A large model-based unmanned aerial vehicle personalized interaction control system and method

Abstract

This invention provides a personalized interactive control system and method for unmanned aerial vehicles (UAVs) based on a large model. The method includes the following steps: in response to interaction events with a user, acquiring current task context information, explicit feedback data, and implicit feedback data; based on the current task context information, determining a target contextualized preference unit from multiple stored contextualized preference units; generating new evidence for target control parameters directly based on the explicit feedback data, or indirectly based on the implicit feedback data through historical trajectory retrieval; and based on the new evidence and its credibility, performing online fusion and updating of the preference estimates and uncertainty measures of the corresponding control parameters in the target contextualized preference unit, so that the updated uncertainty measure is lower than the original. According to this invention, through a dual-channel sparse feedback learning mechanism, the system can effectively learn online from two types of extremely low-frequency natural interaction events.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, specifically to a personalized interactive control system and method for UAVs based on a large model. Background Technology

[0002] Currently, the development and integration of intelligent interaction technology for unmanned aerial vehicles (UAVs) mainly revolves around two major directions. On the one hand, natural language command parsing technology based on large language models can transform high-level, fuzzy task commands issued by users into structured, executable standard task sequences, realizing a semantic bridge from human intentions to machine-operable steps. On the other hand, environmental perception and real-time obstacle avoidance technologies based on the fusion of multiple sensors such as computer vision and lidar enable UAVs to have low-level autonomous capabilities for local reactions and path adjustments in dynamic and complex environments. Together, these technologies constitute the basic technological paradigm for modern UAV intelligent control systems to achieve automated task execution, enabling UAVs to respond to commands and fly safely.

[0003] However, the aforementioned mainstream technological paradigms have inherent limitations and are difficult to meet the deep-seated needs of advanced human-machine collaboration for adaptability, efficiency, and experience. First, existing systems are essentially static and generic, and their decision-making logic is usually based on preset rules or generic models. They completely lack modeling of the user dimension, cannot identify and distinguish different operators, and do not have the ability to learn and adapt to individual unique operating habits and decision-making preferences.

[0004] Secondly, lacking the ability to continuously remember and evolve, each human-computer interaction is treated as an independent event. The system cannot accumulate experience or understand context from historical interaction data, such as the meaning of instructions from the previous interaction, resulting in interaction efficiency failing to improve effectively with increased usage frequency. Finally, general machine learning methods directly transferred from other fields, such as large-scale supervised learning or online reinforcement learning, often struggle to simultaneously meet the extreme engineering constraints faced in actual UAV deployments, including but not limited to: the strong resource limitations of onboard computing platforms, the extreme sparsity and unstructured nature of user feedback signals in real-world scenarios, the strong real-time requirements for flight decisions, and the absolute requirements for flight safety. These factors collectively lead to existing systems exhibiting amnesia and indiscriminate behavior in practical applications. Therefore, this invention researches and designs a personalized interactive control system and method for UAVs based on a large model. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the existing UAV interactive control system in lacking user dimension modeling and continuous learning capabilities, thereby providing a UAV personalized interactive control system and method based on a large model.

[0006] To address the above problems, this invention provides a personalized interactive control method for unmanned aerial vehicles (UAVs) based on a large model, comprising the following steps: S1: In response to user interaction events, obtain current task context information, explicit feedback data, and implicit feedback data; wherein, the explicit feedback data is the parameter correction amount generated when the user manually adjusts the drone flight parameters through the remote controller, and the implicit feedback data is natural language instructions generated with reference to historical execution experience; S2: Based on the current task context information, determine the target contextual preference unit from multiple stored contextual preference units; wherein, each contextual preference unit is associated with the corresponding task context and includes: user preference information for at least one UAV control parameter, the preference information including at least: preference estimate and its uncertainty measure; S3: New evidence for the target control parameters is generated directly from the explicit feedback data or indirectly from the implicit feedback data through historical trajectory retrieval. S4: Based on the new evidence and its credibility, the preference estimate and uncertainty measure of the corresponding control parameter in the target contextualized preference unit are fused and updated online, so that the updated uncertainty measure is lower than that before the update.

[0007] Preferably, the new evidence for indirectly generating target control parameters based on implicit feedback data through historical trajectory retrieval includes: Parse the natural language instructions, extract their spatial behavior semantics, and convert them into query vectors; The query vector is matched with the feature vectors of multiple historical flight trajectories stored in the memory graph. The memory graph includes a semantic layer that records the contextual relationships of the task and a trajectory pattern layer that stores trajectory pattern data. Retrieve semantically similar historical trajectories based on the matching results; The new evidence is generated based on the control parameters used when the similar historical trajectory was successfully executed. The similarity matching uses cosine similarity, which measures the consistency of vector directions.

[0008] Preferably, in step S4, the uncertainty measure is updated online through fusion as follows: The updated preference estimate is obtained by weighting the previous preference estimate and the new evidence, wherein the weight of the new evidence is negatively correlated with its observation noise level, and the weight of the previous preference estimate is negatively correlated with its current uncertainty measure; and the updated, numerically reduced uncertainty measure is calculated based on the previous uncertainty measure and the observation noise level.

[0009] Preferably, after updating the preference information, a personalized flight trajectory decision-making step is also included: Within the solution space that satisfies all inviolable hard safety constraints, an initial safe flight trajectory is generated; Construct an optimization objective function, which includes a first term and a second term. The first term is used to minimize the modification of the initial safe flight trajectory to maintain smoothness and safety, and the second term is used to drive the trajectory parameters closer to the updated preference estimate. The influence weight of the second term is negatively correlated with the updated uncertainty measure. By solving the optimization objective function, the initial safe flight trajectory is adjusted to generate a personalized flight trajectory; The personalized flight trajectory is verified for safety in real time. If the verification fails, a safety rollback mechanism is triggered to output a backup trajectory that meets all hard safety constraints and is closest to the personalized flight trajectory.

[0010] Preferably, the triggering safety rollback mechanism is as follows: among the feasible solutions of the initial safe flight trajectory, the waypoint sequence with the smallest difference from the personalized flight trajectory is selected as the backup trajectory.

[0011] Preferably, the method further includes a training phase: pre-training the initial parameters of the contextualized preference unit by intervening in the historical interaction data of the annotation and context annotation; subsequently, in a simulation environment, end-to-end reinforcement learning is used to jointly fine-tune the online fusion update strategy and the personalized flight trajectory decision strategy with the comprehensive reward consisting of task efficiency and simulated user satisfaction as the optimization objective.

[0012] This invention also provides a large-model-based personalized interactive control system for unmanned aerial vehicles (UAVs), including the large-model-based personalized interactive control method for UAVs described in the preceding claim, comprising: The task context and feedback perception module is used to respond to user interaction events and obtain current task context information, explicit feedback data, and implicit feedback data. The explicit feedback data is the parameter correction amount generated when the user manually adjusts the drone's flight parameters via the remote controller, and the implicit feedback data is natural language commands generated with reference to historical execution experience. The contextualized preference management module is used to store multiple contextualized preference units and determine the target contextualized preference unit based on the current task context information; wherein, each contextualized preference unit is associated with a corresponding task context and includes user preference information for at least one UAV control parameter, the preference information including at least the preference estimate and its uncertainty measure; The online learning module is communicatively connected to the task context and feedback perception module and the contextualized preference management module. It is used to generate new evidence for the target control parameters directly based on the explicit feedback data or indirectly based on the implicit feedback data through historical trajectory retrieval. Based on the new evidence and its credibility, it performs online fusion and update of the preference estimates and uncertainty measures of the corresponding control parameters in the target contextualized preference unit, so that the updated uncertainty measure is lower than that before the update.

[0013] Preferably, it further includes a memory association module, the memory association module comprising: A memory graph storage unit is used to store a hierarchical memory graph, which includes a semantic layer for recording tasks, users, location entities and relationships, and a trajectory pattern layer for storing compressed feature vectors of historical successful flight trajectories. The trajectory retrieval unit is used to respond to the implicit feedback data, map the current instruction into a spatial semantic query vector, perform similarity matching between the query vector and the historical trajectory feature vector in the trajectory pattern layer, and retrieve similar historical trajectories based on the matching results. The trajectory retrieval unit uses a cosine similarity algorithm for matching.

[0014] Preferably, the online learning module is used to: take a weighted average of the previous preference estimate and the new evidence to obtain the updated preference estimate, wherein the weight allocation is inversely proportional to the respective uncertainty level; and calculate a new uncertainty measure whose value is less than the previous uncertainty measure.

[0015] Preferably, it further includes a hierarchical decision-making module, the hierarchical decision-making module comprising: The safety planning unit is used to generate an initial safe flight trajectory within the solution space that satisfies all hard safety constraints. The personalized optimization unit is communicatively connected to the contextualized preference management module and is used to construct and solve an optimization objective function. This function minimizes the modification of the initial safe flight trajectory while adding a personalized driving term weighted by preference confidence to generate a personalized flight trajectory. The safety verification and rollback unit is used to perform rapid simulation verification of the personalized flight trajectory, and to initiate rollback logic to output a backup trajectory that meets all hard safety constraints when verification fails; the safety verification and rollback unit is equipped with a UAV dynamics model to perform forward simulation of the candidate trajectory to predict constraint violations.

[0016] The personalized interactive control method for unmanned aerial vehicles based on a large model provided by this invention has the following beneficial effects: 1. This invention employs a dual-channel sparse feedback learning mechanism. The explicit intervention channel captures occasional manual intervention behaviors by the user, while the implicit semantic channel parses the user's fuzzy reference commands based on historical experience. This enables the system to learn effectively online from two types of extremely low-frequency natural interaction events. The probabilistic data structure of micro-preference clusters models each control parameter preference as a Gaussian distribution. Through the Bayesian update formula, the uncertainty is monotonically reduced, mathematically ensuring that each effective update necessarily reduces cognitive uncertainty, achieving a progressive learning effect of becoming more knowledgeable with use. This overcomes the bottleneck of personalized functions being difficult to deploy due to data sparsity, enabling drones to truly possess the evolutionary ability to accumulate experience from sporadic interaction fragments. 2. This invention also uses probabilistic representation to explicitly link the system output to the internal state: the decision basis can be traced back to the Gaussian distribution parameters under a specific context key, the source identifier of dual-channel evidence, and the historical trajectory of Bayesian updates; transparency enhances user trust in the system, and when the UAV makes a flight decision that conforms to the operator's habits, the system can explain when, where, and what kind of interactive experience the decision originated from; when deviations occur, variance analysis can also be used to locate the parameter dimensions with higher cognitive uncertainty, providing a clear path for error analysis and system debugging, and significantly improving the maintainability of the system and the reliability of human-machine collaboration; 3. This invention also enables the system to understand natural language commands containing historical references, such as "fly as before" and "execute in the manner of checking target A," through implicit semantic channels. This transforms ambiguous semantic descriptions into computable probabilistic update evidence. This capability increases interaction efficiency with usage frequency. Users do not need to repeatedly describe the desired flight parameters in detail. Instead, they can convey their intentions through concise contextual references, significantly reducing the cognitive load of interaction and improving the tacit understanding and fluency of long-term use. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0018] like Figure 1 As shown, this invention provides a personalized interactive control method for unmanned aerial vehicles (UAVs) based on a large model, which includes the following steps: S1: In response to user interaction events, obtain current task context information, explicit feedback data, and implicit feedback data; wherein, the explicit feedback data is the parameter correction amount generated when the user manually adjusts the drone flight parameters through the remote controller, and the implicit feedback data is natural language instructions generated with reference to historical execution experience; S2: Based on the current task context information, determine the target contextual preference unit from multiple stored contextual preference units; wherein, each contextual preference unit is associated with the corresponding task context and includes: user preference information for at least one UAV control parameter, the preference information including at least: preference estimate and its uncertainty measure; S3: New evidence for the target control parameters is generated directly from the explicit feedback data or indirectly from the implicit feedback data through historical trajectory retrieval. S4: Based on the new evidence and its credibility, the preference estimate and uncertainty measure of the corresponding control parameter in the target contextualized preference unit are fused and updated online, so that the updated uncertainty measure is lower than that before the update.

[0019] Specifically, in response to user interaction events, the system acquires current task context information, explicit feedback data, and implicit feedback data. Explicit feedback data consists of parameter corrections generated when the user manually adjusts the drone's flight parameters via remote control, while implicit feedback data consists of natural language commands generated based on historical execution experience. Based on the current task context information, a target contextualized preference unit is determined from multiple stored contextualized preference units. Each contextualized preference unit is associated with a corresponding task context and includes: user preference information for at least one drone control parameter, which includes at least: a preference estimate and its uncertainty measure. New evidence for the target control parameter is generated directly from explicit feedback data or indirectly through historical trajectory retrieval based on implicit feedback data. Based on the new evidence and its credibility, the preference estimate and uncertainty measure of the corresponding control parameter in the target contextualized preference unit are updated online, resulting in an updated uncertainty measure lower than the original one.

[0020] Interactive events are triggered by users at ground control stations or mobile terminals, including but not limited to voice command input, manual takeover by remote control, and task target specification. Current task context information includes three elements: environmental context, task type, and target object. The environmental context is collected in real time by airborne sensors, including indoor and outdoor signs, light intensity levels, and obstacle density levels. The task type is output by the natural language command parsing module, including categories such as static inspection, dynamic tracking, and detailed observation. The target object is determined by the visual recognition module, including entity types such as wind turbine blades, photovoltaic panels, and specific mobile vehicles.

[0021] Explicit feedback data is acquired through the remote control signal acquisition module. When the user presses the manual takeover button on the remote control, the system records the takeover start time t. start and the end time t endDuring this period, flight control status data is sampled at a frequency of 10Hz, and the difference vector Δs between the automatic control parameters and the user-manually set parameters is calculated, including altitude correction Δh, speed correction Δv, and heading angle correction Δψ. Implicit feedback data is obtained by acquiring the raw audio stream through the speech recognition module, which is then converted into text commands by a large language model. The semantic parsing unit identifies the historical reference semantics within these commands, such as expressions like "fly as before" or "execute according to the method for checking target A."

[0022] The contextualized preference unit is implemented using in-memory structured objects. Each unit contains three fields: a context key, a set of preference parameters, and metadata. The context key is generated by a learnable hash function, and the embedding vectors of the environmental context, task type, and target object are concatenated and input into a small neural network f. hash The symbol function generates a ±1 binary hash code, and similar scenarios produce keys with close Hamming distances. The preference parameter set stores the control parameters via a Gaussian distribution N(μ,σ). 2 ), where the mean μ is a floating-point number representing the preference estimate, and the variance σ 2 Uncertainty measures are expressed as floating-point numbers; metadata includes the last update timestamp and access frequency counter, and LRU cache management strategy is supported.

[0023] There are two pathways for generating new evidence. The explicit pathway directly uses the parameter correction Δs as the observation value o. When the implicit pathway is activated, the system parses the spatial behavior semantics in the natural language instructions, converts them into query vectors, retrieves similar historical trajectories in the memory graph, and extracts the control parameters of the historical trajectory when it was successfully executed as indirect evidence. The online fusion update adopts a Bayesian update mechanism, and the update formula is:

[0024]

[0025] In the formula, This is a preset observation noise variance, characterizing the credibility of the evidence. For example, it can be set to a smaller value (e.g., 1) for manual intervention and a larger value (e.g., 25) for indirect evidence. A new uncertainty measure is calculated using this formula. It must be less than the original σ 2 and The system captures sparse feedback signals through a dual-channel mechanism, uses a contextualized Gaussian distribution as a knowledge carrier, and utilizes Bayes' theorem to achieve monotonically decreasing updates of uncertainty. This allows the system's understanding of individual preferences to gradually converge from initial ambiguity to stable certainty, enabling effective learning from extremely sparse and unstructured interactions. Cognitive updates can be generated with a single interaction, solving the bottleneck of personalized functions being difficult to deploy due to data acquisition difficulties, while ensuring extremely low computational overhead and adapting to environments with limited onboard resources.

[0026] In some implementations, the process of indirectly generating new evidence of target control parameters based on implicit feedback data through historical trajectory retrieval includes: parsing the natural language instruction, extracting its spatial behavior semantics and converting it into a query vector; performing similarity matching between the query vector and feature vectors of multiple historical flight trajectories stored in a memory graph, wherein the memory graph includes a semantic layer recording task context relationships and a trajectory pattern layer storing trajectory pattern data; retrieving semantically similar historical trajectories based on the matching results; and generating the new evidence based on the control parameters when the similar historical trajectories were successfully executed; wherein the similarity matching uses cosine similarity calculation, which measures the consistency of vector directions.

[0027] Specifically, natural language instructions are parsed, their spatial behavioral semantics are extracted and converted into query vectors; the query vectors are then matched with feature vectors of multiple historical flight trajectories stored in a memory graph, which includes a semantic layer recording task context relationships and a trajectory pattern layer storing trajectory pattern data; semantically similar historical trajectories are retrieved based on the matching results; new evidence is generated based on the control parameters when similar historical trajectories are successfully executed; the similarity matching uses cosine similarity calculation, which measures the consistency of vector directions.

[0028] Specifically, the natural language instruction parsing uses a pre-trained large-scale language model as the encoder. The model is selected from lightweight variants such as BERT-base or Distil BERT; in practical applications, other models can also be chosen. After the input instruction text T is segmented, the output vector of the CLS marker bits is taken as the semantic representation vcmd∈R^ dcmd Dimension d cmd The value is 768. For commands containing spatial behavior descriptions, such as orbiting, low-altitude flight, and close-range observation, the system further processes them through a lightweight semantic mapping network g. This network has a two-layer fully connected structure with a hidden layer dimension of 256 and an output layer dimension of d. trace The value is 128, using the ReLU activation function and batch normalization layer.

[0029] Specifically, the memory graph is stored in an onboard solid-state drive or memory database using a two-layer heterogeneous graph structure; the semantic layer stores entity relations in the form of triples, with entity types including user, task, location, and object, and relation types including execution, occurrence, and action, etc., implemented using RDF or attribute graph models; the trajectory pattern layer stores compressed representations of historical successful flight trajectories, with each trajectory processed by a lightweight autoencoder to generate a feature vector v. trace ∈R 128 It also stores sparse keyframe sequences for trajectory reconstruction; the two layers are linked by bidirectional pointers, which are 64-bit integer identifiers.

[0030] Specifically, the trajectory autoencoder employs a symmetric fully connected feedforward network. The encoder consists of an input layer, two fully connected layers, and a bottleneck layer. The input layer dimension n corresponds to the flattened features of the trajectory keyframes. Keyframe extraction uses a curvature-based adaptive sampling algorithm, with dense sampling where the trajectory curvature changes drastically and sparse sampling where it is flat. The first fully connected layer outputs a dimension of 64, and the second fully connected layer outputs a dimension of 32, both using the ReLU activation function. The bottleneck layer outputs a dimension of 16, using a linear activation function, and the output is the trajectory feature vector. The decoder is the symmetric inverse of the encoder, and the reconstructed output layer has the same dimension as the input layer. The network training uses a mean squared error loss function, learning a compact representation by minimizing the L2 distance between the input trajectory and the reconstructed trajectory.

[0031] Specifically, in the similarity matching stage, the query vector q is compared with all historical trajectory feature vectors in the trajectory pattern layer. The cosine similarity is calculated using the following formula:

[0032] The calculation results range from [-1, 1]. The closer the value is to 1, the higher the directional consistency and the more similar the spatial behavior patterns. The system returns the K candidate trajectories with the highest similarity. The value of K is usually set to 5 to 10. In practical applications, this parameter can also be selected with other values. After the candidate trajectories are filtered by semantic relations, their associated successful execution parameters, including control parameters such as average altitude, average speed, and maximum acceleration, are extracted as the basic data for generating new evidence. By mapping natural language instructions to a vector space isomorphic to the trajectory, cosine similarity is used to measure the semantic similarity of spatial behavior patterns, bypassing complex symbolic reasoning, and realizing direct retrieval and reuse from fuzzy language descriptions to specific historical experiences. This enables the system to understand contextual reference instructions such as those from the previous time, transforming historical successful flight experiences into references for current decisions, significantly improving the naturalness and efficiency of interaction, and reducing reliance on repetitive and detailed instructions.

[0033] In some implementations, in S4, the online fusion update of the uncertainty measure is performed by: taking a weighted average of the previous preference estimate and the new evidence to obtain an updated preference estimate, wherein the weight of the new evidence is negatively correlated with its observation noise level, and the weight of the previous preference estimate is negatively correlated with its current uncertainty measure; and calculating an updated uncertainty measure with a reduced value based on the previous uncertainty measure and the observation noise level.

[0034] Specifically, the preference estimate before the update and the new evidence are weighted and averaged to obtain the updated preference estimate, where the weight of the new evidence is negatively correlated with its observation noise level, and the weight of the preference estimate before the update is negatively correlated with its current uncertainty measure; and, based on the uncertainty measure before the update and the observation noise level, the updated uncertainty measure with a reduced value is calculated.

[0035] Specifically, the preference estimate μ old It is a floating-point number representing the system's current best estimate of the tendency of a certain user control parameter. The unit is consistent with the physical meaning of the parameter, such as the unit for height parameter being meters, and the unit for speed parameter being meters per second; the uncertainty measure σ old2 This is a floating-point number representing the system's variance in the estimate. The initial value is typically set to a large number, such as 100.0, indicating high uncertainty; observation noise level. To preset hyperparameters and characterize the inherent error of the measurement system, they are typically set to 1.0 to 4.0 for explicit feedback channels and 9.0 to 16.0 for implicit feedback channels, reflecting the relative unreliability of implicit evidence.

[0036] Specifically, in the weighted average calculation, the weighting coefficient for new evidence o is:

[0037] The weighting coefficients of the preference estimates before the update are:

[0038] The weighting coefficients satisfy + =1, and is inversely proportional to their respective uncertainty levels; when When the value is large, the system lacks confidence in historical estimates, and new evidence receives higher weight; when... When the value is large, the observation noise is high, and historical estimates receive higher weight; the updated preference estimate = ·o+ · .

[0039] Specifically, the updated uncertainty measure Calculated by harmonic average:

[0040] Its guarantee <min( , This means that each effective update necessarily reduces uncertainty; this property ensures the monotonic convergence of the system's cognition, achieving a learning effect of increasing confidence with use from a mathematical perspective. The entire update process is implemented on the onboard computing unit, using single-precision floating-point arithmetic, with a computational complexity of O(1), a single update time of less than 1 millisecond, and memory usage involving only a few floating-point numbers; by utilizing the conjugate prior property of Gaussian distribution, Bayesian inference is transformed into a closed-form update formula, realizing the quantitative propagation and monotonic reduction of uncertainty, enabling the system to adaptively integrate multi-source evidence of different reliability; providing a theoretical convergence guarantee for preference learning, avoiding the learning rate parameter tuning problem of traditional gradient descent methods, achieving stable and efficient learning from sparse observations, with extremely low computational overhead, perfectly adapting to onboard resource constraints.

[0041] In some implementations, after updating the preference information, a personalized flight trajectory decision-making step is further included: generating an initial safe flight trajectory within the solution space that satisfies all inviolable hard safety constraints; constructing an optimization objective function, which includes a first term and a second term, wherein the first term is used to minimize the modification of the initial safe flight trajectory to maintain smoothness and safety, and the second term is used to drive the trajectory parameters closer to the updated preference estimate, and the influence weight of the second term is negatively correlated with the updated uncertainty metric; adjusting the initial safe flight trajectory by solving the optimization objective function to generate a personalized flight trajectory; performing real-time safety verification on the personalized flight trajectory, and if the verification fails, triggering a safety backoff mechanism to output a backup trajectory that satisfies all hard safety constraints and is closest to the personalized flight trajectory.

[0042] Specifically, an initial safe flight trajectory is generated within the solution space that satisfies all inviolable hard safety constraints. An optimization objective function is constructed, comprising a first term and a second term. The first term minimizes modifications to the initial safe flight trajectory to maintain smoothness and safety, while the second term drives the trajectory parameters closer to the updated preference estimate. The influence weight of the second term is negatively correlated with the updated uncertainty measure. By solving the optimization objective function, the initial safe flight trajectory is adjusted to generate a personalized flight trajectory. Real-time safety verification is performed on the personalized flight trajectory. If verification fails, a safety backoff mechanism is triggered, outputting a backup trajectory that satisfies all hard safety constraints and is closest to the personalized flight trajectory.

[0043] Specifically, hard safety constraints include three categories: obstacle avoidance constraints, dynamic constraints, and no-fly zone constraints. Obstacle avoidance constraints are achieved through a dense obstacle map M constructed by airborne lidar or binocular vision, requiring that the distance between the trajectory and any obstacle be greater than a safety margin. The limit is typically set between 1.5 meters and 3.0 meters; dynamic constraints include maximum speed. Typically set to 15 meters per second, maximum acceleration Typically set to 5 meters per second squared, minimum turning radius R min The physical limit is usually set to 5 meters; no-fly zone constraints are achieved through pre-loaded geofencing data or real-time updated airspace control information.

[0044] Specifically, the initial safe flight trajectory The algorithm uses the RRT* algorithm, which performs random sampling in the three-dimensional state space and progressively optimizes the path cost through optimal parent node selection and rewiring mechanisms. The algorithm is implemented using a lightweight C++ library. On the NVIDIA Jetson Xavier NX platform, planning within a 30-meter range takes approximately 100 milliseconds. In practical applications, other path planning algorithms such as RRT and PRM can also be selected. The output is an ordered waypoint sequence {p1, p2, ..., p...}. N}, each waypoint p i =(x i ,y i ,z i It contains three-dimensional spatial coordinates.

[0045] Specifically, the objective function J is a quadratic function. Taking the optimization of the height parameter as an example, the expression is:

[0046] The first item is the smoothing and safety item. The weighting coefficient is typically set to 1.0 to minimize the deviation between the personalized trajectory and the safe trajectory, ensuring smooth adjustments. The second item is the personalized fit item. The weighting coefficients are typically set between 0.5 and 2.0. The updated preference estimate is given by w, where w is the confidence weight, calculated as w = exp(-α·σ). h ) Calculation, α is a positive constant, usually set to 0.5, σ h The standard deviation of the updated uncertainty measure; when σ h When w is large, it approaches 0, and the influence of personalized items weakens; when σ is large, it approaches 0. h When w is relatively small, it approaches 1, and the personalized items strongly drive the trajectory toward the preference value.

[0047] Specifically, the optimization problem is solved using efficient convex optimization solvers such as OSQP or ECOS. In practical applications, other quadratic programming solvers can also be selected for this component. For a trajectory containing 10 waypoints, the optimization time is approximately 20 milliseconds. The optimized altitude value is obtained by solving the problem. The x, y coordinates and timestamp of the original waypoint are combined to form a personalized trajectory. .

[0048] Specifically, real-time safety verification employs a lightweight physics simulator, incorporating simplified UAV dynamics models, such as a feedforward plus feedback PID control model or a simplified quadrotor rigid body dynamics model; the simulator operates in 1-millisecond steps... Forward simulation is performed to predict the state sequence during execution and detect whether any hard constraints are violated; the verification process runs in parallel on independent CPU threads or GPU stream processors, with a single verification taking less than 10 milliseconds and a running frequency of up to 100Hz.

[0049] Specifically, the safety fallback mechanism is activated when verification fails, and the fallback function Fallback( , )exist Searching for the waypoint sequence with The subsequence with the smallest mean square error and that satisfies all constraints is selected as the backup trajectory. Specifically, this is implemented through calculation. China's various destinations and By calculating the Euclidean distances between corresponding waypoints, the feasible waypoint combination with the shortest distance is selected, ensuring a safe alternative that closely approximates the user's intent, even in extreme cases. Through a layered architecture of hard constraint isolation and soft constraint injection, security is treated as an insurmountable baseline, while personalization serves as an adjustable optimization objective. Combined with confidence-weighted calculations and real-time verification backoff, a secure and reliable decision-making loop is formed. This approach achieves maximum personalization while strictly guaranteeing physical security, avoiding the trade-off dilemma between security and personalization inherent in traditional methods. Formal security verification and intelligent backoff mechanisms provide reliable security guarantees for close human-machine collaboration.

[0050] In some implementations, the triggering safety rollback mechanism is as follows: among the feasible solutions of the initial safe flight trajectory, the waypoint sequence with the smallest difference from the personalized flight trajectory is selected as the backup trajectory.

[0051] Specifically, the safety rollback mechanism is triggered as follows: among the feasible solutions of the initial safe flight trajectory, the waypoint sequence with the smallest difference from the personalized flight trajectory is selected as the backup trajectory. Initial safe flight trajectory An ordered sequence of waypoints generated by the RRT* algorithm, containing N waypoints, each waypoint recording its 3D position, velocity, heading angle, and timestamp; personalized flight trajectory. The solution is a sequence of waypoints optimized by quadratic programming, containing M waypoints, where M is usually equal to or proportional to N; a feasible solution is defined as... A subset of waypoints that satisfy all hard safety constraints, due to It is generated by the safety planning layer, and its complete sequence is a feasible solution.

[0052] Specifically, the difference measure is calculated using a weighted mean squared error. The i-th waypoint and The j-th waypoint in t Define the difference function: D(i,j)= ·|| - || 2 + ·|| - || 2 + ·( -

[0053] Where pos is the position vector, vel is the velocity vector, and head is the heading angle. , , These are weighting coefficients, set to 1.0, 0.5, and 0.3 respectively, reflecting the priority of position matching.

[0054] Specifically, the waypoint sequence selection employs dynamic programming, starting from the initial waypoint and... Searching for and The next waypoint with the smallest difference from the current waypoint is selected, ensuring time monotonicity constraints. A dynamic programming algorithm constructs the difference matrix and achieves a globally optimal match by finding the minimum cost path, with a computational complexity of O(N·M). In practical applications, other approximate matching algorithms can also be selected, but this application does not limit the specific algorithms used.

[0055] Specifically, the selected backup trajectory Maintain and The same safety guarantees, while approximating the same geometry and kinematic properties as closely as possible. For example, when When the verification fails because the attempt to fly at a low altitude, close to the user's preference, results in the flight passing under an obstacle. Will Instead of simply returning to the default altitude, the system chooses waypoints that are as low as possible but avoid obstacles; it transforms the safety rollback into a constrained trajectory approximation problem, finding the optimal trade-off by minimizing the difference metric, so that the safety backup plan retains the intention of personalized adjustment as much as possible; it avoids a brutal rollback of all or nothing, achieving a smooth transition between safety and personalization, and can provide a suboptimal solution that is close to the user's expectations even in the case of personalization failure, significantly improving the robustness of the system and the continuity of user experience.

[0056] In some implementations, the method further includes a training phase: pre-training the initial parameters of the contextualized preference unit by intervening in historical interaction data of annotation and context annotation; subsequently, in a simulation environment, end-to-end reinforcement learning is used to jointly fine-tune the online fusion update strategy and the personalized flight trajectory decision strategy with the comprehensive reward consisting of task efficiency and simulated user satisfaction as the optimization objective.

[0057] Specifically, the initial parameters of the contextualized preference unit are pre-trained by intervening in the historical interaction data of the annotation and context annotation. Then, in the simulation environment, end-to-end reinforcement learning is used to jointly fine-tune the online fusion update strategy and the personalized flight trajectory decision strategy with the comprehensive reward consisting of task efficiency and simulated user satisfaction as the optimization objective.

[0058] The pre-training phase utilizes a constructed multimodal human-computer interaction dataset, which is acquired through a two-stage acquisition process. The first stage takes place in a high-fidelity simulation environment using platforms such as AirSim or Gazebo. In practical applications, other simulation platforms can also be used; this embodiment does not limit the choice. Diverse task scenarios, ambient lighting, and obstacle layouts are generated programmatically, controlling the virtual operator to issue natural language commands with different syntax and semantics, and recording the virtual drone's state and control flow. The second stage takes place in a real flight scenario, using a drone platform equipped with a forward-looking camera and an onboard computer. Multiple real operators with different operating styles complete the tasks, collecting four types of synchronized data: user voice or text commands, drone downlink video streams, complete flight control status data, and user manual remote control intervention signals.

[0059] Specifically, intervention annotation targets the explicit feedback channel, marking the period when the user manually takes over [t]. start ,t end The flight control data within this time period is sampled at a frequency of 10Hz, and the automatic control parameters covered by the user and their specific values ​​are recorded to form a labeled sequence. Context labeling is applied to each interaction event, labeling the environmental context, task type, target object, and personalized preference score assessed by experienced operators. .

[0060] Specifically, the contextualized preference unit is pre-trained using intervention-labeled data, with the training objective being to maximize the conditional probability of observed user corrections, and the loss function being the negative log-likelihood. The network parameters are updated using the Adam optimizer, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. In practical applications, these hyperparameters can be selected with other values, but this embodiment does not limit this.

[0061] Specifically, end-to-end joint fine-tuning is performed in a high-fidelity simulation environment, employing a near-end strategy to optimize the PPO algorithm; comprehensive reward ,in = The reward is for task efficiency, where T represents task time and E represents energy consumption. and These are the weighting coefficients; To simulate user satisfaction rewards, a pre-trained learner with fixed parameters acts as a virtual user, and the negative distance between the execution trajectory and the mean preference is calculated. Weight coefficients... and The value is usually set to 0.6 and 0.4, but other values ​​can be selected in practical applications.

[0062] Specifically, the PPO algorithm implements a pruning objective function, with the pruning coefficient ε set to 0.2, the discount factor γ set to 0.99, and the generalized advantage estimation parameter λ set to 0.95. The policy network and value network share the parameters of the convolutional and LSTM layers, while the output layers are separated. During training, 16 instances are run in parallel in the simulation environment to improve sampling efficiency, with a total training step count of 10 million steps. In practical applications, these configurations can also be selected with other values, which are not limited in this embodiment. During joint fine-tuning, some parameters of the pre-training module can be frozen or set with a small learning rate, such as a learning rate decay to 0.1 times that of the pre-training stage, to maintain the stability of the learned knowledge while adapting to the global optimization objective. Through a phased training strategy, the professional basic capabilities of each module are first established, and then the global reward signal guides the co-evolution between modules, achieving a gradual transition from local optimum to global optimum. This improves the training stability and final performance of the complex coupled system, avoids the functional imbalance of modules in the early stage of joint training, and enables the system to generate emergent collaborative capabilities that cannot be achieved by individual training.

[0063] This invention also provides a large-model-based personalized interactive control system for unmanned aerial vehicles (UAVs), including the large-model-based personalized interactive control method for UAVs described in the preceding claim, comprising: a task context and feedback perception module, used to respond to interaction events with the user and acquire current task context information, explicit feedback data, and implicit feedback data; wherein, the explicit feedback data is the parameter correction amount generated when the user manually adjusts the UAV flight parameters via a remote controller, and the implicit feedback data is a natural language command generated with reference to historical execution experience; and a contextualized preference management module, used to store multiple contextualized preference units and determine a target contextualized preference unit based on the current task context information; wherein, Each contextualized preference unit is associated with a corresponding task context and includes user preference information for at least one UAV control parameter. The preference information includes at least a preference estimate and its uncertainty measure. An online learning module, communicatively connected to the task context and feedback perception module and the contextualized preference management module, is used to directly generate new evidence for the target control parameters based on the explicit feedback data or indirectly generate new evidence for the target control parameters based on the implicit feedback data through historical trajectory retrieval. Based on the new evidence and its credibility, the preference estimate and its uncertainty measure of the corresponding control parameter in the target contextualized preference unit are fused and updated online, so that the updated uncertainty measure is lower than the original.

[0064] Specifically, the mission context and feedback perception module is deployed in a layered architecture of airborne computing unit and ground control station; the airborne part includes a sensor interface unit that connects flight control sensors such as IMU, GPS, barometer, magnetometer, and optical flow sensor, with a sampling frequency of 100Hz to 1000Hz; and a vision processing unit that connects to a forward-looking camera, a downward-looking camera, or a binocular camera, using an NVIDIA Jetson Xavier NX or Qualcomm RB55G platform to run target detection and tracking algorithms to extract target object information. The ground component includes a voice acquisition unit using an array microphone or headset microphone with a sampling rate of 16kHz, which transmits the data wirelessly to the airborne or cloud-based processing unit; a remote control signal receiving unit using an SBUS or CRSF protocol receiver to parse user-manually controlled channel values; a contextualized preference management module implemented as a hash table structure in memory, with the key being a binary hash code of the context key and the value being a contextualized preference unit object; the hash table uses open addressing or chaining to handle collisions, with a capacity of 1000 units, occupying approximately 50MB of memory; the preference parameter set within each unit is stored using an array structure, supporting random access with O(1) time complexity; the module provides atomic operation interfaces for concurrent read and write operations to ensure consistency in the decision module's reads when the online learning module updates preferences; the online learning module is implemented as an independent computation thread, executing in parallel with the flight control main loop. The module contains dual-channel processing subunits: an explicit feedback processing subunit directly receives parameter corrections from the remote control signal parser; and an implicit feedback processing subunit receives historical trajectory retrieval results from the memory association module via a message queue. The Bayesian update engine uses a lookup table method to perform common mathematical operations, avoiding runtime computational overhead. Update calculations are accelerated on a dedicated mathematical coprocessor or GPU CUDA core. Inter-module communication adopts a publish-subscribe pattern or a shared memory mechanism. Data serialization uses Protocol Buffers or Message Pack format. In practical applications, other data exchange formats can also be selected, and this application embodiment does not limit this, ensuring low-latency and high-throughput data transmission. The task context and feedback perception module publishes context update events, and the contextualized preference management module subscribes to and updates the current context key. The online learning module publishes preference update events, and the contextualized preference management module subscribes to and updates the Gaussian distribution parameters of the corresponding unit. Through a modular hardware-software collaborative architecture, perception, memory, and learning functions are decoupled into independent and maintainable units. Through standardized interfaces, data flow and control flow are organically integrated to form a complete personalized interactive closed loop.

[0065] In some implementations, a memory association module is also included, comprising: a memory graph storage unit for storing a hierarchical memory graph, the memory graph including a semantic layer for recording tasks, users, location entities and relationships, and a trajectory pattern layer for storing compressed feature vectors of historical successful flight trajectories; and a trajectory retrieval unit for responding to the implicit feedback data, mapping the current instruction to a spatial semantic query vector, performing similarity matching between the query vector and historical trajectory feature vectors in the trajectory pattern layer, and retrieving similar historical trajectories based on the matching results. The trajectory retrieval unit uses a cosine similarity algorithm for matching.

[0066] Specifically, the memory graph storage unit is implemented using a graph database or a hybrid storage architecture. The semantic layer uses a graph database such as Neo4j or RocksDB; in practical applications, other models can also be selected for this component, and this embodiment does not limit this choice. It stores entity nodes and relational edges. Node attributes include user ID, task type, location coordinates, object category, etc., while edge attributes include execution time, task result, etc. The trajectory pattern layer uses a vector database such as Faiss or Milvus, or a simple in-memory array structure, to store trajectory feature vectors and keyframe sequences. Each feature vector is associated with the task node ID in the semantic layer as a foreign key.

[0067] The trajectory retrieval unit includes a semantic mapping subunit and a similarity calculation subunit. The semantic mapping subunit is implemented as a lightweight neural network with an input layer dimension of 768 corresponding to the BERT output dimension, hidden layer dimensions of 256 and 128, and an output layer dimension of 128 corresponding to the trajectory feature dimension. The activation function is ReLU, and the network weights are learned through the pre-training stage. The similarity calculation subunit is implemented as a GPU-accelerated batch matrix operation. For 10,000 trajectories in the library, a single batch cosine similarity calculation takes about 5 milliseconds. In practical applications, this calculation can also be implemented using a CPU or other approximate nearest neighbor algorithms such as LSH. This application embodiment does not limit this.

[0068] Specifically, the retrieval process is as follows: after the implicit feedback data arrives, the semantic mapping subunit generates a query vector q; the similarity calculation subunit calculates the similarity between q and all v. i The system uses cosine similarity; sorts and selects Top-K results; retrieves task context through foreign key association queries in the semantic layer; returns identifiers, feature vectors, keyframe sequences, and control parameters of similar historical trajectories; achieves cross-modal retrieval from abstract language instructions to specific spatial behaviors through a two-layer memory structure and vectorized semantic mapping, making historical flight experience a reusable knowledge asset; endows the system with context recall capabilities, supports intelligent decision-making based on historical references, significantly reduces reliance on repeated demonstrations, and improves the context awareness and long-term learning efficiency of interactions.

[0069] In some implementations, the online learning module is used to: take a weighted average of the previous preference estimate and the new evidence to obtain the updated preference estimate, wherein the weights are inversely proportional to their respective uncertainty levels; and calculate a new uncertainty measure that is less than the previous uncertainty measure.

[0070] Specifically, the online learning module includes an evidence generation subunit, a weight calculation subunit, a fusion update subunit, and an uncertainty propagation subunit. The evidence generation subunit receives parameter correction vectors from the explicit channel or historical trajectory control parameters from the implicit channel, performs unit unification and dimension conversion, and generates standardized new evidence o; for altitude parameters, the unit is meters; for speed parameters, the unit is meters per second; for heading angle parameters, the unit is degrees or radians. In practical applications, other standards can also be selected, and this embodiment does not limit this.

[0071] The weight calculation subunit maintains two uncertainty parameters: the uncertainty of the preference estimate before the update. and observation noise level ; Read from the contextualized preference management module. The weighting is dynamically selected based on the source of evidence. Explicit feedback is set to a lower value, such as 4.0, while implicit feedback is set to a higher value, such as 16.0. The weighting coefficients are calculated as follows:

[0072] Single-precision floating-point arithmetic is used to ensure numerical stability.

[0073] Specifically, the merged update subunit performs a weighted average. = ·o+ · The results are written back to the mean field of the corresponding unit in the contextualized preference management module; the uncertainty propagation subunit calculates...

[0074] The result is written back to the variance field; consistency between the two write operations is guaranteed by atomic instructions or transaction mechanisms.

[0075] Specifically, numerical stability processing includes: when or When the precision is less than that of the machine, logarithmic domain calculation is used; when the denominator is close to zero, an exception handling process is triggered, the original value is kept unchanged and logged; variance decay is performed periodically on units that have not been updated for a long time to prevent overconfidence. By using uncertainty as an explicit variable in the calculation, adaptive weighting of evidence reliability is achieved, and the mathematical properties of Bayesian inference are used to ensure the monotonic evolution of cognition.

[0076] In some implementations, a hierarchical decision-making module is also included, comprising: a safety planning unit for generating an initial safe flight trajectory within a solution space that satisfies all hard safety constraints; a personalized optimization unit, communicatively connected to the contextualized preference management module, for constructing and solving an optimization objective function that minimizes modifications to the initial safe flight trajectory while adding a personalized driving term weighted by preference confidence to generate a personalized flight trajectory; and a safety verification and rollback unit for rapidly simulating and verifying the personalized flight trajectory, and, if verification fails, initiating rollback logic to output a backup trajectory that satisfies all hard safety constraints; the safety verification and rollback unit is configured with a UAV dynamics model for forward simulation of candidate trajectories to predict constraint violations.

[0077] Specifically, the safety planning unit is implemented as a path planning server, using an RRT* algorithm library such as OMPL or a self-developed implementation. In practical applications, other path planning libraries can also be selected for this component, and this application does not limit this. The unit receives the task target point, real-time obstacle map, and dynamic constraint set, with a running time limit of 100 milliseconds, and outputs a sequence of safe waypoints. The obstacle map adopts an Octo Map or Voxel Hashing structure, with a resolution of 0.1 meters to 0.5 meters and an update frequency of 10 Hz.

[0078] Specifically, the personalized optimization unit is implemented as a convex optimization solver interface, receiving the safety trajectory and preference parameters, and constructing a quadratic programming problem in sparse matrix form; the objective function matrix H and the vector f are given by λ. smooth , λ pref w, μ h The parameters are constructed, and the constraint matrix A and boundary b are constructed from the safety corridor constraint. The solver can be OSQP, ECOS, or qpOASES. In practical applications, other solvers can also be selected; this embodiment does not limit this choice. The maximum solution time is set to 50 milliseconds, and the maximum number of iterations is 1000. Upon successful solution, the optimized trajectory is output; otherwise, a safe trajectory is directly output.

[0079] The safety verification and rollback unit is implemented as a parallel simulation thread, configured with a quadcopter UAV dynamic model, including parameters such as mass, inertia matrix, thrust coefficient, and torque coefficient. These model parameters are obtained through system identification or manufacturer manuals. The simulator uses Euler's method or Runge-Kutta's method for numerical integration, with a step size of 1 millisecond, and the prediction time domain covers the entire trajectory execution cycle. Verification includes position constraint checks, velocity constraint checks, acceleration constraint checks, and obstacle collision checks; failure in any check triggers rollback.

[0080] The backoff logic is implemented as a trajectory matching algorithm. In the sequence of waypoints on the safe trajectory, the algorithm uses dynamic programming to find the nearest neighbor sequence to the failed personalized trajectory. The distance metric considers the multi-dimensional weighting of position, velocity, and acceleration. The backoff trajectory ensures that it meets 100% of the hard constraints, while the geometry is as close as possible to the personalized intent.

[0081] The overall timing of the hierarchical decision-making module is as follows: safety planning 100ms, personalization optimization 20ms, and safety verification 10ms, with a total decision cycle of less than 130ms, meeting real-time requirements. Each unit coordinates through a state machine to ensure that the system can output a valid trajectory when any unit times out or fails. Through a four-layer architecture of hard constraint isolation, soft constraint injection, real-time verification, and intelligent rollback, security is embedded in every step of the decision-making process, achieving decoupling of security and personalization and reliable assurance.

[0082] This application employs a dual-channel sparse feedback learning mechanism. The explicit intervention channel captures occasional manual takeover behaviors by the user, while the implicit semantic channel parses the user's fuzzy reference commands based on historical experience. This enables the system to learn effectively online from two types of extremely low-frequency natural interaction events. The probabilistic data structure of micro-preference clusters models each control parameter preference as a Gaussian distribution. Through the Bayesian update formula, the uncertainty is monotonically reduced, mathematically ensuring that each effective update will necessarily reduce cognitive uncertainty, achieving a progressive learning effect of becoming more knowledgeable with use. This overcomes the bottleneck of personalized functions being difficult to deploy due to data sparsity, enabling drones to truly possess the evolutionary ability to accumulate experience from sporadic interaction fragments.

[0083] The probabilistic representation of this application makes the system output explicitly associated with the internal state: the decision basis can be traced back to the Gaussian distribution parameters under a specific context key, the source identifier of dual-channel evidence, and the historical trajectory of Bayesian updates; transparency enhances user trust in the system, and when the UAV makes a flight decision that conforms to the operator's habits, the system can explain when, where and what kind of interaction experience the decision originated from; when deviations occur, variance analysis can also be used to locate the parameter dimensions with higher cognitive uncertainty, providing a clear path for error analysis and system debugging, and significantly improving the maintainability of the system and the reliability of human-machine collaboration.

[0084] The micro-preference clusters in this application uniquely correspond to context keys generated by hashing environmental context, task type, and target object features. This binds abstract preferences to specific contexts and objects, enabling the system to learn the complex mapping relationship of which operation mode a user prefers under what circumstances. For example, the same operator may prefer an aggressive style of close observation when inspecting wind turbine blades, but a conservative style when inspecting photovoltaic panels. The system accurately captures this context dependence through independent Gaussian distributions under different context keys, achieving true contextualized personalized adaptation rather than a coarse global average.

[0085] This application enables the system to understand natural language commands containing historical references, such as "fly as last time" and "execute in the manner of checking target A," through implicit semantic channels. It transforms ambiguous semantic descriptions into computable probabilistic update evidence. This capability increases interaction efficiency with usage frequency. Users do not need to repeatedly describe the expected flight parameters in detail. Instead, they can convey their intentions through concise contextual references, significantly reducing the cognitive load of interaction and improving the tacit understanding and fluency of long-term use.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention. The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the protection scope of the present invention.

Claims

1. A personalized interactive control method for unmanned aerial vehicles (UAVs) based on a large model, characterized in that, Includes the following steps: S1: In response to user interaction events, obtain current task context information, explicit feedback data, and implicit feedback data; wherein, the explicit feedback data is the parameter correction amount generated when the user manually adjusts the drone flight parameters through the remote controller, and the implicit feedback data is natural language instructions generated with reference to historical execution experience; S2: Based on the current task context information, determine the target contextual preference unit from multiple stored contextual preference units; wherein, each contextual preference unit is associated with the corresponding task context and includes: user preference information for at least one UAV control parameter, the preference information including at least: preference estimate and its uncertainty measure; S3: New evidence for the target control parameters is generated directly from the explicit feedback data or indirectly from the implicit feedback data through historical trajectory retrieval. S4: Based on the new evidence and its credibility, the preference estimate and uncertainty measure of the corresponding control parameter in the target contextualized preference unit are fused and updated online, so that the updated uncertainty measure is lower than that before the update.

2. The method for personalized interactive control of unmanned aerial vehicles based on a large model according to claim 1, characterized in that, The new evidence for indirectly generating target control parameters based on implicit feedback data through historical trajectory retrieval includes: Parse the natural language instructions, extract their spatial behavior semantics, and convert them into query vectors; The query vector is matched with the feature vectors of multiple historical flight trajectories stored in the memory graph. The memory graph includes a semantic layer that records the contextual relationships of the task and a trajectory pattern layer that stores trajectory pattern data. Retrieve semantically similar historical trajectories based on the matching results; The new evidence is generated based on the control parameters used when the similar historical trajectory was successfully executed. The similarity matching uses cosine similarity, which measures the consistency of vector directions.

3. The method for personalized interactive control of unmanned aerial vehicles based on a large model according to claim 1, characterized in that, In step S4, the uncertainty measure is updated online through fusion as follows: The updated preference estimate is obtained by weighting the original preference estimate and the new evidence, wherein the weight of the new evidence is negatively correlated with its observation noise level, and the weight of the original preference estimate is negatively correlated with its current uncertainty measure. Furthermore, based on the uncertainty metric before the update and the observation noise level, the updated uncertainty metric with a reduced value is calculated.

4. The UAV personalized interactive control method based on a large model according to claim 3, characterized in that, After updating the preference information, a personalized flight trajectory decision-making step is also included: Within the solution space that satisfies all inviolable hard safety constraints, an initial safe flight trajectory is generated; Construct an optimization objective function, which includes a first term and a second term. The first term is used to minimize the modification of the initial safe flight trajectory to maintain smoothness and safety, and the second term is used to drive the trajectory parameters closer to the updated preference estimate. The influence weight of the second term is negatively correlated with the updated uncertainty measure. By solving the optimization objective function, the initial safe flight trajectory is adjusted to generate a personalized flight trajectory; The personalized flight trajectory is verified for safety in real time. If the verification fails, a safety rollback mechanism is triggered to output a backup trajectory that meets all hard safety constraints and is closest to the personalized flight trajectory.

5. The UAV personalized interactive control method based on a large model according to claim 4, characterized in that, The triggering safety rollback mechanism is as follows: among the feasible solutions of the initial safe flight trajectory, the waypoint sequence with the smallest difference from the personalized flight trajectory is selected as the backup trajectory.

6. The method for personalized interactive control of unmanned aerial vehicles based on a large model according to claim 1, characterized in that, The method also includes a training phase: pre-training the initial parameters of the contextualized preference unit by intervening in the historical interaction data of annotation and context annotation; Subsequently, in a simulation environment, end-to-end reinforcement learning is used to jointly fine-tune the online fusion update strategy and the personalized flight trajectory decision strategy, with the comprehensive reward consisting of task efficiency and simulated user satisfaction as the optimization objective.

7. A personalized interactive control system for unmanned aerial vehicles based on a large model, characterized in that: The method for personalized interactive control of unmanned aerial vehicles based on a large model, as described in any one of claims 1-6, includes: The task context and feedback perception module is used to respond to user interaction events and obtain current task context information, explicit feedback data, and implicit feedback data. The explicit feedback data is the parameter correction amount generated when the user manually adjusts the drone's flight parameters via the remote controller, and the implicit feedback data is natural language commands generated with reference to historical execution experience. The contextualized preference management module is used to store multiple contextualized preference units and determine the target contextualized preference unit based on the current task context information; wherein, each contextualized preference unit is associated with a corresponding task context and includes user preference information for at least one UAV control parameter, the preference information including at least the preference estimate and its uncertainty measure; The online learning module is communicatively connected to the task context and feedback perception module and the contextualized preference management module. It is used to generate new evidence for the target control parameters directly based on the explicit feedback data or indirectly based on the implicit feedback data through historical trajectory retrieval. Based on the new evidence and its credibility, it performs online fusion and update of the preference estimates and uncertainty measures of the corresponding control parameters in the target contextualized preference unit, so that the updated uncertainty measure is lower than that before the update.

8. The UAV personalized interactive control system based on a large model according to claim 7, characterized in that, It also includes a memory association module, which includes: A memory graph storage unit is used to store a hierarchical memory graph, which includes a semantic layer for recording tasks, users, location entities and relationships, and a trajectory pattern layer for storing compressed feature vectors of historical successful flight trajectories. The trajectory retrieval unit is used to respond to the implicit feedback data, map the current instruction into a spatial semantic query vector, perform similarity matching between the query vector and the historical trajectory feature vector in the trajectory pattern layer, and retrieve similar historical trajectories based on the matching results. The trajectory retrieval unit uses a cosine similarity algorithm for matching.

9. The UAV personalized interactive control system based on a large model according to claim 7, characterized in that, The online learning module is used to: take a weighted average of the previous preference estimate and the new evidence to obtain the updated preference estimate, wherein the weight allocation is inversely proportional to the respective uncertainty level; and calculate a new uncertainty measure with a value smaller than the previous uncertainty measure.

10. The UAV personalized interactive control system based on a large model according to claim 9, characterized in that, It also includes a hierarchical decision-making module, which includes: The safety planning unit is used to generate an initial safe flight trajectory within the solution space that satisfies all hard safety constraints. The personalized optimization unit is communicatively connected to the contextualized preference management module and is used to construct and solve an optimization objective function. This function minimizes the modification of the initial safe flight trajectory while adding a personalized driving term weighted by preference confidence to generate a personalized flight trajectory. The safety verification and rollback unit is used to perform rapid simulation verification of the personalized flight trajectory, and to initiate rollback logic to output a backup trajectory that meets all hard safety constraints when verification fails; the safety verification and rollback unit is equipped with a UAV dynamics model to perform forward simulation of the candidate trajectory to predict constraint violations.

Images