AI large model control method and device of six-rotor unmanned aerial vehicle

Abstract

The invention relates to the technical field of unmanned aerial vehicle AI control, and discloses an AI large model control method and device for a six-rotor unmanned aerial vehicle, and the method comprises the steps: obtaining a high-precision position and positioning uncertainty; planning an initial operation scheme; the twinborn rehearsal adaptively simulates and evaluates the risk according to the uncertainty; if the risk is low, generating an abstract and presenting a scheme for an operator to confirm; if approving, executing and recording a track, and feeding back deviation to calibrate twinning; if not, executing an expert scheme, and packaging sample feedback to optimize the AI; the device comprises a multi-mode sensing module, a fusion positioning module, an AI decision module, a twinborn rehearsal module, a collaborative supervision module, a flight control module and a man-machine interaction terminal. According to the virtual verification method and device, determinacy rehearsal or robustness rehearsal is dynamically selected and executed by using positioning uncertainty, efficient determinacy rehearsal can be adopted when the credibility is high, robustness rehearsal can be triggered when the credibility is low, and therefore the efficiency and pertinence of virtual verification are improved.

Description

Technical Field

[0001] This application relates to the field of AI control technology for unmanned aerial vehicles (UAVs), specifically to a large-scale AI control method and device for a six-rotor UAV. Background Technology

[0002] In recent years, multi-rotor flight platforms, represented by hexacopter drones, have been increasingly widely used in complex tasks such as power line inspection, emergency search and rescue, and security and bomb disposal due to their high maneuverability and flexibility. These tasks typically require drones to perform high-precision autonomous operations in environments where GPS signals are blocked, such as indoors, in canyons, or between urban buildings. This necessitates improving the autonomous decision-making capabilities and safety of drones in such unstructured, high-risk environments.

[0003] In existing technological applications, a common AI control framework is planning, rehearsing, and then confirming. Specifically, the AI ​​decision-making module first generates an initial operational plan based on data from multimodal sensors (such as LiDAR and visual odometry). Subsequently, to ensure safety, the plan is virtually rehearsed in a digital twin environment to assess its collision or instability risks. Finally, the plan that passes the rehearsal, along with its risk assessment results, is presented to a remote human-machine interface terminal for final supervision and confirmation by the operator before physical execution.

[0004] Existing intelligent dosing technologies for feed production, particularly traditional digital twin simulations, largely rely on deterministic state estimation. This ignores the positioning uncertainty inherent in GPS or BeiDou-denied environments where the fusion positioning module is blocked. When positioning reliability is low, such deterministic simulations can lead to a severe underestimation of risk. An inherent virtual-real discrepancy exists between the digital twin environment and its corresponding physical reality. Inaccurate simulation parameters (such as wind resistance and friction coefficients) can distort the simulation results. When an operator rejects an AI's solution, the expert correction provided is only applicable to that specific instance; the AI ​​decision-making model itself does not learn from this rejection and may repeat the same mistakes in similar future scenarios. Therefore, this invention provides an AI large-scale model control method and device for a six-rotor UAV to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide an AI large-scale model control method and device for hexarotor UAVs. This solves the problem that existing UAV AI control technologies may deviate from physical reality in the digital twin simulation environment of the verification scheme, leading to distorted simulation results. Traditional control systems fail to quantify the uncertainty of the upstream positioning module and transmit it to the downstream decision-making and pre-simulation stages, making it difficult for the system to adjust its safety strategy based on the credibility of its own state.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI large-scale model control method for a hexacoach UAV, comprising the following steps: S1. Obtain high-precision position information and positioning uncertainty of UAV based on multimodal sensors; S2. The AI ​​decision-making module uses the high-precision location information and multimodal sensor data as input to plan the initial operation scheme. S3. The twin simulation module dynamically selects deterministic simulation or robust simulation based on the positioning uncertainty, simulates the initial operation plan, and outputs a risk assessment value. S4. When the risk assessment value is lower than the preset safety threshold, the system generates an interpretable summary and presents the initial work plan to the operator for supervision and confirmation. S5. When the operator approves, the flight control module executes the initial operation plan and records the actual execution trajectory to construct a virtual-real deviation feedback to the twin pre-simulation module for calibrating the simulation parameters of the twin pre-simulation module; S6. When the operator rejects the proposal, the flight control module receives an expert correction plan provided by the operator and executes it. The expert correction plan is then associated with the decision scenario context and used as an expert sample, and fed back to the AI ​​decision module for online optimization of the AI ​​decision module's decision model.

[0007] Preferably, in step S3, the step of the twin pre-simulation module dynamically selecting deterministic pre-simulation or robust pre-simulation based on the positioning uncertainty further includes: The positioning uncertainty is quantified into an uncertainty scalar; The uncertainty scalar is compared with a preset uncertainty threshold; When the uncertainty scalar is lower than the uncertainty threshold, the deterministic pre-simulation is performed; When the uncertainty scalar is greater than or equal to the uncertainty threshold, the robustness pre-simulation is performed.

[0008] Preferably, the step of performing the robustness pre-simulation further includes: Based on the probability distribution formed by the high-precision location information and the positioning uncertainty, the initial state is randomly sampled and the simulation is executed. The risk assessment value is defined as the experience failure rate of the simulation run.

[0009] Preferably, in step S5, the step of constructing the virtual-to-real deviation feedback to the twin pre-simulation module for calibrating the simulation parameters of the twin pre-simulation module further includes: The actual execution trajectory is compared with the simulation trajectory generated by the twin pre-simulation module, and the virtual-real deviation between the actual execution trajectory and the simulation trajectory is calculated. Based on the virtual-real deviation, the simulation parameters inside the digital twin environment are updated in reverse.

[0010] Preferably, the step of associating the expert correction scheme with the decision-making scenario context and using it as an expert sample, and feeding it back to the AI ​​decision-making module for online optimization of the AI ​​decision-making module's decision model, further includes: The expert correction scheme is associated and bound with the decision-making scenario context that leads to the rejection, and the associated and bound data is packaged into the expert sample; The AI ​​decision-making module receives the expert samples and performs online model optimization.

[0011] Preferably, the online model optimization step further includes: The expert samples are added as high-priority examples to the context learning sequence of the large AI model within the AI ​​decision-making module; The expert samples are stored in the expert sample buffer for fine-tuning the model parameters of the large AI model.

[0012] Preferably, in step S3, the step of presenting the initial work plan to the operator for supervision and confirmation further includes: Generate an interpretable summary, which includes the visualization data of the initial work plan, the risk assessment value, and a summary of the rehearsal mode; The initial task plan and the corresponding interpretability summary are sent to the human-computer interaction terminal.

[0013] Preferably, step S2 further includes: The AI ​​decision-making module's large AI model, combined with current task prompts, generates a dynamic risk map; The path planner within the AI ​​decision-making module calculates the optimal path that minimizes the composite cost function based on the dynamic risk map, thereby forming the initial work plan.

[0014] Preferably, step S1 further includes: The fusion localization module uses an extended Kalman filter to fuse multi-source data to obtain a posterior state estimate; The high-precision location information is the location component in the posterior state estimation; Calculate and update the posterior covariance matrix, and use the posterior covariance matrix as the positioning uncertainty.

[0015] This invention also provides an AI large-scale model control device for a hexacoach drone, comprising: The system includes a fusion positioning module for acquiring high-precision location information and positioning uncertainty of the UAV based on multimodal sensors; an AI decision-making module for planning an initial operation plan based on the high-precision location information and multimodal sensor data; and a twin simulation module for acquiring the positioning uncertainty and selecting deterministic or robust simulation based on it, simulating the initial operation plan and outputting a risk assessment value, and receiving the actual execution trajectory to calibrate the simulation parameters of the twin simulation module. The collaborative supervision module is used to submit the initial operation plan to the operator for supervision and confirmation when the risk assessment value is lower than the preset safety threshold, in response to the operator's approval instruction, and to respond to the operator's rejection instruction, in order to receive the expert correction plan provided by the operator, and to feed the expert correction plan back to the AI ​​decision-making module as the expert sample. The flight control module is used to respond to the approval instruction, execute the initial operation plan to generate a real execution trajectory, and to respond to the rejection instruction to execute the expert correction plan.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention obtains the positioning uncertainty output by the fusion positioning module and uses this positioning uncertainty to dynamically select whether to perform deterministic pre-simulation or robust pre-simulation. This adaptive pre-simulation mechanism can use efficient deterministic pre-simulation to save computing power when the positioning confidence is high, and automatically trigger robust pre-simulation to fully assess statistical risks when the positioning confidence is low. Thus, it improves the efficiency and relevance of virtual verification while ensuring security.

[0017] 2. This invention, when an operator rejects the initial work plan planned by the AI, associates the expert correction plan provided by the operator with the decision-making scenario context and packages it into an expert sample, which is then fed back to the AI ​​decision-making module. Through an online learning feedback mechanism, the AI ​​decision-making model can continuously learn from the operator's decision preferences and expert experience, gradually optimizing its decision-making model so that the subsequent work plans it generates are more in line with the intentions of human operators, thereby improving the efficiency of human-machine collaboration and the autonomous decision-making level of the AI ​​model.

[0018] 3. This invention records the actual execution trajectory of the UAV after the plan is approved for implementation and feeds it back to the digital twin pre-simulation module. By calculating the virtual-real deviation between the real trajectory and the simulated trajectory, the simulation parameters inside the digital twin pre-simulation module are calibrated. This self-calibration closed loop ensures that the digital twin environment can iterate itself based on the execution feedback from the physical world, preventing the virtual and real from becoming disconnected, and guaranteeing the long-term accuracy and reliability of the risk assessment performed by the digital twin pre-simulation module. Attached Figure Description

[0019] Figure 1 This is a device architecture diagram of this application; Figure 2 This is a flowchart of the method in this application; Figure 3 This is a closed-loop feedback diagram of this application. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1 -Appendix Figure 3 This application will be described in further detail below.

[0021] See attached document Figure 1 , Figure 1 This is a device architecture diagram according to an embodiment of the present invention. This embodiment provides an AI large-scale model control device for a hexacopter drone, applied to a hexacopter drone or equivalent flight platform, including a multimodal perception module, a fusion positioning module, an AI decision-making module, a twin pre-simulation module, a collaborative supervision module, a flight control module, and a human-machine interaction terminal.

[0022] The multimodal perception module is used to collect multi-source heterogeneous data of the UAV itself and the external environment; the fusion positioning module calculates the high-precision position information of the UAV and the corresponding positioning uncertainty based on the data; the AI ​​decision-making module generates operation plans based on the high-precision position information and perception data.

[0023] The twin simulation module is used to simulate and assess the risks of the operation plan in a virtual environment; the collaborative supervision module is used to present the approved plan to the human-machine interaction terminal and receive operator instructions from the terminal; the flight control module drives the UAV to perform physical operations according to the final instructions from the collaborative supervision module.

[0024] See attached document Figure 2 , Figure 2 This is a flowchart of an AI large-scale model control method for a six-rotor UAV according to an embodiment of the present invention, including the following steps: S1, State Awareness and Fusion Positioning. The UAV acquires multimodal sensor data and uses the fusion positioning module to obtain high-precision position information and reduce positioning uncertainty.

[0025] S2, AI Decision Making and Solution Planning. The AI ​​decision-making module generates a dynamic risk map based on high-precision location information and sensor data, and plans an initial operation plan.

[0026] S3, Adaptive Twin Pre-simulation. The twin pre-simulation module acquires the positioning uncertainty and selects deterministic or robust pre-simulation accordingly, outputting a risk assessment value.

[0027] S4, Human-Machine Collaborative Supervision and Confirmation. If the risk assessment value is lower than the preset threshold, the system will present the work plan to the operator for supervision and confirmation.

[0028] S5, Approval and Execution, and Model Calibration. If the operator approves, the flight control module executes the operation plan; the system records the actual execution trajectory and feeds it back to the twin simulation module for calibrating the simulation parameters of the flight control module.

[0029] S6, Rejection Correction and Online Learning. If the operator rejects the suggestion, the system receives and executes the expert correction plan provided by the operator; simultaneously, the expert correction plan is used as an expert sample and fed back to the AI ​​decision-making module for online optimization of the module's decision-making model.

[0030] See attached document Figure 1 -Appendix Figure 3 For step S1, namely state awareness and fusion positioning, the specific implementation includes: a multimodal perception module acquiring data on the UAV and its operating environment. This multimodal perception module may specifically include an inertial measurement unit (IMU), a visual odometry (VIO), a lidar radar (LiDAR), and a global positioning system (GPS) receiver. The IMU provides high-frequency acceleration and angular velocity measurements, the VIO provides visual motion estimation, the LiDAR provides environmental point cloud data, and the GPS provides global position reference when signal is available.

[0031] The fusion localization module uses an extended Kalman filter (EKF) to fuse the multi-source data to estimate the complete state of the UAV. This state vector x k At time k, it is defined as: x k =[p k ,v k ,q k ] T ; Where, x k This represents the state vector at time k, where k represents time. Let be the position vector of the UAV. For velocity vector, The quaternion used to represent attitude is T, which is a superscript indicating the "transpose" of the vector.

[0032] The fusion positioning module uses IMU measurements (including accelerometer and gyroscope readings) as input and predicts the prior state through a nonlinear state transition function. Simultaneously, the system updates the prior covariance matrix based on the posterior covariance matrix from the previous time step.

[0033] When the fusion positioning module receives external measurements from VIO, LiDAR, or GPS, the system maps the prior state to the measurement space using an observation model. The system calculates the Kalman gain and updates the state estimate using the measurement residuals to obtain the posterior state. This posterior state is the output "high-precision position information" (which contains position components).

[0034] While updating the posterior state, the system calculates the updated posterior covariance matrix P. k∣k : P k∣k =(IK k H k )P k∣k-1 ; Among them, P k∣k Let I be the posterior covariance matrix at time k, and H be the identity matrix. k Is the observation model h(·) in The Jacobian matrix calculated at point K, k P represents the Kalman gain at time k. k Indicates the positional component.

[0035] The posterior covariance matrix is ​​a mathematical quantification of the confidence level of the current posterior state estimate. The diagonal elements (i.e., the variance) of the posterior covariance matrix reflect the individual components (including position p) in the state vector. k The degree of uncertainty of the location is defined as follows. In this embodiment, the posterior covariance matrix is ​​explicitly defined as the "location uncertainty" of the output. This posterior covariance matrix of location uncertainty will serve as a key input for subsequent adaptive twin pre-simulation to determine the pre-simulation strategy.

[0036] For step S2, namely AI decision-making and solution planning, the specific implementation methods include: The AI ​​decision-making module obtains high-precision location information p. k And the acquired multimodal sensor data S k The AI ​​decision-making module integrates a large AI model M. AI For example, multimodal vision-language models. M AI Responsible for parsing S k (including I) cam Image, I ir Infrared, C lidar Environmental semantic information contained in point clouds, etc.

[0037] M AI Based on the current task prompt T prompt (For example, "identify and avoid all obstacles"), analyze the environment, and generate a dynamic risk map R. map The R mapIt is a data structure used to assign a quantified risk rating R to any coordinate point p in the workspace. map (p) This rating takes into account factors such as terrain impassability and proximity of threat sources.

[0038] Based on this dynamic risk map R map The path planner within the AI ​​decision-making module (e.g., A algorithm or RRT algorithm) uses the current position p k Starting from p start With the mission objective point p goal As the endpoint, solve the initial operation plan (π). plan A plan The scheme includes the optimal path π. plan and associated job action sequence A plan The optimal path π plan The goal is to minimize a composite cost function J(π): J(π)=∫ π (w d ·L(s)+w r ·R map (p(s)))ds; Where J(π) is the composite cost function, π is a candidate path; s is the arc length on the path; L(s) is the path length or energy cost at that point; R map (p(s)) is the risk map value corresponding to path point p(s); p(s) represents the point corresponding to arc length s on path π, w d and w r It is a non-negative weighting factor used to balance distance (energy consumption) costs and risk costs, ∫ π …ds represents the line integral along the candidate path π with respect to the arc length s.

[0039] AI decision-making module (especially M) AI It has an online learning interface for receiving "expert demonstration samples". exp The expert demonstrated sample d exp It is a data pair. It links the specific scenario that led to the veto with the "expert correction plan" provided by the operator in that scenario.

[0040] When the AI ​​decision-making module receives d through this interface exp At that time, M AI Perform online model optimization. The specific implementation of this optimization includes: [details of the optimization method]. exp Added to M as a high-priority example AI In the context learning sequence, to guide its decision-making in subsequent reasoning; or, to d exp Store in expert sample buffer Dexp Used for M AI The model parameters are fine-tuned using mini-batch, low-learning-rate gradient tuning. This mechanism ensures that M... AI It can continuously learn the decision-making preferences of operators, so that their planned solutions gradually converge with expert experience.

[0041] For step S3, namely adaptive twin pre-simulation, the specific implementation methods include: The initial operation plan (π) received and output by the twin pre-simulation module plan A plan This twin pre-simulation module operates in a high-fidelity digital twin environment E. DT The solution was virtually verified in China. DT Internally, it constructs an environment model, a drone dynamics model, and a sensor model that are consistent with the real physical world.

[0042] The twin pre-simulation module first obtains the output "localization uncertainty", namely the posterior covariance matrix P. k|k The system quantifies the position-related components of the matrix into an uncertainty scalar U by calculating the trace or determinant. p .

[0043] The system will use the uncertainty scalar U p With a preset uncertainty threshold T uncert A comparison is made. The result of this comparison will dynamically determine the rehearsal strategy to be adopted, i.e., whether to choose "deterministic rehearsal" or "robust rehearsal".

[0044] If U p <T uncert This indicates that the fusion positioning module has a good understanding of the current location p. k The estimate has high reliability. At this point, the system performs a deterministic pre-simulation. The simulation engine Φ of the twin pre-simulation module... sim p k As the sole initial state, perform a high-fidelity simulation to generate the simulation trajectory Z. sim The system is based on this Z. sim Check for risks such as collisions and instability, and calculate a deterministic risk assessment value R. value .

[0045] If U p ≥T uncert This indicates that the current location has low reliability, and a single p k The estimate is insufficient to represent the true state. At this point, the system automatically triggers a robustness simulation, which is specifically implemented as a Monte Carlo simulation.

[0046] In the robustness simulation, the system no longer uses a single p k Instead, it is based on the provided complete probability distribution. Randomly sample N possible initial states {p i} i＝1...N These N samples represent the N most likely actual initial positions of the drone under the current uncertainty.

[0047] Simulation Engine Φ sim For these N different initial states {p i}, execute N simulation runs respectively, and obtain N sets of simulation trajectories {Z sim,i At this point, the risk assessment value R is... value Defined as the empirical failure rate of these N simulations: Among them, R value Here, V(Z) represents the risk assessment value, i is the sample index, and N is the total number of samples. sim,i Let ) be a failure indicator function. If the i-th simulation trajectory Z sim,i If any predefined failure mode is triggered (e.g., collision with an obstacle, flight attitude instability, deviation from the path safety boundary), then V(Z) sim,i If ) = 1, otherwise 0. This R value The work plan (π) was quantified. plan A plan Statistical risks under the current uncertainty of positioning.

[0048] The twin simulation module also has self-calibration capabilities. This twin simulation module is used to receive feedback data from the approval and execution process, namely the "real execution trajectory" Z recorded by the drone after performing the task in the physical world. actual Upon receiving Z actual Subsequently, the twin simulation module compares it with the final approved "simulation trajectory" Z. sim Compare them.

[0049] The system calculates the "virtual-real deviation" Δ between the two trajectories. Z =Z actual -Z sim The deviation Δ Z Used as loss signal L DT (For example, Δ) Z The mean square error (MSE) is used to drive the optimization process.

[0050] This optimization process updates the digital twin environment E in reverse. DT Internal simulation parameters θ DT These parameters θ DT This refers to the physical properties that affect the simulation results, such as the drag coefficient in the environmental model, the motor response time or friction coefficient in the UAV dynamics model. In this way, digital twin E... DTIt can iterate and calibrate itself based on real-world execution feedback, making its simulation of physical reality increasingly accurate.

[0051] The twin simulation module will ultimately calculate the risk assessment value R. value Used for human-machine collaborative supervision and confirmation.

[0052] For step S4, namely human-machine collaborative supervision and confirmation, the specific implementation methods include: The collaborative supervision module receives the final risk assessment value R from the adaptive twin pre-simulation. value The system will assign this risk assessment value R value With the preset safety threshold T risk Compare them.

[0053] R value <T risk ; Among them, T risk This represents the maximum level of risk that the system can autonomously accept.

[0054] If R value ≥T risk This indicates that the twin simulation has identified the initial operational plan (π) plan A plan If the proposed plan poses an unacceptable risk, the collaborative supervision module automatically rejects it and prevents it from being presented to the operator. The system may also choose to send a replanning instruction to the AI ​​decision-making module.

[0055] If R value <T risk This indicates that the solution has passed the system's autonomous security review and has been deemed suitable for human decision-making. At this point, the collaborative supervision module initiates the decision presentation process.

[0056] To assist operator H op Decision-making is carried out, and the collaborative supervision module automatically generates interpretable summaries E. xai This interpretable summary is a composite data package, a comprehensive presentation of AI decision-making and simulation results. Specifically, it includes one or more of the following data: task plan (π plan A plan Visualized data, such as in 3D environment models or dynamic risk maps. map The planned path π is displayed on top. plan ; Output quantitative risk assessment value R value And the uncertainty threshold T corresponding to this quantitative risk assessment value. risk ; The simulation mode summary, for example, clearly informs the operator whether the simulation is based on a high-confidence "deterministic simulation" or a low-confidence "robust simulation" and the number of samples N; key risk points identified during the simulation, such as the minimum safety margin of the path and obstacles, or the area with the highest probability of failure in the robust simulation.

[0057] The collaborative supervision module will use the initial work plan (π) plan A plan Together with its corresponding interpretable summary E xai The data is transmitted to a human-machine interface terminal via a data link. This human-machine interface terminal can be a display interface of a ground control station or a portable control device.

[0058] The system sends data to operator H on the human-computer interaction terminal. op Clearly present the solution content and summary information, and provide "Approval" (U) approve ) and "veto" (U reject Two clear operation options were provided. The system then paused the autonomous operation process and waited for operator H. op Make the final instruction U H .

[0059] For step S5, namely approval execution and model calibration, the specific implementation methods include: When the collaborative supervision module receives a message from the human-computer interaction terminal representing operator H... op Intended approval directive U H =U approve This is triggered at that time. In response to the approval instruction, the collaborative monitoring module confirms the initial work plan (π). plan A plan The final execution plan is then implemented by the collaborative supervision module, which then assigns the plan (π) to the appropriate level. plan A plan The data is sent to the flight control module.

[0060] The flight control module receives the scheme and parses it into a sequence of low-level control commands executable by the UAV's flight dynamics system. For example, the path π plan This is converted into a series of desired velocity and attitude setpoints, and the operation action A is then converted into a series of desired velocity and attitude setpoints. plan This is converted into specific control signals for the robotic arm or bomb disposal tools.

[0061] The flight control module drives the physical actuators (including six or more rotor motors and other operating components) of the hexacopter UAV to begin executing the operation plan in a real physical environment. While the UAV is performing the physical operation, the system activates the "real execution trajectory" recording mechanism.

[0062] This recording mechanism specifically includes: during the UAV's physical flight, the system continuously runs the fusion positioning module in step S100. This fusion positioning module calculates the posterior state estimate in real time during the physical flight. (It contains the drone's real position p at every moment) k Speed ​​v k and posture q k The data is collected, aggregated, and stored in chronological order.

[0063] The aforementioned collected and stored time-series state data constitutes the "actual execution trajectory" Z. actual The Z actual The data structure was designed to work with the "simulation trajectory" Z generated by the twin pre-simulation module. sim Maintain consistency in dimensions and format to facilitate direct comparison later.

[0064] After the physical task is completed (or a key phase is completed), the system will record the complete Z... actual Feedback is transmitted to the twin pre-simulation module, and the Z-axis of this feedback... actual It is explicitly used for its "twin model self-calibration" function. The twin pre-simulation module receives Z... actual Then, it is compared with the Z stored in the pre-rehearsal phase and ultimately approved for execution. sim A mathematical comparison is performed to calculate the "virtual-to-real deviation" Δ between the two. Z This, in turn, drives the development of digital twin environments. DT Internal simulation parameters θ DT Optimization and calibration.

[0065] Regarding step S6, namely, rejecting the amendment and online learning, the specific implementation methods include: When the collaborative supervision module receives a message from the human-computer interaction terminal representing operator H... op Intentional veto instruction U H =U reject It is triggered at that time.

[0066] In response to the veto command, the collaborative monitoring module immediately aborted the initial operational plan (π) originally intended to be submitted to the flight control module. plan A plan To prevent its physical execution, the system activates an interactive correction interface on the human-machine interface terminal. This interface allows operator H... op Modify the AI ​​solution or provide entirely new solutions from human experts.

[0067] The interactive correction interface specifically includes: allowing operators to view the presented visual map (e.g., R...). map Redraw the path π using touch or cursor dragging. humanAlternatively, a menu can be provided for operators to modify key operation actions A. human For example, adjusting the gripping angle of the robotic arm or the hovering height of the drone.

[0068] The system captures the operator's final input on the interface and formats it as an "expert correction scheme". The system executes the expert correction scheme, and the collaborative supervision module sends the expert correction scheme to the flight control module, which drives the UAV to immediately execute the operator's instructions.

[0069] The system associates and binds the expert's revised plan with the decision-making context that led to the rejection. This decision-making context specifically includes p k S k T prompt The system packages the aforementioned bound data into an "expert demonstration sample". exp :d exp =((S) k ,p k ,T prompt )→(π human A human )); Where, d exp Indicates expert demonstration sample, S k p represents part of the context of the decision-making scenario that leads to the veto. k T represents part of the context of the decision-making scenario that led to the veto. prompt This represents part of the context of the decision-making scenario that led to the veto, (S k ,p k ,T prompt ) represents the decision-making scenario (i.e., the problem faced by AI in making decisions), (π) human A human π represents the expert solution provided by the operator (i.e., the answer the operator considers correct). human A represents the path or planning portion of an expert's solution. human This refers to the action portion of the expert solution.

[0070] The system will use the expert demonstration sample d exp The feedback is transmitted to the AI ​​decision-making module. This feedback's d exp It is explicitly used in its "online model optimization" function. AI Decision Module M AI Receive the sample and store it in the expert sample buffer D. exp This is used to update M through in-context learning or mini-batch model fine-tuning. AI The decision preferences are considered to make the subsequent work plans more in line with the operator's decision-making patterns and expert experience.

Claims

1. A large-scale AI model control method for a six-rotor unmanned aerial vehicle, characterized in that, Includes the following steps: S1. Obtain high-precision position information and positioning uncertainty of UAV based on multimodal sensors; S2. The AI ​​decision-making module uses the high-precision location information and multimodal sensor data as input to plan the initial operation scheme. S3. The twin simulation module dynamically selects deterministic simulation or robust simulation based on the positioning uncertainty, simulates the initial operation plan, and outputs a risk assessment value. S4. When the risk assessment value is lower than the preset safety threshold, the system generates an interpretable summary and presents the initial work plan to the operator for supervision and confirmation. S5. When the operator approves, the flight control module executes the initial operation plan and records the actual execution trajectory to construct a virtual-real deviation feedback to the twin pre-simulation module for calibrating the simulation parameters of the twin pre-simulation module; S6. When the operator rejects the proposal, the flight control module receives an expert correction plan provided by the operator and executes it. The expert correction plan is then associated with the decision scenario context and used as an expert sample, and fed back to the AI ​​decision module for online optimization of the AI ​​decision module's decision model.

2. The AI ​​large-scale model control method for a hexacopter UAV according to claim 1, characterized in that, In step S3, the step of the twin pre-simulation module dynamically selecting deterministic pre-simulation or robust pre-simulation based on the positioning uncertainty further includes: The positioning uncertainty is quantified into an uncertainty scalar; The uncertainty scalar is compared with a preset uncertainty threshold; When the uncertainty scalar is lower than the uncertainty threshold, the deterministic pre-simulation is performed; When the uncertainty scalar is greater than or equal to the uncertainty threshold, the robustness pre-simulation is performed.

3. The AI ​​large-scale model control method for a hexarotor UAV according to claim 2, characterized in that, The step of performing the robust pre-simulation further includes: Based on the probability distribution formed by the high-precision location information and the positioning uncertainty, the initial state is randomly sampled and the simulation is executed. The risk assessment value is defined as the experience failure rate of the simulation run.

4. The AI ​​large-scale model control method for a hexacopter UAV according to claim 1, characterized in that, In step S5, the step of constructing the virtual-to-real deviation feedback to the twin pre-simulation module for calibrating the simulation parameters of the twin pre-simulation module further includes: The actual execution trajectory is compared with the simulation trajectory generated by the twin pre-simulation module, and the virtual-real deviation between the actual execution trajectory and the simulation trajectory is calculated. Based on the virtual-real deviation, the simulation parameters inside the digital twin environment are updated in reverse.

5. The AI ​​large-scale model control method for a hexacopter UAV according to claim 1, characterized in that, The step of associating the expert correction scheme with the decision-making scenario context and using it as an expert sample, and feeding it back to the AI ​​decision-making module for online optimization of the AI ​​decision-making module's decision model, further includes: The expert correction scheme is associated and bound with the decision-making scenario context that leads to the rejection, and the associated and bound data is packaged into the expert sample; The AI ​​decision-making module receives the expert samples and performs online model optimization.

6. The AI ​​large-scale model control method for a six-rotor UAV according to claim 5, characterized in that, The online model optimization step further includes: The expert samples are added as high-priority examples to the context learning sequence of the large AI model within the AI ​​decision-making module; The expert samples are stored in the expert sample buffer for fine-tuning the model parameters of the large AI model.

7. The AI ​​large-scale model control method for a hexarotor UAV according to claim 1, characterized in that, In step S3, the step of presenting the initial work plan to the operator for supervision and confirmation further includes: Generate an interpretable summary, which includes the visualization data of the initial work plan, the risk assessment value, and a summary of the rehearsal mode; The initial task plan and the corresponding interpretability summary are sent to the human-computer interaction terminal.

8. The AI ​​large-scale model control method for a hexarotor UAV according to claim 1, characterized in that, Step S2 further includes: The AI ​​decision-making module's large AI model, combined with current task prompts, generates a dynamic risk map; The path planner within the AI ​​decision-making module calculates the optimal path that minimizes the composite cost function based on the dynamic risk map, thereby forming the initial work plan.

9. The AI ​​large-scale model control method for a hexarotor UAV according to claim 1, characterized in that, Step S1 further includes: The fusion localization module uses an extended Kalman filter to fuse multi-source data to obtain a posterior state estimate; The high-precision location information is the location component in the posterior state estimation; Calculate and update the posterior covariance matrix, and use the posterior covariance matrix as the positioning uncertainty.

10. An AI large-scale model control device for a hexacopter unmanned aerial vehicle (UAV), applied to the AI ​​large-scale model control method for a hexacopter UAV as described in any one of claims 1-9, characterized in that, include: The fusion positioning module is used to acquire high-precision position information and reduce positioning uncertainty of the UAV based on multimodal sensors; The AI ​​decision-making module is used to plan the initial operation scheme based on the high-precision location information and multimodal sensor data; The twin simulation module is used to acquire the positioning uncertainty and select deterministic simulation or robust simulation accordingly, simulate the initial operation plan and output risk assessment value, and receive the actual execution trajectory to calibrate the simulation parameters of the twin simulation module. The collaborative supervision module is used to submit the initial operation plan to the operator for supervision and confirmation when the risk assessment value is lower than the preset safety threshold, in response to the operator's approval instruction, and to respond to the operator's rejection instruction, in order to receive the expert correction plan provided by the operator, and to feed the expert correction plan back to the AI ​​decision-making module as the expert sample. The flight control module is used to respond to the approval instruction, execute the initial operation plan to generate a real execution trajectory, and to respond to the rejection instruction to execute the expert correction plan.

Images