Unmanned aerial vehicle navigation method based on visual language model and related equipment

Abstract

The invention discloses an unmanned aerial vehicle navigation method and related equipment based on a visual language model, and the method comprises the steps: carrying out the thinking chain construction of an initial training data set, generating high-quality thinking chain data, and constructing a target training data set according to the initial training data set and the high-quality thinking chain data; based on a supervised fine tuning method and a reinforcement learning method, training the initial visual language model according to the target training data set to obtain a candidate visual language model; deploying the target visual language model passing the model verification to an unmanned aerial vehicle navigation control system; and generating a reasoning result and a target position according to the natural language instruction of the user through the target visual language model, and predicting a target action sequence according to the target position by adopting a foresight mechanism, so that the unmanned aerial vehicle executes the target action sequence. Complex natural language instructions can be accurately understood, spatial reasoning is carried out in combination with environmental semantics, accurate target positioning and path planning are achieved, and the method can be widely applied to the technical field of unmanned aerial vehicle control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle control, and in particular to an unmanned aerial vehicle navigation method based on a vision language model and related equipment. BACKGROUND

[0002] Unmanned aerial vehicles (UAVs) are increasingly widely used in disaster response, city inspection, logistics distribution, and security patrol, and the degree of intelligence directly determines the efficiency and safety of the task. Although current vision-and-language navigation (VLN) methods for unmanned aerial vehicles (UAVs) have made some progress, there are still some defects in the related technology: due to insufficient multi-modal data fusion, the unmanned aerial vehicle is prone to understanding deviation and navigation failure when facing long instructions, ambiguous descriptions, or multiple landmark associations; the performance of the model is greatly reduced in unknown environments such as Out-of-Distribution (OOD); and the related technology basically directly outputs a navigation action sequence, making it difficult for users to understand the basis for the decision.

[0003] In summary, the technical problems in the related technology need to be improved. SUMMARY

[0004] Embodiments of the present application aim to at least partially solve one of the technical problems in the related art. To this end, the main purpose of the embodiments of the present application is to propose an unmanned aerial vehicle navigation method based on a vision language model and related equipment, which can accurately understand complex natural language instructions and combine environmental semantics for spatial reasoning, achieving accurate target positioning and path planning.

[0005] To achieve the above-mentioned purpose, one aspect of an embodiment of the present application proposes an unmanned aerial vehicle navigation method based on a vision language model, the method comprising the following steps: Collecting multi-modal data generated by the unmanned aerial vehicle during task execution; wherein the multi-modal data includes natural language instruction text, environmental images, semantic maps, and unmanned aerial vehicle action sequences; Performing data cleaning processing on the multi-modal data to obtain an initial training data set; Using a thought chain generation model to construct a thought chain for the initial training data set, generate high-quality thought chain data, and construct a target training data set according to the initial training data set and the high-quality thought chain data; Based on a supervised fine-tuning method and a reinforcement learning method, model training is performed on an initial vision language model according to the target training data set to obtain a candidate vision language model; The candidate visual language model is subjected to model verification, a candidate visual language model passing the model verification is taken as a target visual language model, and the target visual language model is deployed in the unmanned aerial vehicle navigation control system; The current user natural language instruction is input into the unmanned aerial vehicle navigation control system, a current reasoning result and a current target position are generated according to current real environment information by the target visual language model, and a target action sequence is predicted according to the current target position by using a forward-looking mechanism, so that the current unmanned aerial vehicle executes the target action sequence.

[0006] In some embodiments, the data cleaning processing on the multi-modal data to obtain an initial training data set comprises: The natural language instruction text and the environment image are subjected to data cleaning, wherein the data cleaning comprises noise elimination processing and outlier removal processing; The natural language instruction text after the data cleaning is subjected to standardization processing, and the environment image after the data cleaning and the semantic map are subjected to semantic annotation processing; The initial training data set is constructed according to the natural language instruction text after the standardization processing, the environment image after the semantic annotation processing, the semantic map after the semantic annotation processing, and the unmanned aerial vehicle action sequence.

[0007] In some embodiments, the thought chain generation model is used to construct a thought chain on the initial training data set to generate high-quality thought chain data, and a target training data set is constructed according to the initial training data set and the high-quality thought chain data, comprising: The thought chain generation model is used to construct a thought chain on the initial training data set to generate initial thought chain data; The initial thought chain data is subjected to quality screening by using a data quality screening mechanism to obtain the high-quality thought chain data; The target training data set is constructed according to the initial training data set and the high-quality thought chain data.

[0008] In some embodiments, the initial visual language model is subjected to model training according to the target training data set based on a supervised fine-tuning method and a reinforcement learning method to obtain a candidate visual language model, comprising: The initial visual language model is subjected to supervised fine-tuning according to the target training data set based on the supervised fine-tuning method to obtain a fine-tuned visual language model; wherein the fine-tuned visual language model has preliminary reasoning ability and preliminary navigation prediction ability; Based on the reinforcement learning method, the fine-tuned visual language model is subjected to model optimization processing according to the target training data set, to obtain the candidate visual language model.

[0009] In some embodiments, based on the reinforcement learning method, the fine-tuned visual language model is subjected to model optimization processing according to the target training data set, to obtain the candidate visual language model, including: inputting the target training data set into the fine-tuned visual language model, and outputting a model inference result and a model predicted position; According to the model inference result and the model predicted position, a reward value is calculated by using a composite reward function; wherein the composite reward function includes a format reward function, a target distance reward function and an intermediate process reward function; According to the reward value, a target model loss for the fine-tuned visual language model is determined; According to the target model loss, the fine-tuned visual language model is optimized to obtain the candidate visual language model.

[0010] In some embodiments, after the current user natural language instruction is input into the unmanned aerial vehicle navigation control system, and the current inference result and the current target position are generated by the target visual language model according to the current real environment information, the method further includes: outputting the current inference result to a visual interface for display.

[0011] To achieve the above-mentioned purpose, another aspect of the embodiment of the present application proposes an unmanned aerial vehicle navigation device based on a visual language model, which comprises the following modules: A data acquisition module is configured to acquire multi-modal data generated by an unmanned aerial vehicle during task execution; wherein the multi-modal data includes natural language instruction text, environment image, semantic map and unmanned aerial vehicle action sequence; A data cleaning module is configured to perform data cleaning processing on the multi-modal data to obtain an initial training data set; A thought chain construction module is configured to use a thought chain generation model to construct a thought chain for the initial training data set, to generate high-quality thought chain data, and to construct a target training data set according to the initial training data set and the high-quality thought chain data; A model training module is configured to perform model training on an initial visual language model according to the target training data set based on a supervised fine-tuning method and a reinforcement learning method, to obtain a candidate visual language model; A model verification module is configured to perform model verification on the candidate visual language model, take the candidate visual language model passing the model verification as a target visual language model, and deploy the target visual language model to the UAV navigation control system. A model prediction module is configured to input a current user natural language instruction into the UAV navigation control system, generate a current reasoning result and a current target position according to current real environment information by using the target visual language model, and predict a target action sequence according to the current target position by using a forward-looking mechanism, so as to make the current UAV execute the target action sequence.

[0012] To achieve the above object, another aspect of the embodiments of the present application provides an electronic device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the above method when executing the computer program.

[0013] To achieve the above object, another aspect of the embodiments of the present application provides a computer readable storage medium, which stores a computer program, and the computer program implements the above method when executed by a processor.

[0014] To achieve the above object, another aspect of the embodiments of the present application provides a computer program product, which comprises a computer program, and the computer program implements the above method when executed by a processor.

[0015] The embodiments of the present application at least have the following beneficial effects: The present application provides a UAV navigation method based on a visual language model and related equipment, which acquires multi-modal data generated by a UAV during task execution; wherein the multi-modal data includes natural language instruction text, environment images, semantic maps, and UAV action sequences; the multi-modal data is subjected to data cleaning processing to obtain an initial training data set; a thought chain generation model is used to construct a thought chain for the initial training data set, generate high-quality thought chain data, and construct a target training data set according to the initial training data set and the high-quality thought chain data; based on a supervised fine-tuning method and a reinforcement learning method, the initial visual language model is subjected to model training according to the target training data set to obtain a candidate visual language model; the candidate visual language model is subjected to model verification, the candidate visual language model that passes the model verification is used as a target visual language model, and the target visual language model is deployed in a UAV navigation control system; a current user natural language instruction is input into the UAV navigation control system, a current reasoning result and a current target position are generated by the target visual language model according to current real environment information, and a target action sequence is predicted according to the current target position by using a forward-looking mechanism, so that the current UAV executes the target action sequence. The embodiments of the present application acquire multi-modal data generated by a UAV during task execution, use a visual language large model as a base, effectively integrate visual information and language information through deep multi-modal fusion, accurately understand complex natural language instructions and combine environment semantics for spatial reasoning, solve the problems of related technologies in semantic analysis and visual understanding, and thus realize accurate target positioning and path planning; through the two-stage training process of supervised fine-tuning and reinforcement learning, the model can perform stronger generalization and robustness in unknown environments; through the introduction of a chain reasoning mechanism, an intermediate reasoning process is generated before the output of a navigation result, so that the navigation decision is more transparent and interpretable, which not only facilitates user understanding and verification of system behavior, but also is beneficial to operation and maintenance and safety supervision, and improves the controllability and reliability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a step flowchart of a UAV navigation method based on a visual language model provided by the embodiments of the present application; Figure 2 is a model supervised fine-tuning flowchart provided by the embodiments of the present application; Figure 3 is a model reinforcement learning flowchart provided by the embodiments of the present application; Figure 4 is a model navigation algorithm flowchart provided by the embodiments of the present application; Figure 5 is a system framework flowchart of a UAV navigation method based on a visual language model provided by the embodiments of the present application; Figure 6 is a structural schematic diagram of a UAV navigation device based on a visual language model provided by an embodiment of the present application; Figure 7 is a hardware structure schematic diagram of an electronic device. DETAILED DESCRIPTION

[0017] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all the implementations consistent with the embodiments of the present application, but are only examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the appended claims.

[0018] It can be understood that the terms "first", "second", and the like used in the present application can be used herein to describe various concepts, but unless specifically stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of the present application, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information. Depending on the context, the word "if" as used herein can be interpreted as "when" or "when" or "in response to determining".

[0019] The terms "at least one", "multiple", "each", "any" and the like used in the present application include one, two or more than two, multiple includes two or more than two, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0021] Before the embodiments of the present application are described in detail, first, some nouns and terms involved in the embodiments of the present application are described, and the nouns and terms involved in the embodiments of the present application are applicable to the following explanations.

[0022] (1) Vision-and-Language Navigation (abbreviated as VLN), which is a technology that integrates visual perception and natural language instructions to drive an intelligent agent to complete target positioning and path planning.

[0023] (2) Vision-and-Language Model (VLM), which is a model that simultaneously processes images / text to form multi-modal representations and perform downstream tasks.

[0024] (3) Semantic Map, which is a scene graph containing drone pose, orientation, first-person view projection, and landmark annotations.

[0025] (4) Landmark / landmark_bbox, which is a key region / bounding box annotation for localization inference.

[0026] (5) target_location, which is the target two-dimensional coordinate predicted by the vision-language model.

[0027] (6) Action space, which is the discrete action that the drone can perform, such as ascending, descending, stopping, etc.

[0028] (7) Chain-of-Thought (CoT), which refers to the explicit output of the intermediate reasoning process ( <think>) with the final answer ( <answer>), and further improve the interpretability and controllability of visual language models.

[0029] (8) Supervised Fine-Tuning (SFT), which refers to optimizing the model using real label data.

[0030] (9) Reinforcement Learning (RL), which optimizes the model through reward and punishment strategies.

[0031] (10) Group Relative Policy Optimization (GRPO), which is a reinforcement learning algorithm that trains the model by designing a rule-based reward function.

[0032] (11) <think> / <answer>Tags, structured output conventions, respectively carry inference chain and target coordinates.

[0033] (12) CityNav dataset, which is a city-level UAV-VLN benchmark constructed based on real city three-dimensional scanning, contains multiple cities, multiple targets and human demonstrations.

[0034] (13) Intersection over Union (IoU), which is a commonly used evaluation index in target detection and positioning tasks, used to measure the degree of overlap between two regions (usually predicted box and real box). The calculation formula of Intersection over Union is: .

[0035] Currently, unmanned aerial vehicles (UAVs) are increasingly widely used in disaster response, city inspection, logistics distribution, and security patrol, and the degree of intelligence directly determines the efficiency and safety of the task. Traditional UAV navigation mostly relies on GPS (Global Positioning System) positioning combined with a pre-set path. Although it can meet the needs in open environments, it often faces the following problems in urban canyons, complex neighborhoods, or post-disaster scenarios: (1) GPS error and obstruction: positioning information is unreliable in high-rise dense or harsh environments; (2) complex and variable environment: dynamic obstacles (vehicles, crowds, falling objects) make the pre-set path invalid; (3) limited human-machine interaction: operators can usually only control UAVs through coordinates or simple instructions, making it difficult to achieve natural and efficient human-machine interaction. Therefore, the industry has gradually proposed using visual-linguistic navigation technology to enhance the autonomy of UAVs. This technology enables UAVs to understand natural language instructions (such as "fly to the gray building next to the playground") and combine visual maps or camera inputs for reasoning and path planning, thereby completing tasks in complex environments. This approach not only reduces dependence on human operation but also significantly improves the generalization ability and flexibility of UAVs. To achieve visual-linguistic navigation (VLN) of UAVs, there are currently mainly the following methods: (1) Seq2Seq model: Early research uses the Seq2Seq framework to encode natural language instructions into fixed representations and then generates action sequences from the decoder. This method realizes the direct mapping of language to action, but has the problems of insufficient semantic expression and poor generalization performance; (2) Cross-Modal Attention (CMA) mechanism: Subsequent research introduces cross-modal attention to align language and visual features, improving the accuracy of UAV task execution. However, this method mostly stays at the level of shallow fusion and has limited ability to reason about complex semantics and space; (3) Map-based Goal Predictor (MGP): The latest research, such as the CityNav dataset and its accompanying methods, proposes predicting target locations using maps and then planning paths. This method has made progress in scene structure perception, but still relies on task-specific datasets and performance significantly decreases when the environment exceeds the training distribution.

[0036] Therefore, although the UAV visual-linguistic navigation (VLN) method in the related art has made some progress, it still has the following main defects: (1) Insufficient multi-modal fusion: Most methods only combine visual and linguistic features through simple concatenation or shallow attention, making it difficult to fully capture complex semantic relationships and spatial layouts. This leads to understanding biases and navigation failures when UAVs face long instructions, ambiguous descriptions, or multiple landmark associations.

[0037] (2) Limited generalization ability: the model of the related technology performs well in the training environment, but the performance is greatly reduced in the out-of-distribution environment. For example, the Seq2Seq or map-based method relies on the training set scene, and once it encounters a strange city block or dynamic obstacle, the navigation success rate is significantly reduced, lacking robustness in practice.

[0038] (3) Lack of explainability: the related technology directly outputs the navigation action sequence, and lacks the intermediate reasoning chain proposed in the application, making it difficult for users to understand the basis for its decision. Once there is a deviation or target recognition error, it is difficult to trace the source of the problem, and it is also not conducive to the system optimization and trust establishment by the operation and maintenance personnel.

[0039] Therefore, in the embodiments of the application, a UAV navigation method based on a visual language model and related equipment are provided. The scheme collects multi-modal data generated by the UAV during task execution; wherein the multi-modal data includes natural language instruction text, environment image, semantic map and UAV action sequence; the multi-modal data is subjected to data cleaning processing to obtain an initial training data set; a thought chain generation model is used to construct a thought chain for the initial training data set, generate high-quality thought chain data, and construct a target training data set according to the initial training data set and the high-quality thought chain data; based on a supervised fine-tuning method and a reinforcement learning method, the initial visual language model is trained according to the target training data set to obtain a candidate visual language model; the candidate visual language model is subjected to model verification, the candidate visual language model that passes the model verification is used as a target visual language model, and the target visual language model is deployed in a UAV navigation control system; the current user natural language instruction is input into the UAV navigation control system, the target visual language model generates a current reasoning result and a current target position according to the current real environment information, and a forward-looking mechanism is used to predict a target action sequence according to the current target position, so that the current UAV executes the target action sequence. In the embodiments of the application, multi-modal data generated by the UAV during task execution is collected, a visual language large model is used as a base, and through deep multi-modal fusion, visual information and language information can be effectively integrated, complex natural language instructions can be accurately understood and spatial reasoning can be combined with environmental semantics, solving the problems of the related technology in semantic analysis and visual understanding, thereby realizing accurate target positioning and path planning; through the two-stage training process of supervised fine-tuning and reinforcement learning, the model can perform better generalization and robustness in unknown environments; by introducing a chain reasoning mechanism, an intermediate reasoning process is generated before the navigation result is output, making the navigation decision more transparent and explainable, which not only facilitates user understanding and verification of system behavior, but also is conducive to operation and maintenance and safety supervision, improving the controllability and credibility of the system.

[0040] The unmanned aerial vehicle navigation method based on a visual language model provided in the embodiments of the present application relates to the technical field of unmanned aerial vehicle control. The unmanned aerial vehicle navigation method based on a visual language model provided in the embodiments of the present application can be applied to a terminal, can also be applied to a server, and can further be software running in the terminal or the server. In some embodiments, the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, a vehicle-mounted terminal, and the like, but is not limited thereto; the server end can be configured as a stand-alone physical server, can also be configured as a server cluster or a distributed system formed by multiple physical servers, can further be configured as a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and big data and artificial intelligence platforms, and the server can also be a node server in a blockchain network; and the software can be an application program for implementing the unmanned aerial vehicle navigation method based on a visual language model, and the like, but is not limited to the above forms.

[0041] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment, in which tasks are performed by remote processing devices connected by a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0042] Please refer to Figure 1 , Figure 1 is an optional flowchart of the unmanned aerial vehicle navigation method based on a visual language model provided in the embodiments of the present application, Figure 1 The method in the flowchart can include, but is not limited to, steps S101 to S106.

[0043] Step S101, collecting multi-modal data generated by the unmanned aerial vehicle during task execution; wherein the multi-modal data includes natural language instruction text, environment images, semantic maps, and unmanned aerial vehicle action sequences; In a specific implementation, first, multi-modal data is collected from a UAV task scene, the multi-modal data including natural language instruction texts, environment images, semantic maps, and UAV action sequences, and the multi-modal data is used as a basis for subsequent model training.

[0044] In step S102, data cleaning processing is performed on the multi-modal data to obtain an initial training data set. In some embodiments, step S102 can include data cleaning of the natural language instruction texts and the environment images; the data cleaning includes noise elimination processing and outlier removal processing; standardization processing is performed on the natural language instruction texts after data cleaning, and semantic annotation processing is performed on the environment images and the semantic maps after data cleaning; and the initial training data set is constructed according to the natural language instruction texts after standardization processing, the environment images after semantic annotation processing, the semantic maps after semantic annotation processing, and the UAV action sequences.

[0045] Before model training, the original multi-modal data collected needs to be processed to ensure data quality and consistency. In this embodiment of the application, the original multi-modal data is cleaned by using a hierarchical processing method. In this embodiment of the application, data cleaning mainly includes two links of data cleaning and data preprocessing, and the specific content is as follows: (1) Data cleaning: comprehensively checking the collected original data (including natural language instruction texts and environment images): 1) deleting or eliminating noise data in the natural language instruction texts and the environment images, such as invalid instructions, garbled texts, damaged images, and severely blurred images, etc. 2) removing outlier samples in the natural language instruction texts and the environment images, such as obviously inconsistent with the task target or lacking necessary information, to ensure the overall quality of the data set.

[0046] (2) Data preprocessing: after data cleaning, the remaining data is further standardized according to the task: 1) natural language instruction text standardization: using the same set of prompt templates for all natural language instruction texts to ensure that the sentence structure conforms to the task template; it should be noted that the construction of the prompt template can be set according to the actual situation, and this embodiment of the application does not limit this; 2) environment image and semantic map annotation: manually or automatically annotating the environment images or semantic maps, such as supplementing the boundary box information of buildings, roads, and key landmarks, etc., so as to enable the model to learn the corresponding relationship between natural language and visual semantics.

[0047] Step S103, constructing a thought chain for the initial training data set by using a thought chain generation model to generate high-quality thought chain data, and constructing a target training data set according to the initial training data set and the high-quality thought chain data. In some embodiments, step S103 can include: constructing a thought chain for the initial training data set by using a thought chain generation model to generate initial thought chain data; performing quality screening on the initial thought chain data by using a data quality screening mechanism to obtain high-quality thought chain data; and constructing a target training data set according to the initial training data set and the high-quality thought chain data.

[0048] In the embodiments of the present application, an external large model (such as GPT-4o) is introduced to generate high-quality thought chains (Chain-of-Thought, CoT) to enhance the reasoning dimension of the training data. Specifically, first, a thought chain generation model is used to construct a thought chain for the initial training data set to generate initial thought chain data. In this process, the embodiments of the present application not only generate initial thought chain data, but also design a data quality screening mechanism: the generated initial thought chain data is checked one by one, and samples with unclear, insufficient reasoning process or incorrect reasoning result are removed, and only high-confidence data (i.e., high-quality thought chain data) with logical and correct conclusions are retained. Through this screening step, the interference of possible errors introduced by the external model is effectively avoided, and the accuracy and interpretability of the reasoning data are guaranteed. Finally, a target training data set is constructed according to the initial training data set and the high-quality thought chain data. The target training data set is no longer a binary form of "input-label", but a ternary structure containing input, label and high-quality thought chain data. The subsequent model supervised fine-tuning (SFT) is performed using this data set, so that the model not only learns the mapping relationship between input and output in the first stage (supervised fine-tuning stage), but also learns the human-like reasoning chain, thereby having stronger explainability and generalization ability in subsequent task execution.

[0049] Step S104, performing model training on the initial visual language model according to the target training data set based on a supervised fine-tuning method and a reinforcement learning method to obtain a candidate visual language model. In some embodiments, step S104 can include: performing supervised fine-tuning on the initial visual language model according to the target training data set based on a supervised fine-tuning method to obtain a fine-tuned visual language model; wherein the fine-tuned visual language model has preliminary reasoning ability and preliminary navigation prediction ability; and performing model optimization processing on the fine-tuned visual language model according to the target training data set based on a reinforcement learning method to obtain a candidate visual language model.

[0050] In some embodiments, based on the reinforcement learning method, the step of performing model optimization on the fine-tuned visual language model according to the target training data set to obtain a candidate visual language model can include: inputting the target training data set into the fine-tuned visual language model to output a model inference result and a model predicted position; calculating a reward value using a composite reward function according to the model inference result and the model predicted position; wherein the composite reward function includes a format reward function, a target distance reward function and an intermediate process reward function; determining a target model loss for the fine-tuned visual language model according to the reward value; and optimizing the fine-tuned visual language model according to the target model loss to obtain the candidate visual language model.

[0051] In the embodiments of the present application, the training process of the visual language model includes a supervised fine-tuning (SFT) stage and a reinforcement learning (RL) stage.

[0052] For the supervised fine-tuning (SFT) stage, please refer to Figure 2 , Figure 2 is a model supervised fine-tuning process diagram provided by the embodiments of the present application, as Figure 2 shown, before performing supervised fine-tuning of the model, the inference dimension of the training data needs to be enhanced, specifically, Figure 2 The training data in the step S102 refers to the multi-modal data (i.e., the initial data set) obtained after data cleaning in the step S102. The basic structure of the initial data set includes input (picture and text information) and label (target position information and landmark position information). Different from the related art in which the unmanned aerial vehicle (UAV) navigation task only relies on the input and label, the embodiment of the present application introduces a method for generating high-quality thought chains by an external large model (such as GPT-4o) to enhance the reasoning dimension of the training data. Specifically, first, a thought chain generation model is used to construct the initial training data set to generate initial thought chain data. In this process, the embodiment of the present application not only generates the initial thought chain data, but also designs a data quality screening mechanism: the generated initial thought chain data is checked one by one, and samples with unclear, insufficient or incorrect reasoning results are removed, and only high-confidence data (i.e., high-quality thought chain data) with logical and correct conclusions are retained. Through this screening step, the interference of possible errors introduced by the external model is effectively avoided, and the accuracy and interpretability of the reasoning data are ensured. Further, according to the initial training data set and the high-quality thought chain data, a target training data set is constructed. The target training data set is no longer a binary form of "input-label", but a ternary structure including input, label and high-quality thought chain data. Finally, the target training data set is used for supervised fine-tuning (SFT) of the model, so that the model not only learns the mapping relationship between the input and the output in the first stage (supervised fine-tuning stage), but also learns the human-like reasoning chain, thereby having stronger interpretability and generalization ability in subsequent task execution.

[0053] For the reinforcement learning (RL) stage, please refer to Figure 3 , Figure 3 is a model reinforcement learning process schematic diagram provided by the embodiment of the present application, as shown in Figure 3 , in the reinforcement learning stage, the embodiment of the present application continues to optimize the model based on the model that has completed the supervised fine-tuning (SFT). Since the model trained by the supervised fine-tuning (SFT) has a preliminary reasoning and navigation prediction ability, the reinforcement learning stage can further improve the performance on the basis of the supervised fine-tuning (SFT). Specifically, in the reinforcement learning (RL) stage, after the visual language model receives the input and label (i.e., the target data set of the reinforcement learning stage only needs input and label), the visual language model outputs the reasoning process <think>(i.e., model inference results) and final predicted coordinates <answer>The output will be used as the basis for reward calculation. To improve the format specification, target accuracy and reasoning interpretability of the visual language model at the same time, the embodiment of the application proposes a composite reward function design method, which specifically includes the following three parts: (1) Format reward: ensure that the reasoning process and the final coordinates of the model are strictly derived from the structured results of the model predicted position (i.e. <think>With <answer>The label is wrapped to ensure the output structure and controllability; (2) Target distance reward: dynamically calculate the reward based on the Euclidean distance between the predicted target position and the real target position. The closer the distance, the higher the reward, thereby directly optimizing the navigation accuracy of the UAV. (3) Intermediate process reward: the model is required to explicitly output the position of the landmark during the inference process, and is represented in the form of a landmark bounding box (landmark_bbox). The intersection over union (IoU) between the predicted bounding box and the real landmark bounding box is used as a reward signal to guide the model to generate more reasonable and interpretable semantic information during the intermediate inference step.

[0054] The embodiments of the present application form a composite reward mechanism for a UAV (UAV) visual-language navigation task through the weighted combination of the above three rewards (format reward, target distance reward, and intermediate process reward). The mechanism not only ensures the correctness of the output format, but also simultaneously includes the final target prediction and the intermediate inference process in the optimization target, thereby effectively improving the comprehensive performance of the model. Compared with related technologies, the composite reward of the embodiments of the present application not only can significantly improve the accuracy of target prediction, but also can enhance the rationality and interpretability of the inference chain, achieving the efficiency, robustness and transparency of the navigation task.

[0055] In step S105, the candidate visual language model is subjected to model verification, the candidate visual language model passing the model verification is taken as a target visual language model, and the target visual language model is deployed in a UAV navigation control system; After the model training is completed, the model is verified in an independent test set and a simulation environment to evaluate the success rate, generalization and inference interpretability of the model in different task scenarios. If the model does not achieve the expected effect, the training process is returned to continue optimization until the model achieves the expected effect. The model passing the model verification can be applied to an actual UAV navigation task. It should be noted that the definition of the expected effect can be set according to actual conditions, and the embodiments of the present application do not limit this.

[0056] The model passing the verification is saved and deployed in a UAV navigation control system (such as a UAV on-board computing unit or a ground station control device) for real-time task execution.

[0057] In step S106, a current user natural language instruction is input into the UAV navigation control system, a current inference result and a current target position are generated by the target visual language model according to current real environment information, and a target action sequence is predicted according to the current target position by using a forward-looking mechanism, so that the current UAV executes the target action sequence.

[0058] In some embodiments, after the step of inputting the current user natural language instruction into the unmanned aerial vehicle navigation control system, generating the current reasoning result and the current target position according to the current real environment information by the target visual language model, the method can further include: outputting the current reasoning result to a visual interface for display.

[0059] Please refer to Figure 4 , Figure 4 is a model navigation algorithm flowchart provided by the embodiment of the application, as Figure 4 shown, the specific implementation process of the model navigation algorithm in real-time task execution is as follows: (1) User input: the user provides a natural language instruction for the unmanned aerial vehicle, and the natural language instruction is input in text form into the visual language model verified by the model.

[0060] (2) Real environment information is input into the model, and the real environment information includes the unmanned aerial vehicle and the environment (environment map) in which the unmanned aerial vehicle is located. The real environment information specifically includes the following contents: 1) Environment map: the model is provided with a picture, and the picture is marked with the entire block map, the current position of the unmanned aerial vehicle and the landmark. 2) Unmanned aerial vehicle state: the coordinates, altitude and orientation angle of the unmanned aerial vehicle are provided in text form.

[0061] (3) Model reasoning: after the model receives the above inputs (user input and real environment information), the reasoning result (i.e. reasoning process) and target position are generated: 1) Predicted target position: the coordinates that the unmanned aerial vehicle should reach in the current stage.

[0062] 2) Output reasoning process: explicitly show the reasoning steps of the model, so that researchers and users can observe the intermediate process and enhance the explainability and controllability of navigation. The reasoning process output by the model can be directly displayed in the back end, or the reasoning process can be directly output to the client, and the embodiment of the application does not limit this.

[0063] (4) Prospective mechanism: in the embodiment of the application, the action space of the unmanned aerial vehicle is a limited discrete action such as forward, left turn, right turn, rise, fall and stop. The prospective mechanism is a rule-based relative position inference. The predicted action sequence is obtained, and the specific implementation process is as follows: 1) Relative position calculation between target and current position: Suppose the model outputs the two-dimensional coordinates (xt, yt) of the predicted target, and the current position and orientation of the unmanned aerial vehicle are (x0, y0, θ); the relative vector Δx = xt–x0, Δy = yt–y0 of the target point and the current position can be calculated, and combined with the orientation θ to be converted to the local coordinate system of the unmanned aerial vehicle itself.

[0064] 2) Direction determination: If the target is in front of the UAV (within 30° of the y-axis in the local coordinate system), choose to move forward; If the target is in the front left (x < 0 and y > 0), first choose to turn left (fixed 30° at a time), until the target meets the "UAV front" condition, then execute forward; If the target is in the front right (x > 0 and y > 0), first choose to turn right (fixed 30° at a time), until the target meets the "UAV front" condition, then execute forward; If the target is directly below or directly above, execute descent or ascent; When the UAV reaches the target area (Euclidean distance to the target is less than the threshold), execute stop.

[0065] 3) Action output According to the above direction determination rule, determine all the actions that the current UAV should execute; plan all the actions of the current stage at a time, guide the UAV to fly to the vicinity of the current predicted target point, and then execute the flight of the next stage.

[0066] (5) UAV executes action, the UAV executes corresponding operation according to the received UAV flight instruction according to the action sequence provided by the model, after executing the action, the real environment will also change: the UAV position in the map is updated, the corresponding picture is also updated; the UAV's own state (coordinates, altitude, orientation) is also updated.

[0067] (6) Loop iteration, repeat steps (1) - step (5) of the model navigation algorithm in real-time task execution until the system determines that the navigation task is completed, and the whole navigation process is over.

[0068] The steps S101 to S106 shown in the embodiments of the present application are as follows: multi-modal data generated by the UAV during task execution is collected; the multi-modal data includes natural language instruction text, environment images, semantic maps, and UAV action sequences; the multi-modal data is subjected to data cleaning processing to obtain an initial training data set; a thought chain generation model is used to construct a thought chain for the initial training data set, generate high-quality thought chain data, and construct a target training data set according to the initial training data set and the high-quality thought chain data; based on a supervised fine-tuning method and a reinforcement learning method, the initial visual language model is trained according to the target training data set to obtain a candidate visual language model; the candidate visual language model is subjected to model verification, the candidate visual language model that passes the model verification is used as a target visual language model, and the target visual language model is deployed in a UAV navigation control system; a current user natural language instruction is input into the UAV navigation control system, the target visual language model is used to generate a current reasoning result and a current target position according to current real environment information, and a look-ahead mechanism is used to predict a target action sequence according to the current target position, so that the current UAV executes the target action sequence. The embodiments of the present application collect multi-modal data generated by the UAV during task execution, use a visual language large model as a base, effectively integrate visual information and language information through deep multi-modal fusion, accurately understand complex natural language instructions and combine environment semantics for spatial reasoning, solve the problems of related technologies in semantic analysis and visual understanding, and thus realize accurate target positioning and path planning; through the two-stage training process of supervised fine-tuning and reinforcement learning, the model can exhibit stronger generalization and robustness in unknown environments; through the introduction of a chain reasoning mechanism, an intermediate reasoning process is generated before the output of the navigation result, making the navigation decision more transparent and interpretable, which not only facilitates user understanding and verification of system behavior, but also is conducive to operation and maintenance and safety supervision, and improves the controllability and reliability of the system.

[0069] To explain the principle of the technical scheme of the present application in detail, the overall process of the present application will be described below in conjunction with some specific embodiments. It should be easily understood that the following is an explanation of the technical principle of the present application and cannot be regarded as a limitation of the present application.

[0070] The unmanned aerial vehicle navigation method based on the visual language model provided in the embodiments of the present application mainly aims at the following three points: (1) improving semantic understanding and spatial reasoning ability by deeply fusing image and text information; (2) designing a two-stage training process (supervised fine-tuning + reinforcement learning optimization), using high-quality demonstrations and a composite reward mechanism to enable the model to have stronger generalization and adaptation ability in complex, dynamic or even unknown environments; (3) introducing a chain-of-thought (CoT) mechanism to enable the system to output intermediate reasoning steps before decision-making, thereby enhancing the explainability and controllability of the navigation process. Through these contents, the embodiments of the present application can significantly improve the autonomous navigation success rate and reliability of the unmanned aerial vehicle (UAV) in a complex environment, and realize efficient, stable and interpretable unmanned aerial vehicle visual-language navigation capability.

[0071] Please refer to Figure 5 , Figure 5 is a system framework process schematic diagram of the unmanned aerial vehicle navigation method based on the visual language model provided in the embodiments of the present application, as Figure 5 indicated, the overall running process of the embodiments of the present application includes six steps of data acquisition, data cleaning, model training, model verification, model saving and task execution, and the specific implementation process of each step is as follows: First step, data acquisition: In the specific implementation, first, multi-modal data is collected from the unmanned aerial vehicle task scene, the multi-modal data including natural language instruction text, environment image, semantic map and unmanned aerial vehicle action sequence, and these multi-modal data are used as the basis for subsequent model training.

[0072] Second step, data cleaning: Before model training, the collected original multi-modal data needs to be processed to ensure data quality and consistency. The embodiments of the present application process the original multi-modal data through a hierarchical processing manner for data cleaning. In the embodiments of the present application, data cleaning mainly includes two links of data cleaning and data preprocessing, and the specific contents are as follows: (1) Data cleaning, comprehensively checking the collected original data (including natural language instruction text and environment image): 1) deleting or rejecting noise data in the natural language instruction text and the environment image, such as invalid instructions, garbled text, damaged images and severely blurred images, etc.; 2) removing abnormal value samples in the natural language instruction text and the environment image, such as information obviously inconsistent with the task target or lacking necessary labels, to ensure the overall quality of the data set.

[0073] (2) Data preprocessing, after completing data cleaning, further normalizing the data related to the task for the remaining data: 1) Natural language instruction text standardization: use the same set of prompt templates for all natural language instruction texts to ensure that the sentence structure conforms to the task template; 2) Environment image and semantic map labeling: manually or automatically label environment images or semantic maps, such as supplementing building, road, and key landmark bounding box information, to enable the model to learn the correspondence between natural language and visual semantics.

[0074] Finally, based on the standardized natural language instruction text, the environment image and semantic map labeled with semantics, and the UAV action sequence, an initial training data set is constructed.

[0075] Third step, model training, model training is divided into two stages: supervised fine-tuning (SFT) stage and reinforcement learning (RL) stage: (1) Supervised fine-tuning (SFT): use high-quality labeled data to perform supervised learning on the model to enable it to master the basic mapping ability from input (instruction + semantic map) to output (reasoning process + target position); Specifically, first, a thought chain generation model is used to construct the initial training data set obtained by data cleaning in the second step to generate initial thought chain data, and in this process, the embodiments of the present application not only generate initial thought chain data, but also design a data quality screening mechanism: the generated initial thought chain data is checked one by one, and samples with unclear, insufficient reasoning process or incorrect reasoning result are removed, and only high-confidence data (i.e. high-quality thought chain data) with logical and correct conclusions are retained. Through this screening step, the interference of possible errors introduced by external models is effectively avoided, ensuring the accuracy and interpretability of the reasoning data; further, based on the initial training data set and the high-quality thought chain data, a target training data set is constructed, wherein the target training data set is no longer a binary form of "input—label", but a ternary structure containing input, label and high-quality thought chain data; finally, the target training data set is used to perform supervised fine-tuning (SFT) of the model, so that the model not only learns the mapping relationship between input and output in the first stage (supervised fine-tuning stage), but also learns the human-like reasoning chain, thereby having stronger interpretability and generalization ability in subsequent task execution.

[0076] (2) Reinforcement Learning (Reinforcement Learning, RL): On the basis of supervised fine-tuning, further introduce a compound reward mechanism (including target accuracy reward (i.e. target distance reward), reasoning rationality reward (i.e. intermediate process reward) and format specification reward (i.e. format reward) to optimize the generalization ability and task execution stability of the model.

[0077] For the reinforcement learning (RL) stage, the present application continues to optimize based on the model that has completed supervised fine-tuning (SFT). Since the model trained through supervised fine-tuning (SFT) already has preliminary reasoning and navigation prediction capabilities, the reinforcement learning stage can further improve performance on the basis of supervised fine-tuning (SFT). Specifically, in the reinforcement learning (RL) stage, after the visual language model receives the input and label, it will output a reasoning process <think>(i.e., model inference results) and final predicted coordinates <answer>The output will be used as the basis for reward calculation. To improve the format specification, target accuracy and reasoning interpretability of the visual language model at the same time, the embodiment of the application proposes a composite reward function design method, which specifically includes the following three parts: 1) Format reward: ensure that the reasoning process and final coordinates of the model are strictly derived from <think>With <answer>Label wrapping, thereby ensuring the output structure and controllability; 2) Target distance reward: dynamically calculate the reward based on the Euclidean distance between the predicted target position and the real target position, the closer the distance, the higher the reward, thereby directly optimizing the navigation accuracy of the UAV; 3) Intermediate process reward: the model is required to explicitly output the position of the landmark during the inference process, and is represented in the form of a landmark bounding box (landmark_bbox). The intersection over union (IoU) between the predicted bounding box and the real landmark bounding box is used as a reward signal to guide the model to generate more reasonable and interpretable semantic information during the intermediate inference step.

[0078] The embodiments of the present application form a composite reward mechanism for the UAV visual-language navigation task through the weighted combination of the above three rewards (format reward, target distance reward, and intermediate process reward). The mechanism not only ensures the correctness of the output format, but also simultaneously includes the final target prediction and the intermediate inference process in the optimization target, thereby effectively improving the comprehensive performance of the model. Compared with related technologies, the composite reward of the embodiments of the present application not only can significantly improve the accuracy of target prediction, but also can enhance the rationality and interpretability of the inference chain, achieving the efficiency, robustness and transparency of the navigation task.

[0079] Fourth step, model verification: After the model training is completed, the model is verified in an independent test set and a simulated environment to evaluate its success rate, generalization and inference interpretability in different task scenarios. If the model does not achieve the expected effect, return to the training link for continuous optimization, until the model achieves the expected effect, then the model that passes the model verification can be applied to the actual UAV navigation task.

[0080] Fifth step, model saving: The verified model will be saved and deployed to the UAV navigation control system (such as the UAV on-board computing unit or the ground station control device) for real-time task execution.

[0081] Sixth step, task execution: In actual application, the user inputs a natural language instruction, the model generates an inference result and a target position according to the current environment information, and then predicts an action sequence according to the target position based on the look-ahead mechanism, and the UAV flight control module controls the UAV to execute the predicted action sequence, realizing autonomous navigation and task completion.

[0082] It should be noted that the embodiments only briefly and schematically illustrate the overall process of the UAV navigation method based on the visual language model, and the detailed description of each step can refer to the related content in the foregoing embodiments, which will not be repeated here. It can be understood that the present application does not limit this.

[0083] The embodiment of the application collects multi-modal data generated by the unmanned aerial vehicle in the process of performing a task, wherein the multi-modal data includes natural language instruction text, environment images, semantic maps, and unmanned aerial vehicle action sequences; the multi-modal data is subjected to data cleaning processing to obtain an initial training data set; a thought chain generation model is used to construct a thought chain for the initial training data set, generate high-quality thought chain data, and construct a target training data set according to the initial training data set and the high-quality thought chain data; based on a supervised fine-tuning method and a reinforcement learning method, an initial visual language model is trained according to the target training data set to obtain a candidate visual language model; the candidate visual language model is subjected to model verification, the candidate visual language model that passes the model verification is used as a target visual language model, and the target visual language model is deployed in an unmanned aerial vehicle navigation control system; a current user natural language instruction is input into the unmanned aerial vehicle navigation control system, a current reasoning result and a current target position are generated by the target visual language model according to current real environment information, and a target action sequence is predicted according to the current target position by using a forward-looking mechanism, so that the current unmanned aerial vehicle performs the target action sequence. The embodiment of the application collects multi-modal data generated by the unmanned aerial vehicle in the process of performing a task, uses a visual language large model as a base, effectively integrates visual information and language information through deep multi-modal fusion, accurately understands complex natural language instructions and combines environment semantics for spatial reasoning, solves the problem of insufficient semantic analysis and visual understanding in related technologies, and thus realizes accurate target positioning and path planning; through the two-stage training process of supervised fine-tuning and reinforcement learning, the model can perform stronger generalization and robustness in unknown environments; by introducing a chain reasoning mechanism, an intermediate reasoning process is generated before the navigation result is output, so that the navigation decision is more transparent and interpretable, which not only facilitates user understanding and verification of system behavior, but also is beneficial to operation and maintenance and safety supervision, and improves the controllability and reliability of the system.

[0084] In summary, the key points of the embodiment of the application mainly lie in the following aspects: (1) Two-stage training framework (supervised fine-tuning + reinforcement learning optimization): In the supervised fine-tuning stage, high-quality labeled data is used to make the model learn the mapping relationship from "natural language and environment information" to "process reasoning and target prediction"; through this stage, the model not only obtains basic reasoning capability in the unmanned aerial vehicle navigation scene, but also can output <think>With <answer>the structured results of the first stage, thus laying a solid foundation for the subsequent optimization stage.

[0085] In the reinforcement learning stage, the Group Relative Policy Optimization (GRPO) method is introduced for the model trained in the previous supervised fine-tuning stage, and a composite reward function is designed to further optimize the model. This stage encourages the model to continuously improve the reasoning chain and the target prediction results during exploration, effectively improving the model's generalization ability and stability in executing tasks in out-of-distribution environments.

[0086] (2) Design and application of composite reward mechanism: The reward function not only considers the accuracy of the predicted target position, but also introduces reasoning completeness reward and format specification reward. Through multi-dimensional reward signals, the model is guided to output a reasoning chain that is more understandable and verifiable by humans, enhancing the system's explainability.

[0087] (3) Introduction of chain reasoning mechanism: The model automatically generates intermediate reasoning steps before outputting the final navigation result <think>re-outputs the predicted target position <answer>The mechanism makes the UAV navigation process transparent and traceable, facilitating researchers and users to understand, verify, and supervise.

[0088] Compared with the optimal UAV visual-language navigation technology in the related art (such as the CityNav method based on map prediction or the large-scale VLM direct inference scheme), the embodiments of the present application have obvious advantages in the following aspects: (1) First, the embodiments of the present application use a visual-language large model (VLM) as a base, effectively integrate visual and language information through deep multi-modal fusion, can accurately understand complex natural language instructions and combine environmental semantics for spatial reasoning, and solve the shortcomings of existing methods in semantic analysis and visual understanding.

[0089] (2) Secondly, the embodiments of the present application adopt a two-stage training process (supervised fine-tuning + reinforcement learning), and design a composite reward function combining target accuracy, reasoning rationality and format standardization, so that the model performs better generalization and robustness in unknown environments, which is obviously superior to traditional methods relying on a single scene or dataset.

[0090] (3) Finally, the present application introduces a chain reasoning mechanism to generate an intermediate reasoning process before outputting the navigation result, making the navigation decision more transparent and interpretable. This not only facilitates user understanding and verification of system behavior, but also helps operation and safety supervision, improving the controllability and credibility of the system.

[0091] Therefore, the embodiments of the present application have the following advantages: (1) Improve semantic understanding and spatial reasoning ability: by introducing a visual-language large model (VLM) and designing a deep fusion mechanism, complex natural language instructions can be more accurately parsed, and precise target positioning and path planning can be achieved by combining scene semantic information.

[0092] (2) Enhance cross-environment generalization ability: adopt a two-stage training process (supervised fine-tuning + reinforcement learning) and combine a composite reward mechanism to effectively improve the navigation success rate and robustness of the model in out-of-distribution environments, adapting to complex dynamic environments such as post-disaster and unfamiliar urban blocks.

[0093] (3) Realize interpretable navigation: output intermediate reasoning steps through the chain reasoning mechanism, making the decision-making process of the UAV transparent and traceable, facilitating verification, debugging and safety supervision of the task execution process.

[0094] (4) Improve task execution efficiency and safety: in a variable environment, the embodiments of the present application can reduce the failure rate caused by GPS errors, environmental occlusion and instruction ambiguity, thereby improving the efficiency of UAV task completion and reducing manual intervention and potential risks.

[0095] (5) Expanding the application range: the embodiments of the application can be applied in disaster emergency, city inspection, security patrol, logistics distribution and other scenes, and can significantly improve the intelligent level and value of the unmanned aerial vehicle in actual application.

[0096] Please refer to Figure 6 The embodiment of the application also provides an unmanned aerial vehicle navigation device 600 based on a visual language model, which can implement the above method, and the device comprises the following modules: A data acquisition module 601 is configured to acquire multi-modal data generated by the unmanned aerial vehicle during task execution; wherein the multi-modal data comprises natural language instruction text, environment images, semantic maps and unmanned aerial vehicle action sequences. A data cleaning module 602 is configured to perform data cleaning processing on the multi-modal data to obtain an initial training data set. A thought chain construction module 603 is configured to use a thought chain generation model to construct a thought chain for the initial training data set, generate high-quality thought chain data, and construct a target training data set according to the initial training data set and the high-quality thought chain data. A model training module 604 is configured to perform model training on an initial visual language model according to the target training data set based on a supervised fine-tuning method and a reinforcement learning method, to obtain a candidate visual language model. A model verification module 605 is configured to perform model verification on the candidate visual language model, use the candidate visual language model that passes the model verification as a target visual language model, and deploy the target visual language model in an unmanned aerial vehicle navigation control system. A model prediction module 606 is configured to input a current user natural language instruction into the unmanned aerial vehicle navigation control system, generate a current reasoning result and a current target position according to current real environment information through the target visual language model, and predict a target action sequence according to the current target position using a forward-looking mechanism, so that the current unmanned aerial vehicle executes the target action sequence.

[0097] It can be understood that the contents in the above method embodiments are applicable to the device embodiments, the device embodiments specifically implement the functions of the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0098] The embodiment of the application also provides an electronic device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the above method when executing the computer program. The electronic device can be any intelligent terminal including a tablet computer, a vehicle-mounted computer, etc.

[0099] It can be understood that the contents in the above method embodiments are all applicable to the present device embodiments, the present device embodiments specifically implement the functions same as those of the above method embodiments, and achieve the same beneficial effects as those of the above method embodiments.

[0100] Please refer to Figure 7 , Figure 7 The hardware structure of the electronic device of another embodiment is illustrated, and the electronic device comprises: The processor 701 can be implemented in a general-purpose CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits, and is used to execute related programs to implement the technical solutions provided by the embodiments of the present application. The memory 702 can be implemented in the form of a ROM (Read Only Memory), a static storage device, a dynamic storage device, or a RAM (Random Access Memory). The memory 702 can store an operating system and other application programs. When the technical solutions provided by the embodiments of the present application are implemented by software or firmware, the related program codes are stored in the memory 702 and are called and executed by the processor 701 to implement the above-mentioned methods of the embodiments of the present application. The input / output interface 703 is used to realize information input and output. The communication interface 704 is used to realize the communication interaction between the present device and other devices. The communication can be realized by a wired manner (for example, a USB, a network cable, etc.) or a wireless manner (for example, a mobile network, WIFI, Bluetooth, etc.). The bus 705 is used to transmit information between the components (for example, the processor 701, the memory 702, the input / output interface 703, and the communication interface 704) of the device. The processor 701, the memory 702, the input / output interface 703, and the communication interface 704 are connected to each other through the bus 705 to realize the communication connection between them in the device.

[0101] The embodiments of the present application further provide a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the above-mentioned method.

[0102] It can be understood that the contents in the above method embodiments are all applicable to the present storage medium embodiments, the present storage medium embodiments specifically implement the functions same as those of the above method embodiments, and achieve the beneficial effects same as those of the above method embodiments.

[0103] The present application also provides a computer program product, comprising a computer program which is executed by a processor to implement the above method.

[0104] It can be understood that the contents in the above method embodiments are all applicable to the present program product embodiments, the present program product embodiments specifically implement the functions same as those of the above method embodiments, and achieve the beneficial effects same as those of the above method embodiments.

[0105] The memory, as a non-transitory computer readable storage medium, can be used to store non-transitory software programs and non-transitory computer executable programs. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory remotely arranged relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0106] The unmanned aerial vehicle navigation method and device based on a visual language model provided by the embodiments of the present application, the electronic device, the storage medium and the program product, collect multi-modal data generated by the unmanned aerial vehicle during task execution, wherein the multi-modal data includes natural language instruction text, environment images, semantic maps and unmanned aerial vehicle action sequences, perform data cleaning processing on the multi-modal data to obtain an initial training data set, use a thought chain generation model to construct a thought chain for the initial training data set, generate high-quality thought chain data, and construct a target training data set according to the initial training data set and the high-quality thought chain data, perform model training on an initial visual language model according to the target training data set based on a supervised fine-tuning method and a reinforcement learning method, obtain a candidate visual language model, perform model verification on the candidate visual language model, use the candidate visual language model that passes the model verification as a target visual language model, deploy the target visual language model on an unmanned aerial vehicle navigation control system, input a current user natural language instruction into the unmanned aerial vehicle navigation control system, generate a current reasoning result and a current target position according to current real environment information through the target visual language model, and predict a target action sequence according to the current target position by using a forward-looking mechanism, so that the current unmanned aerial vehicle executes the target action sequence. The embodiments of the present application collect multi-modal data generated by the unmanned aerial vehicle during task execution, use a visual language large model as a base, effectively integrate visual information and language information through deep multi-modal fusion, accurately understand complex natural language instructions and perform spatial reasoning combined with environment semantics, solve the problems of related technologies in semantic analysis and visual understanding, and thus realize accurate target positioning and path planning. Through the two-stage training process of supervised fine-tuning and reinforcement learning, the model can perform stronger generalization and robustness in unknown environments. By introducing a chain reasoning mechanism, an intermediate reasoning process is generated before the output of the navigation result, so that the navigation decision is more transparent and interpretable, which not only facilitates user understanding and verification of system behavior, but also is beneficial to operation and maintenance and safety supervision, and improves the controllability and reliability of the system.

[0107] The embodiments described in the embodiments of the present application are used to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that, with the evolution of technology and the appearance of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0108] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and can include more or fewer steps than those shown in the figures, or combine certain steps or different steps.

[0109] The apparatus embodiments described above are merely exemplary, and units described as separate units can or can not be physically separate, i.e., can be located in one place, or can be distributed over multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.

[0110] Those skilled in the art can understand that all or some of the steps in the method disclosed above, the functional modules / units in the system and the device can be implemented as software, firmware, hardware and appropriate combinations thereof.

[0111] The terms "first", "second", "third", "fourth" and the like in the description of the application and in the claims of the foregoing drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so

[0112] It should be understood that in this application, "at least one" means one or more, "multiple" means two or more. "And / or", used to describe the relationship between associated objects, means that there can be three relationships, for example, "A and / or B" can mean: only A, only B, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally represents a "or" relationship between the associated objects. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c, can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0113] In several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can be implemented by other manners. For example, the apparatus embodiments described above are merely illustrative, for example, the division of the above units is merely a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or components shown or discussed can be indirect coupling or communication connection through some interfaces, apparatuses or units, and can be electrical, mechanical or other forms.

[0114] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they can be located in one place or distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0115] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0116] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that makes a contribution or the whole or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method of each embodiment of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program storage media.

[0117] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, but this does not limit the scope of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the embodiments of the present application.< / answer> < / think> < / answer> < / think> < / answer> < / think> < / answer> < / think> < / answer> < / think> < / answer> < / think> < / answer> < / think> < / answer> < / think>

Claims

1. A method for unmanned aerial vehicle navigation based on visual language model, characterized in that, The method comprises the following steps: Collecting multi-modal data generated by the unmanned aerial vehicle during task execution; wherein the multi-modal data comprises natural language instruction text, environment images, semantic maps, and unmanned aerial vehicle action sequences; Performing data cleaning processing on the multi-modal data to obtain an initial training data set; Using a thought chain generation model to construct a thought chain for the initial training data set, generate high-quality thought chain data, and construct a target training data set according to the initial training data set and the high-quality thought chain data; Based on a supervised fine-tuning method and a reinforcement learning method, model training is performed on an initial visual language model according to the target training data set to obtain a candidate visual language model; Model verification is performed on the candidate visual language model, the candidate visual language model that passes the model verification is used as a target visual language model, and the target visual language model is deployed in an unmanned aerial vehicle navigation control system; A current user natural language instruction is input into the unmanned aerial vehicle navigation control system, a current reasoning result and a current target position are generated according to current real environment information by using the target visual language model, and a target action sequence is predicted according to the current target position by using a forward-looking mechanism, so that the current unmanned aerial vehicle executes the target action sequence.

2. The method of claim 1, wherein, The data cleaning processing on the multi-modal data to obtain an initial training data set comprises: Data cleaning is performed on the natural language instruction text and the environment images; wherein the data cleaning comprises noise elimination processing and outlier removal processing; Standardization processing is performed on the natural language instruction text after data cleaning, and semantic annotation processing is performed on the environment images and the semantic maps after data cleaning; The initial training data set is constructed according to the natural language instruction text after standardization processing, the environment images after semantic annotation processing, the semantic maps after semantic annotation processing, and the unmanned aerial vehicle action sequences.

3. The method of claim 1, wherein, The thought chain generation model is used to construct a thought chain for the initial training data set to generate initial thought chain data; Data quality screening is performed on the initial thought chain data by using a data quality screening mechanism to obtain the high-quality thought chain data; The target training data set is constructed according to the initial training data set and the high-quality thought chain data. Based on the supervised fine-tuning method, the initial visual language model is supervised and fine-tuned according to the target training data set to obtain a fine-tuned visual language model; wherein the fine-tuned visual language model has preliminary reasoning ability and preliminary navigation prediction ability; 4. The method of claim 1, wherein, ​ ​ Based on the reinforcement learning method, the fine-tuned visual language model is subjected to model optimization processing according to the target training data set, to obtain the candidate visual language model.

5. The method of claim 4, wherein, The method comprises the following steps: inputting the target training data set into the fine-tuned visual language model, and outputting a model inference result and a model predicted position; According to the model inference result and the model predicted position, a reward value is calculated by using a composite reward function; wherein the composite reward function comprises a format reward function, a target distance reward function and an intermediate process reward function; According to the reward value, a target model loss for the fine-tuned visual language model is determined; According to the target model loss, the fine-tuned visual language model is optimized to obtain the candidate visual language model.

6. The method of claim 1, wherein, After the current user natural language instruction is input into the unmanned aerial vehicle navigation control system, and the current inference result and the current target position are generated by the target visual language model according to the current real environment information, the method further comprises: outputting the current inference result to a visual interface for display.

7. A visual language model based unmanned aerial vehicle navigation device, characterized in that, The device comprises the following modules: a data acquisition module for acquiring multi-modal data generated by the unmanned aerial vehicle during task execution; wherein the multi-modal data comprises natural language instruction text, environment images, semantic maps and unmanned aerial vehicle action sequences; a data cleaning module for performing data cleaning processing on the multi-modal data to obtain an initial training data set; a thought chain construction module for constructing a thought chain by using a thought chain generation model to generate high-quality thought chain data, and constructing a target training data set according to the initial training data set and the high-quality thought chain data; a model training module for training an initial visual language model based on a supervised fine-tuning method and a reinforcement learning method according to the target training data set, to obtain a candidate visual language model; a model verification module for verifying the candidate visual language model, using the candidate visual language model that passes the model verification as a target visual language model, and deploying the target visual language model in an unmanned aerial vehicle navigation control system; a model prediction module for inputting a current user natural language instruction into the unmanned aerial vehicle navigation control system, generating a current inference result and a current target position by the target visual language model according to current real environment information, and predicting a target action sequence according to the current target position by using a forward-looking mechanism, so that the current unmanned aerial vehicle executes the target action sequence.

8. An electronic device, comprising: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method of any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement the method of any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1 to 6.

Images