Unmanned aerial vehicle risk adaptive model predictive control method based on visual language perception

Abstract

The application discloses a UAV risk adaptive model predictive control method based on visual language perception. The method acquires multi-source observation of a UAV, natural language task instruction and flight state, analyzes task semantics and safety constraints; fuses task semantics and multi-source observation for visual-language collaborative perception and geometric coding, constructs a semantic world state containing targets, obstacles, free space, task stages and perception uncertainty; generates semantic sub-targets and candidate trajectories based on the state and determines a reference trajectory; uses a risk adaptive model predictive controller to roll and solve control quantity, and adjusts safety distance, speed and control parameters according to perception uncertainty and execution state; updates the semantic world state and the reference trajectory according to subsequent observation, execution result and optimization feasibility, and executes a preset safety strategy when a safety degradation condition is met. The application improves the dynamic feasibility and flight safety of natural language task execution in a complex low-altitude scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, computer vision, intelligent planning, and autonomous control of unmanned aerial vehicles (UAVs), specifically to a method for autonomous UAV mission execution that couples vision-language collaborative perception, semantic world state construction, hierarchical trajectory planning, and risk adaptive model predictive control in a closed loop. Background Technology

[0002] As drones are increasingly used in scenarios such as low-altitude inspection, disaster relief, security monitoring, logistics transportation, target tracking, and autonomous operation, actual tasks are gradually shifting from pre-programmed fixed-route flights to open task execution driven by natural language commands. For example, an operator might instruct the drone to "avoid the building ahead, approach the red tent from the right, and hover above the entrance," a task that simultaneously includes target recognition, spatial relationship understanding, path selection, flight constraints, and termination conditions.

[0003] Existing AI-based UAV methods typically utilize object detection, semantic segmentation, visual language models, or visual-language-action models to achieve scene understanding and action prediction. While these methods can improve the flexibility of task expression, when directly outputting discrete actions, speed commands, or waypoints, they often lack explicit modeling of UAV dynamics, actuator saturation, flight envelope, wind disturbance, and communication latency. This can lead to issues such as unexecutable actions, non-smooth control, or insufficient safety margins in the predicted actions.

[0004] Existing UAV control methods typically employ proportional-integral-derivative (PID) control, linear quadratic control, model predictive control (MMC), or control barrier functions to achieve trajectory tracking and obstacle avoidance. Among these, MMC can simultaneously consider dynamics and multiple constraints in the prediction time domain, but its reference trajectory, cost weights, and safety distance are mostly preset manually, making it difficult to understand task semantics in natural language such as "approach from the right," "keep the target in the center of the field of view," or "prioritize avoiding crowds."

[0005] Existing perception, planning, and control systems often employ a unidirectional serial structure, where the perception module outputs the target or obstacle, the planning module outputs the trajectory, and the control module passively tracks the trajectory. When the front-end AI model experiences decreased recognition confidence, increased depth estimation error, or task ambiguity, the uncertainty cannot be effectively transmitted to the controller. Furthermore, when the controller detects that the trajectory is infeasible, constraints remain active, or the target is about to leave the field of view, there is a lack of a mechanism to provide feedback to the semantic planner and modify the task phase.

[0006] In low-altitude scenarios such as dynamic obstacles, weak textures, unstable satellite positioning, target occlusion, and gusts of wind, the aforementioned problems can easily lead to target loss, path oscillation, frequent emergency braking, or increased collision risk. Therefore, there is a need for an autonomous UAV mission execution method that can unify mission semantic understanding, environmental perception, trajectory planning, dynamic constraints, and safety control into a closed-loop framework. Summary of the Invention

[0007] The purpose of this invention is to provide a risk-adaptive model predictive control method for unmanned aerial vehicles (UAVs) based on visual language perception. This method addresses problems encountered in existing autonomous UAV mission execution processes, such as the difficulty in converting high-level task semantics into executable flight constraints, the mismatch between AI-generated planning results and UAV dynamic capabilities, and the lack of two-way feedback based on risk state between perception, planning, and control. By establishing a closed-loop relationship between task semantic parsing, visual-language collaborative perception, semantic world state, trajectory planning, and risk-adaptive model predictive control, the UAV can autonomously generate and correct its flight trajectory based on natural language task and environmental changes, and dynamically adjust its flight strategy according to perception uncertainty and control feasibility.

[0008] To achieve the above objectives, the present invention provides the following technical solution: First, acquire natural language task instructions, UAV multi-source observations, and the current flight state, and obtain the task object, action primitives, spatial relationships, termination conditions, and safety constraints through task semantic parsing; then, utilize visual-language collaborative perception and geometric encoding to obtain target, obstacle, and free space information related to the current task, and construct a semantic world state by combining temporal memory; generate semantic sub-targets and multiple candidate trajectories based on the semantic world state and task stage, and determine a reference trajectory by comprehensively considering task consistency, environmental risk, target visibility, and control feasibility; input the reference trajectory and the current flight state into a risk adaptive model predictive controller, and solve the UAV control variables by combining the flight dynamics model and risk state; during control execution, perform bidirectional closed-loop updates on the semantic world state, semantic sub-targets, reference trajectories, and control parameters based on new perception results, trajectory tracking errors, control constraint states, and optimization solution states; when the system risk exceeds preset conditions, execute the corresponding safety degradation strategy.

[0009] Preferably, the natural language task instructions are provided by text, speech-to-text, structured task description, or remote operator interaction information. The task semantic encoder outputs word-level and sentence-level task features, and the constraint parser converts category, attribute, orientation, distance, altitude, speed, observation direction, no-fly zone requirements, and completion conditions into a set of structured constraints.

[0010] Preferably, the multi-source observations include one or more of the following: monocular or multi-view images, depth maps, laser point clouds, inertial measurement data, global navigation satellite system data, visual odometry, flight controller status, gimbal status, wind speed estimation, and optional prior maps. Different data are unified to the body coordinate system, local world coordinate system, or mission coordinate system through timestamp alignment and coordinate calibration.

[0011] Preferably, the visual-language collaborative perception includes cross-modal attention of visual features to task text and reverse spatial anchoring of task text to visual regions to highlight regions related to task objects, spatial relationships, and action conditions; scene geometry branch outputs depth, occupancy probability, flyable area, and obstacle motion state; temporal memory is used to maintain consistency of target identity, obstacle trajectory, and task stage.

[0012] Preferably, the semantic world state is denoted as It includes at least a target set, a static obstacle set, a dynamic obstacle set, a free space representation, a task stage, target visibility, state estimation, and corresponding uncertainties. The free space representation can be a semantic occupancy grid, a sparse topological graph, a local point cloud, bird's-eye view features, or a continuous neural field.

[0013] Preferably, the hierarchical intelligent planner first determines the current task stage based on the task completion rate, and then generates sub-goals with semantic meaning. The sub-target includes the target's relative position, desired observation azimuth, desired altitude, desired heading, desired speed, and gimbal view. Subsequently, multiple candidate trajectories are generated around the sub-target, and forward simulation is performed using the UAV dynamics model to eliminate candidate trajectories that are obviously uncontrollable or violate the flight envelope.

[0014] Preferably, the joint evaluation cost of the candidate trajectory can be expressed as a weighted sum of task semantic bias, security risk, target visibility, energy consumption, trajectory smoothing, terminal error and control feasibility, and the weights can be dynamically adjusted according to the task type and perceived uncertainty.

[0015] Preferably, the risk adaptive model predictive controller reads the current state in each control cycle. and reference trajectory In length of Solving the control sequence in the prediction time domain It executes only the first control variable, and then re-solves based on the new observations in the next control cycle, thus forming a rolling time-domain closed-loop control.

[0016] Preferably, the risk adaptation includes: increasing the obstacle safety expansion radius based on the obstacle position covariance or occupancy uncertainty; reducing the maximum speed based on the target and free space confidence; increasing the control smoothing weight based on communication delay and wind disturbance estimation; and increasing the target visibility weight based on the degree to which the target is about to leave the field of view.

[0017] Preferably, the controller feeds back the solution status, prediction terminal error, active constraint ratio, remaining safe distance, and target field of view margin to the intelligent planner. When the feedback indicates that the reference trajectory remains infeasible, the intelligent planner changes the sub-target azimuth, adds intermediate waypoints, adjusts the mission phase, or triggers a safety degradation.

[0018] The present invention also provides an autonomous mission execution system for unmanned aerial vehicles, including a multi-source input and synchronization module, a mission semantic parsing module, a vision-language collaborative perception and geometric encoding module, a semantic world state module, a trajectory planning module, a risk adaptive model prediction and control module, and a feedback and safety management module.

[0019] S1 receives natural language task instructions, UAV sensor observations and flight controller status, and performs time synchronization, distortion correction, coordinate calibration and outlier processing on image, depth, point cloud, inertial navigation, positioning and gimbal data; S2. Input the natural language task instructions into the task semantic encoder to extract the task object, action primitives, attributes, spatial relationships, target distance or height, observation orientation, termination conditions and safety restrictions, and form a set of task constraints. S3. Utilize visual-language collaborative perception and geometric coding to extract multi-scale semantic features, target state, depth information, obstacle occupancy probability, obstacle speed, and flyable area, and establish the correspondence between task semantics and visual regions through cross-modal attention; S4. Construct a semantic world state by combining current perception results, historical time sequence memory, and flight status. And estimate the target confidence, depth variance, obstacle occupancy entropy and state covariance; S5. Generate semantic sub-targets based on task stage and semantic world state. The semantic sub-target specifies at least one of the following: target relative position, desired altitude, desired heading, speed range, and target observation orientation; S6. Based on semantic sub-target sampling or prediction of candidate waypoints and candidate trajectories, a joint score is performed using task semantic bias, safety risk, target visibility, energy consumption, trajectory smoothness, and control feasibility to select a reference trajectory. ; S7. Construct a model to predict control optimization problems based on the current flight status, flight dynamics model, reference trajectory, obstacle status and perception uncertainty, and dynamically set the safety distance, speed limit, risk weight and target visibility weight; S8. Solve for the optimal control sequence under state constraints, control quantity constraints, geofence constraints, obstacle safety distance constraints, and control barrier function constraints, and send the first thrust, attitude, angular velocity, acceleration, or velocity control quantity to the flight controller. S9. Obtain the observation and flight status at the next moment, calculate the trajectory tracking error, target field of view margin, minimum obstacle distance, constraint activation ratio and optimization solution status, and update the temporal memory; S10. Output the mission completion status when the mission termination condition is met; regenerate the sub-target and reference trajectory when the tracking error or risk exceeds the threshold; execute deceleration observation, hovering, ascent obstacle avoidance, return to safe waypoint, automatic return to home or manual takeover when there is no feasible solution or the safety condition is triggered.

[0020] Compared with the prior art, the present invention has the following beneficial effects: First, by using visual-language collaborative perception, objects, actions, and spatial relationships in natural language tasks are transformed into structured task constraints that can be directly used for trajectory planning and UAV control, thereby improving the consistency between open task instructions and actual flight behavior.

[0021] Second, by using a unified semantic world state to represent the target, obstacles, free space, task stages, and perception uncertainty, the perception results can be directly used to participate in trajectory planning and control parameter adjustment, reducing the safety risks caused by the direct propagation of perception errors to the control end.

[0022] Third, a hierarchical planning control structure is formed by using semantic sub-objectives, candidate trajectories, and risk adaptive model predictive control. This ensures that the generated reference trajectory simultaneously meets the semantic requirements of the mission and the flight dynamics constraints of the UAV, thereby improving the executability of the trajectory.

[0023] Fourth, by establishing a two-way feedback mechanism between planning and control, when perceived uncertainty increases, reference trajectory becomes infeasible, or environmental risks change, the trajectory, safety distance, and control strategy are adaptively adjusted, and a safety degradation state is entered when necessary, thereby improving the safety and robustness of mission execution in complex low-altitude environments. Attached Figure Description

[0024] Figure 1 This is the overall flowchart of the visual-language collaborative perception-semantic planning-risk adaptive control closed loop of the present invention; Figure 2 This is a schematic diagram of the visual-language collaborative perception and semantic world state construction structure of the present invention; Figure 3 This is a schematic diagram illustrating the semantic sub-targets and candidate trajectory generation and evaluation of the present invention; Figure 4 This is a schematic diagram of the risk adaptive model predictive control structure of the present invention; Figure 5 This is a schematic diagram of the perception-planning-control bidirectional feedback and safety degradation mechanism of the present invention; Figure 6 This is a schematic diagram illustrating the model training, parameter tuning, and system deployment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.

[0026] Please see Figures 1 to 6 This invention provides a risk adaptive model predictive control method for unmanned aerial vehicles (UAVs) based on visual language perception. The system takes natural language task commands, multi-source observations, and current flight status as input, and outputs UAV control variables, task trajectory, task stage, risk status, and execution log.

[0027] In one implementation, the drone state can be represented as: ,in For location, For speed, For attitude quaternions, This refers to the airframe angular velocity. Control variables can be expressed via the flight control interface as total thrust and airframe torque, desired attitude and total thrust, desired angular velocity and thrust, or three-dimensional velocity and yaw angular velocity.

[0028] Task semantic encoder for instructions The encoding process outputs task object embeddings, action primitive embeddings, spatial relation embeddings, and global sentence-level embeddings. The constraint parser can employ rule parsing, sequence labeling, structured generation, or large language model constraint decoding to convert instructions into a set of task constraints. For the instruction "stay 12 meters behind and above the white vehicle", the system resolves the target to be the white vehicle, the action primitive to be follow, the relative position to be behind and above, the desired distance to be 12 meters, and adds continuous visibility and non-collision constraints.

[0029] The vision-language collaborative perception module extracts multi-scale features from images and calculates the correspondence between task text and image regions through cross-modal attention. The scene geometry branch generates local occupancy representations based on depth maps, point clouds, or depth estimation results. Dynamic obstacle states can be obtained through continuous frame target association, optical flow, scene flow, or Kalman filtering.

[0030] Semantic world state It can be represented as ,in The state of the task objective. It is a static obstacle. For dynamic obstacles, For free space, For task phases and semantic sub-goals, The uncertainty is denoted by . The system updates historical memory using gated loop units, temporal attention, or explicit trajectory tables.

[0031] Candidate Trajectory The joint evaluation cost can be expressed as:

[0032] in, This indicates the deviation between the trajectory terminal state and the relationship in the task space. This indicates the risk of collision along the trajectory. This indicates the degree to which the target is off-field or obstructed. Indicates thrust and range consumption. Indicates velocity, acceleration, and curvature changes. This represents the feasibility estimate of the model predicting the controller. The weights can be dynamically adjusted based on the task type, target confidence level, and obstacle uncertainty.

[0033] In the trajectory control phase, the risk adaptive model predictive control optimization problem can be expressed as:

[0034] The optimization satisfies and satisfy the state boundary Control Boundaries Obstacle safety distance And control barrier function constraints. For the safety function corresponding to the obstacle. Discrete control barrier function constraints can be written as ,in This is the safe convergence coefficient.

[0035] The obstacle safety distance can be adaptively set according to the uncertainty. ,in Based on the safe distance, The uncertainty in estimating the location or occupancy of the obstacle. The coefficient of thermal expansion. The maximum permissible speed can be set to... ,in To comprehensively perceive uncertainty, the drone automatically reduces its speed and increases its safety margin when the scene is unidentifiable or the depth estimation is unstable.

[0036] Example 1: In this embodiment, the UAV performs a power transmission line inspection task. The input command is "fly along the left side of the line at a distance of eight meters, observe the insulators ahead, and detour around the towers from the outside." The task semantic parser converts "left side of the line," "maintain eight meters," "observe the insulators," and "detour around the towers from the outside" into relative orientation, distance, target visibility, and obstacle avoidance constraints.

[0037] The vision-language collaborative perception module identifies lines, insulators, and towers, and constructs a local occupancy map using depth information. A hierarchical planner generates semantic corridors located to the left of the line and generates outer bypass sub-targets when approaching towers. A risk-adaptive model predictive controller incorporates visibility costs into the line observation orientation while limiting lateral distance error, maximum roll angle, and tower safety distance.

[0038] When gusts cause lateral drift and tower safety distance constraints remain active, the controller feeds back the constraint activation ratio to the planner, which then expands the detour radius and adds intermediate waypoints. Compared to tracking only a fixed route, this closed loop can simultaneously maintain mission observation direction and dynamic feasibility.

[0039] Example 2: In this embodiment, the drone flies over the disaster site and receives the instruction "Avoid the collapsed building, approach the red tent from the right and hover above the entrance". The task semantic parser identifies the red tent as the target, defines the collapsed building as a priority avoidance area, and converts approaching from the right and hovering above the entrance into intermediate sub-target and terminal state, respectively.

[0040] Due to smoke and obstruction at the disaster site, the confidence level of visual perception and the stability of depth estimation decrease. The system expands the building safety expansion area based on obstacle occupancy entropy, reduces the maximum speed, and increases the weight of candidate trajectory risk terms. When the red tent is briefly obscured by smoke, the temporal memory maintains the target semantic position, and the planner enters a deceleration observation phase instead of immediately changing the target.

[0041] Once the target becomes visible again, the system regenerates a reference trajectory approaching from the right, and the risk-adaptive model predicts that the controller will smoothly enter the hovering point above the entrance under velocity, acceleration, and safe distance constraints. If a feasible solution cannot be obtained for several consecutive control cycles, the state manager triggers a vertical ascent to a safe height and then replans.

[0042] Example 3: In this embodiment, the UAV performs a dynamic vehicle following task. The input command is "keep following the white vehicle 12 meters behind and above, and keep the vehicle in the center of the frame." The visual-language collaborative perception module locks onto the target vehicle based on color and category, and estimates the vehicle's speed, heading, and target visibility.

[0043] The planner defines the twelve meters above and behind as a dynamic semantic sub-target and generates multiple following trajectories based on the vehicle's predicted trajectory. The candidate trajectory evaluation simultaneously considers relative position error, target field of view deviation, obstacles above the road, turning smoothness, and the feasibility of model predictive control.

[0044] When a vehicle makes a sharp turn and the target approaches the edge of the frame, the system increases the visibility cost and adjusts the UAV's heading and gimbal view. When the target vehicle intersects with a similar vehicle, causing a decrease in confidence, the system reduces its flight speed, maintains a safe altitude, and uses time-series target memory to reconfirm the target, preventing the controller from making a large turn based on an incorrect target.

[0045] Example 4: During the training phase, data samples can be constructed that include UAV videos, natural language commands, target and obstacle annotations, depth or occupancy annotations, semantic sub-targets, reference trajectories, flight states, and control variables. The training loss for the artificial intelligence component includes one or more of the following: task semantic parsing loss, target detection or segmentation loss, depth estimation loss, occupancy prediction loss, trajectory imitation loss, uncertainty calibration loss, and time consistency loss.

[0046] The total training loss can be expressed as:

[0047] The risk-adaptive model predictive control section obtains mass, inertia, thrust coefficient, drag coefficient, and execution delay through flight dynamics parameter identification, and tunes tracking weights, control weights, safety distance, and control barrier function parameters in a simulation environment. Domain randomization can be used to modify illumination, texture, wind field, sensor noise, and communication delay to verify the stability of the vision-language collaborative perception and control closed loop under different conditions.

[0048] During the inference phase, the visual-language collaborative perception model can be deployed on the UAV's onboard computing unit, with the risk adaptive model prediction controller running on the same computing unit or the flight controller's accompanying computer; alternatively, the high-performance visual-language collaborative perception model can be deployed on a ground station, while the onboard unit retains lightweight perception, safety control, and disconnection degradation modules.

[0049] This invention is adaptable to multi-rotor UAVs, vertical takeoff and landing fixed-wing UAVs, or other low-altitude flight platforms with autonomous control interfaces. For different platforms, only the corresponding flight dynamics model, state variables, and control quantity definitions need to be replaced; the semantic world state, task parsing, and risk feedback mechanism can remain unchanged.

[0050] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Any equivalent substitutions, variations, or improvements made within the spirit and principles of this invention should fall within the scope of protection defined by the claims.

Claims

1. A risk adaptive model predictive control method for unmanned aerial vehicles based on visual language perception, characterized in that: The system receives multi-source observations from the UAV, natural language task commands, and flight status; performs task semantic parsing on the task commands; combines the task semantics with visual-language collaborative perception and geometric encoding of the multi-source observations to construct a semantic world state; generates semantic sub-targets and candidate trajectories based on the semantic world state, and determines a reference trajectory; based on the reference trajectory and flight status, performs predictive control rolling solution and executes control variables through a risk adaptive model; updates the semantic world state and reference trajectory in a closed loop based on subsequent observations and control execution results, and executes a safety strategy when safety degradation conditions are met.

2. The method according to claim 1, characterized in that: The task semantic parsing includes identifying target categories or target instances, action primitives, spatial relationships, termination conditions, and safety restrictions in natural language task instructions; the action primitives include one or more of approaching, moving away, crossing, detouring, following, hovering, patrolling, and landing; the spatial relationships include one or more of left, right, front, rear, above, below, adjacent, maintaining distance, and maintaining target visibility; and converting one or more of maximum speed, minimum altitude, no-fly zone, target visibility, arrival accuracy, and task time limit into structured task constraints.

3. The method according to claim 1, characterized in that: The vision-language collaborative perception and geometric coding includes extracting multi-scale visual features from airborne images, establishing a correspondence between task semantics and visual regions through cross-modal feature interaction, obtaining scene geometric information using one or more of depth sensors, binocular vision, monocular depth estimation, laser point clouds, or dense matching, and performing coordinate transformations by combining inertial measurement unit, global navigation satellite system, visual odometry, or flight controller status to obtain target status, obstacle status, free space, and corresponding perception confidence.

4. The method according to claim 1, characterized in that: The semantic world state is updated through temporal memory; the temporal memory includes one or more of the following: historical target state, target appearance or semantic features, obstacle movement trajectory, completed task stages, historical waypoints, and control execution results, and is updated through one or more of the following: recurrent neural network, temporal attention, state filtering, Kalman filtering, or explicit trajectory table; the perceived uncertainty is determined by one or more of the following: target confidence, depth variance, obstacle occupancy entropy, state estimation covariance, and communication delay.

5. The method according to claim 1, characterized in that: The semantic sub-objectives include one or more of the following: target relative position, desired altitude, desired heading, desired speed, target observation azimuth, gimbal view, and mission phase identifier. Based on the semantic sub-objectives, multiple candidate trajectories are generated using a sampling planner, a learning planner, a trajectory optimizer, or a combination thereof. The candidate trajectories are evaluated based on one or more of the following: mission consistency, safety risk, target visibility, energy consumption, trajectory smoothness, terminal state error, and control feasibility, and the reference trajectory is determined.

6. The method according to claim 1, characterized in that: The risk adaptive model predictive controller uses one or more of the UAV's position, velocity, attitude, and angular velocity as state variables, and one or more of thrust, torque, desired attitude, desired angular velocity, desired acceleration, or desired velocity as control variables. In the prediction time domain, it predicts the future state based on the UAV's flight dynamics model and solves the optimal control sequence by minimizing one or more of the following: reference trajectory tracking error, control change, collision risk, and target visibility loss. The solution satisfies one or more of the following: flight state constraints, control variable constraints, and geofencing constraints. After executing at least one control variable in the optimal control sequence in each control cycle, it is re-optimized.

7. The method according to claim 6, characterized in that: The risk adaptive model predictive control dynamically adjusts one or more of the following based on perceived uncertainty: obstacle safety distance, permissible flight speed, control execution step size, and optimization cost weight. The obstacle safety distance is determined based on the basic safety distance and obstacle perceived uncertainty. The risk adaptive model predictive controller sets control barrier function constraints to restrict the predicted state from entering the obstacle safety zone or no-fly zone. When the obstacle perceived uncertainty increases, the state estimation error increases, the dynamic obstacle approaches, or the control solution delay increases, the safety distance is increased or the permissible flight speed is reduced.

8. The method according to claim 1, characterized in that: One or more of the following factors are fed back into the semantic sub-target generation and candidate trajectory planning process: control tracking error, optimization feasibility, constraint activation state, target field of view state and obstacle safety state. When the reference trajectory is infeasible, the tracking error exceeds a preset threshold, the target is about to leave the field of view, or the environmental risk changes, the semantic sub-target or reference trajectory is regenerated; when the perception confidence is lower than a preset threshold, an obstacle intrudes into the safe area, communication is interrupted, the flight status exceeds the safety boundary, or the risk adaptive model predictive controller has no feasible solution, the system enters a safety degradation state and performs at least one of the following: deceleration observation, hovering, ascent obstacle avoidance, backward disengagement, return to a safe waypoint, automatic return to home, or manual takeover.

9. An autonomous mission execution system for unmanned aerial vehicles (UAVs), characterized in that: The system includes a multi-source input and synchronization module for acquiring multi-source observations, natural language task commands, and flight status of the UAV; a task semantic parsing module for extracting task semantics and safety constraints from the natural language task commands; a vision-language collaborative perception and geometric encoding module for combining task semantics and multi-source observations to obtain target state, obstacle state, free space, and perception uncertainty; a semantic world state module for constructing and updating the semantic world state; a trajectory planning module for generating semantic sub-targets, candidate trajectories, and reference trajectories; and a risk adaptive model predictive control module for continuously solving UAV control variables based on the reference trajectory and flight status. The feedback and safety management module is used to update the trajectory planning and control status based on the perception results, control execution results, and optimization feasibility, and to execute safety policies when the safety degradation conditions are met; wherein, each module is configured to execute the method described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it causes the processor to perform the method described in any one of claims 1 to 8.

Images