A method and system for dynamic adaptation of hyper-realistic intelligent robot human-computer interaction models for multiple scenarios

Abstract

This invention belongs to the field of robot interaction management technology, and provides a method and system for dynamic adaptation of hyper-realistic intelligent robot human-computer interaction models for multiple scenarios. The method includes: a hyper-realistic intelligent robot terminal collecting feature data and uploading it to the cloud; the cloud pre-setting a general large model, a heart-chain cloud bridge model, and feature vectors for each scenario; the cloud automatically matching the heart-chain cloud bridge model as the target model based on the feature data; loading the target model and pushing it to the hyper-realistic intelligent robot terminal; the terminal deploying the model and performing runtime verification; if verification fails, re-matching until successful. This invention also provides a corresponding system. Through a two-layer architecture of a general large model and a heart-chain cloud bridge model, dynamic matching based on feature vectors, edge node caching, and a closed-loop runtime verification, this invention solves the technical problems of a single model being difficult to adapt to multiple scenarios and rigid deployment, achieving accurate, efficient, and stable adaptation of human-computer interaction models for multiple scenarios.

Description

Technical Field

[0001] This invention belongs to the field of robot interaction management technology. Specifically, it relates to a method and system for dynamic adaptation of hyper-simulation intelligent robot human-computer interaction models for multiple scenarios. Background Technology

[0002] With the rapid development of artificial intelligence technology, human-computer interaction systems based on hyper-simulation intelligent robots have been widely used in scenarios such as health care, family care, and rehabilitation medicine. Current intelligent robots are usually equipped with pre-trained large language models to understand and respond to users' interaction needs. They mainly have the following problems: (1) Most existing robots are equipped with a single model, which is difficult to adapt to the interaction needs of multiple scenarios, such as health monitoring scenarios and rehabilitation guidance scenarios. Moreover, with a fixed number of parameters and training data distribution, a single model is difficult to meet the differentiated needs of multiple scenarios at the same time. On the other hand, if a general large model is used to process all scenarios, it will lead to poor interaction experience in specific scenarios. (2) Existing solutions usually pre-set a fixed model for the robot. This static deployment method cannot dynamically switch models according to the real-time changes of the current interaction scenario. For example, when the robot switches from the health monitoring task in the morning to the companion chat task in the afternoon, it still needs to use the same model, which cannot achieve the real-time optimal matching of the model corresponding to the scenario. Summary of the Invention

[0003] The purpose of this invention is to provide a dynamic adaptation method for human-computer interaction models of hyper-simulated intelligent robots for multiple scenarios, so as to solve the technical problems existing in the prior art.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A dynamic adaptation method for a hyper-realistic intelligent robot human-computer interaction model for multiple scenarios includes the following steps: Step S1: The ultra-realistic intelligent robot terminal collects feature data during the human-computer interaction process and uploads the feature data to the cloud platform. The cloud platform is pre-set with a general large model, multiple heart chain cloud bridge models adapted to different application scenarios, and scene feature vectors corresponding to each heart chain cloud bridge model. Step S2: Based on the received feature data, the cloud platform automatically matches the HeartChain Cloud Bridge model that is suitable for the current application scenario as the target model; Step S3: The cloud platform loads the target model and pushes the target model to the hyper-realistic intelligent robot terminal; Step S4: The ultra-realistic intelligent robot terminal receives the target model and deploys it locally; Step S5: Run the verification of the loaded target model. If the verification fails, return to step S2 to rematch until the verification passes.

[0005] Preferably, the feature data includes interaction object feature data and application scenario feature data, wherein the interaction object feature data includes age information, behavioral habit information, and demand preference information, and the application scenario feature data includes health monitoring scenario data, companion chat scenario data, and rehabilitation guidance scenario data.

[0006] Preferably, in step S1, the hyper-simulation intelligent robot terminal performs feature processing on the raw data of the collected feature data to generate a feature vector representing the current interaction state, and uploads the feature vector to the cloud platform.

[0007] Preferably, the HeartChain Cloud Bridge model is a customized model, which is generated by training a general large model through transfer learning, and each HeartChain Cloud Bridge model is configured with a unique parameter set, which includes language style parameters, response latency parameters, and interaction accuracy threshold parameters.

[0008] Preferably, the scene feature vector is obtained by clustering analysis of historical interaction data in each application scenario or by expert annotation.

[0009] Preferably, in step S2, the cloud platform calculates the similarity between the received feature vector and the feature vector of each scene, and selects the HeartChain Cloud Bridge model corresponding to the scene feature vector with the highest similarity as the target model; if the highest similarity is lower than a preset threshold, it is determined that the current application scenario does not belong to the adaptation range of any HeartChain Cloud Bridge model, and the general large model described in step S1 is matched as the target model.

[0010] Preferably, the metadata of the HeartChain Cloud Bridge model is stored on the edge node. In step S2, during automatic matching, the HeartChain Cloud Bridge model that is suitable for the current application scenario is retrieved first from the edge node. If the edge node does not have a target model, it is downloaded from the cloud platform master node.

[0011] Preferably, in step S5, the verification indicators include response time, interaction accuracy, and data transmission stability. If any indicator fails to meet the preset threshold, the verification is deemed unsuccessful.

[0012] To achieve the above objectives, the present invention also provides a dynamic adaptation system for human-computer interaction models in multiple scenarios, comprising: a cloud platform and a hyper-simulation intelligent robot terminal. The cloud platform includes a model management unit and an adaptation decision unit, and the hyper-simulation intelligent robot terminal includes an information acquisition unit, a model running unit, and a model verification unit. The cloud platform is pre-configured with a general large model, multiple heart-chain cloud bridge models adapted to different application scenarios, and scene feature vectors corresponding to each heart-chain cloud bridge model. The information acquisition unit is used to collect feature data during human-computer interaction and upload the feature data to the cloud platform. The adaptation decision unit is used to calculate the similarity between the received feature vector and the feature vector of each scene, and select the HeartChain Cloud Bridge model corresponding to the scene feature vector with the highest similarity as the target model; if the highest similarity is lower than the preset threshold, it is determined that the current application scene does not belong to the adaptation range of any HeartChain Cloud Bridge model, and the general large model is matched as the target model. The model management unit is used to load the target model and push the target model to the hyper-simulation intelligent robot terminal; The model execution unit is used to receive the target model and deploy it locally; The model verification unit performs runtime verification on the loaded target model. If the verification fails, it re-matches until the verification passes.

[0013] Preferably, it also includes at least one edge node, which is deployed in a network location close to the hyper-simulation intelligent robot terminal for caching metadata and / or model parameters of the HeartChain Cloud Bridge model.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention adopts a two-layer architecture combining a general large model and a multi-heart chain cloud bridge model. The general large model ensures basic interactive capabilities, while the heart chain cloud bridge model adopts a customized model. Multiple heart chain cloud bridge models are generated through transfer learning and are optimized for specific scenarios such as health monitoring, companion chat, and rehabilitation guidance. At the same time, each heart chain cloud bridge model is configured with a unique parameter set to achieve fine-grained adaptation of one model and one parameter for one scenario, significantly improving the interactive experience in various scenarios.

[0015] This invention collects feature data such as the age, behavioral habits, and needs preferences of the interactive object, as well as scene data. After feature processing, feature vectors are generated and uploaded to the cloud. The cloud calculates the similarity between the received feature vectors and the pre-stored scene feature vectors, and matches the optimal Heart Chain Cloud Bridge model in real time. When there is no matching scene, it automatically switches to a general large model, thus realizing the dynamic adaptation capability of robot interaction.

[0016] This invention stores the metadata of the HeartChain Cloud Bridge model to edge nodes. When matching, it prioritizes retrieving from the edge nodes. If the model is not found, it is downloaded from the cloud master node. The introduction of edge nodes significantly reduces the model acquisition latency. At the same time, the hyper-simulation intelligent robot terminal only deploys the HeartChain Cloud Bridge model required for the current scenario, without needing to resident a general large model, which greatly saves the terminal's storage space and computing resources. Attached Figure Description

[0017] Figure 1 This is a flowchart of Embodiment 1 of the present invention.

[0018] Figure 2 This is a system block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to have a clearer understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described below are merely for illustrative purposes and to facilitate understanding. The technical solutions provided by the present invention are not limited to those provided in the following embodiments, nor should they limit the scope of protection of the present invention. Example 1

[0020] like Figure 1 As shown, this embodiment provides a dynamic adaptation method for the human-computer interaction heart-chain cloud bridge model for multiple scenarios, including the following steps: 1. Step S1: Data Acquisition and Preprocessing The highly realistic intelligent robot terminal (such as a home companion robot or a rehabilitation training robot) collects feature data in real time during interaction with the user. In this embodiment, the feature data includes two types: Interaction object characteristic data: collected through cameras, microphones, infrared sensors, etc. on the robot. Age information can be obtained through facial image analysis or user registration information; behavioral habit information includes the user's daily activity patterns, common interaction periods, interaction frequency, etc.; demand preference information includes the types of questions users frequently ask (such as health consultation, emotional expression, rehabilitation movement guidance), preferred language style, etc.

[0021] Application scenario feature data: This embodiment pre-sets three application scenarios: health monitoring scenario (such as measuring blood pressure and blood sugar, recording medication status), companion chat scenario (such as emotional companionship, news broadcast, entertainment interaction), and rehabilitation guidance scenario (such as rehabilitation movement demonstration, training plan follow-up). Scenario data can be obtained through user active selection, time regularity (such as morning and evening monitoring, afternoon chat), and device status (such as connecting a blood pressure monitor and determining it as health monitoring).

[0022] The hyper-simulated intelligent robot terminal performs feature processing on the collected raw data to generate feature vectors representing the current interaction state. Taking a health monitoring scenario as an example, the raw data includes: the user's age is 75 years old, the current behavior is measuring blood pressure, and the historical preference is for concise voice prompts. After feature processing, a feature vector V=[0.85,0.32,0.91,...] is formed (the specific dimensions are determined according to the model design). Feature processing can use mature methods such as normalization, one-hot encoding, and embedding layer mapping. This embodiment does not limit the specific feature engineering method, as long as it can convert the raw information into a vector form that can be used for similarity calculation.

[0023] The cloud platform is pre-deployed with: (1) a general large model: such as a model based on the Transformer architecture, which has general language understanding and generation capabilities; (2) multiple HeartChain Cloud Bridge models: the HeartChain Cloud Bridge model is a customized model, which is generated by transfer learning training based on the general large model. It is a model that is lightweight and professionally adjusted for specific scenarios. In this example, it is optimized for health monitoring, companion chat and rehabilitation guidance scenarios respectively. Each HeartChain Cloud Bridge model is configured with a dedicated parameter set, including language style parameters (such as friendly type, professional type), response latency parameters (such as real-time type, tolerance latency type), and interaction accuracy threshold parameters (such as fault tolerance rate); Example of transfer training: Based on the general large model, training datasets are constructed for health monitoring, companion chat and rehabilitation guidance scenarios respectively. Among them, the health monitoring dataset contains dialogue samples such as blood pressure and blood sugar reading interaction, medication reminder, and health report query, and the language style is labeled as "concise, objective, caring". The companion chat dataset contains samples such as emotional expression, news discussion, and entertainment interaction, and the language style is labeled as friendly, humorous, Empathy, rehabilitation guidance dataset: includes samples of rehabilitation action explanation, training plan adjustment, action correction feedback, etc., labeled with language style as professional, patient, and encouraging; during training, most of the underlying parameters of the general large model are frozen, and only the top layer (task-related layer) is fine-tuned, while the exclusive parameter set is adjusted (e.g., language style parameter: controls the word choice and sentence style of the generated text, implemented by style embedding vector; response delay parameter: controls the delay threshold of the model generating the first character, the health monitoring scenario requires low latency, the companion chat scenario can tolerate slightly higher latency, implemented by adjusting the decoder bundle search width; interaction accuracy threshold parameter: controls the confidence requirement of the model for user input, the rehabilitation guidance scenario requires high accuracy (e.g., refusing to answer rehabilitation questions with confidence below 0.9), implemented by adjusting the softmax temperature parameter). After training, three heart chain cloud bridge models (i.e. customized models, with about 1 / 10 of the parameters of the general large model) are generated to adapt to different scenarios; (3) scene feature vectors corresponding to each heart chain cloud bridge model: for example, health monitoring scenario feature vector V h Feature vector V of companionship chat scenario c , Rehabilitation guidance scenario feature vector V r The dimension is consistent with the feature vector uploaded by the robot.

[0024] 2. Step S2: Dynamically match the target model The cloud platform receives the feature vector V uploaded by the hyper-realistic intelligent robot terminal. cThe scene feature vectors pre-set on the cloud platform are obtained by clustering analysis of historical interaction data in various application scenarios, or they can be obtained by manual annotation by experts (e.g., experts define the ideal feature distribution for health monitoring scenarios). Preferably, this embodiment uses cosine similarity for matching calculation. Sim h =cosine(Vc,Vh); Sim c =cosine(Vc,Vc);Sim r =cosine(Vc,Vr).

[0025] It should be noted that, for those skilled in the art, other similarity measurement methods such as Euclidean distance, Manhattan distance, and Pearson correlation coefficient can also be used in practical applications, and these are all simple variations of the present invention.

[0026] Based on the similarity calculation results, the Heart Chain Cloud Bridge model corresponding to the scene with the highest similarity is selected as the candidate target model. For example, in this embodiment, a preset threshold T=0.7 is set. If the highest similarity is sim max If the value is ≥0.7, then the Heart Chain Cloud Bridge model corresponding to this scenario is matched; if the value is ≥0.7, then the Heart Chain Cloud Bridge model is matched. max If the similarity is less than 0.7, the current interaction scenario is determined to be outside the adaptation range of any HeartChain Cloud Bridge model. In this case, a general large model is matched as the target model. For example, if the cosine similarity between the current feature vector and the feature vector of the health monitoring scenario is 0.92, the similarity with the companion chat scenario is 0.45, and the similarity with the rehabilitation guidance scenario is 0.38, then the current scenario is determined to be health monitoring, and the target model is the health monitoring HeartChain Cloud Bridge model.

[0027] 3. Step S3: Model Loading and Pushing The model management unit of the cloud platform loads the matched target model. To improve distribution efficiency, the metadata of the Xinlian Cloud Bridge model (including model structure description, parameter download address, version information, etc.) is pre-distributed to edge nodes (such as regional servers geographically close to the robot). After the matching is completed in step S2, the target model is retrieved from the edge node first: if the edge node has the complete parameters of the model cached, it is downloaded directly from the edge node; if the edge node only stores the metadata and does not have the model parameters, the complete model is downloaded from the master node of the cloud platform. After the download is completed, the cloud platform pushes the target model to the hyper-simulation intelligent robot terminal that initiated the request.

[0028] 4. Step S4: Local Deployment After receiving the target model, the hyper-realistic intelligent robot terminal deploys it locally. The deployment process includes: loading model parameters into memory, initializing the model's operating environment, and establishing interfaces with the robot's underlying hardware (such as the speech synthesis module and motion control module). Once deployed, the robot uses the new model for subsequent human-computer interaction.

[0029] 5. Step S5: Run Verification and Feedback After the model is deployed and run, the model verification unit of the hyper-simulation intelligent robot terminal verifies the loaded target model and collects verification indicators: response time: the delay from the end of user input to the start of model output, which is required to be ≤500ms; interaction accuracy: the correct response rate to the preset test set (such as 50 scene-related queries), which is required to be ≥95%; data transmission stability: the network packet loss rate during model operation, which is required to be ≤1%.

[0030] If any of the above indicators fails to meet the preset threshold, the verification is deemed unsuccessful. The hyper-realistic intelligent robot terminal sends a re-matching request to the cloud platform and returns to step S2 to re-execute the matching process. If multiple re-matching attempts still fail the verification, the cloud platform can mark the terminal or the scenario as abnormal, triggering manual intervention or switching to offline emergency mode. Example 2

[0031] like Figure 2 As shown, this embodiment provides a dynamic adaptation system for human-computer interaction models in multiple scenarios, including a cloud platform and a hyper-realistic intelligent robot terminal.

[0032] In this embodiment, the cloud platform includes: Model storage unit: Used to store general large models, multiple HeartChain Cloud Bridge models (e.g., health monitoring model, companion chat model, rehabilitation guidance model), and scene feature vectors corresponding to each HeartChain Cloud Bridge model. The model storage unit adopts a distributed file system and supports high-concurrency reading.

[0033] Adaptation Decision Unit: Receives the feature vector uploaded by the hyper-simulation intelligent robot terminal, calculates the cosine similarity between it and the feature vectors of each scene in the model storage unit, selects the Heart Chain Cloud Bridge model corresponding to the scene with the highest similarity as the target model, and if the highest similarity is lower than the preset threshold (such as 0.7), then matches the general large model as the target model. The adaptation decision unit sends the matching result to the model management unit.

[0034] Model Management Unit: Based on the matching results of the adaptation decision unit, the target model is loaded. The model management unit maintains an edge node cache table and prioritizes retrieving models from edge nodes. If there is no target model at an edge node, it is downloaded from the model storage unit (master node) and pushed to the hyper-simulation intelligent robot terminal.

[0035] The ultra-realistic intelligent robot terminal includes: Information Acquisition Unit: Collects feature data of interactive objects (age, behavioral habits, needs and preferences) and application scenario feature data (health monitoring, companionship chat, rehabilitation guidance) through various sensors, performs feature processing on the collected raw data, generates feature vectors of preset dimensions, and uploads the feature vectors to the adaptation decision unit on the cloud platform.

[0036] Model execution unit: Receives the target model pushed by the cloud platform, completes the deployment and loading locally, and executes subsequent human-computer interaction tasks based on the model. Model execution unit 202 supports dynamic switching of models and can release the resources of the old model when a new model is received.

[0037] Model Validation Unit: Validates the target model loaded by the model execution unit, and collects response time, interaction accuracy, and data transmission stability metrics. If any metric fails to meet the preset threshold, a re-matching request is sent to the adaptation decision unit on the cloud platform until the validation is successful or the maximum number of retries is reached.

[0038] This embodiment also includes edge nodes: deployed in network locations close to the hyper-realistic intelligent robot terminal, these nodes cache metadata and model parameters of frequently used Heartlink Cloud Bridge models. When the model management unit retrieves a model, it prioritizes routing to the edge node, reducing download latency. It should be noted that edge nodes can be regional servers, base station edge computing units, or content delivery network nodes. The frequently used Heartlink Cloud Bridge models can be popular models based on historical call frequency statistics, or models frequently requested by hyper-realistic intelligent robot terminals in geographically proximate areas. Edge nodes can choose to cache only the model's metadata (for quick location and verification) or cache both metadata and complete model parameters, depending on actual storage capacity and network policies. When only metadata is cached, the hyper-realistic intelligent robot terminal can obtain the complete model from the cloud or other edge nodes based on the address information in the metadata. When the complete parameters are cached, model download services can be directly provided, achieving model distribution with minimal latency.

[0039] The above description represents the preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements without departing from the design principles and technical solutions of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic adaptation method for a hyper-realistic intelligent robot human-computer interaction model for multiple scenarios, characterized in that, Includes the following steps: Step S1: The ultra-realistic intelligent robot terminal collects feature data during the human-computer interaction process and uploads the feature data to the cloud platform. The cloud platform is pre-set with a general large model, multiple heart chain cloud bridge models adapted to different application scenarios, and scene feature vectors corresponding to each heart chain cloud bridge model. Step S2: Based on the received feature data, the cloud platform automatically matches the HeartChain Cloud Bridge model that is suitable for the current application scenario as the target model; Step S3: The cloud platform loads the target model and pushes the target model to the hyper-realistic intelligent robot terminal; Step S4: The hyper-realistic intelligent robot terminal receives the target model and deploys it locally; Step S5: Run the verification of the loaded target model. If the verification fails, return to step S2 to rematch until the verification passes.

2. The method for dynamic adaptation of hyper-simulation intelligent robot human-computer interaction model for multiple scenarios as described in claim 1, characterized in that: The feature data includes interactive object feature data and application scenario feature data. The interactive object feature data includes age information, behavioral habit information, and demand preference information. The application scenario feature data includes health monitoring scenario data, companion chat scenario data, and rehabilitation guidance scenario data.

3. The method for dynamic adaptation of hyper-simulation intelligent robot human-computer interaction model for multiple scenarios as described in claim 2, characterized in that: In step S1, the hyper-simulation intelligent robot terminal performs feature processing on the raw data of the collected feature data to generate a feature vector that represents the current interaction state, and uploads the feature vector to the cloud platform.

4. The method for dynamic adaptation of hyper-simulation intelligent robot human-computer interaction model for multiple scenarios as described in claim 3, characterized in that: The HeartChain Cloud Bridge model is a customized model, which is generated by training a general large model through transfer learning. Each HeartChain Cloud Bridge model is configured with a unique parameter set, which includes language style parameters, response latency parameters, and interaction accuracy threshold parameters.

5. The method for dynamic adaptation of hyper-simulation intelligent robot human-computer interaction model for multiple scenarios as described in claim 4, characterized in that: The scene feature vectors are obtained by clustering analysis of historical interaction data in various application scenarios or by expert annotation.

6. The method for dynamic adaptation of hyper-simulation intelligent robot human-computer interaction model for multiple scenarios as described in claim 5, characterized in that: In step S2, the cloud platform calculates the similarity between the received feature vector and the feature vector of each scene, and selects the HeartChain Cloud Bridge model corresponding to the scene feature vector with the highest similarity as the target model; if the highest similarity is lower than the preset threshold, it is determined that the current application scenario does not belong to the adaptation range of any HeartChain Cloud Bridge model, and the general large model described in step S1 is matched as the target model.

7. The method for dynamic adaptation of hyper-simulation intelligent robot human-computer interaction model for multiple scenarios as described in claim 1, characterized in that: The metadata of the HeartChain Cloud Bridge model is stored on the edge node. In step S2, during automatic matching, the HeartChain Cloud Bridge model that is suitable for the current application scenario is retrieved first from the edge node. If the edge node does not have a target model, it is downloaded from the cloud platform master node.

8. The method for dynamic adaptation of hyper-simulation intelligent robot human-computer interaction model for multiple scenarios as described in claim 1, characterized in that: In step S5, the verification indicators include response time, interaction accuracy, and data transmission stability. If any indicator fails to meet the preset threshold, the verification is deemed unsuccessful.

9. A dynamic adaptation system for a hyper-realistic intelligent robot human-computer interaction model for multiple scenarios, characterized in that, include: The cloud platform includes a model management unit and an adaptation decision unit, and the hyper-simulation intelligent robot terminal includes an information acquisition unit, a model running unit, and a model verification unit. The cloud platform is pre-configured with a general large model, multiple HeartChain Cloud Bridge models adapted to different application scenarios, and scene feature vectors corresponding to each HeartChain Cloud Bridge model. The information collection unit is used to collect feature data during human-computer interaction and upload the feature data to the cloud platform. The adaptation decision unit is used to calculate the similarity between the received feature vector and the feature vector of each scene, and select the HeartChain Cloud Bridge model corresponding to the scene feature vector with the highest similarity as the target model; if the highest similarity is lower than the preset threshold, it is determined that the current application scene does not belong to the adaptation range of any HeartChain Cloud Bridge model, and the general large model is matched as the target model. The model management unit is used to load the target model and push the target model to the hyper-simulation intelligent robot terminal; The model execution unit is used to receive the target model and deploy it locally; The model verification unit performs runtime verification on the loaded target model. If the verification fails, it re-matches until the verification passes.

10. The dynamic adaptation system for human-computer interaction models of hyper-simulation intelligent robots oriented towards multiple scenarios as described in claim 9, characterized in that: It also includes at least one edge node, which is deployed in a network location close to the hyper-simulation intelligent robot terminal, for caching metadata and / or model parameters of the HeartChain Cloud Bridge model.

Images