AgenticAI and MCP protocol-based three-dimensional reconstruction rehabilitation training auxiliary method

Abstract

The invention relates to the technical field of rehabilitation training, and discloses a three-dimensional reconstruction rehabilitation training auxiliary method based on AgenticAI and MCP protocols, which is used for providing personalized rehabilitation training support for a user and improving the body function state. The method comprises the steps that current state description information of a user is obtained through a natural language interaction interface, and a decision scheme containing multiple candidate rehabilitation actions is generated through fusion analysis of an AgenticAI decision module in combination with historical rehabilitation training records and physiological index data, obtained by a context engineering engine, of the user. And a corresponding three-dimensional reconstruction training tool is matched by means of an MCP protocol server, personalized calling parameters are generated according to a user capability evaluation result and preference setting, the tool is executed, real-time operation data are collected, the data are stored, and a user capability evaluation portrait is updated. According to training effect feedback data, weight parameters of the decision module are adjusted through a federated learning optimization algorithm, a cross-cycle personalized rehabilitation training plan is finally generated, and the personalization and systematicness of rehabilitation training are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rehabilitation training, in particular to a three-dimensional reconstruction rehabilitation training auxiliary method based on AgenticAI and MCP protocol, which is used to provide personalized rehabilitation training support for users and improve the physical function state. BACKGROUND

[0002] In the current field of rehabilitation training, the traditional training method generally has the problem of insufficient interactivity. In most cases, users can only passively follow the preset training assistance process, and it is difficult for them to feed back their real-time state to the system in a convenient way, resulting in a disconnection between the training assistance scheme and the actual needs of the users. In terms of obtaining user information, traditional methods can only obtain the current training state, and cannot effectively associate the user's past rehabilitation training records and physiological index data, making the judgment of the user's overall rehabilitation condition lack comprehensiveness, and thus affecting the rationality of the training assistance scheme.

[0003] When generating a training assistance scheme, the traditional system lacks intelligent analysis capability and is difficult to effectively fuse and process multi-dimensional user data, usually only providing fixed training action options, and cannot generate diversified candidate schemes according to individual differences of users, limiting the degree of personalization of training. In the matching of training tools, due to the lack of unified and efficient protocol support, the system is difficult to quickly access various training tool resources, and cannot accurately match the corresponding three-dimensional reconstruction training tools according to specific training actions, resulting in low efficiency of training tool calling and affecting the smoothness of the training assistance process.

[0004] The traditional rehabilitation training assistance system is relatively fixed in parameter setting and cannot dynamically adjust the calling parameters of the training tools in combination with the ability level and personal preferences of the users, resulting in poor user experience in the training assistance process, and some users may be unable to complete the training task due to inappropriate parameters, or the rehabilitation effect may be affected due to the mismatch between the training difficulty and their own ability. In terms of data storage and updating, the traditional system does not make full use of real-time operation data generated during training, and cannot update the user's ability assessment profile in a timely manner based on these data, resulting in a lack of basis for adjusting the subsequent training assistance scheme and making it difficult to achieve continuous optimization of training.

[0005] The traditional system lacks effective algorithm support to optimize the decision module, and cannot dynamically adjust the decision-making process based on training assistance effect feedback data, making the adaptability of the decision module poor and difficult to continuously improve the accuracy of scheme generation as the user's rehabilitation process progresses. When formulating a long-term training assistance plan, the traditional method cannot combine user ability changes and decision optimization results to generate a personalized plan across cycles, resulting in a lack of coherence and systematicness in training, and making it difficult to meet the needs of long-term rehabilitation of users. SUMMARY

[0006] The purpose of this invention is to provide a three-dimensional reconstruction rehabilitation training assistance method based on AgenticAI and MCP protocols, serving as an auxiliary tool in the rehabilitation training process to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides a three-dimensional reconstruction rehabilitation training assistance method based on AgenticAI and the MCP protocol, the method comprising: Obtain the current state description information input by the user through a natural language interaction interface; Based on the current state description information, the user's historical rehabilitation training records and physiological index data are queried and obtained through the context engineering engine; The AgenticAI decision-making module is used to fuse and analyze the current state description information, the historical rehabilitation training records, and the physiological index data to generate a decision scheme containing multiple candidate rehabilitation actions. Access the registered tool library through the MCP protocol server, and match the corresponding three-dimensional reconstruction training tool according to the candidate rehabilitation actions in the decision scheme; Based on the user's ability assessment results and preference settings, the personalized calling parameters of the 3D reconstruction training tool are dynamically generated; The 3D reconstruction training tool, configured with the personalized calling parameters, is executed to collect real-time operation data of the user during the training process. The context engineering engine stores the real-time operation data and updates the user's capability assessment profile. Based on the feedback data after training, the decision weight parameters of the AgenticAI decision module are adjusted using a federated learning optimization algorithm. Based on the adjusted decision weight parameters and the updated ability assessment profile, a personalized rehabilitation training plan is generated across cycles.

[0008] Preferably, obtaining the current state description information input by the user through the natural language interaction interface includes: Receive functional status descriptions from users via voice or text input; Semantic parsing is performed on the functional state description to extract key state feature vectors; The key state feature vectors are matched with a predefined state classification model to determine the user's current state category identifier.

[0009] Preferably, the step of querying and obtaining the user's historical rehabilitation training records and physiological indicator data through the context engineering engine includes: Based on the state category identifier, retrieve the associated historical training performance data from the time series database; obtaining historical assessment report data matching the state category identifier from a document database; integrating the historical training effect data and the historical assessment report data into a multi-dimensional context feature vector.

[0010] Preferably, the fusion analysis using the Agentic AI decision module includes: spatiotemporal alignment processing of the key state feature vector and the multi-dimensional context feature vector; calculating the expected utility value of different rehabilitation training actions through a weighted multi-objective optimization algorithm; ranking candidate rehabilitation actions according to the expected utility value to form a priority call sequence.

[0011] Preferably, the access to the registered tool library through the MCP protocol server includes: sending a tool discovery request to the MCP server according to the rehabilitation action type in the priority call sequence; receiving the available tool list and its function description information returned by the MCP server; selecting the optimal three-dimensional reconstruction training tool based on the matching degree of the function description information and the rehabilitation action requirement.

[0012] Preferably, the dynamically generated personalized call parameters of the three-dimensional reconstruction training tool include: extracting the user's historical performance data from the multi-dimensional context feature vector; calculating the current recommended difficulty level parameter according to the historical performance data; combining the user's device configuration information to generate adaptive resolution parameters and rendering quality parameters.

[0013] Preferably, the execution of the three-dimensional reconstruction training tool includes: sending a call request and the personalized call parameters to the selected three-dimensional reconstruction training tool through the MCP protocol; receiving the training scene configuration data returned by the three-dimensional reconstruction training tool; loading the training scene configuration data and starting the interactive training process.

[0014] Preferably, the collection of real-time operation data of the user during the training process includes: obtaining the user's body motion trajectory data through a motion capture device; recording the user's interactive response time data in the three-dimensional scene; collecting physiological signal monitoring data during the training process.

[0015] Preferably, the adjustment of the decision weight parameter through the federated learning optimization algorithm includes: differentially analyze the real-time operation data and the expected training target; generate a local model update gradient based on the differential analysis result; fuse the local model update gradient and global model parameters through a secure aggregation algorithm.

[0016] Preferably, the generation of the cross-cycle personalized rehabilitation training plan comprises: determining a key improvement target for the next training cycle according to the updated ability assessment profile; selecting a plurality of complementary training tools based on the key improvement target from a tool library, arranging the time sequence and intensity parameters of different training tools, and forming a multi-day training assistance scheme.

[0017] Compared with the prior art, the present application has the following advantages: The natural language interaction interface is used to obtain the current state description information of the user, which provides a convenient and intuitive interaction method for the user, and the user can easily feed back the real-time situation to the system, breaking the limitation of passive participation in traditional training, and making the formulation of the training assistance scheme more in line with the actual needs of the user at the moment. With the help of the context engineering engine, the historical rehabilitation training records and physiological index data of the user are queried and obtained, the multi-dimensional user information is integrated, the cognition of the system to the rehabilitation condition of the user is more comprehensive, and more rich reference information is provided for the generation of the subsequent training assistance scheme, avoiding the deviation of the scheme caused by single information.

[0018] The AgenticAI decision module is used to fuse and analyze the multi-dimensional data and generate a decision scheme containing a plurality of candidate rehabilitation actions, which fully utilizes the advantages of intelligent analysis, can provide diversified training choices for the user from different angles, changes the single training action of the traditional system, meets the individual needs of different users in different rehabilitation stages, and enables the user to select more suitable training actions according to the actual situation, improving the flexibility of training. The MCP protocol server is used to access the registered tool library and match the corresponding three-dimensional reconstruction training tools, and a unified and efficient protocol is used to ensure the rapid access of the training tool resources, and the precise docking between the training actions and the tools is realized, which effectively improves the calling efficiency of the training tools, ensures the smooth progress of the training process, and reduces the training interruption or delay caused by tool matching problems.

[0019] The personalized calling parameters of the three-dimensional reconstruction training tool are dynamically generated based on the user's ability evaluation results and preference settings, so that the parameter settings of the training tool can be highly adapted to the individual situation of the user, and the user can obtain a more comfortable experience in the training process, neither too high parameter settings nor too low parameter settings can be difficult to complete the training, nor can the expected training effect be achieved, which helps the user to better adhere to the training task. The training tool with the configured personalized parameters is executed and real-time operation data is collected, and at the same time, the data is stored and the user ability evaluation portrait is updated through the context engineering engine, realizing real-time utilization of training data, so that the user's ability evaluation portrait can timely reflect the latest rehabilitation progress, providing real-time basis for subsequent adjustment of the training assistance scheme, so that the training assistance scheme can be dynamically optimized with the change of the user's ability.

[0020] According to the training effect feedback data, the decision weight parameters of the AgenticAI decision module are adjusted through the federated learning optimization algorithm, which can continuously optimize the analysis ability of the decision module, so that the decision process is more in line with the training effect requirements of the user, and with the continuous advancement of the training, the ability of the decision module to generate reasonable schemes is continuously improved, further enhancing the accuracy and applicability of the training assistance scheme. Based on the adjusted decision weight parameters and the updated ability evaluation portrait, a cross-cycle personalized rehabilitation training plan is generated, so that the training plan can run through different rehabilitation cycles of the user, maintaining the continuity and systematicness of the training, and the user can obtain training guidance continuously adapted to his own situation in the long-term rehabilitation process, which helps to gradually achieve the rehabilitation goal and improve the overall rehabilitation effect. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A working principle diagram of the three-dimensional reconstruction rehabilitation training assistance method based on AgenticAI and MCP protocol described in the application; Figure 2 A flowchart for querying and obtaining the historical rehabilitation training records and physiological index data of the user through the context engineering engine; Figure 3 A flowchart for dynamically generating personalized calling parameters of the three-dimensional reconstruction training tool. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0023] Please refer to Figure 1The application provides a three-dimensional reconstruction rehabilitation training auxiliary method based on Agentic AI and MCP protocol, which comprises the following steps: obtaining current state description information input by a user through a natural language interaction interface, the interface supporting voice or text input and being capable of receiving real-time description of the functional state of the user; based on the current state description information, a context engineering engine is activated to query and obtain historical rehabilitation training records and physiological index data of the user, the data being stored in a distributed database and comprising time series data and document data; then, an Agentic AI decision module is used to perform fusion analysis on the current state description information, the historical rehabilitation training records and the physiological index data, a multi-source data alignment and optimization algorithm is adopted to generate a decision scheme containing multiple candidate rehabilitation actions; after the decision scheme is formed, an MCP protocol server is used to access a registered tool library, the tool library being pre-stored with multiple three-dimensional reconstruction training tools, and the system matches corresponding tools according to the candidate rehabilitation actions in the decision scheme; during the matching process, the system dynamically generates individualized calling parameters of the three-dimensional reconstruction training tools in combination with the ability evaluation results and preference settings of the user, the parameters including difficulty levels, resolution and the like; the system executes the three-dimensional reconstruction training tools configured with the individualized calling parameters, and the tools collect real-time operation data of the user during the training process, such as motion trajectories and physiological signals, when running; the collected data are stored through the context engineering engine and used to update the ability evaluation portrait of the user to reflect the latest training effect; after the training is completed, the decision weight parameters of the Agentic AI decision module are adjusted based on the effect feedback data through a federated learning optimization algorithm, the process involving local gradient calculation and global parameter aggregation; according to the adjusted decision weight parameters and the updated ability evaluation portrait, a cross-cycle individualized rehabilitation training plan is generated, the plan covering multi-day training arrangements and intensity parameters; through the closed-loop feedback mechanism, the method realizes continuous optimization and individualized adaptation of the rehabilitation training.

[0024] Example 1: see Figure 2, natural language interaction interface as the system and the user directly communicate with the channel, its design focuses on the user daily, non-structured description of the understanding and transformation, the interface is usually integrated in the user familiar terminal device application, such as tablet computer or smart home terminal, the user can describe their own daily physical feeling through voice dialogue or text input box, for example, “today morning feel wrist more stiff than yesterday, some difficult when hand to shoulder height” or “knee pain feeling lightened, but still not stable when standing for a long time”. Voice input signal after front end pretreatment, including environmental noise suppression and acoustic feature enhancement, then by automatic speech recognition engine into the original text, text input is directly into the subsequent processing pipeline. Semantic analysis link assumes the task of extracting structured information from free text, it uses pre-trained language model to analyze the input text, the model first performs word segmentation and part-of-speech tagging, identifies the entity related to the rehabilitation state, such as “wrist”, “stiff”, “difficult”, “pain”, “unstable” and so on, and further analyzes the semantic relationship between these entities, for example, the subject-state relationship between “wrist” and “stiff”, the action-difficulty relationship between “hold hand” and “difficult”. The analysis process is not simply keyword matching, but combined with the knowledge graph in the field of rehabilitation medicine, so as to accurately judge that “difficult” describes the limited joint range of motion, while “unstable” reflects the balance ability problem. Finally, the analysis output is a set of key state feature vectors, these vectors represent the core elements of user description in numerical form, for example, a feature dimension may represent “upper limb joint range of motion subjective score”, its value is quantified according to the degree of “difficult”, another dimension may represent “lower limb stability subjective score”, the value comes from the intensity evaluation of “unstable” description.

[0025] The extraction of the key state feature vector provides the basis for subsequent classification, and the pre-defined state classification model is a classifier trained by a large amount of labeled data, which comes from the historical accumulation of user state descriptions annotated by rehabilitation experts and their corresponding clinical state categories. This model maps the input feature vector to a discrete state category identifier, and the classification process can be understood as a pattern matching process. The model calculates the similarity between the input feature vector and each pre-defined state category prototype vector, for example, a feature vector characterized by "significant decrease in joint activity" and "with slight pain" will have a high similarity with the prototype vector of the "moderate disability" category, and thus be assigned to this identifier. The state category identifier adopts a standardized coding system, for example, "M01" may represent "upper limb mild motor dysfunction period", "L02" may represent "lower limb stability recovery middle period", and this standardized identifier facilitates accurate and error-free exchange of information between different modules within the system, and also provides a unified query key for subsequent retrieval of relevant information from the database. Once the state category identifier is determined, the context engineering engine starts working, and its primary task is to retrieve associated historical training effect data from the time series database according to this identifier. The time series database is specially used to store data points arranged in chronological order, in this application scenario, it records the detailed data sequence of each rehabilitation training of the user, and each record has an accurate time stamp, and the content may include the identifier of the training action, the number of repetitions of each group of actions, the amplitude, speed and system automatically determined action standard degree score of the completed action. When the engine performs a query, it will use the current state category identifier as the screening condition, and focus on the historical training project data associated with the current state, for example, if the current state identifier points to wrist dysfunction, the query will preferentially retrieve all historical data records related to wrist training. The query result may be a series of time series data, showing the trend of the user's completion and quality when performing wrist flexion, rotation and other training in the past few weeks or months.

[0026] The contextual engineering engine concurrently retrieves historical assessment report data matching the same state category identifier from a document database. The document database stores semi-structured or unstructured document materials, such as periodic rehabilitation assessment reports, doctor or therapist's comment notes, and user self-filled functional scale results. These documents contain rich qualitative information and clinical observations. The engine locates and extracts relevant assessment reports related to the current user state through document indexing and content retrieval techniques. For example, a last week's assessment report might record a text description of "patient's wrist joint extension activity has improved, but endurance is still insufficient" and a scale score result. After obtaining these text and score data, the engine needs to perform information extraction to convert the text description into quantifiable features, such as extracting the "endurance indicator" as a "to be improved" label from "endurance is still insufficient". The historical training effect data obtained from the time series database and the historical assessment report data obtained from the document database are fused into a unified multi-dimensional contextual feature vector. This process involves data cleaning, format standardization, and feature alignment. Data cleaning handles missing or abnormal values, for example, if a training session is missing due to sensor failure, interpolation or ignore strategy is used. Format standardization ensures that all numerical features have the same dimension and range, facilitating subsequent model processing. Feature alignment ensures that features from different sources can be corresponded in the time dimension, for example, associating Tuesday's training effect data with the conclusion in Wednesday's assessment report. The fused multi-dimensional contextual feature vector is a high-dimensional numerical array, each dimension representing a feature, such as "past seven-day average training completion", "wrist joint activity historical best value", "last assessment report endurance score", "pain frequency trend coefficient", etc. This vector comprehensively and multi-dimensionally depicts the user's historical background and ability baseline related to the current state, providing rich and structured contextual information for the Agentic AI decision module to perform fusion analysis, enabling subsequent decisions to be based on a deep understanding of the user's historical progress and current situation.

[0027] Embodiment 2: The Agentic AI decision module receives the key state feature vector and the multi-dimensional context feature vector from the pre-processing pipeline, which respectively carry the user's immediate state awareness and long-term historical background. The first operation of the module after starting is to perform spatio-temporal alignment on the two heterogeneous data sets, because the key state feature vector reflects the user's state snapshot at the current time point, while the multi-dimensional context feature vector contains historical sequence data across different time granularities. The alignment operation aims to eliminate the bias caused by inconsistent time scales and mismatched feature dimensions. The spatio-temporal alignment process uses a variant of the dynamic time warping algorithm to perform nonlinear stretching and transformation on the historical context feature sequence, so that it can be optimally aligned with the time reference point represented by the current state feature on the time axis. At the same time, for feature dimensions, feature projection technology is used to map features from different sources to a unified hidden space, ensuring their comparability at the semantic level.

[0028] After spatio-temporal alignment, the fused feature set is sent to a weighted multi-objective optimization algorithm for decision analysis. The algorithm internally defines multiple optimization objectives that need to be considered and balanced, including maximizing training effect, minimizing user fatigue risk, promoting training compliance, and satisfying user personal preferences. Each optimization objective is quantified as a mathematical function, and its function value is calculated based on the input features. For example, the training effect objective function may be positively correlated with the user's historical progress speed and current ability baseline, while the risk objective function may be negatively correlated with the user's recent fatigue index and past injury record. The algorithm assigns a dynamic weight to each objective function, which is not fixed but dynamically adjusted based on the user's last training feedback, long-term rehabilitation goals, and system safety policies. Based on these weighted objectives, the algorithm calculates the expected utility value for each candidate rehabilitation action in the pre-set action library. The calculation process simulates multi-attribute decision analysis and predicts the contribution of a specific rehabilitation action, such as "shoulder joint abduction resistance training," to each optimization objective within a future training period. Finally, a scalarized expected utility value is synthesized. The calculation of the expected utility value provides a quantitative basis for the ranking of candidate rehabilitation actions. The decision module uses an efficient ranking algorithm to sort all candidate actions in descending order based on the calculated utility values. Actions with higher utility values mean that their overall benefit-risk ratio is better under the current user state. The ranking result is structured into a priority call sequence, which not only contains the identifiers of the actions, but also may include recommended intensity ranges, pre-set durations, and other metadata. The structure of the sequence ensures that in the case of limited resources or time constraints, the system can prioritize the most beneficial action scheme at the top of the sequence. This priority call sequence constitutes the output of the decision module, which is a preliminary decision recommendation containing multiple optional schemes.

[0029] The execution flow of the system turns to the MCP protocol server to materialize the rehabilitation action suggested in the decision plan. The MCP protocol server serves as a system-neutral function invocation hub, maintaining a dynamically updated tool registry where various 3D reconstruction training tools implementing standardized interfaces are registered. The system initiates a structured tool discovery request to the MCP server based on the generated prioritized invocation sequence, especially the top-ranked rehabilitation action type in the sequence, which encapsulates the action type code, required input parameter format, and expected output specification within the request message. Upon receiving the tool discovery request, the MCP server performs a query match in its registry, which can be imagined as a service directory where each registered tool publishes its metadata, including the tool's unique identifier, version number, describable functionality, required parameter list, and communication endpoint address. The matching engine within the server parses the action type requirement in the request and performs a semantic-level similarity comparison with the functionality descriptions of all tools in the registry, which might utilize domain ontology-based semantic matching techniques to ensure the understanding of the functional correlation between "upper limb joint range of motion training" and "arm stretching exercise". After the matching is completed, the MCP server generates a list of available tools as a response to return to the invoker, and each item in the list contains sufficient information for the system to make the next judgment.

[0030] Upon receiving the list of available tools returned by the MCP server, the system enters the tool selection phase. At this point, the system needs to perform a secondary matching degree evaluation based on the returned tool functionality description information and the precise requirements of the current rehabilitation action, aiming to select the optimal 3D reconstruction training tool. This matching degree evaluation is more detailed than the preliminary matching on the server side, as it needs to consider the degree of functional fit, the performance indicators of the tool, compatibility with the user terminal device, and historical usage effect data of the tool. For example, for the same "shoulder joint abduction resistance training" action, there may be two tools that both claim to support it, but one tool focuses on precise motion trajectory tracking and feedback, while the other provides more interactive fun in a virtual scene. The system needs to make a final selection based on the specific profile of the current user, whether they need more precise correction or more motivation maintenance. The evaluation process may involve a simple scoring card model, scoring each candidate tool in several key dimensions, then summing the weighted scores to select the tool with the highest total score as the final execution carrier. Once the optimal tool is selected, its identifier and invocation specification are determined, and the system is ready for subsequent parameter configuration and tool execution. The entire process from intelligent decision-making to tool discovery and selection embodies the seamless integration between the autonomous decision-making capability of AgenticAI and the flexibility of the MCP protocol in resource scheduling.

[0031] Example 3: refer toFigure 3 The generation of personalized calling parameters is a fine-tuning process relying on data-driven and rule-based reasoning, with the core input derived from the multi-dimensional context feature vector constructed in the previous step. This vector, like a compressed data archive, encapsulates the user's historical performance, ability baseline, and trend related to rehabilitation training. The system extracts the user's historical performance data from this feature vector, not simply by reading the original values, but by purposefully selecting and aggregating features. For example, the system focuses on the success rate sequence, average reaction time, error type distribution statistics (such as the frequency of excessive angle deviation), and fatigue decay curve of the user performing specific types of actions within the last training period (e.g., the last seven times). These historical performance data are pre-processed, including removing possible outliers and smoothing fluctuations, to obtain a stable and representative set of historical ability features.

[0032] Based on the extracted historical performance data, the system enters the calculation phase of the difficulty level parameter, which directly relates to the safety and effectiveness of training. If the difficulty level is too high, it may lead to frustration or injury risk, and if it is too low, it may not produce enough training stimulation. The calculation process can use an adaptive logistic regression model based on historical performance data, which takes historical success rate, reaction time, and other features as input and outputs a continuous difficulty coefficient. This coefficient is then mapped to a discrete difficulty level. For example, the difficulty level parameter can be determined as follows: ; where: represents the average success rate of recent training tasks (dimensionless), represents the average reaction time (dimension: seconds), is a weight coefficient (dimensionless) preset according to the task type, used to balance the importance of speed and accuracy, is a bias term based on the user's long-term progress trend (dimensionless), is the ideal success rate benchmark value (dimensionless), is the maximum difficulty level number preset by the system (dimensionless). The symbol represents the ceiling function. is a standardization function used to map to a dimensionless value in the interval 0~1, with the inverse normalization (the shorter the reaction time, the closer the normalized value to 1, consistent with the success rate trend): ; is the maximum allowed reaction time for this training task (system preset, such as 5 seconds); is the ideal reaction time of this training task (system default, such as 0.5 seconds); if , take ; if , take .

[0033] The meaning of this calculation rule is that when the user's average success rate is high, the reaction time is short, and there is a trend of progress, the value will increase accordingly, and the system will recommend a higher difficulty level to provide new challenges; otherwise, it will lower the difficulty to ensure the feasibility.

[0034] The system needs to combine the user's device configuration information to generate adaptive technical parameters to ensure that the three-dimensional reconstruction training content can be smoothly and high-quality presented on the user's terminal. Device configuration information usually includes the screen resolution of the display device, the model and performance of the graphics processor, the size of the memory, and the network bandwidth, etc. The system will query the pre-stored device profile or detect the current device status in real time. When generating the resolution parameter, the system needs to balance between rendering effect and performance overhead. It will select an optimal rendering resolution according to the physical resolution of the device screen and the processing power of the GPU. For example, for high-performance graphics workstations, the native resolution may be directly used for rendering, while for mobile terminals, dynamic resolution scaling technology may be used. The rendering quality parameter involves the setting of model detail level, texture filtering quality, shadow effect, and anti-aliasing level, etc. The combination of these parameters directly affects the visual fidelity and computing load. The system selects the most suitable one from the pre-set several rendering quality schemes according to the ability rating of the device GPU, or performs more fine-grained parameter adjustment. After the personalized invocation parameter generation is completed, the system starts the invocation and execution process of the tool. This process communicates with the selected three-dimensional reconstruction training tool through the MCP protocol. The system constructs a structured request message containing tool invocation instructions and all personalized invocation parameters according to the message format specified in the MCP protocol. This request message is sent to the running environment or service endpoint of the tool through the network.

[0035] After receiving the invocation request, the target three-dimensional reconstruction training tool will first parse and verify the incoming parameters to ensure their validity and compatibility. Then, the tool dynamically configures its internal training scene according to these parameters. For example, according to the difficulty level parameter , the tool loads a three-dimensional model of corresponding complexity, sets the physical properties of the virtual objects (such as weight, friction), adjusts the distance or precision requirements of the task objectives. According to the resolution parameters and rendering quality parameters, the tool initializes the graphics rendering engine, sets the viewport size, texture precision, and light and shadow effect levels. After completing the internal configuration, the tool generates a training scene configuration data, which may include scene description files, resource indexes, initialization scripts, etc., and returns it to the main system to confirm readiness. After receiving the training scene configuration data, the main system loads these data and officially starts the interactive training process. The loading process involves loading scene resources into memory, initializing the graphics rendering pipeline, and establishing a connection with user input devices (such as data gloves, motion capture cameras, force feedback devices). After the training starts, the system enters a real-time interaction loop, and the three-dimensional reconstructed training scene is presented to the user. The tool guides the user to perform actions according to the preset logic and personalized parameters, such as grabbing a floating ball in a virtual environment to train hand-eye coordination, or maintaining the balance of a virtual character to train core muscle groups. During the entire training execution period, the system continuously runs, not only rendering the scene, but also processing user interaction inputs in real time, and judging the correctness of the action according to the built-in algorithm, providing immediate feedback such as visual cues (color changes), auditory signals (prompt sounds), or tactile stimuli (vibrations). This process continues until the training task is completed or the user actively interrupts, and all real-time operation data generated during this period is recorded, providing a data basis for subsequent analysis and optimization. The entire implementation process embodies a closed loop from intelligent parameter generation to precise tool invocation, ensuring the personalization and adaptability of rehabilitation training.

[0036] Example 4: The entire details of the user's interaction with the three- dimensional reconstructed training scene are recorded objectively and quantitatively by a multi-modal sensor system deployed in the training environment. Taking a specific upper limb rehabilitation training scene as an example, Ms. Li is performing a virtual "reach and touch moving target" task, which aims to improve her shoulder joint range of motion and stability. The training is conducted in a dedicated space equipped with multiple sensors. Data acquisition begins at the moment the training starts and continues until the end of the training session. Motion capture devices form the core of data acquisition. A high-precision optical motion capture system (such as the Vicon series) continuously tracks the three-dimensional coordinates of reflective markers attached to Ms. Li's major joints (such as the acromion, elbow joint, wrist joint) through infrared cameras arranged around the training site. These coordinate data are recorded at a frequency of more than 100 frames per second, reconstructing her upper limb skeletal motion trajectory in real time. The trajectory data contain the position sequence (X, Y, Z coordinates) of each marker point in three-dimensional space, allowing calculation of joint angles, motion speed, acceleration, and motion smoothness. For example, in the "reach" motion, the system can accurately calculate the flexion angle of the shoulder joint as a function of time and the path of the hand end effector (wrist marker point) to the target position. At the same time, inertial measurement unit (IMU) sensors are fixed to her forearm and upper arm, providing angular velocity and angular acceleration data, which are fused with optical data to further improve the accuracy and robustness of motion measurement, especially when optical markers are temporarily obscured. In addition to macroscopic motion trajectories, the user's interaction response time in the three-dimensional scene is also accurately recorded. The system presents a visual or auditory stimulus in the virtual environment, and the time interval from the precise time point of stimulus presentation to the user's effective motion response towards the target is recorded as the reaction time. The total completion time of the task, from the start of the training task to the successful touching of the target or the end of the task timeout, is also recorded. These time data reflect the user's cognitive processing speed and motion initiation efficiency.

[0037] Physiological signal monitoring data reflect the physiological load of training from another dimension. Ms. Li wears a comfortable heart rate monitoring chest strap or wrist strap that continuously collects her electrocardiogram signals and extracts heart rate, heart rate variability, and other indicators from them. Electromyography (EMG) sensor surface electrodes are attached to the skin surfaces of her main muscle groups involved in the action, such as the deltoid and supraspinatus muscles, to record the changes in muscle electrical signals during activity. These signals, after processing, can reflect muscle activation, fatigue state, and muscle coordination. All these data streams: motion trajectory, interaction time, physiological signal, are stamped with uniform high-precision timestamps, ensuring strict time alignment and correlation analysis in subsequent analysis. See Table 1 for a display of the fragmented real-time operation data collected in a training task.

[0038] Table 1: Real-time operational data record table (shoulder joint "reach for object" task) ; The massive amount of real-time operational data collected is transmitted to the contextual engineering engine for storage, which uses a time-series database format to facilitate efficient querying and aggregated analysis by time range. After storage, the system immediately initiates a differential analysis of the effectiveness of this training. This analysis compares the real-time operational data with the expected training objectives preset for this training task. The expected objectives can include: shoulder joint range of motion reaching a certain angle range, motion trajectory smoothness exceeding a certain threshold, reaction time being shorter than a certain period, heart rate being maintained within a certain safety interval, etc. The analysis calculates the differences between actual performance and expected objectives, for example, the average absolute error between actual motion trajectory and ideal trajectory (shortest path), or the difference between actual reaction time and expected reaction time. For physiological signals, it checks for abnormal fluctuations, such as whether the heart rate suddenly rises above the safety threshold. Based on the results of the difference analysis, the system begins to generate local model update gradients. This process occurs on the client side (i.e., the training terminal of user Ms. Li), aiming to optimize the global model without exposing raw data. The current decision weight parameters of the AgenticAI decision module are used to perform a forward propagation calculation on the data of this training, obtaining a predicted value of the effectiveness of this training. Then this predicted value is compared with the real effectiveness evaluation value calculated based on the actual collected data (for example, a comprehensive score combining task completion, motion quality score, and physiological load index), and the prediction error (loss) is calculated. This error is then calculated through the backpropagation algorithm to obtain the gradient relative to the current model decision weight parameters. This gradient vector indicates how the model's weight parameters should be fine-tuned to reduce the prediction error on this user's data. This gradient contains the knowledge learned from this individual training data, but does not directly contain any identifiable raw terminal data of the individual.

[0039] After generating the local gradient, the core step of the federated learning optimization algorithm: secure aggregation, comes into play. The local model update gradient computed on Ms. Li's terminal is derived from the real-time operational data collected by the user during the training process through the backpropagation algorithm, and does not directly contain identifiable user privacy. To further strengthen the privacy protection of the original information closed loop and avoid the risk of indirect disclosure of privacy after the gradient data is combined with other associated information, this local model update gradient still needs to be encrypted before being sent to the central aggregation server. The server also receives model update gradients computed on their own local data from many other user terminals in the system (all of which have been anonymized). Using secure aggregation protocols (such as homomorphic encryption-based techniques), the aggregation server weights and averages or sums the local gradients from multiple sources without decrypting individual gradient vectors. The global gradient update obtained by aggregation is used to adjust the decision weight parameters of the central global AgenticAI decision module. The decision weight parameters adjusted by the federated learning optimization algorithm allow the global model to absorb the common knowledge contained in the training data of all participating users, such as which action parameter settings are more likely to achieve the goal of improving physical function (such as expanding the range of shoulder joint activity and enhancing motion stability) with lower risk for user groups with similar characteristics to Ms. Li (such as shoulder joint activity to be improved). The updated model parameters are then distributed to each user terminal, including Ms. Li's terminal, to guide the next cycle of personalized decision-making. This means that Ms. Li's next training is based not only on her own historical data, but also implicitly benefits from the collective experience of the entire anonymous user group, while the secure handling of gradient transmission and aggregation further consolidates the privacy protection effect of the original information, ensuring that user privacy is only retained on the local terminal and not leaked during the model optimization process.

[0040] Example 5: The generation of a plan begins with a deep analysis of the updated capability assessment profile, which is not a static report but a dynamic data structure that integrates real-time operational data from the user's most recent training, historical performance trends, and model insights optimized through federated learning. For example, Mr. Wang, a user recovering from a shoulder injury, has an updated capability assessment profile that shows significant improvement in his shoulder joint flexion range of motion, from 90 degrees initially to 140 degrees in the past four weeks. However, the profile also reveals that he has made slow progress in shoulder joint abduction endurance, with noticeable declines in the quality of his performance in the later sets of repetitive abduction resistance training. Additionally, the motion control analysis module in the profile indicates that Mr. Wang exhibits a slight compensatory shrugging phenomenon when his shoulder joint approaches maximum range of motion. Based on this detailed profile, the system initiates a process to determine key improvement targets, which involves a multi-objective decision analysis that weighs the current level of each capability, the gap to the ideal target, the importance of the capability to overall function, and the potential difficulty of improvement. After calculation, the system may identify "improving shoulder joint abduction muscle endurance" and "improving motion control at shoulder joint end range" as the key improvement targets for the next training cycle, while adjusting "shoulder joint flexion range of motion," which has already made good progress, to a maintenance training target. After the key improvement targets are determined, the system enters the stage of selecting complementary training tools from the tool library. The selection process is not random or simple enumeration, but is based on a complex tool-target matching model. This model embeds knowledge in the field of rehabilitation medicine, understanding the specific effects of different training tools on certain physiological functions and the synergistic or interfering effects of tool combinations. For the goal of "improving shoulder joint abduction muscle endurance," the system may select an "isokinetic abduction resistance training tool" from the tool library, which can provide constant resistance and accurately record the fatigue curve. At the same time, to "improve motion control at shoulder joint end range," the system may match a "virtual reality trajectory tracking tool," which renders a fine path in three-dimensional space that the user needs to control with their arm, with a special emphasis on stability control at the end of shoulder joint abduction. The system will evaluate the combined effect of these two tools, confirming that they are complementary in function rather than overlapping, with endurance training providing a foundation for motion control, and motion control training improving the quality of endurance training and preventing compensation. The system may also add a "scapular stability activation tool" as an auxiliary tool, as the profile analysis suggests that compensatory shrugging may be related to insufficient scapular stability, thus forming a targeted and hierarchical tool combination.

[0041] Once the tool combination is selected, the most crucial step is to arrange the time series and intensity parameters of different training tools to construct a multi-day training support plan with internal logic. The time series arrangement follows the principles of supercompensation and periodic training, avoiding overtraining the same muscle group or pattern. The system may be designed with a "train one day, rest one day" or "alternating high and low intensity" rhythm. For example, during a 14-day training cycle, the system might schedule the following: Monday, moderate-intensity isokinetic abduction resistance training and high-repetition scapular stability activation training, focusing on building endurance; Tuesday, low-intensity virtual reality trajectory tracking training, emphasizing neuromuscular control and movement pattern re-education, allowing muscles to recover from Monday's load while practicing technique; Wednesday could be an active recovery day, with the system recommending some gentle stretching or training another body part; Thursday, isokinetic abduction resistance training again, but this time the resistance or repetitions might be slightly increased from the previous week's intensity to achieve progressive overload; Friday, virtual reality trajectory tracking training again, with appropriately increased path complexity or stability requirements. This cycle repeats, forming a wave-like intensity curve.

[0042] The setting of intensity parameters is fully personalized and equipped with dynamic adjustment rules. The system will preset initial intensity parameters for each training day and each tool called. For "isokinetic external rotation resistance training", the initial resistance value will be set according to Mr. Wang's performance in the last cycle of endurance atlas, for example, set to the load he can complete 12-15 times of repetition in standard posture. At the same time, the system will preset a strength progression rule, for example, "if the user can complete more than 15 repetitions in standard posture in the last training, increase the resistance by 5% in the next training". For "virtual reality track tracking training", the intensity parameter may be manifested as the width of the path (tolerance), the moving speed and the length of time that needs to be maintained, and these parameters are also based on historical performance data to set the baseline and contain advanced logic. All these decisions: goals, tools, time series, intensity parameters and their adjustment rules are integrated to generate a structured, cross-cycle personalized rehabilitation training plan. This plan may be presented to the user and therapist in the form of a calendar view or a structured document, clearly listing the training theme, the tools used, the preset intensity indicators and the expected training focus for each day in the future. The plan is not only an execution list, but also a living document containing a feedback loop. During subsequent execution, the system will continuously collect Mr. Wang's training data and compare it with the expected plan. According to his actual completion and physical response, the system will dynamically fine-tune the arrangement of subsequent dates, for example, automatically reduce the intensity of the next day when the user shows unexpected fatigue, or start higher-level training in advance when the progress exceeds expectations. This cross-cycle planning based on data-driven and continuous optimization makes rehabilitation training break away from the rigidity of fixed schemes and truly realizes dynamic personalized management synchronized with the user's recovery curve.

[0043] Example 6: The application can be further extended to a comprehensive rehabilitation scenario in the implementation process, where the user not only has physical function recovery needs, but also has cognitive or emotional improvement goals. For example, Mr. Zhang, a user recovering from the sequelae of a stroke, inputs his current state description information through a natural language interaction interface, including physical states such as "left arm lifting still has a pulling sensation" and cognitive states such as "recent memory and attention have declined somewhat." After the system receives these voice or text inputs, the semantic analysis module extracts key state feature vectors, including not only physical indicators such as joint range of motion limitations, but also subjective ratings of cognitive function, and matches these vectors with a pre-defined state classification model to determine the user's current state category as "upper limb motor dysfunction with mild cognitive impairment." Based on this identification, the context engineering engine retrieves Mr. Zhang's historical rehabilitation training records from the time series database, such as the success rate curve of left arm flexion and extension training in the past month, and retrieves his cognitive assessment report data from the document database, including digital span test results and subjective fatigue scores, and finally integrates them into a multi-dimensional context feature vector, covering the historical performance of both physical and cognitive dimensions.

[0044] The Agentic AI decision module performs fusion analysis on multi-source data, aligns the key state feature vectors with the multi-dimensional context feature vectors in space and time, ensuring the consistency of physical rehabilitation data and cognitive assessment data on the time axis. Through a weighted multi-objective optimization algorithm, the expected utility values of different rehabilitation actions are calculated, considering not only the improvement effect of physical training such as "shoulder joint abduction resistance training" on motor function, but also the potential of cognitive training such as "virtual environment attention tracking task" on cognitive ability, and generating a priority call sequence according to the utility value ranking, which may contain both physical and cognitive rehabilitation actions. When accessing the registered tool library through the MCP protocol server, the system sends a tool discovery request to the server according to the priority call sequence, receives the available tool list, and matches the optimal three-dimensional reconstruction training tool based on the function description; for example, for cognitive training needs, the tool library may register a "three-dimensional reconstruction attention strengthening tool", which simulates dynamic scenes in a virtual environment and requires users to track moving targets to train attention, with input parameters including difficulty level and scene complexity, and output responses including accuracy and reaction time data.

[0045] Based on the ability assessment results and preference settings of Mr. Zhang, the system dynamically generates personalized calling parameters; from the multi-dimensional context feature vector, it extracts historical performance data such as the average completion of left arm training and reaction time of cognitive tasks, calculates the current recommended difficulty level parameter, and combines the VR device configuration information used to generate adaptive resolution parameters and rendering quality parameters, ensuring the smoothness and immersion of the three-dimensional scene. When executing the three-dimensional reconstruction training tool with the configured parameters, send a calling request to the selected tool through the MCP protocol, and after the tool returns the training scene configuration data, the system loads the scene and starts the interactive training process; for example, Mr. Zhang is conducting left arm resistance training and attention tracking tasks at the same time, in the virtual environment, he needs to control the virtual arm to complete the action, while responding to the visual stimuli appearing in the environment. During the training process, the motion capture device collects body motion trajectory data, records interactive response time data in the three-dimensional scene, and synchronously collects physiological signal monitoring data such as heart rate variability, all real-time operation data are stored in the context engineering engine for updating the user's ability assessment portrait, and a new cognitive function dimension index is added.

[0046] According to the effect feedback data after the training is completed, the system adjusts the decision weight parameters of the AgenticAI decision module through the federated learning optimization algorithm; compares the real-time operation data with the expected training target, such as comparing the improvement of left arm activity and the attention improvement amplitude, generates local model update gradient based on the difference, and fuses the gradient with the global model parameters through the secure aggregation algorithm, optimizes the adaptability of the decision module to the compound rehabilitation needs. Based on the adjusted weight parameters and the updated ability assessment portrait, the system generates a cross-cycle personalized rehabilitation training plan; according to the portrait, determine the key improvement target of the next training cycle, such as prioritizing left arm endurance and consolidating attention stability, select multiple complementary three-dimensional reconstruction training tools from the tool library, arrange the time sequence and intensity parameters of physical training and cognitive training, form a multi-day training assistance plan, such as high-intensity joint activity training in the morning and low-load cognitive relaxation tasks in the evening, to ensure the synergy and sustainability of the training. This embodiment integrates multi-dimensional state recognition and tool collaborative calling, while maintaining the core framework of three-dimensional reconstruction rehabilitation training, it expands the breadth of personalized support, and is suitable for complex rehabilitation scenarios that require physical and mental comprehensive intervention.

[0047] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A three-dimensional reconstruction rehabilitation training assistance method based on Agentic AI and MCP protocol, characterized in that, Comprise: Obtain the current state description information of user input through natural language interaction interface; Based on the current state description information, query and obtain the user's historical rehabilitation training records and physiological index data through the context engineering engine; Utilize the AgenticAI decision module to fuse and analyze the current state description information, the historical rehabilitation training records and the physiological index data, and generate a decision scheme containing multiple candidate rehabilitation actions; Access the registered tool library through the MCP protocol server, and match the corresponding three-dimensional reconstruction training tools according to the candidate rehabilitation actions in the decision scheme; Based on the ability evaluation results and preference settings of the user, dynamically generate the personalized calling parameters of the three-dimensional reconstruction training tools; Execute the three-dimensional reconstruction training tools configured with the personalized calling parameters, and collect real-time operation data of the user during the training process; Store the real-time operation data through the context engineering engine, and update the user's ability evaluation portrait; According to the effect feedback data after training, adjust the decision weight parameters of the AgenticAI decision module through the federated learning optimization algorithm; Based on the adjusted decision weight parameters and the updated ability evaluation portrait, generate a cross-cycle personalized rehabilitation training plan.

2. The method of claim 1, wherein the method is based on an Agentic AI and MCP protocol. The current state description information of user input obtained through the natural language interaction interface comprises: Receive the functional state description input by the user through voice or text; Perform semantic analysis on the functional state description to extract a key state feature vector; Match the key state feature vector with a pre-defined state classification model to determine the user's current state category identifier.

3. The method of claim 2, wherein the method is based on an Agentic AI and MCP protocol. The context engineering engine queries and obtains the user's historical rehabilitation training records and physiological index data, which comprises: According to the state category identifier, retrieve the associated historical training effect data from the time series database; Obtain the historical evaluation report data matching the state category identifier from the document database; Integrate the historical training effect data and the historical evaluation report data into a multi-dimensional context feature vector.

4. The method of claim 3, wherein the method is based on an Agentic AI and MCP protocol. The fusion analysis by the AgenticAI decision module comprises: Perform spatio-temporal alignment processing on the key state feature vector and the multi-dimensional context feature vector; Calculate the expected utility value of different rehabilitation training actions through a weighted multi-objective optimization algorithm; Sort the candidate rehabilitation actions according to the expected utility value to form a priority calling sequence.

5. The method of claim 4, wherein the method is based on an Agentic AI and MCP protocol. Access the registered tool library through the MCP protocol server, which comprises: According to the rehabilitation action type in the priority calling sequence, send a tool discovery request to the MCP server; Receive the available tool list and its function description information returned by the MCP server; Based on the matching degree of the function description information and the rehabilitation action requirement, select the optimal three-dimensional reconstruction training tool.

6. The method of claim 5, wherein the method is based on an Agentic AI and MCP protocol. The dynamic generation of personalized calling parameters of three-dimensional reconstruction training tools comprises: Extract the user's historical performance data from the multi-dimensional context feature vector; Calculate the current recommended difficulty level parameter according to the historical performance data; Combine the user's device configuration information to generate adaptive resolution parameters and rendering quality parameters.

7. The method of claim 6, wherein the method is based on an Agentic AI and MCP protocol. The execution of the three-dimensional reconstruction training tool includes: Sending a call request and the personalized call parameters to the selected three-dimensional reconstruction training tool through the MCP protocol; Receiving the training scene configuration data returned by the three-dimensional reconstruction training tool; Loading the training scene configuration data and starting an interactive training process.

8. The method of claim 7, wherein the method is based on an Agentic AI and MCP protocol. The collection of real-time operation data of the user during the training process includes: Obtaining body movement trajectory data of the user through a motion capture device; Recording interactive response time data of the user in the three-dimensional scene; Collecting physiological signal monitoring data during the training process.

9. The method of claim 8, wherein the method is based on an Agentic AI and MCP protocol. The adjustment of the decision weight parameter through the federated learning optimization algorithm includes: Differential comparison analysis of the real-time operation data and the expected training target; Generating a local model update gradient based on the differential analysis result; Fusing the local model update gradient and the global model parameter through a secure aggregation algorithm.

10. The method of claim 9, wherein the method is based on an Agentic AI and MCP protocol. The generation of the cross-cycle personalized rehabilitation training plan includes: Determining a key improvement target for the next training cycle according to the updated ability assessment profile; Selecting multiple complementary training tools from the tool library based on the key improvement target, arranging the time sequence and intensity parameters of different training tools, and forming a multi-day training assistance scheme.

Images