A child teaching method and apparatus for intelligent robots

Abstract

This invention relates to the field of robotics, and more particularly to a teaching method and apparatus for children using intelligent robots. This invention can quantitatively assess children's cognitive understanding and emotional engagement during interactive learning in real time, thereby driving a teaching strategy generation model to dynamically output goal-oriented, progressive teaching scripts. This invention not only achieves precise, personalized, and scenario-based guidance for interactive learning, significantly improving children's immersive learning experience and cognitive construction efficiency, but also provides educational administrators with scientific and systematic decision support through structured, encrypted child development reports, powerfully promoting the intelligence and humanization of robot-assisted teaching. Furthermore, it solves the technical problem in existing technologies where it is difficult to achieve deep semantic understanding of teaching scenarios and children's multimodal behavioral feedback, leading to an inability to flexibly adjust the pace and content of teaching.

Description

Technical Field

[0001] This invention relates to the field of robotics, and more particularly to a teaching method and apparatus for children using intelligent robots. Background Technology

[0002] With the rapid development of artificial intelligence and robotics, intelligent robots have been gradually applied to children's education, becoming a new teaching medium. Compared to traditional multimedia courseware or tablet-based teaching, intelligent robots possess characteristics such as physical interaction, multimodal perception, and high anthropomorphism, enabling them to interact with children more naturally through voice dialogue, visual recognition, and body movements. Currently, educational robots capable of basic question-and-answer sessions, storytelling, or simple programming introductions have appeared on the market, aiming to stimulate children's learning interest through human-computer interaction and assist parents or teachers in completing early childhood education.

[0003] However, existing educational robot solutions for children rely on pre-set, fixed interaction scripts, making it difficult to achieve a deep semantic understanding of teaching scenarios and children's multimodal behavioral feedback, resulting in an inability to flexibly adjust the teaching pace and content. Summary of the Invention

[0004] The purpose of this invention is to provide a teaching method and device for children using intelligent robots, which solves the technical problem in the prior art that it is difficult to achieve deep semantic understanding of teaching scenarios and children's multimodal behavioral feedback, resulting in the inability to flexibly adjust the teaching pace and content.

[0005] To achieve the above objectives, the present invention provides a teaching method for children using intelligent robots, comprising: Acquire multimodal perception data containing the current interaction scenario. The multimodal perception data includes visual image data and auditory speech data. Multimodal perception data is input into a pre-trained cross-modal semantic understanding model, which includes a visual Transformer branch, a speech Transformer branch, and a cross-modal attention fusion layer. Multimodal semantic features are extracted and fused, and a dynamic scene knowledge graph is constructed in real time based on the fused semantic features. Real-time capture and analysis of multimodal behavioral feedback data of target child users, including facial micro-expressions, eye movement trajectories, vocal emotions, and body posture, to quantitatively assess the child's current cognitive understanding and emotional engagement. Based on a dynamic scene knowledge graph and the children's cognitive and emotional states, a teaching strategy generation model is used to dynamically generate goal-oriented progressive teaching scripts. The teaching scripts contain multiple teaching interaction nodes that are progressive in logic and difficulty. Following a progressive teaching script, the robot generates and outputs multimodal interactive behaviors that match the current teaching interaction node in sequence, in order to guide children to complete a complete scenario-based cognitive learning session. After the entire progressive teaching script is completed, a structured children's cognitive development report is generated; and the children's cognitive development report is encrypted and pushed to the management terminal.

[0006] The constructed dynamic scene knowledge graph is a dynamic scene knowledge graph that includes objects, events, relationships between people, and implicit storylines mined through temporal causal reasoning. The specific method for encrypting and pushing children's cognitive development briefings to the management terminal is as follows: After generating children's cognitive development briefings, the briefings are encrypted using the encryption key generated by the key generator to generate ciphertext. Then, the ciphertext is transmitted to the management terminal using a secure transmission protocol based on asymmetric encryption. When the management terminal receives the ciphertext, it decrypts the ciphertext using the decryption key generated by the key generator to obtain the original data. After compression using a lossless compression algorithm, the original data is fragmented and distributed across multiple cloud storage nodes, and an attribute-based access control policy is established.

[0007] Specifically, a smart contract-based consortium blockchain is used to store key teaching nodes, changes in children's cognitive states, and model adjustments. Graph neural network data mining algorithms are used to mine anonymized cross-user teaching records stored on the blockchain, uncovering potential correlation patterns between different teaching strategies and children's cognitive development paths. The high-value knowledge patterns discovered are used as positive feedback to optimize and iterate the teaching strategy generation model. A federated learning framework is used for distributed optimization of the teaching strategy generation model, with each terminal robot only uploading encrypted gradients and not sharing the original data, thus achieving model iteration while protecting user privacy.

[0008] Among them, when guiding children to complete scenario-based cognitive learning, the child's problem-solving behavior for the preset cognitive conflict goal of the node is monitored in real time. The problem-solving behavior includes the correctness of the answer, the exploration of the operation path, and the timing of seeking help. If the overall score of the resolution behavior meets the preset threshold, the script will be triggered to advance to the next node; Otherwise, the teaching strategy generation model is triggered to reduce the difficulty of the current node, or to retrieve relevant concept explanations or analogies from the knowledge graph to generate new explanatory interactions, so as to guide children to complete a complete, contextualized cognitive learning based on cognitive conflict resolution.

[0009] Specifically, the quantitative assessment of children's current cognitive understanding and emotional engagement includes: The collected facial micro-expression, eye movement trajectory, voice emotion and body posture data are respectively input into the corresponding pre-trained deep learning models to extract their respective high-dimensional features; The multi-modal fusion encoder is used to fuse the above-mentioned multi-source high-dimensional features to generate a hidden state vector that can comprehensively represent the child's current state; The latent state vector is input into a cognitive-emotion classifier. The classifier outputs quantitative scores for multiple dimensions, including attention concentration, comprehension confusion, emotional valence, and engagement, thereby enabling real-time quantitative assessment of children's cognitive comprehension and emotional engagement.

[0010] The process of distributed optimization of the teaching strategy generation model by the federated learning framework includes: The central server initializes the global parameters of the teaching strategy generation model and distributes the global parameters to each terminal robot; Each terminal robot uses locally stored children's interaction data to calculate the model gradient locally, and the local interaction data does not leave the terminal device; Each terminal robot encrypts and performs differential privacy processing on the calculated local gradient, and adds random noise that conforms to the Laplace distribution to the local gradient. Each terminal robot uploads the processed encrypted gradient to the central server; The central server securely aggregates the collected encrypted gradients and updates the global model parameters using a federated averaging algorithm. The updated global model parameters are redistributed to each terminal robot, and the above process is repeated until the model converges.

[0011] The present invention also provides a children's teaching device for intelligent robots, used to perform the children's teaching method for intelligent robots as described above. include: The module includes a multimodal perception module, a dynamic knowledge graph construction module, a child status assessment module, a teaching strategy generation module, an interaction execution module, a cognitive conflict monitoring module, a data security module, and a model optimization module. The multimodal perception module is used to acquire multimodal perception data containing the current interaction scenario; the dynamic knowledge graph construction module is configured to perform deep semantic understanding on the multimodal perception data and construct a dynamic scene knowledge graph; the child state assessment module is used to capture and analyze the multimodal behavioral feedback data of the target child user in real time, and quantitatively assess the child's current cognitive understanding state and emotional investment; the teaching strategy generation module is used to dynamically generate progressive teaching scripts based on the dynamic scene knowledge graph and the child's cognitive and emotional state through a teaching strategy generation model; the interaction execution module is used to generate and output the robot's multimodal interactive behavior according to the progressive teaching script; the cognitive conflict monitoring module is used to monitor the child's problem-solving behavior for cognitive conflict goals in real time, and dynamically adjust the teaching pace according to the problem-solving behavior score; the data security module is used to encrypt, fragment, and store the child's cognitive development report on blockchain; the model optimization module is used to mine cross-user teaching records using graph neural networks and iteratively optimize the teaching strategy generation model under a federated learning framework.

[0012] This invention discloses a teaching method and apparatus for children using intelligent robots. It can quantitatively assess children's cognitive understanding and emotional engagement during interactive learning in real time, thereby driving a teaching strategy generation model to dynamically output goal-oriented, progressive teaching scripts. This invention not only achieves precise, personalized, and scenario-based guidance for interactive learning, significantly improving children's immersive learning experience and cognitive construction efficiency, but also provides educational administrators with scientific and systematic decision support through structured, encrypted child development reports, powerfully promoting the intelligence and humanization of robot-assisted teaching. Furthermore, it solves the technical problem in existing technologies where it is difficult to achieve deep semantic understanding of teaching scenarios and children's multimodal behavioral feedback, leading to an inability to flexibly adjust the pace and content of teaching. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0014] Figure 1 This is a flowchart of the child teaching method for intelligent robots according to the present invention. Detailed Implementation

[0015] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0016] Please refer to Figure 1 , Figure 1This is a flowchart of a children's teaching method for intelligent robots according to the present invention. An embodiment of the present invention provides a children's teaching method for intelligent robots, including: S1. Acquire multimodal perception data containing the current interaction scenario. The multimodal perception data includes visual image data and auditory speech data. S2. Input the multimodal perception data into the pre-trained cross-modal semantic understanding model. The cross-modal semantic understanding model includes a visual Transformer branch, a speech Transformer branch, and a cross-modal attention fusion layer. Extract and fuse multimodal semantic features, and construct a dynamic scene knowledge graph in real time based on the fused semantic features. In this specific implementation, the constructed dynamic scene knowledge graph is a dynamic scene knowledge graph that includes objects, events, relationships between people, and implicit storylines mined through temporal causal reasoning. S3. Capture and analyze multimodal behavioral feedback data of target child users in real time. Multimodal behavioral feedback data includes facial micro-expressions, eye movement trajectory, voice emotion and body posture, which are used to quantitatively assess the child's current cognitive understanding and emotional engagement. In this specific implementation method, the quantitative assessment of children's current cognitive understanding and emotional engagement includes: The collected facial micro-expression, eye movement trajectory, voice emotion and body posture data are respectively input into the corresponding pre-trained deep learning models to extract their respective high-dimensional features; The multi-modal fusion encoder is used to fuse the above-mentioned multi-source high-dimensional features to generate a hidden state vector that can comprehensively represent the child's current state; The latent state vector is input into a cognitive-emotion classifier. The classifier outputs quantitative scores for multiple dimensions, including attention concentration, comprehension confusion, emotional valence, and engagement, thereby enabling real-time quantitative assessment of children's cognitive comprehension and emotional engagement.

[0017] S4. Based on the dynamic scene knowledge graph and the child's cognitive and emotional state, a teaching strategy generation model is used to dynamically generate a goal-oriented progressive teaching script. The teaching script contains multiple teaching interaction nodes that are progressive in logic and difficulty. S5. Following the progressive teaching script, generate and output the robot's multimodal interactive behaviors that match the current teaching interaction node in sequence, so as to guide children to complete a complete scenario-based cognitive learning session. In this specific implementation, when guiding children to complete scenario-based cognitive learning, the child's problem-solving behavior for the preset cognitive conflict goal of the node is monitored in real time. The problem-solving behavior includes the correctness of the answer, the exploratory nature of the operation path, and the timing of seeking help. If the overall score of the resolution behavior meets the preset threshold, the script will be triggered to advance to the next node; Otherwise, the teaching strategy generation model is triggered to reduce the difficulty of the current node, or to retrieve relevant concept explanations or analogies from the knowledge graph to generate new explanatory interactions, so as to guide children to complete a complete, contextualized cognitive learning based on cognitive conflict resolution.

[0018] S6. After the entire progressive teaching script is completed, generate a structured children's cognitive development report; and encrypt and push the children's cognitive development report to the management terminal.

[0019] In this specific implementation, the encrypted delivery of the children's cognitive development report to the management terminal is carried out as follows: After the children's cognitive development report is generated, the report is encrypted using the encryption key generated by the key generator to generate ciphertext. Then, the ciphertext is transmitted to the management terminal using a secure transmission protocol based on asymmetric encryption. When the management terminal receives the ciphertext, it decrypts the ciphertext using the decryption key generated by the key generator to obtain the original data. After compression using a lossless compression algorithm, the original data is fragmented and distributed across multiple cloud storage nodes, and an attribute-based access control policy is established.

[0020] A smart contract-based consortium blockchain is used to record key teaching nodes, changes in children's cognitive states, and model adjustments. Graph neural network data mining algorithms are then used to mine anonymized cross-user teaching records stored on the blockchain to uncover potential correlation patterns between different teaching strategies and children's cognitive development paths. The high-value knowledge patterns discovered are used as positive feedback to optimize and iterate the teaching strategy generation model. A federated learning framework is employed for distributed optimization of the teaching strategy generation model, with each terminal robot only uploading encrypted gradients and not sharing the original data, thus achieving model iteration while protecting user privacy.

[0021] The federated learning framework's distributed optimization process for the instruction strategy generation model includes: The central server initializes the global parameters of the teaching strategy generation model and distributes the global parameters to each terminal robot; Each terminal robot uses locally stored children's interaction data to calculate the model gradient locally, and the local interaction data does not leave the terminal device; Each terminal robot encrypts and performs differential privacy processing on the calculated local gradient, and adds random noise that conforms to the Laplace distribution to the local gradient. Each terminal robot uploads the processed encrypted gradient to the central server; The central server securely aggregates the collected encrypted gradients and updates the global model parameters using a federated averaging algorithm. The updated global model parameters are redistributed to each terminal robot, and the above process is repeated until the model converges.

[0022] The construction of the dynamic scene knowledge graph includes: Perform object detection, scene graph generation, and spatiotemporal relationship modeling on visual image data to extract visual entities and their relationships; Perform speech recognition, semantic role labeling, and intent understanding on auditory speech data to extract language entities and their relationships; Visual entities and linguistic entities are aligned and fused through a cross-modal entity alignment mechanism; A temporal causal reasoning model is introduced, and implicit causal chains and storylines are mined based on event sequences, organizing discrete event nodes into implicit storyline structures with narrative logic; The structured knowledge described above is stored in a graph database, and a multidimensional index of entity-relationship-attribute is created.

[0023] The collected facial micro-expression, eye movement, vocal emotion, and body posture data are input into their respective pre-trained deep learning models. When extracting their respective high-dimensional features, among them... Facial micro-expression feature extraction employs a 3D-CNN network structure combined with an attention mechanism; Eye movement trajectory feature extraction uses an LSTM encoder to perform temporal modeling of the gaze point sequence; Speech emotion feature extraction uses a wav2vec 2.0 pre-trained model to extract emotion embedding vectors; Limb pose feature extraction uses graph convolutional networks to perform spatial-temporal modeling of the skeletal keypoint sequence; The multi-source high-dimensional features mentioned above are fused using a multi-modal fusion encoder. The multi-modal fusion encoder adopts a cross-modal Transformer structure and learns the interaction weights between different modalities through a multi-head attention mechanism to generate a hidden state vector that can comprehensively represent the child's current state. The latent state vector is input into the cognitive-emotion classifier. The output of the classifier includes quantitative scores in multiple dimensions, including attention concentration, comprehension confusion, emotional valence, and engagement, thereby enabling real-time quantitative assessment of children's cognitive comprehension and emotional engagement. The cognitive sentiment classifier is trained using a multi-task learning framework, which simultaneously optimizes the prediction accuracy across the aforementioned multiple dimensions. The steps for dynamically generating progressive instructional scripts include: Key information (such as objects, events, and relationships) in dynamic scene knowledge graphs can be used as candidate teaching materials. The children's cognitive understanding and emotional engagement are quantified as a teaching initiation coefficient and a difficulty adjustment coefficient. The teaching strategy generation model aims to maximize children's cognitive development and emotional engagement. It selects and sorts candidate teaching materials to generate a series of teaching interaction nodes. The types of teaching interaction nodes include: guided observation, stimulating questioning, causal reasoning, emotional resonance, and knowledge expansion.

[0024] This invention discloses a child teaching method for intelligent robots. This invention can quantitatively assess children's cognitive understanding and emotional engagement during teaching interactions in real time, thereby driving a teaching strategy generation model to dynamically output goal-oriented, progressive teaching scripts. This invention not only achieves precise, personalized, and scenario-based guidance for teaching interactions, significantly improving children's immersive learning experience and cognitive construction efficiency, but also provides educational administrators with scientific and systematic decision support through structured, encrypted child development reports, powerfully promoting the intelligence and humanization of robot-assisted teaching. Furthermore, it solves the technical problem in existing technologies where it is difficult to achieve deep semantic understanding of teaching scenarios and children's multimodal behavioral feedback, resulting in an inability to flexibly adjust the pace and content of teaching.

[0025] The present invention also provides a children's teaching device for intelligent robots, used to perform the children's teaching method for intelligent robots as described above. include: The module includes a multimodal perception module, a dynamic knowledge graph construction module, a child status assessment module, a teaching strategy generation module, an interaction execution module, a cognitive conflict monitoring module, a data security module, and a model optimization module. The multimodal perception module is used to acquire multimodal perception data containing the current interaction scenario; the dynamic knowledge graph construction module is configured to perform deep semantic understanding on the multimodal perception data and construct a dynamic scene knowledge graph; the child state assessment module is used to capture and analyze the multimodal behavioral feedback data of the target child user in real time, and quantitatively assess the child's current cognitive understanding state and emotional investment; the teaching strategy generation module is used to dynamically generate progressive teaching scripts based on the dynamic scene knowledge graph and the child's cognitive and emotional state through a teaching strategy generation model; the interaction execution module is used to generate and output the robot's multimodal interactive behavior according to the progressive teaching script; the cognitive conflict monitoring module is used to monitor the child's problem-solving behavior for cognitive conflict goals in real time, and dynamically adjust the teaching pace according to the problem-solving behavior score; the data security module is used to encrypt, fragment, and store the child's cognitive development report on blockchain; the model optimization module is used to mine cross-user teaching records using graph neural networks and iteratively optimize the teaching strategy generation model under a federated learning framework.

[0026] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A teaching method for children using intelligent robots, characterized in that, include: Acquire multimodal perception data containing the current interaction scenario. The multimodal perception data includes visual image data and auditory speech data. Multimodal perception data is input into a pre-trained cross-modal semantic understanding model, which includes a visual Transformer branch, a speech Transformer branch, and a cross-modal attention fusion layer. Multimodal semantic features are extracted and fused, and a dynamic scene knowledge graph is constructed in real time based on the fused semantic features. Real-time capture and analysis of multimodal behavioral feedback data of target child users, including facial micro-expressions, eye movement trajectories, vocal emotions, and body posture, to quantitatively assess the child's current cognitive understanding and emotional engagement. Based on a dynamic scene knowledge graph and the children's cognitive and emotional states, a teaching strategy generation model is used to dynamically generate goal-oriented progressive teaching scripts. The teaching scripts contain multiple teaching interaction nodes that are progressive in logic and difficulty. Following a progressive teaching script, the robot generates and outputs multimodal interactive behaviors that match the current teaching interaction node in sequence, in order to guide children to complete a complete scenario-based cognitive learning session. After the entire progressive teaching script is completed, a structured children's cognitive development report is generated; and the children's cognitive development report is encrypted and pushed to the management terminal.

2. The child teaching method for intelligent robots as described in claim 1, characterized in that, The constructed dynamic scene knowledge graph is a dynamic scene knowledge graph that includes objects, events, character relationships, and implicit storylines mined through temporal causal reasoning; The specific method for encrypting and pushing children's cognitive development briefings to the management terminal is as follows: After generating children's cognitive development briefings, the briefings are encrypted using the encryption key generated by the key generator to generate ciphertext. Then, the ciphertext is transmitted to the management terminal using a secure transmission protocol based on asymmetric encryption. When the management terminal receives the ciphertext, it decrypts the ciphertext using the decryption key generated by the key generator to obtain the original data. After compression using a lossless compression algorithm, the original data is fragmented and distributed across multiple cloud storage nodes, and an attribute-based access control policy is established.

3. The child teaching method for intelligent robots as described in claim 2, characterized in that, A smart contract-based consortium blockchain is used to record key teaching nodes, changes in children's cognitive states, and model adjustments. Graph neural network data mining algorithms are then used to mine anonymized cross-user teaching records stored on the blockchain to uncover potential correlation patterns between different teaching strategies and children's cognitive development paths. The high-value knowledge patterns discovered are used as positive feedback to optimize and iterate the teaching strategy generation model. A federated learning framework is employed for distributed optimization of the teaching strategy generation model, with each terminal robot only uploading encrypted gradients and not sharing the original data, thus achieving model iteration while protecting user privacy.

4. The child teaching method for intelligent robots as described in claim 3, characterized in that, When guiding children to complete scenario-based cognitive learning, the system monitors in real time the children's problem-solving behaviors in response to the preset cognitive conflict goals of that node. The problem-solving behaviors include the correctness of the answer, the exploratory nature of the operation path, and the timing of seeking help. If the overall score of the resolution behavior meets the preset threshold, the script will be triggered to advance to the next node; Otherwise, the teaching strategy generation model is triggered to reduce the difficulty of the current node, or to retrieve relevant concept explanations or analogies from the knowledge graph to generate new explanatory interactions, so as to guide children to complete a complete, contextualized cognitive learning based on cognitive conflict resolution.

5. The child teaching method for intelligent robots as described in claim 4, characterized in that, Quantitative assessment of a child's current cognitive understanding and emotional engagement specifically includes: The collected facial micro-expression, eye movement trajectory, voice emotion and body posture data are respectively input into the corresponding pre-trained deep learning models to extract their respective high-dimensional features; The multi-modal fusion encoder is used to fuse the above-mentioned multi-source high-dimensional features to generate a hidden state vector that can comprehensively represent the child's current state; The latent state vector is input into a cognitive-emotion classifier. The classifier outputs quantitative scores for multiple dimensions, including attention concentration, comprehension confusion, emotional valence, and engagement, thereby enabling real-time quantitative assessment of children's cognitive comprehension and emotional engagement.

6. The child teaching method for intelligent robots as described in claim 5, characterized in that, The federated learning framework's distributed optimization process for the instruction strategy generation model includes: The central server initializes the global parameters of the teaching strategy generation model and distributes the global parameters to each terminal robot; Each terminal robot uses locally stored children's interaction data to calculate the model gradient locally, and the local interaction data does not leave the terminal device; Each terminal robot encrypts and performs differential privacy processing on the calculated local gradient, and adds random noise that conforms to the Laplace distribution to the local gradient. Each terminal robot uploads the processed encrypted gradient to the central server; The central server securely aggregates the collected encrypted gradients and updates the global model parameters using a federated averaging algorithm. The updated global model parameters are redistributed to each terminal robot, and the above process is repeated until the model converges.

7. A children's teaching device for intelligent robots, used to perform the children's teaching method for intelligent robots as described in claim 6, characterized in that, include: The module includes a multimodal perception module, a dynamic knowledge graph construction module, a child status assessment module, a teaching strategy generation module, an interaction execution module, a cognitive conflict monitoring module, a data security module, and a model optimization module. The multimodal perception module is used to acquire multimodal perception data containing the current interaction scenario; the dynamic knowledge graph construction module is configured to perform deep semantic understanding on the multimodal perception data and construct a dynamic scene knowledge graph; the child state assessment module is used to capture and analyze the multimodal behavioral feedback data of the target child user in real time, and quantitatively assess the child's current cognitive understanding state and emotional investment; the teaching strategy generation module is used to dynamically generate progressive teaching scripts based on the dynamic scene knowledge graph and the child's cognitive and emotional state through a teaching strategy generation model; the interaction execution module is used to generate and output the robot's multimodal interactive behavior according to the progressive teaching script; the cognitive conflict monitoring module is used to monitor the child's problem-solving behavior for cognitive conflict goals in real time, and dynamically adjust the teaching pace according to the problem-solving behavior score; the data security module is used to encrypt, fragment, and store the child's cognitive development report on blockchain; the model optimization module is used to mine cross-user teaching records using graph neural networks and iteratively optimize the teaching strategy generation model under a federated learning framework.

Images