Artificial intelligence general identification education method and system based on bimodal closed loop feedback

Abstract

The invention relates to the technical field of intelligent education, in particular to an artificial intelligence general identification education method and system based on bimodal closed-loop feedback. The method comprises the steps that a project task is acquired, students are guided to acquire situational data through an entity AI teaching aid, and a terminal provides real-time feedback for the acquisition process; entity verification is carried out on an entity AI teaching aid in a real environment, if a task is judged to be failed, the intelligent routing unit classifies problems based on multi-dimensional classification, and a feedback mode is dynamically selected; if the rule feedback mode is selected, matching a predefined rule base to generate a structured instruction, pushing the structured instruction to the terminal, guiding the student to take additional data and performing incremental training; if an Agent collaboration mode is selected, calling a cloud large model API to generate a Scotch bottom type guide dialogue, guiding students to analyze problems and select solutions, and performing entity verification again after data supplementary shooting and incremental training are completed; according to the invention, the practice efficiency and thinking cultivation depth of K12 artificial intelligence education are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to an artificial intelligence general education method and system based on bimodal closed-loop feedback. Background Technology

[0002] With the deep penetration of artificial intelligence technology into various industries, AI literacy education has become an important component of core competency development in the K-12 stage. The Ministry of Education's "Compulsory Education Information Technology Curriculum Standards (2022 Edition)" explicitly requires the cultivation of students' computational thinking and intelligent application abilities. However, existing K-12 AI education products face significant systemic deficiencies in practical teaching.

[0003] Existing technical solutions mainly fall into three categories:

[0004] 1. Graphical programming platforms (such as Scratch and Kitten): provide a visual AI model training interface, but the training results are only on screen simulation and lack the verification process in the physical world, making it difficult for students to establish a complete cognitive link from data to intelligent behavior.

[0005] 2. Educational robot kits (such as Makeblock and LEGO SPIKE): Equipped with sensors and actuators, but the AI ​​model training and hardware execution are disconnected, usually requiring manual migration between different platforms. The operation process is complex and not suitable for classroom time constraints.

[0006] 3. Online AI learning platforms (such as Baidu PaddlePaddle EasyDL Education Edition): provide cloud-based model training services, but the feedback mechanism is limited. When the model fails, it only provides abstract metrics such as accuracy and recall, and cannot provide actionable optimization paths for primary and secondary school students.

[0007] The core defects of the existing technical solutions are as follows:

[0008] 1. Disconnect between learning and application: Model training and entity verification are separated, and students cannot experience the complete closed loop of AI from data to decision-making.

[0009] 2. Ineffective feedback: After failure, only statistical indicators or general suggestions are provided, lacking precise guidance based on specific task contexts.

[0010] 3. Iteration efficiency issues: The traditional "training → exporting → deploying → testing → analysis → retraining" process requires cross-platform operation, and each iteration takes more than 30 minutes, making it difficult to complete effective optimization within class time.

[0011] 4. Guidance mismatch: Different types of problems require different guidance strategies. The existing system adopts a single feedback mode, which cannot balance efficiency and depth. Summary of the Invention

[0012] To address the aforementioned technical issues, this invention provides an AI general education method and system based on bimodal closed-loop feedback. By innovatively constructing a rule-driven and agent-coordinated bimodal adaptive feedback mechanism, the AI ​​model iteration cycle is compressed to classroom-operable time, achieving an efficient closed loop of practice-verification-diagnosis-optimization.

[0013] On one hand, embodiments of the present invention provide a method for general education in artificial intelligence based on bimodal closed-loop feedback, the method comprising the following steps:

[0014] S100: Obtain project tasks and guide students to collect contextualized data through physical AI teaching aids. The terminal provides real-time feedback on the collection process. The contextualized data includes images of target objects. The real-time feedback includes shooting quality detection, diversity reminders, and quantity progress bars.

[0015] S200: Receives category labeling data of target object images from students, uploads the labeling data to the cloud training engine for visual model training, and downloads the trained model to the physical AI teaching aid. The visual model training displays the real-time accuracy curve, the recognition status of each category, and the estimated completion time.

[0016] S300 performs physical verification of physical AI teaching aids in a real environment. The behavior monitoring unit records the recognition results, execution actions and environmental parameters to determine whether the task is successful or failed.

[0017] If the task fails, the behavior monitoring unit reports the failure data packet to the cloud feedback engine, the context analysis unit extracts key features, and the intelligent routing unit classifies the problem based on multi-dimensional classification and dynamically selects the feedback mode.

[0018] In the S500, if the rule feedback modality is selected, the rule feedback module matches a predefined rule base to generate structured instructions and pushes them to the terminal to guide students to supplement data and perform incremental training. If the agent collaboration modality is selected, the agent collaboration module calls the cloud-based large model API to generate a Socratic guided dialogue to guide students to analyze problems and select solutions. After completing data supplementation and incremental training, entity verification is performed again.

[0019] Optionally, in S100, the step of acquiring project tasks and guiding students to collect contextualized data through physical AI teaching aids, with the terminal providing real-time feedback on the collection process, includes the contextualized data including images of target objects. The real-time feedback includes image quality detection, diversity reminders, and a quantity progress bar, including:

[0020] S110, Receive a project task description sent by the teacher, the project task description includes teaching objectives, target object categories and class time limits;

[0021] S120, the terminal interface displays a task guidance animation, prompting students to take pictures of the target object through the camera of the physical AI teaching aid. The target object images cover different angles, lighting conditions and appearance features.

[0022] S130, the edge computing unit detects the image clarity, whether it contains target objects and whether there are repeated features in real time. If the clarity is lower than the threshold, it triggers a prompt to adjust the shooting distance. If the features of multiple consecutive images are repeated, it triggers a prompt to adjust the shooting angle.

[0023] S140: The terminal interface updates the progress bar in real time, displaying the ratio of the number of collected samples to the target number of samples. When the preset number is reached, a prompt indicating that the collection is complete is triggered.

[0024] Optionally, in step S200, receiving the category annotation data of the target object image from the student, uploading the annotation data to the cloud training engine for visual model training, and downloading the trained model to the physical AI teaching aid includes:

[0025] S210, the terminal interface displays a queue of images to be labeled, prompting students to complete the labeling by clicking on preset category labels, obtain the labeled data, and record the label timestamp and label confidence level; the queue of images to be labeled includes all acquired target object images;

[0026] S220: After the labeled data is verified for integrity, it is packaged and uploaded to the cloud. The cloud training engine calls the pre-trained basic model and uses the transfer learning method to perform incremental training, generating a visual log in real time that includes training rounds and changes in loss value.

[0027] S230 dynamically displays the proportion of samples in each category, confusion matrix, and typical error cases during training. Training is automatically terminated when the accuracy is stable above the threshold for multiple consecutive rounds.

[0028] S240 generates a model performance report after training, including overall accuracy, scene adaptation scores and optimization suggestions, and encrypts and transmits the lightweight model file to the local storage unit of the physical AI teaching aid.

[0029] Optionally, in S300, the step of verifying the physical AI teaching aid in a real environment, with the behavior monitoring unit recording the recognition results, executed actions, and environmental parameters, and determining whether the task is successful or failed, includes:

[0030] S310, the terminal sends a start verification command to the physical AI teaching aid, and the physical AI teaching aid enters the task execution mode and moves in the real environment according to the preset path or command;

[0031] S320, the sensor unit of the physical AI teaching aid collects environmental parameters in real time, including light intensity, distance to obstacles and its own posture data. The recognition module identifies the target object in front in real time and outputs the recognition result including category and confidence level.

[0032] S330, the execution unit drives the mechanical structure to complete the preset action according to the recognition result, and the behavior monitoring unit records the actual execution result simultaneously. The actual execution result includes recognition time, recognition result, type of execution action, degree of action completion and corresponding environmental parameters.

[0033] S340, the result determination unit compares the actual execution result with the task target. If the action is successfully identified and executed correctly multiple times in a row, the task is determined to be successful; otherwise, the task is determined to be a failure and the failure recording process is triggered.

[0034] Optionally, in S400, the behavior monitoring unit reports failure data packets to the cloud feedback engine, the context analysis unit extracts key features, and the intelligent routing unit classifies the problem based on multi-dimensional classification and dynamically selects the feedback modality, including:

[0035] S410, the behavior monitoring unit encapsulates the failed data packets into structured data containing recognition results, executed actions, environmental parameters and timestamps, and uploads them to the cloud feedback engine through an encrypted channel;

[0036] S420, The context analysis unit extracts features from the failed data packets and generates a feature vector, which includes error type, environmental interference factor, data distribution deviation and student operation habit label;

[0037] S430, the intelligent routing unit calls the intelligent routing decision algorithm to match the feature vector with the question type template in the rule base. When the matching degree is higher than the preset threshold, it is determined to be a clear question and the rule feedback mode is selected. When the matching degree is lower than the threshold, it is determined to be an open question and the Agent collaboration mode is selected.

[0038] S440 pushes the feedback mode selection result and corresponding processing priority to the task scheduling unit, which then allocates computing resources based on the current system load.

[0039] Optionally, in S500, if the Agent collaborative modality is selected, the Agent collaborative module calls the cloud-based large model API to generate a Socratic guided dialogue, guiding students to analyze the problem and select a solution. After completing data re-enhancing and incremental training, entity verification is performed again, including:

[0040] S510, the context analysis unit extracts environmental parameters, error types and student historical operation records from the failed data packet, and generates context information containing failure scene images and key feature labels;

[0041] S520, the Agent collaboration module loads the preset teaching Prompt project framework, calls the cloud-based large model API, and passes in context information to generate the first round of guided dialogue. The framework includes project task description, current problem location, student's attempted operations, and a set of optional solutions.

[0042] S530: The terminal receives the analysis results of the questions input by the students. The Agent performs intent recognition on the answer content. If it is determined to be a valid analysis, it pushes several differentiated solutions for selection. If it is determined to be off-topic, it generates follow-up questions to guide the dialogue.

[0043] S540 generates an operation guide that includes supplementary data collection requirements and parameter adjustment suggestions based on the solution selected by the student. It guides the student to complete targeted data supplementation and incremental model training. After training is completed, it jumps to the entity verification stage.

[0044] Optionally, in S520, the Agent collaboration module loads a preset teaching Prompt project framework. This framework includes a project task description, current problem location, student-tried operations, and a set of possible solutions. It then calls the cloud-based large model API to generate the first round of guided dialogue, including:

[0045] S521, Load the Prompt template corresponding to the current project task from the teaching resource library. The Prompt template includes role definition, dialogue flow control and feedback style parameters.

[0046] S522, populate the context information generated by the context analysis unit into the template variables, the template variables including the URL of the failure scene image, the error identification category, the student's historical annotation accuracy rate and the abnormal values ​​of environmental parameters;

[0047] S523, set the dialogue generation parameters, including the maximum number of characters in a single reply, the question guidance coefficient, and the solution display format, and call the cloud-based large model API to send a generation request;

[0048] S524 receives the dialogue text returned by the large model, the content filtering unit performs educational compliance checks, replaces expressions that do not conform to teaching standards, and then pushes it to the terminal for display.

[0049] Optionally, in S540, based on the solution selected by the student, an operation guide is generated that includes supplementary data collection requirements and parameter adjustment suggestions to guide the student in completing targeted data supplementation and incremental model training. After training is completed, the process automatically jumps to the entity verification stage, including:

[0050] S541, the solution analysis unit extracts the corresponding data collection supplementary rules from the knowledge base according to the solution selected by the student. The supplementary rules include the sample categories, angle ranges, lighting conditions and quantity requirements to be supplemented.

[0051] S542, Generate a list of parameter adjustment suggestions, which includes specific settings for model learning rate, training epochs and data augmentation strategies, and a brief explanation of the basis for the adjustment;

[0052] S543, the terminal displays operation instructions in a combination of text and graphics, including a diagram of the target to be reshot, the path to the parameter adjustment interface, and a description of the expected effect of each step of the operation;

[0053] S544: After the student completes the reshoot and parameter settings, the incremental training process is triggered. After the training is completed, a "Verification ready" prompt is pushed, and the physical AI teaching aid is switched to physical verification.

[0054] On the other hand, embodiments of the present invention provide an artificial intelligence general education system based on bimodal closed-loop feedback, characterized in that it includes:

[0055] At least one processor;

[0056] At least one memory for storing at least one program;

[0057] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0058] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the above-described method.

[0059] The beneficial effects of this invention are as follows: This invention discloses an artificial intelligence general education method and system based on bimodal closed-loop feedback. This invention constructs a rule-driven and agent-collaborative bimodal adaptive feedback mechanism, automatically selecting the feedback method according to the question type: for high-frequency, clearly defined questions, a rule-based feedback module generates structured instructions for rapid error correction; for complex, open questions, an agent-collaborative module guides Socratic dialogue to cultivate metacognitive abilities. The two modalities are dynamically switched through an intelligent routing unit, balancing the efficiency of basic skills training with the depth of higher-order thinking development. Simultaneously, this invention constructs a cloud-edge-device collaborative architecture, integrating the training platform, model deployment, hardware execution, and behavior monitoring into a unified system for automatic flow. This effectively solves core problems in existing technologies such as the disconnect between learning and application, feedback failure, low iteration efficiency, and guidance mismatch, achieving a minute-level teaching closed loop and meeting the teaching needs of standard class hours. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 A flowchart illustrating an artificial intelligence general education method based on bimodal closed-loop feedback, provided in an embodiment of the present invention.

[0062] Figure 2 This is a general architecture diagram of an artificial intelligence general education system based on bimodal closed-loop feedback, provided as an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0064] It is understood that the terms "first," "second," etc., used in this invention may be used to describe various concepts herein, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are used only to distinguish one concept from another. For example, first information may also be referred to as second information without departing from the scope of embodiments of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the words "if" or "when" as used herein may be interpreted as "when," "in response to a determination," or "in the event of a determination."

[0065] The terms "at least one", "multiple", "each", "any", etc., used in this invention, "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiple, and "any" refers to any one of the multiple.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0067] It should be noted that in various specific embodiments of the present invention, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, the permission or consent of the user's guardian will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards of the relevant countries and regions. In addition, when embodiments of the present invention require access to sensitive personal information of users, separate permission or consent from the guardian will be obtained through methods such as a guardian confirmation page. Only after obtaining the separate permission or consent of the guardian will the necessary user-related data for the normal operation of the embodiments of the present invention be obtained.

[0068] The core idea of ​​this invention is to dynamically select feedback strategies based on problem type by constructing a rule-driven and agent-coordinated dual-modal adaptive feedback mechanism; to achieve integrated data flow and minute-level iteration cycle through a cloud-edge collaborative architecture; and to provide precise and personalized teaching guidance through context-aware diagnostic technology.

[0069] refer to Figure 1 ,like Figure 1 The image shows an embodiment of an artificial intelligence general education method based on bimodal closed-loop feedback provided by the present invention. The method includes the following steps:

[0070] S100: Obtain project tasks and guide students to collect contextualized data through physical AI teaching aids. The terminal provides real-time feedback on the collection process. The contextualized data includes images of target objects. The real-time feedback includes shooting quality detection, diversity reminders, and quantity progress bars.

[0071] S200: Receives category labeling data of target object images from students, uploads the labeling data to the cloud training engine for visual model training, and downloads the trained model to the physical AI teaching aid. The visual model training displays the real-time accuracy curve, the recognition status of each category, and the estimated completion time.

[0072] S300 performs physical verification of physical AI teaching aids in a real environment. The behavior monitoring unit records the recognition results, execution actions and environmental parameters to determine whether the task is successful or failed.

[0073] If the task fails, the behavior monitoring unit reports the failure data packet to the cloud feedback engine, the context analysis unit extracts key features, and the intelligent routing unit classifies the problem based on multi-dimensional classification and dynamically selects the feedback mode.

[0074] In the S500, if the rule feedback modality is selected, the rule feedback module matches a predefined rule base to generate structured instructions and pushes them to the terminal to guide students to supplement data and perform incremental training. If the agent collaboration modality is selected, the agent collaboration module calls the cloud-based large model API to generate a Socratic guided dialogue to guide students to analyze problems and select solutions. After completing data supplementation and incremental training, entity verification is performed again.

[0075] In the embodiments provided by this invention, a complete teaching loop of "data acquisition - model training - entity verification - feedback optimization" is constructed by combining abstract AI technology principles with concrete physical teaching aids. This guides students to independently discover the root causes of problems and try to add shape features for secondary training, thereby deepening their understanding of AI technology principles, model limitations, and problem-solving methods in practice, and effectively improving students' AI literacy and engineering practice capabilities.

[0076] Specifically, this invention achieves a close integration of data collection, model training, and physical verification through physical AI teaching aids. Students can personally participate in the complete closed-loop process of AI from data to decision-making, significantly enhancing the immersion and practicality of learning. The dual-modal feedback mechanism intelligently selects between rule-based feedback and agent-based collaborative feedback based on the problem type. The rule-based feedback modality, by matching a predefined rule base, quickly provides students with structured operational instructions, guiding them to supplement data and perform incremental training, suitable for explicit problems such as insufficient data or labeling errors. The agent-based collaborative modality, through Socratic guided dialogue, stimulates students' independent thinking, helping them to deeply analyze the reasons for model failures and explore diverse solutions, making it more suitable for handling open-ended problems such as data distribution bias and complex environmental interference. This dynamically selected feedback method ensures both the efficiency of problem-solving and the cultivation of students' deep thinking abilities. Meanwhile, this invention optimizes and integrates the traditional AI model iteration process. Through seamless integration of the cloud-based training engine and physical AI teaching aids, and the rapid deployment of lightweight models, the time required for a single iteration is significantly reduced from over 30 minutes to within the timeframe feasible in class. This allows students to complete multiple iterations of practice-verification-diagnosis-optimization within limited class time, greatly improving learning efficiency and the depth of knowledge acquisition. Furthermore, the real-time feedback provided during the data collection phase, the visual display during model training, and the detailed behavioral records during physical verification all offer students comprehensive and intuitive learning support. This helps students develop a correct understanding of artificial intelligence technology, cultivate their computational thinking and intelligent application abilities, and provides strong technical support for the effective implementation of general AI education in the K-12 stage.

[0077] In some embodiments, in S100, the step of acquiring project tasks and guiding students to collect contextualized data through physical AI teaching aids, with the terminal providing real-time feedback on the collection process, wherein the contextualized data includes images of target objects, and the real-time feedback includes image quality detection, diversity reminders, and a quantity progress bar, including:

[0078] S110, Receive a project task description sent by the teacher, the project task description includes teaching objectives, target object categories and class time limits;

[0079] S120, the terminal interface displays a task guidance animation, prompting students to take pictures of the target object through the camera of the physical AI teaching aid. The target object images cover different angles, lighting conditions and appearance features.

[0080] S130, the edge computing unit detects the image clarity, whether it contains target objects and whether there are repeated features in real time. If the clarity is lower than the threshold, it triggers a prompt to adjust the shooting distance. If the features of multiple consecutive images are repeated, it triggers a prompt to adjust the shooting angle.

[0081] S140: The terminal interface updates the progress bar in real time, displaying the ratio of the number of collected samples to the target number of samples. When the preset number is reached, a prompt indicating that the collection is complete is triggered.

[0082] In this embodiment, a progress bar that updates in real time on the terminal interface visually displays the ratio of the number of collected samples to the target number of samples, helping students clearly grasp the overall progress of data collection. When the number of collected samples reaches the preset minimum effective training threshold, a prompt indicating completion of collection is automatically triggered. The prompt is presented in a prominent pop-up window with a "Start Annotation" shortcut button, guiding students naturally to the next step. Simultaneously, to encourage students to collect richer data to improve model robustness, a "Recommended Collection Amount" indicator and the current diversity score (calculated based on dimensions such as angle, illumination, and pose of the collected images) are displayed next to the progress bar. When the diversity score falls below a good threshold, specific optimization suggestions are given, such as "It is recommended to supplement samples under XX angle / illumination conditions."

[0083] In some embodiments, S200, receiving the category annotation data of the target object image from the student, uploading the annotation data to the cloud training engine for visual model training, and downloading the trained model to the physical AI teaching aid, includes:

[0084] S210, the terminal interface displays a queue of images to be labeled, prompting students to complete the labeling by clicking on preset category labels, obtain the labeled data, and record the label timestamp and label confidence level; the queue of images to be labeled includes all acquired target object images;

[0085] S220: After the labeled data is verified for integrity, it is packaged and uploaded to the cloud. The cloud training engine calls the pre-trained basic model and uses the transfer learning method to perform incremental training, generating a visual log in real time that includes training rounds and changes in loss value.

[0086] S230 dynamically displays the proportion of samples in each category, confusion matrix, and typical error cases during training. Training is automatically terminated when the accuracy is stable above the threshold for multiple consecutive rounds.

[0087] S240 generates a model performance report after training, including overall accuracy, scene adaptation scores and optimization suggestions, and encrypts and transmits the lightweight model file to the local storage unit of the physical AI teaching aid.

[0088] Through this example, students can intuitively observe changes in model performance during visualization training. For instance, in the "garbage sorting" project, when students find that the confusion matrix values ​​for plastic bottles and aluminum cans are too high, the system will automatically mark typical incorrect samples and display their common characteristics (such as silver reflective material), helping students understand the model's limitations. The model performance report uses a star-rating system, providing specific fit scores for different environmental scenarios (such as strong light and shadow), and uses a heatmap in the optimization suggestion section to mark the sample categories that need to be supplemented, enabling students to improve their data collection strategies in a targeted manner.

[0089] In some embodiments, S300, the step of performing entity verification on the physical AI teaching aid in a real environment, with the behavior monitoring unit recording the recognition results, executed actions, and environmental parameters, and determining whether the task is successful or failed, includes:

[0090] S310, the terminal sends a start verification command to the physical AI teaching aid, and the physical AI teaching aid enters the task execution mode and moves in the real environment according to the preset path or command;

[0091] S320, the sensor unit of the physical AI teaching aid collects environmental parameters in real time, including light intensity, distance to obstacles and its own posture data. The recognition module identifies the target object in front in real time and outputs the recognition result including category and confidence level.

[0092] S330, the execution unit drives the mechanical structure to complete the preset action according to the recognition result, and the behavior monitoring unit records the actual execution result simultaneously. The actual execution result includes recognition time, recognition result, type of execution action, degree of action completion and corresponding environmental parameters.

[0093] S340, the result determination unit compares the actual execution result with the task target. If the action is successfully identified and executed correctly multiple times in a row, the task is determined to be successful; otherwise, the task is determined to be a failure and the failure recording process is triggered.

[0094] In this embodiment, when the physical AI teaching aid is verified in a real environment, the behavior monitoring unit records the recognition results, execution actions and environmental parameters with millisecond-level precision, thereby achieving accurate determination of whether the task is successful or not.

[0095] In some embodiments, in S400, the behavior monitoring unit reports failed data packets to the cloud feedback engine, the context analysis unit extracts key features, and the intelligent routing unit classifies the problem based on multi-dimensional classification and dynamically selects the feedback modality, including:

[0096] S410, the behavior monitoring unit encapsulates the failed data packets into structured data containing recognition results, executed actions, environmental parameters and timestamps, and uploads them to the cloud feedback engine through an encrypted channel;

[0097] S420, The context analysis unit extracts features from the failed data packets and generates a feature vector, which includes error type, environmental interference factor, data distribution deviation and student operation habit label;

[0098] S430, the intelligent routing unit calls the intelligent routing decision algorithm to match the feature vector with the question type template in the rule base. When the matching degree is higher than the preset threshold, it is determined to be a clear question and the rule feedback mode is selected. When the matching degree is lower than the threshold, it is determined to be an open question and the Agent collaboration mode is selected.

[0099] S440 pushes the feedback mode selection result and corresponding processing priority to the task scheduling unit, which then allocates computing resources based on the current system load.

[0100] This embodiment enables the intelligent and dynamic matching of the most suitable feedback strategy to different teaching scenarios based on the type and complexity of the failure problem. Through the multi-dimensional classification system of the intelligent routing decision algorithm, the system can accurately identify the essence of the problem. The algorithm first distinguishes between identification errors (such as misclassification), execution errors (such as robotic arm grasping deviation), and environmental interference (such as sudden changes in lighting) based on the error type dimension; then, it analyzes whether the problem is related to insufficient sample size, labeling errors, or missing features based on the data correlation dimension; finally, it combines the student's operation history dimension to determine whether the problem is a common issue encountered for the first time or a personalized operational error. This provides a precise basis for subsequent feedback modality selection.

[0101] In some embodiments, in S500, if the Agent collaborative modality is selected, the Agent collaborative module calls the cloud-based large model API to generate a Socratic guided dialogue, guiding students to analyze the problem and select a solution. After completing data re-enhancing and incremental training, entity verification is performed again, including:

[0102] S510, the context analysis unit extracts environmental parameters, error types and student historical operation records from the failed data packet, and generates context information containing failure scene images and key feature labels;

[0103] S520, the Agent collaboration module loads the preset teaching Prompt project framework, calls the cloud-based large model API, and passes in context information to generate the first round of guided dialogue. The framework includes project task description, current problem location, student's attempted operations, and a set of optional solutions.

[0104] S530: The terminal receives the analysis results of the questions input by the students. The Agent performs intent recognition on the answer content. If it is determined to be a valid analysis, it pushes several differentiated solutions for selection. If it is determined to be off-topic, it generates follow-up questions to guide the dialogue.

[0105] S540 generates an operation guide that includes supplementary data collection requirements and parameter adjustment suggestions based on the solution selected by the student. It guides the student to complete targeted data supplementation and incremental model training. After training is completed, it jumps to the entity verification stage.

[0106] In this embodiment, the constructed Socratic guided dialogue framework can effectively stimulate students' deep thinking.

[0107] In some embodiments, in S520, the Agent collaboration module loads a preset teaching Prompt project framework, which includes a project task description, current problem location, student-tried operations, and a set of optional solutions. It then calls the cloud-based large model API to generate the first round of guided dialogue, including:

[0108] S521, Load the Prompt template corresponding to the current project task from the teaching resource library. The Prompt template includes role definition, dialogue flow control and feedback style parameters.

[0109] S522, populate the context information generated by the context analysis unit into the template variables, the template variables including the URL of the failure scene image, the error identification category, the student's historical annotation accuracy rate and the abnormal values ​​of environmental parameters;

[0110] S523, set the dialogue generation parameters, including the maximum number of characters in a single reply, the question guidance coefficient, and the solution display format, and call the cloud-based large model API to send a generation request;

[0111] S524 receives the dialogue text returned by the large model, the content filtering unit performs educational compliance checks, replaces expressions that do not conform to teaching standards, and then pushes it to the terminal for display.

[0112] In this embodiment, a balance between standardization and personalization in teaching guidance is achieved through a preset Prompt template. Dynamic filling of template variables allows for the precise integration of specific failure scenario data (such as "in an environment where the light intensity is below 200 lux, the confidence level for identifying the 'glass' category remains below 0.5") into the guiding dialogue, avoiding vague generalities. Simultaneously, the flexible configuration of dialogue generation parameters can adapt to the cognitive levels of students at different grade levels. For example, for upper elementary school students, the "question guidance coefficient" can be adjusted to 0.8 to generate more closed-ended questions (such as "Do you think the model misclassifying a glass bottle as a plastic bottle might be related to the lighting during filming?"); for middle school students, the coefficient can be adjusted to 0.5 to increase open-ended questions (such as "Observe the cross-data of 'metal' and 'plastic' in the confusion matrix, and combine it with the characteristics of the samples you collected, analyze what might be the reasons for misclassification?"). The content filtering unit not only filters out inappropriate expressions, but also provides educational translations of technical terms. For example, it translates "overfitting" into "the model has 'memorized' the details of the training samples, and is prone to making mistakes when encountering new situations," ensuring the educational appropriateness and comprehensibility of the dialogue content.

[0113] In some embodiments, in S540, the step of generating operation guidelines containing supplementary data collection requirements and parameter adjustment suggestions based on the solution selected by the student, guiding the student to complete targeted data supplementation and incremental model training, and automatically jumping to the entity verification stage after training is completed includes:

[0114] S541, the solution analysis unit extracts the corresponding data collection supplementary rules from the knowledge base according to the solution selected by the student. The supplementary rules include the sample categories, angle ranges, lighting conditions and quantity requirements to be supplemented.

[0115] S542, Generate a list of parameter adjustment suggestions, which includes specific settings for model learning rate, training epochs and data augmentation strategies, and a brief explanation of the basis for the adjustment;

[0116] S543, the terminal displays operation instructions in a combination of text and graphics, including a diagram of the target to be reshot, the path to the parameter adjustment interface, and a description of the expected effect of each step of the operation;

[0117] S544: After the student completes the reshoot and parameter settings, the incremental training process is triggered. After the training is completed, a "Verification ready" prompt is pushed, and the physical AI teaching aid is switched to physical verification.

[0118] In this embodiment, the generated operation guide adopts a "three-dimensional precise navigation" design to ensure that students can efficiently implement the solution.

[0119] refer to Figure 2The system architecture diagram disclosed in this invention, a K12 artificial intelligence general education method based on bimodal closed-loop feedback, includes the following steps:

[0120] I. System Overall Architecture;

[0121] The system of this invention consists of the following four main modules:

[0122] 1. Cloud service platform;

[0123] Model training engine: Provides a visual training interface for AI capabilities such as image classification and object detection, and supports transfer learning and incremental training;

[0124] Agent Collaborative Service: A teaching assistant built on a large language model, capable of multi-turn dialogue, problem analysis, and strategy suggestion;

[0125] Data storage service: Unified storage of student project data, training models, operation logs, and knowledge base resources;

[0126] Dual-modal feedback engine (core): includes behavior monitoring unit, intelligent routing unit, rule feedback module, and agent collaboration module.

[0127] 2. User interaction terminal;

[0128] Hardware form factor: Tablet or laptop computer, equipped with a dedicated teaching client;

[0129] Functional modules:

[0130] Data acquisition tools (camera acquisition, file import, real-time annotation);

[0131] Visualized training interface (drag-and-drop model configuration, real-time training progress display);

[0132] Feedback and interaction interface (rule command display, Agent dialogue window, iteration guidance panel);

[0133] Project management tools (progress tracking, history recording, and results saving).

[0134] 3. Physical AI teaching aids;

[0135] Hardware configuration:

[0136] Sensor components (camera, infrared sensor, ultrasonic sensor, ambient light sensor);

[0137] Actuator components (servo motor, motor, air pump, LED indicator);

[0138] Edge computing unit (embedded chip, supporting model inference and behavior monitoring);

[0139] Communication module (WiFi / Bluetooth, for communication with the terminal and cloud);

[0140] Typical product forms: garbage sorting robots, intelligent tracking vehicles, gesture recognition robotic arms, and voice assistant devices.

[0141] 4. Dual-modal closed-loop feedback engine (system core);

[0142] Behavior monitoring and context analysis unit: Receives hardware sensor data and execution results in real time, identifies failure behavior patterns, extracts context features, and generates structured failure reports;

[0143] Intelligent routing unit: Based on decision tree and rule engine, it classifies problems in multiple dimensions (clarity, complexity, frequency) and dynamically selects the feedback mode;

[0144] Rule Feedback Module (Modal A): Maintains a predefined rule base and instruction template library, and generates structured feedback JSON for frequently asked and clearly defined questions;

[0145] Agent Collaboration Module (Modal B): Calls the cloud-based large model API, passes in context information, and generates a Socratic guided dialogue;

[0146] Project Status Synchronization Unit: Maintains the global status map of the project in real time, including dataset status, model status, and hardware status.

[0147] II. Work Process;

[0148] Phase 1: Contextualized data collection;

[0149] Students receive project tasks (such as "training a robot to sort plastic bottles and crumpled paper") via a terminal. The system guides students to use the camera of a physical AI teaching aid to capture images of target objects in a real-world environment. The terminal provides real-time feedback: image quality detection (blurry, overexposed, incomplete target → prompt for retake), diversity reminders (3 images of blue bottles have been captured; it is recommended to capture images of other colors), and a quantity progress bar (currently captured: 5 plastic bottles, 3 crumpled paper; it is recommended to capture 10 images of each type).

[0150] Phase 2: Visualized model training;

[0151] Students label the captured images on the terminal (drag and drop to the "plastic bottle" or "paper ball" classification box). Clicking the "Start Training" button uploads the data to the cloud and invokes the training engine. The training process is visualized: real-time accuracy curves, recognition rates for each category (plastic bottle 95%, paper ball 88%), and estimated completion time (approximately 2 minutes). After training, the model is automatically downloaded to the hardware device.

[0152] Phase 3: Entity Deployment and Verification;

[0153] Students place a test object (such as a plastic bottle with a red cap) in a real-world environment and press the "Start Execution" button on the hardware. The physical AI teaching aid is activated: the camera captures an image of the object, the edge computing unit runs model inference, and controls a robotic arm to put the object into the corresponding trash can based on the recognition result. The behavior monitoring unit records the recognition result, the executed actions, and environmental parameters throughout the process, and finally determines whether it is a success or failure.

[0154] Phase 4: Bimodal Feedback and Iteration (Core Component);

[0155] (A) Intelligent diagnosis and routing decision-making;

[0156] If the task fails, the behavior monitoring unit immediately reports the failure data packet to the cloud feedback engine. The context analysis unit extracts key features: characteristics of the failed object, error type, current status of the dataset, and environmental factors. The intelligent routing unit makes decisions based on multi-dimensional judgment logic.

[0157] If the problem is "clear + simple + high frequency" (e.g., insufficient data diversity) → route to mode A;

[0158] If the problem is "open + complex + low frequency" (such as environment adaptation strategy) → route to mode B;

[0159] If a student fails the same question three times in a row → switch from mode A to mode B;

[0160] If the remaining class time is less than 10 minutes, the route will be prioritized to Mode A.

[0161] (B) Execution of Modal A: Rule feedback;

[0162] The rule feedback module matches predefined rules (e.g., "Category X misidentified AND dataset lacks this appearance feature → prompt to supplement samples with this feature"), generates structured instructions, and pushes them to the terminal interface. The instructions include four parts: recognition update, cause of the problem, optimization suggestions, and expected results, accompanied by visual prompts (highlighting the type of sample that needs to be supplemented). Students supplement the data according to the instructions (approximately 1 minute), click the "Incremental Training" button, and the system quickly retrains the model based on the original model (3 minutes). The new model is automatically deployed to the hardware, and students test and verify it again.

[0163] (C) Modal B execution: Agent collaboration;

[0164] The Agent collaboration module calls the cloud-based large-scale model API, passing in context such as project task description, current failure status, student operation history, and alternative solutions. It then generates multiple rounds of Socratic guided dialogue.

[0165] Round 1: Describe the problem phenomenon and ask students about the details they observed;

[0166] Round 2: Propose 2-3 possible solutions, guide students to analyze the advantages and disadvantages and make a choice;

[0167] Round 3: Provide specific implementation suggestions based on student choices;

[0168] Summary round: Guide students to summarize the reasons for the problem and think about how to apply them to new scenarios;

[0169] The entire dialogue should be kept within 5 minutes to ensure a reasonable allocation of class time.

[0170] III. Key Technical Details;

[0171] 1. Intelligent routing decision algorithm;

[0172] Input: Failure behavior characteristics F, project status S, student's historical operations H

[0173] Output: Feedback mode selection M∈{A,B}

[0174] IF F.problem_type IN [missing data, incorrect labeling, class imbalance] AND F.frequency > threshold 1:

[0175] M=A / / High-frequency, clearly defined questions → Rule feedback

[0176] ELSE IF F.complexity_score>threshold2 AND S.student_level>="advanced":

[0177] M=B / / Complex problem with student's ability to explore → Agent collaboration

[0178] ELSE IF H.consecutive_failures_on_same_issue>=3:

[0179] M=B / / Student fails multiple times, requires in-depth guidance → Switch to Agent

[0180] ELSE IF S.remaining_class_time < 10 minutes:

[0181] M=A / / Time is of the essence, prioritize quick resolution → Rule Feedback

[0182] ELSE:

[0183] M=A / / Default priority rule feedback (efficiency priority).

[0184] 2. Rule base design (partial examples);

[0185] Rule ID: R001;

[0186] Triggering condition: Category X number of samples < 5;

[0187] Feedback template: "Please provide at least 10 photos of {Category X}".

[0188] Priority: High;

[0189] Rule ID: R002;

[0190] Triggering conditions: Identification error AND lack of the appearance feature in the training set;

[0191] Feedback template: "Please provide additional samples showing {missing features}";

[0192] Priority: High;

[0193] Rule ID: R003;

[0194] Triggering conditions: Ambient light > 800 lux AND recognition failure;

[0195] Feedback template: "The current light is too strong. We suggest adjusting the hardware angle or taking additional photos in strong light conditions."

[0196] Priority: Medium;

[0197] Rule ID: R004;

[0198] Triggering condition: Three consecutive failures on the same sample;

[0199] Feedback template: "This sample may contain special interference. We suggest checking the background or re-labeling."

[0200] Priority: Medium.

[0201] 3. Agent Prompt project framework;

[0202] Core keyword structure:

[0203] "You are a teaching assistant for a K-12 artificial intelligence course, helping students optimize AI models."

[0204] [Project Task] {task_description};

[0205] [Current Issue] {failure_description};

[0206] [Students have already tried] {previous_actions};

[0207] [Optional solutions] {possible_solutions};

[0208]

Teaching Requirements

[0209] Example 1: Application of the rule feedback module in the "garbage sorting robot" project;

[0210] Project Background: Fifth grade elementary school student, 45-minute class period, task is to train a robot to identify and sort "plastic bottles" and "paper balls".

[0211] Operating procedures:

[0212] 1. Data collection (8 minutes): Students collect 5 plastic bottles (all with blue caps) and 8 crumpled pieces of paper (white and yellow);

[0213] 2. Model Training (5 minutes): Training is completed in the cloud, and the model is deployed to the robot;

[0214] 3. First physical verification (2 minutes): Test a red plastic bottle with a cap. The robot misidentifies it as a crumpled piece of paper (confidence level 62%), and the robotic arm incorrectly disposes of it.

[0215] 4. Dual-modal feedback - Modal A trigger (instantaneous): Intelligent routing determines a clear problem of "insufficient data diversity";

[0216] 5. Rule feedback module generates instructions: A pop-up interface prompts "The plastic bottles in the training data all have blue caps. Please add 2-3 photos of bottle caps of different colors. The expected accuracy will be improved to 92%."

[0217] 6. Iterative optimization (5 minutes): Students take two more photos of red, green, and transparent bottle caps, and then click incremental training (3 minutes).

[0218] 7. Second physical verification (2 minutes): Successfully identified the red-capped plastic bottle, and the robotic arm correctly delivered it;

[0219] Technical results: Iteration cycle of 5 minutes (traditional solutions require 30 minutes), feedback accuracy of 95%, and completion of a complete closed loop within 45 minutes.

[0220] Example 2: Application of the Agent Collaboration Module in the "Intelligent Line-Following Car" Project;

[0221] Project background: Seventh grade junior high school, 45-minute class period, the task is to train a car to drive along the black track line.

[0222] Operating procedures:

[0223] 1. Data Acquisition (8 minutes): Acquire 50 trajectory images (30 with the black line centered, 10 slightly to the left, and 10 slightly to the right).

[0224] 2. Model Training (5 minutes): Training complete and deployed to the vehicle;

[0225] 3. First physical verification (3 minutes): The car deviated from the trajectory in the reflective area and failed;

[0226] 4. Dual-modal feedback - Modal B trigger (instantaneous): Intelligent routing determines the open issue of "environment adaptation strategy";

[0227] 5. Agent-led collaborative guidance (5 minutes, 3 rounds of dialogue):

[0228] Round 1: Show the image of the failed location and ask, "Why can't I see the black line clearly anymore?"

[0229] Round 2: After students choose "the light is too strong and causes glare", they are given two options to compare: "take a detour" and "learn to adapt".

[0230] Round 3: After students choose Option 2, guide them to take additional photos of the reflective environment (12 photos, from different locations).

[0231] 6. Iterative optimization (8 minutes): Retake data and incremental retraining (5 minutes);

[0232] 7. Second physical verification (3 minutes): The car successfully completes the entire trajectory.

[0233] Technical effects: The iteration cycle is 10 minutes (including thinking dialogue), which cultivates the ability to decompose problems and select strategies. Students understand the principle of the solution and can transfer it to new scenarios such as shadows.

[0234] According to the above embodiments, compared with the prior art, the present invention has the following significant advantages:

[0235] 1. It pioneered a dual-modal adaptive feedback mechanism, and for the first time in the industry, it collaboratively integrated rule-driven and large-model agent into a unified feedback engine;

[0236] 2. Construct a "minute-level" teaching loop to compress the AI ​​model iteration cycle from the industry average of 30-60 minutes to 5-10 minutes;

[0237] 3. Engineering implementation of the cloud-device-edge collaborative architecture: rule feedback is completed on the terminal (millisecond response), and agent collaboration is completed in the cloud (3-5 second response).

[0238] 4. Promote deep learning and the development of higher-order thinking, enabling students to not only learn "how to do it", but also understand "why to do it this way";

[0239] 5. Strong context awareness and accurate diagnostic capabilities; feedback content closely matches the current task context, avoiding generalized suggestions.

[0240] 6. An extensible knowledge base and rule base architecture that supports custom extensions by teachers and adapts to different subject projects;

[0241] 7. Reduce teachers' teaching burden; the system automates the diagnosis and feedback process.

[0242] 8. Supports large-scale deployment and data accumulation; cloud architecture supports thousands of concurrent users.

[0243] 9. Complies with educational ethics and data security requirements, with student data stored locally and transmitted encrypted to the cloud.

[0244] This invention also provides an artificial intelligence general education system based on bimodal closed-loop feedback, comprising:

[0245] At least one processor;

[0246] At least one memory for storing at least one program;

[0247] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0248] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0249] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including a server, a teacher's terminal, or a student tablet.

[0250] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0251] Those skilled in the art will understand that all or some of the steps and systems disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is known to those skilled in the art that communication media typically include computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and can include any information delivery medium.

[0252] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method for general education in artificial intelligence based on bimodal closed-loop feedback, characterized in that, The method includes the following steps: S100: Obtain project tasks and guide students to collect contextualized data through physical AI teaching aids. The terminal provides real-time feedback on the collection process. The contextualized data includes images of target objects. The real-time feedback includes shooting quality detection, diversity reminders, and quantity progress bars. S200: Receives category labeling data of target object images from students, uploads the labeling data to the cloud training engine for visual model training, and downloads the trained model to the physical AI teaching aid. The visual model training displays the real-time accuracy curve, the recognition status of each category, and the estimated completion time. S300 performs physical verification of physical AI teaching aids in a real environment. The behavior monitoring unit records the recognition results, execution actions and environmental parameters to determine whether the task is successful or failed. If the task fails, the behavior monitoring unit reports the failure data packet to the cloud feedback engine, the context analysis unit extracts key features, and the intelligent routing unit classifies the problem based on multi-dimensional classification and dynamically selects the feedback mode. In the S500, if the rule feedback modality is selected, the rule feedback module matches a predefined rule base to generate structured instructions and pushes them to the terminal to guide students to supplement data and perform incremental training. If the agent collaboration modality is selected, the agent collaboration module calls the cloud-based large model API to generate a Socratic guided dialogue to guide students to analyze problems and select solutions. After completing data supplementation and incremental training, entity verification is performed again.

2. The method according to claim 1, characterized in that, In S100, the process of acquiring project tasks and guiding students to collect contextualized data using physical AI teaching aids, with the terminal providing real-time feedback on the collection process, includes contextualized data such as images of target objects. The real-time feedback includes image quality detection, diversity reminders, and a quantity progress bar. S110, Receive project task description sent by the teacher, the project task description includes teaching objectives, target object category and class time limit; S120, the terminal interface displays a task guidance animation, prompting students to take pictures of the target object through the camera of the physical AI teaching aid. The target object images cover different angles, lighting conditions and appearance features. S130, the edge computing unit detects the image clarity, whether it contains target objects and whether there are repeated features in real time. If the clarity is lower than the threshold, it triggers a prompt to adjust the shooting distance. If the features of multiple consecutive images are repeated, it triggers a prompt to adjust the shooting angle. S140: The terminal interface updates the progress bar in real time, displaying the ratio of the number of collected samples to the target number of samples. When the preset number is reached, a prompt indicating that the collection is complete is triggered.

3. The method according to claim 1, characterized in that, In step S200, receiving the category annotation data of the target object image from the student, uploading the annotation data to the cloud training engine for visual model training, and downloading the trained model to the physical AI teaching aid includes: S210, the terminal interface displays a queue of images to be labeled, prompting students to complete the labeling by clicking on preset category labels, obtain the labeled data, and record the label timestamp and label confidence level; the queue of images to be labeled includes all acquired target object images; S220: After the labeled data is verified for integrity, it is packaged and uploaded to the cloud. The cloud training engine calls the pre-trained basic model and uses the transfer learning method to perform incremental training, generating a visual log in real time that includes training rounds and changes in loss value. S230 dynamically displays the proportion of samples in each category, confusion matrix, and typical error cases during training. Training is automatically terminated when the accuracy is stable above the threshold for multiple consecutive rounds. S240 generates a model performance report after training, including overall accuracy, scene adaptation scores and optimization suggestions, and encrypts and transmits the lightweight model file to the local storage unit of the physical AI teaching aid.

4. The method according to claim 1, characterized in that, In S300, the physical verification of the AI ​​teaching aid in a real environment, the behavior monitoring unit records the recognition results, executed actions, and environmental parameters, and determines whether the task is successful or failed, including: S310, the terminal sends a start verification command to the physical AI teaching aid, and the physical AI teaching aid enters the task execution mode and moves in the real environment according to the preset path or command; S320, the sensor unit of the physical AI teaching aid collects environmental parameters in real time, including light intensity, distance to obstacles and its own posture data. The recognition module identifies the target object in front in real time and outputs the recognition result including category and confidence level. S330, the execution unit drives the mechanical structure to complete the preset action according to the recognition result, and the behavior monitoring unit records the actual execution result simultaneously. The actual execution result includes recognition time, recognition result, type of execution action, degree of action completion and corresponding environmental parameters. S340, the result determination unit compares the actual execution result with the task target. If the action is successfully identified and executed correctly multiple times in a row, the task is determined to be successful; otherwise, the task is determined to be a failure and the failure recording process is triggered.

5. The method according to claim 1, characterized in that, In S400, the behavior monitoring unit reports failed data packets to the cloud feedback engine, the context analysis unit extracts key features, and the intelligent routing unit classifies the problem based on multi-dimensional classification and dynamically selects the feedback modality, including: S410, the behavior monitoring unit encapsulates the failed data packets into structured data containing recognition results, executed actions, environmental parameters and timestamps, and uploads them to the cloud feedback engine through an encrypted channel; S420, The context analysis unit extracts features from the failed data packets and generates a feature vector, which includes error type, environmental interference factor, data distribution deviation and student operation habit label; S430, the intelligent routing unit calls the intelligent routing decision algorithm to match the feature vector with the question type template in the rule base. When the matching degree is higher than the preset threshold, it is determined to be a clear question and the rule feedback mode is selected. When the matching degree is lower than the threshold, it is determined to be an open question and the Agent collaboration mode is selected. S440 pushes the feedback mode selection result and corresponding processing priority to the task scheduling unit, which then allocates computing resources based on the current system load.

6. The method according to claim 1, characterized in that, In S500, if the Agent collaborative modality is selected, the Agent collaborative module calls the cloud-based large model API to generate a Socratic guided dialogue, guiding students to analyze the problem and select a solution. After data re-enhancing and incremental training, entity verification is performed again, including: S510, the context analysis unit extracts environmental parameters, error types and student historical operation records from the failed data packet, and generates context information containing failure scene images and key feature labels; S520, the Agent collaboration module loads the preset teaching Prompt project framework, calls the cloud-based large model API, and passes in context information to generate the first round of guided dialogue. The framework includes project task description, current problem location, student's attempted operations, and a set of optional solutions. S530: The terminal receives the analysis results of the questions input by the students. The Agent performs intent recognition on the answer content. If it is determined to be a valid analysis, it pushes several differentiated solutions for selection. If it is determined to be off-topic, it generates follow-up questions to guide the dialogue. S540 generates an operation guide that includes supplementary data collection requirements and parameter adjustment suggestions based on the solution selected by the student. It guides the student to complete targeted data supplementation and incremental model training. After training is completed, it jumps to the entity verification stage.

7. The method according to claim 6, characterized in that, In S520, the Agent collaboration module loads a preset teaching Prompt project framework. This framework includes a project task description, current problem location, student-tried operations, and a set of possible solutions. It then calls the cloud-based large model API to generate the first round of guided dialogue, including: S521, Load the Prompt template corresponding to the current project task from the teaching resource library. The Prompt template includes role definition, dialogue flow control and feedback style parameters. S522, populate the context information generated by the context analysis unit into the template variables, the template variables including the URL of the failure scene image, the error identification category, the student's historical annotation accuracy rate and the abnormal values ​​of environmental parameters; S523, set the dialogue generation parameters, including the maximum number of characters in a single reply, the question guidance coefficient, and the solution display format, and call the cloud-based large model API to send a generation request; S524 receives the dialogue text returned by the large model, the content filtering unit performs educational compliance checks, replaces expressions that do not conform to teaching standards, and then pushes it to the terminal for display.

8. The method according to claim 6, characterized in that, In S540, based on the solution selected by the student, an operation guide is generated that includes supplementary data collection requirements and parameter adjustment suggestions. This guides the student to complete targeted data supplementation and incremental model training. After training is completed, the system automatically jumps to the entity verification stage, including: S541, the solution analysis unit extracts the corresponding data collection supplementary rules from the knowledge base according to the solution selected by the student. The supplementary rules include the sample categories, angle ranges, lighting conditions and quantity requirements to be supplemented. S542, Generate a list of parameter adjustment suggestions, which includes specific settings for model learning rate, training epochs and data augmentation strategies, and a brief explanation of the basis for the adjustment; S543, the terminal displays operation instructions in a combination of text and graphics, including a diagram of the target to be reshot, the path to the parameter adjustment interface, and a description of the expected effect of each step of the operation; S544: After the student completes the reshoot and parameter settings, the incremental training process is triggered. After the training is completed, a "Verification ready" prompt is pushed, and the physical AI teaching aid is switched to physical verification.

9. An artificial intelligence general education system based on bimodal closed-loop feedback, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images