning logic. Analyze the deviation between the output of the large language model and the expert results, and automatically adjust the internal parameters (such as path weight, obstacle avoidance priority). Input the specific aerial pruning task description (natural language or structured text) into the optimized large language model, and the LLM combines the real-time map structured data (textual environmental information) transmitted by the API to call the internal planning logic, and outputs the task steps in natural language or coordinate format (such as "take off → obstacle avoidance point 1 → pruning point (x, y, z) → evacuation point") and the corresponding flight path points. The specific implementation process of the large language model extracting unmanned aerial vehicle aerial pruning skills and generating task path points can be referred to the description in related technologies, which will not be repeated here.

[0034] Optionally, after the output containing the task execution step of cutting the target coordinate information and the corresponding flight path point, the method further comprises: obtaining the shape structure parameters of the unmanned aerial vehicle, converting the overall shape of the unmanned aerial vehicle into a three-dimensional convex polyhedron model for collision detection using a fast convex hull generation algorithm; calling the structured data from the application program interface, extracting the obstacle feature information, and arranging the three-dimensional convex polyhedron model, the task execution step and the corresponding flight path point together into a collision detection data set; inputting the collision detection data set and the collision detection instruction into the large language model, simulating the whole process of the unmanned aerial vehicle performing the cutting task along the flight path point based on the obstacle feature information, the three-dimensional convex polyhedron model and the flight path point, checking whether the unmanned aerial vehicle as a whole collides with the obstacle; the large language model analyzes the relative relationship between the spatial position of the unmanned aerial vehicle moving along each flight path point and the obstacle based on the input collision detection data set and collision detection instruction, judges whether there is a collision risk, and generates a collision detection result; when the collision detection result is collision, the collision path segment information and the obstacle feature information are fed back to the large language model, and the large language model re-plans the flight path point combined with the task constraint set and the skill set until the collision detection result is no collision.

[0035] In a specific implementation, the shape structure parameters of the UAV are obtained, including the dimensions (length, width, height), contour shapes (such as the cylindrical section of the fuselage, the connecting rod structure of the mechanical arm) and other data of the UAV body, mechanical arm and on-board scissors tool; using the QuickHull algorithm, the parameters are input to convert the overall shape of the UAV into a three-dimensional convex polyhedral model (composed of multiple triangular facets or convex polygonal faces, covering the spatial contours of all components of the UAV) for collision detection. The specific implementation process of the QuickHull algorithm can be referred to the description in the related art, which will not be repeated here. The structured data is called from the application programming interface to extract obstacle feature information (such as the three-dimensional coordinate range of the tree, the plane position and height of the building wall, the spatial direction of the power line, etc.); the obstacle feature information is sorted together with the three-dimensional convex polyhedral model, the task execution steps (such as the action sequence of taking off, hovering, moving to the point directly above the cutting point) and the corresponding flight path points (the three-dimensional coordinates of each path point, the time stamp) into a collision detection data set. The collision detection data set is input into the large language model, and the collision detection instruction (the instruction content is: based on the provided obstacle feature information, the three-dimensional convex polyhedral model of the UAV and the flight path points, simulate the whole process of the UAV performing the cutting operation along the path points, check the spatial overlap between the UAV (including the body, the mechanical arm and the scissors tool) and the obstacles (trees, walls, power lines, etc.) in each segment, and judge whether there is a collision risk”) is input at the same time; based on the input data set and instruction, the large language model simulates the dynamic process of the UAV moving along each flight path point, and analyzes the relative relationship between the spatial position of each path point and path segment and the obstacles. The large language model judges whether there is a collision risk through spatial geometric operations (such as convex polyhedron intersection detection, distance calculation): if the convex polyhedral model of the UAV overlaps with the spatial area of a certain obstacle, it is determined that there is a collision, and the specific collision path segment (such as the interval from path point A to path point B) and the corresponding obstacle type (such as trees) are marked; if there is no overlap throughout the process, it is determined that there is no collision, and the collision detection result is finally generated. If the collision detection result is that there is a collision, the collision path segment information (such as the starting / ending path point coordinates, the collision interval length), the obstacle feature information (such as the spatial range of the obstacle, the material hardness) are fed back to the large language model; the large language model combines the task constraint set (such as the cutting accuracy requirement, the flight time limit) and the skill set (such as the UAV obstacle avoidance skill, the path flying strategy) to re-plan the flight path points (such as adjusting the path point coordinates to avoid obstacles, adding hovering points to change the motion timing); then the detection step is repeated until the collision detection result is no collision. If the collision detection result is no collision, the task execution steps and the corresponding flight path points at this time are output.

[0036] S103, convert the task path points into a Bézier curve flight trajectory of a convex polyhedron safety channel through the fast thinking layer, control the UAV to move to the cutting target directly above along the Bézier curve flight trajectory.

[0037] Specifically, the convex polyhedron safety channel is a three-dimensional safety area constructed around the UAV flight trajectory, taking a convex polyhedron (such as a cuboid, a prism, etc.) as a spatial form, for guaranteeing the safety distance of the UAV during flight from obstacles. The Bézier curve flight trajectory is a continuous and smooth flight path converted from discrete task path points, generated based on a Bézier curve.

[0038] In a specific implementation, the conversion of the task path points into a Bézier curve flight trajectory of a convex polyhedron safety channel through the fast thinking layer includes:

[0039] (1) receive the task path points output by the slow thinking layer and the structured data transmitted through the application program interface, and extract the obstacle position information from the structured data.

[0040] Specifically, the fast thinking layer receives the task path points output by the slow thinking layer through a preset data interface, and simultaneously calls an application program interface adapted to the aerial cutting task to read the structured data (containing obstacle ID, three-dimensional coordinates, size, and target branch information) transmitted in real time. The structured data is field parsed, and the three-dimensional spatial coordinate information (such as the X, Y, and Z coordinates of tree branches and power lines) of all obstacles is filtered and extracted to form an obstacle position list.

[0041] (2) sequentially connect the task path points to form an initial convex polyhedron, take the initial convex polyhedron as a reference, gradually expand the volume along the direction of the obstacles until the convex polyhedron boundary contacts the obstacle surface, and splice the multiple convex polyhedrons after continuous expansion to form a safety channel flyable by the UAV.

[0042] Specifically, read the task path points output by the slow thinking layer, connect adjacent task path points in pairs in order, generate an initial convex polyhedron (such as a cuboid) with the center axis of the connection line and the size (such as length, width and height) of the UAV body, and ensure that the UAV has no collision risk within the initial range. Call the obstacle position information in the structured data to determine the orientation of each obstacle relative to the initial convex polyhedron (such as left, above, front, etc.). Expand the volume of the initial convex polyhedron in the direction of the obstacle (such as increasing the boundary distance on the side of the obstacle), and check the distance between the convex polyhedron boundary and the obstacle surface after each expansion. When the convex polyhedron boundary contacts (distance is 0) a certain obstacle surface, stop expanding in that direction and record the current convex polyhedron shape. Perform the above expansion operation on the initial convex polyhedron corresponding to each task path point connection line, and concatenate the expanded multiple convex polyhedrons in the order of the path points to form a continuous three-dimensional space channel, i.e. a safe channel for the UAV to fly.

[0043] (3) Fit the flight path in the safe channel based on the multi-segment Bezier curve to obtain a Bezier curve flight trajectory.

[0044] In specific implementation, based on the spatial range of the safe channel, set the segmentation nodes of the multi-segment Bezier curve, which correspond one-to-one to the inflection points of the safe channel; configure a control point vector for each segment of the Bezier curve, construct a convex constraint condition in combination with the maximum flight speed of the UAV and the dynamics parameters of the mechanical arm movement acceleration; substitute the convex constraint condition into the Bezier curve parameter solving process, determine the order, time scaling factor and basis vector parameters of each segment of the Bezier curve through convex optimization operation, fit to form a smooth flight trajectory that adapts to the safe channel; verify whether the flight trajectory is completely within the convex polyhedron safe channel and meets the dynamics constraint requirements, if not, adjust the control point vector and the constraint parameters, and re-execute the fitting operation until the flight trajectory meets the requirements.

[0045] Specifically, the three-dimensional coordinate data of the convex polyhedron safety passage is traversed to identify spatial locations where the passage direction changes (i.e., inflection points, such as turning points at the junction of adjacent convex polyhedra). The three-dimensional coordinates of these inflection points are set as segment nodes of multiple Bézier curves. For each Bézier curve segment, its two corresponding segment nodes are used as the start and end points. Two to three intermediate control points are added inside the safety passage between the start and end points to form the control point vector of the curve segment (containing the X / Y / Z coordinates of the start, intermediate control points, and end point). The positions of the intermediate control points must avoid local narrow areas within the passage. Dynamic parameters are read from the UAV flight control system and robotic arm controller, including the UAV's maximum flight speed (e.g., ≤5m / s), maximum acceleration (e.g., ≤2m / s²), and maximum motion acceleration of the robotic arm joints (e.g., ≤1rad / s²). These parameters are converted into convex constraint conditions (e.g., the velocity derivative of each point on the curve ≤ maximum acceleration, and the magnitude of the velocity vector ≤ maximum flight speed). Based on the parametric equations of Bézier curves, the convex constraints are transformed into a system of inequalities concerning the curve order, the time scaling factor (which controls the stretching ratio of the curve on the time axis), and the basis vector parameters. Convex optimization is then performed using the interior point method to solve for the parameter values ​​that satisfy the constraints, and each smooth curve segment is fitted. The fitted complete flight trajectory is then verified point-by-point: densely sampled points on the trajectory (one point every 0.1 seconds) are substituted into the boundary equations of the convex polyhedral safety channel to check if all points are inside the channel (without exceeding any boundary plane); the velocity and acceleration values ​​of each point on the trajectory are calculated to check if they meet the preset maximum threshold; if any verification fails, the coordinates of the intermediate control points are adjusted (e.g., shifted towards the center of the channel) or the time scaling factor is relaxed (the flight time of that curve segment is extended), and the parameters are resubmitted into the convex optimization operation to solve for the parameters. The fitting and verification steps are repeated until the trajectory simultaneously satisfies both spatial and dynamic constraints.

[0046] For example, in one embodiment, the flight trajectory is The segment of the Bézier curve. Let For the flight trajectory, then:

[0047] ;

[0048] in, ; It is a scaling factor that transforms the time domain from [0, 1] to... ; Let be the order of the Bézier curve; It is the first Control point vector of a segment of a Bézier curve; It is the first Segment Bessel basis vectors, and .

[0049] Further, the unmanned aerial vehicle flight control system receives the fitted Bezier curve flight trajectory data, and analyzes the target three-dimensional coordinates (X, Y, Z) and speed instructions corresponding to each time on the trajectory. The flight control system collects the current position and attitude information of the unmanned aerial vehicle in real time, calculates the deviation between the current position and the target position at the corresponding time on the trajectory. Based on the deviation value, the output instructions of each power unit (such as the rotor) of the unmanned aerial vehicle are generated through the PID control algorithm, and the rotor speed is adjusted to correct the position deviation, so that the unmanned aerial vehicle moves smoothly along the trajectory. In the flight process, the current position is compared with the coordinate directly above the cutting target in real time, and when the unmanned aerial vehicle enters the preset error range (such as ±5cm) of the coordinate directly above the cutting target, the hovering instruction is triggered. After hovering, the position is calibrated again through high-precision positioning to ensure that the unmanned aerial vehicle is stably parked directly above the cutting target, and the moving process is completed.

[0050] S104, based on the cutting target image fed back by the unmanned aerial vehicle on-board camera, the speed instruction of the end of the mechanical arm is issued through the force feedback hand controller, the position instruction is obtained by integrating the speed instruction, the control parameters are solved by combining the quadratic programming model containing the rotation matrix of the unmanned aerial vehicle and the angular velocity compensation, the mechanical arm carrying the on-board scissors tool is positioned to the cutting point, and the on-board scissors tool is controlled to open and close for air cutting.

[0051] Specifically, the cutting target image is image data containing the object to be cut (such as branches) and fed back by the unmanned aerial vehicle on-board camera (such as binocular vision camera, depth camera) in real time, including two-dimensional visual features, spatial position information and relative relationship with the surrounding environment of the cutting target. The speed instruction of the end of the mechanical arm is an instruction signal for controlling the movement speed of the end of the mechanical arm (the part carrying the on-board scissors) issued by the operator through the force feedback hand controller, which is used to directly guide the end of the mechanical arm to move to the cutting point, and the force feedback function can let the operator perceive the contact force between the mechanical arm and the target to avoid overshoot.

[0052] Optionally, the objective function of the quadratic programming model is to minimize the weighted sum of the optimization variables composed of the movement speed of the unmanned aerial vehicle in three directions and the movement speed of each joint of the mechanical arm after being weighted by the positive diagonal matrix; the constraint conditions include the equality constraint that the product of the Jacobian matrix and the optimization variables is equal to the control instruction vector, and the constraint that the optimization variables are in the feasible region containing the movement speed range of the unmanned aerial vehicle, the movement speed range of the joints of the mechanical arm and the joint angle range.

[0053] Specifically, the optimization variables of the quadratic programming model are determined as a vector composed of the movement speeds of the unmanned aerial vehicle in three directions (X, Y, Z axes) and the movement speeds of each joint of the mechanical arm. A positive diagonal matrix is configured for the optimization variables, the optimization variables are weighted through the matrix, and the objective function of the quadratic programming model is set as minimizing the square sum of the weighted optimization variables. Further, an equality constraint is constructed, the product of the Jacobian matrix and the optimization variables is calculated, and is equal to the control command vector; a feasible region constraint is constructed: the value range of the optimization variables is limited, which needs to contain the maximum / minimum value of the movement speed of the unmanned aerial vehicle, the allowable range of the movement speed of each joint of the mechanical arm, and the speed constraint boundary mapped from the joint angle range.

[0054] It should be noted that the Jacobian matrix is a matrix describing the mapping relationship between the movement speed of the unmanned aerial vehicle, the movement speed of each joint of the mechanical arm, and the operating speed of the end of the mechanical arm. The Jacobian matrix quantifies the linear relationship between the input vector "the movement speed of the unmanned aerial vehicle in three directions and the movement speed of each joint of the mechanical arm" and the output vector "the actual movement speed of the end of the mechanical arm (the part carrying the scissors) in space" through mathematical form, that is, through the product of the Jacobian matrix and the optimization variables (input vector), the speed of the end of the mechanical arm (i.e. the speed target corresponding to the control command vector) can be directly obtained, which is a key conversion tool connecting joint space movement and operating space movement. Since the structural parameters (such as the length of each joint and the connection mode) of the mechanical arm are fixed, the basic expression of the Jacobian matrix (containing joint angle variables) can be established based on these parameters. In actual operation, the specific numerical value of the Jacobian matrix will dynamically change with the current joint angle of the mechanical arm and the real-time attitude of the unmanned aerial vehicle (such as the orientation represented by the rotation matrix). The joint angle and attitude data of the unmanned aerial vehicle are collected in real time by sensors, and are substituted into the basic expression to calculate the Jacobian matrix at the current time.

[0055] In specific implementation, the on-board camera of the unmanned aerial vehicle continuously captures the clipping target image; the operator determines the clipping point position based on the clipping target image, and issues a speed command of the end of the mechanical arm to the unmanned aerial vehicle through the force feedback hand controller; after receiving the speed command, the unmanned aerial vehicle performs integral operation on the speed signal to generate a desired position command of the end of the mechanical arm; the rotation matrix of the body coordinate system relative to the world coordinate system and the real-time angular velocity data of the unmanned aerial vehicle are obtained through the attitude sensor of the unmanned aerial vehicle; the rotation matrix, the angular velocity data, and the real-time position of the end of the mechanical arm, the clipping target position, and the reaction force of the clipping target are combined to construct a speed desired command formula of the end of the mechanical arm, and the speed desired command is solved; the movement speeds of the unmanned aerial vehicle in three directions and the movement speeds of each joint of the mechanical arm are taken as optimization variables, a quadratic programming model is constructed, the quadratic programming model is solved, and the desired position parameters of the unmanned aerial vehicle in three directions and the desired joint vector of the mechanical arm are obtained.

[0056] Specifically, the UAV on-board camera is started to continuously collect real-time images of the pruning target, and the images are transmitted to the ground station display screen in real time. Based on the real-time images displayed on the ground station, the operator determines the pruning point (such as the most suitable cutting point of the branch diameter) through visual judgment, operates the rocker of the force feedback hand controller, and sends a three-dimensional velocity command (including the velocity values in the X-axis forward and backward, Y-axis left and right, and Z-axis up and down directions) of the end of the mechanical arm to the UAV. The hand controller simultaneously feeds back the contact force simulation signal of the end of the mechanical arm and the target in real time. After receiving the velocity command, the flight control system of the UAV performs integral operation (using trapezoidal integral method) on the velocity signal, and accumulates the continuous velocity values into the expected position command of the end of the mechanical arm. The UAV obtains the rotation matrix (3x3 matrix representing the attitude angle conversion relationship of the UAV) of the body coordinate system relative to the world coordinate system and the real-time angular velocity data (unit rad / s, sampling frequency 100 Hz) of the UAV in the roll, pitch and yaw directions in real time through the built-in IMU (inertial measurement unit) and GPS module. The current angles of each joint are read from the mechanical arm joint encoder, and the real-time position (three-dimensional coordinates) of the end is calculated by combining the DH parameter model of the mechanical arm; the position coordinates of the pruning target in the world coordinate system are extracted through the on-board camera. The rotation matrix, angular velocity data, real-time position of the end, target position, and reaction force of the pruning target are substituted into the preset velocity expectation command formula to obtain the real-time velocity expectation command of the end of the mechanical arm. The X, Y and Z three-direction motion velocities of the UAV and the motion velocities of each joint of the mechanical arm constitute the optimization variable vector; the objective function of the quadratic programming model is the square sum of the product of the vector and the positive diagonal weight matrix (the diagonal elements are set according to the control priority, such as the high precision weight of the mechanical arm higher than the UAV movement); the constraint condition is that the product of the Jacobian matrix and the optimization variable vector is equal to the velocity expectation command (equality constraint), and each variable value is within the preset range. The interior point method is used to solve the quadratic programming problem, and the three-dimensional expected position parameters of the UAV (coordinate sequence changing with time) and the expected joint vector of the mechanical arm (joint angle sequence changing with time) are obtained.

[0057] Optionally, in combination with the rotation matrix, angular velocity data, and the real-time position of the end of the mechanical arm, the cutting target position, the reaction force of the cutting target, a mechanical arm end velocity expectation instruction formula is constructed, including: through the three-axis force sensor on the airborne scissors tool, the interactive reaction force generated by the cutting target on the airborne scissors tool during the cutting process is collected in real time; a first positive diagonal matrix and a second positive diagonal matrix are set for adjusting the position error and the reaction force suppression; the difference between the cutting target position and the real-time position of the end of the mechanical arm is calculated, and the difference is multiplied by the first positive diagonal matrix to form an adjustment term; the rotation matrix, the real-time angular velocity of the unmanned aerial vehicle, and the real-time position of the end of the mechanical arm are subjected to vector cross multiplication operation to form a compensation term; the interactive reaction force is multiplied by the second positive diagonal matrix to form a suppression term; and the sum of the mechanical arm end velocity instruction, the adjustment term, the compensation term, and the suppression term is determined as the velocity expectation instruction.

[0058] In a specific implementation, a three-axis force sensor mounted on the airborne scissors tool is started, and the interactive reaction force generated by the cutting target on the airborne scissors tool is collected in real time during the cutting process. The sensor converts the force signal into an electrical signal and transmits it to the unmanned aerial vehicle control module. Two positive diagonal matrices are set in the unmanned aerial vehicle control module: the first positive diagonal matrix is used to adjust the position error between the end of the mechanical arm and the cutting target, and the diagonal element value is set according to the position accuracy requirement of the aerial cutting task; the second positive diagonal matrix is used to suppress the interference of the interactive reaction force of the cutting target on the operation of the mechanical arm, and the diagonal element value is set according to the requirement of the reaction force suppression strength; the real-time angles of each joint are collected by the joint encoder of the mechanical arm, and the real-time position of the end of the mechanical arm in the unmanned aerial vehicle body coordinate system is calculated by combining the DH parameter model of the mechanical arm; at the same time, the real-time position of the cutting target in the world coordinate system is extracted through the visual recognition algorithm of the unmanned aerial vehicle onboard camera; the rotation matrix of the body coordinate system relative to the world coordinate system is obtained by means of the IMU inertial measurement unit of the unmanned aerial vehicle, and the real-time position of the end of the mechanical arm in the world coordinate system is converted into the real-time position of the end of the mechanical arm in the world coordinate system through matrix operation; the difference between the real-time position of the end of the mechanical arm in the world coordinate system and the real-time position of the cutting target in the world coordinate system is calculated, and the difference is multiplied by the first positive diagonal matrix to obtain the adjustment term for correcting the position deviation; the real-time angular velocity of the unmanned aerial vehicle in the roll, pitch and yaw directions is obtained from the IMU inertial measurement unit of the unmanned aerial vehicle, and the rotation matrix, the real-time angular velocity and the real-time position of the end of the mechanical arm in the unmanned aerial vehicle body coordinate system are subjected to vector cross multiplication operation to obtain the compensation term for offsetting the disturbance of the unmanned aerial vehicle attitude change on the position of the end of the mechanical arm; the interactive reaction force collected is multiplied by the second positive diagonal matrix to obtain the suppression term for suppressing the operation deviation of the mechanical arm caused by excessive reaction force; the end of the mechanical arm speed command issued by the operator through the force feedback hand controller is retrieved from the unmanned aerial vehicle control module (the command is the initial speed command in the body coordinate system), and the end of the mechanical arm speed command, the adjustment term, the compensation term and the suppression term are summed to obtain the final sum value as the end of the mechanical arm speed command (the speed command in the world coordinate system).

[0059] For example, in an embodiment, the image information of the target is fed back in real time by the onboard camera carried by the unmanned aerial vehicle, and the image information is transmitted to the ground station through WiFi. The operator can view the image information and the state information of the unmanned aerial vehicle in real time. The operator issues the end of the mechanical arm speed command to the unmanned aerial vehicle through the force feedback hand controller. After receiving the end of the mechanical arm speed command, the unmanned aerial vehicle integrates it to obtain the corresponding expected position command . The formula for solving the end of the mechanical arm speed command is:

[0060] ;

[0061] in, The desired speed command for the robotic arm's end effector; This is the speed command for the robotic arm's end effector. These represent the positions of the robotic arm's end effector and the target, respectively. These are the rotation matrix and real-time angular velocity of the UAV, respectively. , It is a diagonal matrix; It is a reaction force; This refers to the position of the robotic arm's end effector in the drone's body coordinate system.

[0062] Solving the quadratic programming model yields the expected three-axis position parameters of the UAV and the expected joint vectors of the robotic arm:

[0063] ;

[0064] in, , These represent the desired positions of the drone in three directions. Let be the desired joint angle of the robotic arm; It is a diagonal matrix; It is a Jacobian matrix; The desired speed command for the robotic arm's end effector; Let be the feasible region of the variable.

[0065] Furthermore, regarding Integrating, we get This allows the desired positions of the drone in three directions and the desired joint angles of the robotic arm to be obtained and sent to the drone and robotic arm. Once the end effector of the robotic arm reaches the designated area, the ground operator presses a button on the pressure feedback hand controller to open and close the scissors, thus completing the cutting.

[0066] Optionally, after controlling the airborne scissors tool to open and close for aerial cutting, the method further includes: controlling the UAV's airborne camera to acquire an image of the cut target, the image including the cut section and surrounding area features; transmitting the image and user-preset cutting expectation parameters to a large language model; the large language model generating an evaluation result based on the image features and the cutting expectation parameters; transmitting the evaluation result and the acquired image to a ground station computer; the user judging the effect based on the evaluation result and the image; and re-executing the robotic arm positioning and cutting operation when the user determines that the cutting effect does not meet expectations.

[0067] Specifically, after the aerial cutting operation is completed, the unmanned aerial vehicle on-board camera is controlled to capture the image of the target after cutting to ensure that the image clearly contains the visual features of the cutting section (such as the branch cutting position) and the surrounding area (such as the remaining branch and the shape of the cutting section). The user's preset cutting expected effect parameters (such as the cutting section flatness threshold and the upper limit of the remaining branch length) are read from the ground station computer, and the parameters are transmitted to the large language model together with the collected image. The large language model identifies the cutting section and surrounding features in the image (such as whether the cutting section is flat and whether the remaining length meets the requirements), and compares them with the preset expected effect parameters to generate evaluation results containing "meets expectations", "cutting section is not flat", "remaining length is too long", etc. The large language model sends the evaluation results and the original collected image to the ground station computer, and the user views the image and the evaluation results through the ground station to manually determine whether the cutting effect meets the expectations. If the user determines that it does not meet the expectations, the ground station sends instructions to the unmanned aerial vehicle to restart the mechanical arm positioning program (adjust the position based on real-time images) and the scissors opening and closing operation until the user confirms that the cutting effect meets the requirements.

[0068] The method provided by the embodiment, in the first aspect, first, based on the task path points output by the slow thinking layer, an initial convex polyhedron is formed by connecting adjacent path points, and is gradually expanded along the direction of the obstacle until it contacts the surface of the obstacle, and then a plurality of expanded convex polyhedrons are spliced to form a continuous three-dimensional safe channel, which strictly isolates the unmanned aerial vehicle from the obstacle through the concrete spatial boundary, and eliminates the collision risk from the physical space layer; then, based on the inflection points of the safe channel, the nodes of the Bezier curve are set, the dynamics parameters such as the maximum speed of the unmanned aerial vehicle and the acceleration of the mechanical arm are combined to construct convex constraint conditions, and a smooth trajectory is fitted through convex optimization operation, so that the trajectory is completely within the safe channel and can match the movement characteristics of the unmanned aerial vehicle, avoiding the body shaking caused by sudden stop and sudden turn. This "first boundary and then trajectory" method not only builds a flight safety line through the safe channel, but also guarantees the stability of the unmanned aerial vehicle movement through the smooth trajectory, laying a foundation for subsequent precise operation of the mechanical arm.

[0069] Secondly, from building a quadratic programming model to solving control parameters, firstly, the end speed instruction of the manipulator issued by the operator is acquired through the force feedback hand controller, and the position instruction is obtained by integration, and then the rotation matrix (correcting the deviation between the body coordinate system and the world coordinate system) and the angular velocity data (compensating the body attitude disturbance) acquired by the attitude sensor of the unmanned aerial vehicle are fused to build a speed expectation instruction formula containing a compensation term, and then the three-dimensional motion speed of the unmanned aerial vehicle and the joint speed of the manipulator are taken as optimization variables, and the model is solved by minimizing the weighted sum of squares objective function under the equality constraint of the Jacobian matrix mapping relationship and the inequality constraint of the feasible region of motion parameters, which not only offsets the interference of the dynamic attitude of the unmanned aerial vehicle on the manipulator through real-time compensation, but also finds the optimal control parameters under the premise of meeting multiple constraints through the optimization algorithm, so that the manipulator can accurately position to the cutting point, and high-precision operation under man-machine cooperation is realized.

[0070] Thirdly, after the cutting is completed, the image containing the cutting section and the surrounding area is collected by the on-board camera, the image and the expected effect parameter preset by the user are transmitted to the large language model, the evaluation result is generated by the model comparison and fed back to the ground station, the user combines the image and the evaluation result to judge the effect, and if it does not meet the expectation, the positioning and cutting operation is re-executed. This verification method of machine preliminary evaluation and manual final judgment not only improves efficiency by quickly completing feature comparison with the help of the large language model, but also avoids the limitations of machine evaluation through manual intervention, forming a closed-loop verification mechanism to ensure that the final cutting effect meets the actual demand.

[0071] Corresponding to the foregoing embodiment of the air cutting method based on the end control of quadratic programming, the present application also provides an embodiment of an air cutting device based on the end control of quadratic programming.

[0072] Figure 2 The structure schematic diagram of the air cutting device based on the end control of quadratic programming provided for the second embodiment of the present application is shown in Figure 2 The device provided in the embodiment comprises a construction module 210, a generation module 220 and a control module 230.

[0073] The construction module 210 is configured to construct a bidirectional cognitive task planning architecture, wherein the bidirectional cognitive task planning architecture comprises a slow thinking layer and a fast thinking layer.

[0074] The generation module 220 is configured to extract the air cutting skill of the unmanned aerial vehicle through the slow thinking layer and generate a task path point.

[0075] The control module 230 is configured to convert the task path point into a Bezier curve flight trajectory of a convex polyhedron safety channel through the fast thinking layer, and control the unmanned aerial vehicle to move to the cutting target directly above along the Bezier curve flight trajectory.

[0076] The control module 230 is further configured to, based on the cropped target image fed back by the camera on the UAV in real time, issue an end speed instruction of the mechanical arm through the force feedback hand controller, obtain a position instruction by integrating the speed instruction, solve a control parameter by combining a quadratic programming model containing a rotation matrix of the UAV and an angular velocity compensation, drive the mechanical arm to position the onboard scissors tool to a cutting point, and control the onboard scissors tool to open and close to perform aerial cutting.

[0077] The device of the embodiment can be used to execute the method. Figure 1 The steps of the method embodiment are similar to the specific implementation principles and implementation processes, and thus will not be described here.

[0078] The implementation processes of the functions and roles of the units in the device are specifically described in the implementation processes of the corresponding steps in the above method, and thus will not be described here.

[0079] For the device embodiment, it basically corresponds to the method embodiment, and thus the related parts can be understood by referring to the parts of the method embodiment. The device embodiments described above are merely illustrative, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. According to actual needs, some or all of the modules can be selected to achieve the purposes of the present application. Those skilled in the art can understand and implement it without creative labor.

[0080] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An aerial cropping method based on quadratic programming end control, characterized in that, The method comprises: constructing a bidirectional cognitive task planning architecture, the bidirectional cognitive task planning architecture comprising a slow thinking layer and a fast thinking layer; extracting a UAV air cutting skill through the slow thinking layer and generating a task path point; converting the task path point into a Bezier curve flight trajectory containing a convex polyhedron safety channel through the fast thinking layer, controlling the UAV to move along the Bezier curve flight trajectory to directly above a cutting target; based on a cutting target image fed back in real time by a UAV on-board camera, issuing an end-of-arm speed instruction through a force feedback hand controller, after integrating the speed instruction to obtain a position instruction, combining a quadratic programming model containing a UAV rotation matrix and angular velocity compensation to solve control parameters, driving a mechanical arm to position a on-board cutting tool to a cutting point, and controlling the on-board cutting tool to open and close to perform air cutting; the extracting a UAV air cutting skill through the slow thinking layer and generating a task path point comprises: constructing an air cutting task description file; performing semantic analysis on the description file to extract a UAV air cutting skill; performing feature abstraction on the UAV air cutting skill to form a task constraint set and a skill set, and constructing an application program interface for adapting to an air cutting task based on the skill set; reading a point cloud map through the application program interface, converting the point cloud map into structured data, constructing a cutting scene task planning test set based on the structured data and cutting task requirements, the test set containing multiple sets of cutting task input conditions and corresponding expert planning results; inputting the test set into a task planning operation unit, optimizing the planning logic of the unit through iterative training, adjusting the internal parameters of the unit using feedback data of the expert planning results, and updating the planning reasoning capability; inputting an air cutting task description into the optimized task planning operation unit, combining real-time map structured data transmitted through the application program interface, and outputting a task execution step containing cutting target coordinate information and corresponding flight path points through internal planning logic operation of the unit.

2. The method of claim 1, wherein, based on a cutting target image fed back in real time by a UAV on-board camera, issuing an end-of-arm speed instruction through a force feedback hand controller, after integrating the speed instruction to obtain a position instruction, combining a quadratic programming model containing a UAV rotation matrix and angular velocity compensation to solve control parameters, comprises: controlling the UAV on-board camera to continuously capture cutting target images; an operator determines a cutting point position based on the cutting target image and issues an end-of-arm speed instruction to the UAV through a force feedback hand controller; after the UAV receives the speed instruction, the speed signal is integrated to generate a desired position instruction for the end of the mechanical arm; acquiring a rotation matrix of a body coordinate system relative to a world coordinate system and real-time angular velocity data of the UAV through a built-in attitude sensor of the UAV; combining the rotation matrix, angular velocity data, and real-time position of the end of the mechanical arm, cutting target position, and reaction force of the cutting target to construct an end-of-arm speed desired instruction formula, and solving the speed desired instruction; The three-direction motion speed of the unmanned aerial vehicle and the joint motion speed of the mechanical arm are taken as optimization variables, a quadratic programming model is constructed, the quadratic programming model is solved, and the three-direction expected position parameters of the unmanned aerial vehicle and the expected joint vector of the mechanical arm are obtained.

3. The method of claim 2, wherein, Combined with the rotation matrix, the angular velocity data, and the real-time position of the mechanical arm end, the cutting target position, and the reaction force of the cutting target, a mechanical arm end speed expected instruction formula containing a compensation term is constructed, including: Through the three-axis force sensor mounted on the airborne shears tool, the interactive reaction force generated by the cutting target on the airborne shears tool during the cutting process is collected in real time; A first positive diagonal matrix and a second positive diagonal matrix are set for adjusting the position error and the reaction force suppression; The difference between the cutting target position and the real-time position of the mechanical arm end is calculated, the difference is multiplied by the first positive diagonal matrix and then negated to form an adjustment term; The rotation matrix, the real-time angular velocity of the unmanned aerial vehicle, and the real-time position of the mechanical arm end are subjected to vector cross multiplication operation to form a compensation term; The interactive reaction force is multiplied by the second positive diagonal matrix and then negated to form a suppression term; The sum of the mechanical arm end speed instruction, the adjustment term, the compensation term, and the suppression term is determined as the speed expected instruction.

4. The method of claim 2, wherein, The objective function of the quadratic programming model takes the vector composed of the motion speed of the unmanned aerial vehicle in three directions and the joint motion speed of the mechanical arm as the optimization variable, and the weighted sum of the weighted optimization variables is minimized; the constraint conditions include the equality constraint that the product of the Jacobian matrix and the optimization variable is equal to the control instruction vector, and the constraint that the optimization variable is in the feasible region containing the motion speed range of the unmanned aerial vehicle, the joint motion speed range of the mechanical arm, and the joint angle range mapping.

5. The method of claim 1, wherein, After the output containing the task execution steps and the corresponding flight path points of the cutting target coordinate information, the method further includes: Obtain the shape structure parameters of the unmanned aerial vehicle, and convert the overall shape of the unmanned aerial vehicle into a three-dimensional convex polyhedron model for collision detection using a fast convex hull generation algorithm; Retrieve structured data from the application programming interface, extract obstacle feature information, and organize the three-dimensional convex polyhedron model, task execution steps, and corresponding flight path points into a collision detection data set; Input the collision detection data set and the collision detection instruction into the large language model, simulate the whole process of the unmanned aerial vehicle performing cutting work along the flight path points based on the obstacle feature information, the three-dimensional convex polyhedron model, and the flight path points, and check whether the unmanned aerial vehicle as a whole collides with the obstacles; The large language model analyzes the relative relationship between the spatial position of the unmanned aerial vehicle moving along each flight path point and the obstacles based on the input collision detection data set and collision detection instruction, determines whether there is a collision risk, and generates a collision detection result; When the collision detection result is collision, the collision path segment information and the obstacle feature information are fed back to the large language model, and the large language model re-plans the flight path points in combination with the task constraint set and the skill set until the collision detection result is no collision.

6. The method of claim 1, wherein, The conversion of the task path points into a Bézier curve flight trajectory containing a convex polyhedron safe channel through the fast thinking layer includes: Receive the task path points output by the slow thinking layer and the structured data transmitted through the application programming interface, and extract obstacle location information from the structured data; The task path points are connected sequentially to form an initial convex polyhedron. Using the initial convex polyhedron as a reference, the volume is gradually expanded along the direction of the obstacle until the boundary of the convex polyhedron contacts the surface of the obstacle. The multiple convex polyhedra after continuous expansion are spliced ​​together to form a safe passage for the drone to fly. The flight path within the safety passage is fitted using multiple Bézier curves to obtain the Bézier curve flight trajectory.

7. The method of claim 6, wherein, The process of fitting the flight path within the safety passage based on multiple Bézier curves to obtain the Bézier curve flight trajectory includes: Based on the spatial range of the safety passage, segment nodes of multiple Bézier curves are set, and each segment node corresponds one-to-one with the inflection point of the safety passage. Configure control point vectors for each Bézier curve segment, and construct convex constraint conditions by combining the dynamic parameters of the UAV's maximum flight speed and the robotic arm's motion acceleration; Substitute the convex constraints into the Bézier curve parameter solution process, determine the order, time scaling factor and basis vector parameters of each Bézier curve segment through convex optimization operation, and fit to form a smooth flight trajectory that adapts to the safe passage. Verify whether the flight trajectory is completely within the convex polyhedron safety channel and meets the dynamic constraints. If not, adjust the control point vector and constraint parameters, and re-execute the fitting calculation until the flight trajectory meets the requirements.

8. The method of claim 1, wherein, After controlling the opening and closing of the airborne scissors tool for aerial cutting, the method further includes: The drone's onboard camera is controlled to acquire images of the cropped target, the images including the cropped section and features of the surrounding area; The image and the user-preset cropping expectation parameters are transmitted to the large language model; The large language model generates an evaluation result based on the comparison between image features and the expected cropping effect parameters. The evaluation result and the collected image are transmitted to the ground station computer, and the user makes an effect judgment based on the evaluation result and the image. If the user determines that the cutting effect does not meet expectations, the robotic arm positioning and cutting operations are re-executed.

9. An aerial cutting device based on quadratic programming end control, characterized by The device includes a construction module, a generation module, and a control module; The construction module is used to construct a bidirectional cognitive task planning architecture, which includes a slow thinking layer and a fast thinking layer. The generation module is used to extract drone aerial cropping skills through the slow thinking layer and generate task path points; The control module is used to convert the task path point into a Bézier curve flight trajectory containing a convex polyhedron safe channel through the fast thinking layer, and control the UAV to move along the Bézier curve flight trajectory to directly above the trimming target. The control module is also used to send a speed command to the end of the robotic arm through a force feedback hand controller based on the real-time feedback of the target image from the UAV's onboard camera. After integrating the speed command to obtain the position command, the control parameters are solved by combining a quadratic programming model containing the UAV rotation matrix and angular velocity compensation. The control module drives the robotic arm to carry the onboard scissors tool to the cutting point and controls the onboard scissors tool to open and close for aerial cutting. The unmanned aerial vehicle air cutting skill is extracted through the slow thinking layer, and a task path point is generated, including: constructing an air cutting task description file; performing semantic analysis on the description file to extract an unmanned aerial vehicle air cutting skill; performing feature abstraction on the unmanned aerial vehicle air cutting skill to form a task constraint set and a skill set, and constructing an application program interface suitable for an air cutting task based on the skill set; reading a point cloud map through the application program interface, converting the point cloud map into structured data, constructing a cutting scene task planning test set based on the structured data and cutting task requirements, and the test set containing multiple sets of cutting task input conditions and corresponding expert planning results; inputting the test set into a task planning operation unit, optimizing the planning logic of the unit through iterative training, adjusting the internal parameters of the unit by using the feedback data of the expert planning results, and updating the planning reasoning capability; inputting an air cutting task description into the optimized task planning operation unit, combining the real-time map structured data transmitted by the application program interface, and outputting a task execution step containing cutting target coordinate information and a corresponding flight path point through internal planning logic operation.

Images