An AI-driven entity robot user profile dynamic updating method
By using an AI-driven approach, we acquire and process local modal data from multiple robots, configure temporal and spatial parameters, calculate attenuation weights and constraint penalties, and utilize a reinforcement learning decision model to schedule physical interactions between the robots. This solves the problem of poor file correlation in multi-robot collaborative environments, achieves high-precision data correlation and consistency, and improves the accuracy and security of user files.
Patent Information
- Application Number
- CN202610756597.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-29
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-05-29
AI Technical Summary
In a multi-robot collaborative environment, the poor correlation and low consistency of files due to the heterogeneity of perception data affect the accuracy of user file records and cross-terminal consistency.
By using an AI-driven approach, local modal data of the target user is acquired, time and space parameters are configured, the data is stored in a memory pool, historical feature labels are traced, decay weights and constraint penalties are calculated, and a reinforcement learning decision model is used to schedule physical robots to perform physical interactions, thereby achieving dynamic updates of cross-modal fusion confidence.
It achieves high-precision data association and spatiotemporal calibration in multi-robot scenarios, ensuring the consistency of global user profiles, improving the accuracy of perception data and interactive experience, avoiding noise-induced overwriting, and ensuring data security.
Smart Images

Figure CN122615264B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of robot user profile updating technology, specifically to an AI-driven method for dynamically updating the user profiles of physical robots. Background Technology
[0002] Physical robots are automated intelligent devices with physical movement, multimodal environmental perception, and interaction capabilities. They are widely used in home and office spaces. Through sensors such as vision, hearing, and distance perception, they capture the surrounding environment and human behavior in real time. In a multi-robot collaborative environment, the digital file management process for target users refers to the use of multi-source modal data streams for real-time analysis, state feature extraction, and file value iteration to maintain an accurate digital record of the target user's long-term behavioral preferences and physiological state, supporting the robot's personalized services and collaborative scheduling. Currently, in multi-robot collaborative scenarios, each physical robot typically uses its own independent sensors to independently collect local modal data and attempt to correct the values of file nodes.
[0003] However, due to the significant heterogeneity of modal data acquired by different robots and the lack of effective cross-modal and cross-terminal correlation mechanisms, when multiple robots are sensing simultaneously, it is easy to produce different judgments or logical conflicts regarding the state of the same target user, affecting the overall accuracy of user profile records and cross-terminal consistency. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an AI-driven method for dynamically updating user profiles of physical robots, solving the problems of poor profile correlation and low consistency caused by heterogeneous perception data in multi-robot collaborative environments.
[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-driven method for dynamically updating user profiles of physical robots, applied to a target scenario containing multiple physical robots, comprising: S1. Obtain the local modal data of the target user and extract the corresponding basic feature labels. The local modal data includes: visual data, auditory data, and distance perception data. S2. Configure time and space parameters for the basic feature labels and store them in the memory pool; S3. In response to the newly written basic feature tags in the memory pool, based on the time parameters and spatial parameters, trace and extract the historical basic feature tags belonging to the target user, and extract the benchmark constraint file; S4. Calculate the attenuation weight based on the time difference between the new and old basic feature labels, generate a constraint penalty term based on the deviation between the newly written basic feature labels and the benchmark constraint file, and calculate the cross-modal fusion confidence level in combination with the modality type. S5. If the cross-modal fusion confidence is within the detection trigger range, then the candidate entity robot is scheduled to perform physical interaction actions based on the active detection strategy, and supplementary feature labels are extracted to update the cross-modal fusion confidence. S6. When the cross-modal fusion confidence meets the file update conditions, determine the dynamic overwrite weight and write the data after fusing the new and old basic feature tags into the global user file of the target user.
[0006] Preferably, in step S2, configuring temporal and spatial parameters for the basic feature labels and storing them in a memory pool includes: Record the timestamp of the moment when the corresponding physical robot acquires the local modal data, and use the timestamp as the time parameter; Extract the three-dimensional global topological coordinates when the corresponding physical robot collects the local modal data, and the field of view direction vector of the corresponding sensing sensor on the physical robot, and combine the three-dimensional global topological coordinates and the field of view direction vector to form the spatial parameters; The basic feature labels are bound to the corresponding time parameters and spatial parameters using data structures to generate a multidimensional feature data package, and the multidimensional feature data package is stored in the memory pool. The timestamp is based on the Network Time Protocol to achieve clock synchronization among the multiple physical robots, and the three-dimensional global topological coordinates are obtained based on a fusion system of ultra-wideband positioning, real-time positioning and map construction.
[0007] Preferably, in step S3, based on the time and space parameters, historical basic feature tags belonging to the same target user are extracted, including: Extract the target timestamp and the target's three-dimensional global topological coordinates corresponding to the newly written basic feature labels; The target timestamp, the target three-dimensional global topological coordinates, and the time and spatial parameters of the candidate historical tags in the memory pool are used as input items and input into the spatial trajectory continuity determination model. If the spatial trajectory continuity determination model calculates that the physical motion trajectory of the target user between the target timestamp and the time parameters corresponding to the candidate historical tags satisfies the preset spatial continuity distribution characteristics, then the candidate historical tags are determined to belong to the target user and are extracted as the historical basic feature tags. The preset spatial continuity distribution feature includes: the estimated movement speed of the target user between the target timestamp and the corresponding time parameter of the candidate historical tag is within the preset reasonable human movement speed threshold range.
[0008] Preferably, the extraction of the baseline constraint file in S3 includes: The frequency and duration of each of the multiple physical robots within the target scene interacting with the target user in the past are statistically analyzed. Based on the frequency of historical valid interactions and the duration of the interactions, the master node entity robot with the highest historical confidence is selected from the plurality of entity robots. The profile data, which contains the long-term deterministic state preferences of the target user and is constructed locally by the master node entity robot based on long-term interaction data, is extracted as the benchmark constraint profile.
[0009] Preferably, in step S4, calculating the attenuation weight based on the time difference between the old and new basic feature labels includes: Extract the content attribute categories represented by the historical basic feature tags, and configure corresponding state volatility indices for different content attribute categories; The time difference and the state variability index are input into the decay calculation model as common inputs, and the decay weight is output. The attenuation weight is inversely mapped to both the time difference and the state variability index.
[0010] Preferably, a constraint penalty term is generated based on the deviation between the newly written basic feature labels and the baseline constraint file, and a cross-modal fusion confidence score is calculated in conjunction with the modality type, including: The state value represented by the newly written basic feature label is compared with the value of the deterministic archive node of the corresponding dimension in the benchmark constraint archive to calculate the state deviation. If the deviation of the state is greater than the preset constraint threshold, a negative adjustment value is generated as the constraint penalty item; Determine whether the modal type of the new and old basic feature labels belongs to positive interactive complementarity modality. If so, generate a positive complementary gain value. The cross-modal fusion confidence score is calculated by weighting and summing the attenuation weight, the constraint penalty term, and the complementary gain value.
[0011] Preferably, in step S5, scheduling candidate entity robots to perform physical interaction actions based on the active detection strategy includes: Obtain the physical movement energy consumption data of each candidate physical robot among the multiple physical robots, as well as the disturbance risk index corresponding to the execution of the detection action, and use the physical movement energy consumption data and the disturbance risk index as penalty parameters; Calculate the expected probability of obtaining the missing supplementary modal data, and use the expected probability as a revenue parameter; The penalty term parameter and the benefit term parameter are input into the reinforcement learning decision model, and the reinforcement learning decision model outputs physical interaction action instructions assigned to specific candidate entity robots. The reward function of the reinforcement learning decision model is positively correlated with the expected probability and negatively correlated with the physical mobility energy consumption data and the disturbance risk index.
[0012] Preferably, the physical interaction action instructions output by the reinforcement learning decision model and assigned to a specific candidate entity robot include: Extract the current physiological or environmental state data of the target user recorded in the baseline constraint file and input it as boundary constraint conditions into the reinforcement learning decision model; Based on the boundary constraints, the reinforcement learning decision model removes intrusive interaction actions with state conflicts from the preset set of physical interaction instructions, and outputs the physical interaction action instructions from the remaining interaction instructions.
[0013] Preferably, in step S6, determining the dynamic overwrite weight based on the cross-modal fusion confidence, and then using the dynamic overwrite weight to fuse the new and old basic feature labels and write them into the global user profile of the target user, includes: The cross-modal fusion confidence score is converted into a learning rate parameter calculated by moving average, and the learning rate parameter is set as the dynamic overwrite weight. Using the dynamic overwrite weight, a moving average calculation is performed on the feature values represented by the new basic feature labels and the historical basic feature labels to generate fused feature data with smooth transition characteristics. Using the fused feature data, a non-mutational numerical overwrite update is performed on the corresponding node in the global user profile of the target user.
[0014] Preferably, when the cross-modal fusion confidence fails to meet the file update conditions, the following fault-tolerant correction process is performed: Extract the newly written basic feature labels that fail to meet the target due to the negative adjustment of the constraint penalty term, and store them in a temporary queue; The cumulative truncation frequency of basic feature labels representing the same feature value in the temporary queue is counted within a sliding time window of fixed length. When it is determined that the cumulative truncation frequency reaches the preset wall breaking threshold, the restriction on the reduction of the cross-modal fusion confidence by the constraint penalty term is lifted to trigger the numerical overwrite update, and a reverse feature correction instruction is generated simultaneously and sent to the master node entity robot that provides the benchmark constraint file to overwrite the corresponding expired file node. The multiple entity robots adopt a federated learning architecture. The original stream of local modal data acquired by each entity robot is kept locally, and only the extracted basic feature labels and their corresponding spatiotemporal parameters are uploaded to the memory pool.
[0015] This invention provides an AI-driven method for dynamically updating user profiles of physical robots. It offers the following advantages: 1. This invention calculates cross-modal fusion confidence by weighted summation attenuation weights, constraint penalty terms, and complementary gain values. Combined with historical feature tracing and benchmark constraints, it achieves high-precision association and spatiotemporal calibration of heterogeneous data in multi-robot scenarios, ensuring the dynamic consistency of global user profiles under multi-device collaboration and improving the accuracy of perceived data.
[0016] 2. This invention utilizes a reinforcement learning decision model to schedule entity robots to actively explore. By using benchmark constraints as boundary conditions, intrusive actions are eliminated. Under the premise of ensuring user experience, missing modalities are actively acquired, effectively overcoming the recognition bottleneck caused by data deficiency, realizing a closed loop of perception and decision-making, and significantly improving the perception efficiency and interactive experience of human-machine collaboration.
[0017] 3. This invention is based on a fault-tolerant process of temporary queue, truncation frequency statistics and reverse feature correction. Combined with a federated learning architecture, it removes the downsizing restriction and updates the file through the barrier threshold logic. This avoids accidental overwriting caused by noise and protects the privacy of local modal data, thereby achieving long-term accuracy and data security of user profiles in a multi-machine environment. Attached Figure Description
[0018] Figure 1 This is a flowchart of an AI-driven method for dynamically updating user profiles of physical robots according to the present invention. Figure 2 This is a flowchart illustrating the process of generating physical interaction action commands based on a reinforcement learning decision model according to the present invention. Figure 3 This is a schematic diagram of the fault tolerance correction and global user profile update process based on the cell wall breaking threshold of the present invention. Detailed Implementation
[0019] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] Please see the appendix Figure 1 This invention provides an AI-driven method for dynamically updating user profiles of physical robots, applicable to a target scenario containing multiple physical robots, including: S1. Obtain the local modal data of the target user and extract the corresponding basic feature labels. The local modal data includes: visual data, auditory data, and distance perception data. Specifically, step S1 is executed by various physical robots within the target scene. These robots utilize their heterogeneous sensing hardware, such as RGB cameras, microphone arrays, infrared or ToF distance sensors, to acquire local modal data of the target user in real time. This local modal data includes visual data, auditory data, and distance perception data. Subsequently, the physical robots can call locally deployed feature extraction algorithms or lightweight pre-trained models, such as object detection networks or acoustic feature extraction models, to perform semantic parsing and dimensionality reduction on the aforementioned unstructured local modal data, thereby extracting the corresponding basic feature labels. In one embodiment, a robot vacuum cleaner patrolling the living room captures the outline of a user lying on the sofa (visual data) through its camera, combines it with the steady breathing sounds collected by the microphone (auditory data) and the physical distance between the user and the robot measured by the ToF sensor (distance perception data), and after comprehensive analysis by the local model, extracts and outputs the basic feature label "current posture: lying flat," thus completing the conversion of local modal data into structured feature labels.
[0021] S2. Configure the time and space parameters for the basic feature labels and store them in the memory pool; Furthermore, in S2, temporal and spatial parameters are configured for the basic feature labels and stored in the memory pool, including: Record the timestamp of the moment when the corresponding physical robot acquires local modal data, and use the timestamp as a time parameter; Extract the three-dimensional global topological coordinates when the corresponding physical robot collects local modal data, as well as the field of view direction vector of the corresponding sensor on the physical robot, and combine the three-dimensional global topological coordinates and the field of view direction vector into spatial parameters; The basic feature labels are bound to the corresponding time and space parameters in a data structure to generate a multidimensional feature data package, which is then stored in a memory pool. The timestamp is based on the network time protocol to achieve clock synchronization among multiple physical robots, and the three-dimensional global topological coordinates are obtained based on a fusion system of ultra-wideband positioning, real-time positioning and map building.
[0022] Specifically, after the structured extraction of local modal data is completed, a unified time reference benchmark needs to be assigned to the generated basic feature labels. Specifically, the timestamp of the moment when the corresponding physical robot obtains the local modal data is recorded and used as a time parameter. As a preferred implementation method, the timestamp is based on the network time protocol to realize the underlying clock synchronization between multiple physical robots in the target scene. For example, when a robot vacuum cleaner extracts a label representing a lying posture in the living room, it will record a specific millisecond-level timestamp value. That is, when there are multiple heterogeneous devices such as robot vacuum cleaners and companion robots in the scene that report perception data alternately or simultaneously, it can effectively reduce the timing error caused by the drift of their respective independent hardware clocks and provide an accurate time judgment basis for subsequent judgment of the evolution relationship of user behavior status. In addition to recording the time attribute, it is also necessary to determine the corresponding physical spatial location information for this perception action. Specifically, the three-dimensional global topological coordinates when the corresponding physical robot collects the local modal data are extracted, as well as the field of view direction vector of the corresponding visual or ranging perception sensor, and the two are combined into spatial parameters. In a preferred embodiment, the three-dimensional global topological coordinates are obtained based on the positioning steps that integrate ultra-wideband positioning technology with real-time positioning and mapping technology. Continuing with the aforementioned scenario, the robot vacuum cleaner is recorded as being at a specific living room coordinate point defined by three-dimensional values, and its sensor orientation vector is simultaneously recorded as pointing to the sofa area. This overcomes the deficiency of conventional recommendation methods in lacking spatial orientation perception capabilities, clarifies the three-dimensional position of the physical robot and the actual detection coverage area of its sensors, and provides objective physical judgment conditions for subsequent identification of different users due to unreasonable spatial span. After acquiring the corresponding time and space parameters, the data structure encapsulation and storage actions are then performed. Through data processing methods such as serialization conversion, the extracted basic feature labels are bound to their corresponding time and space parameters to generate multi-dimensional feature data packets. These multi-dimensional feature data packets are then transmitted to a preset memory pool for caching or persistent storage. For example, the aforementioned lying posture label, the corresponding timestamp value, and the spatial parameters containing coordinates and orientation are packaged into a standard data frame and written into the memory pool. This storage step transforms the single perception results of each physical robot into structured historical data with time and space attributes. This allows for conditional retrieval and reuse of this historical data in subsequent steps based on specific time slices or spatial topological ranges, thus providing the necessary data support foundation for the next step of tracing and verifying the continuity of spatial trajectories.
[0023] S3. In response to newly written basic feature tags in the memory pool, based on time and space parameters, trace and extract historical basic feature tags belonging to the same target user, and extract the baseline constraint file; Furthermore, in S3, based on time and spatial parameters, historical basic feature tags belonging to the same target user are extracted, including: Extract the target timestamp corresponding to the newly written basic feature labels, as well as the target's three-dimensional global topological coordinates; The target timestamp, the target's three-dimensional global topological coordinates, and the temporal and spatial parameters of the candidate historical tags in the memory pool are used as inputs into the spatial trajectory continuity determination model. If the spatial trajectory continuity determination model calculates that the physical motion trajectory of the target user between the target timestamp and the time parameters corresponding to the candidate historical tags satisfies the preset spatial continuity distribution characteristics, then the candidate historical tags are determined to belong to the target user and are extracted as historical basic feature tags. Among them, the preset spatial continuity distribution characteristics include: the estimated movement speed of the target user between the target timestamp and the corresponding time parameters of the candidate historical tags is within the preset reasonable human movement speed threshold range.
[0024] Specifically, in response to the newly written basic feature labels in the memory pool, it is necessary to perform attribute separation on the newly written data node, extract the target timestamp corresponding to the newly written basic feature label, and the target three-dimensional global topological coordinates. Extracting the target timestamp and target three-dimensional global topological coordinates establishes a benchmark physical anchor point for subsequent spatiotemporal continuity calculations, ensuring that the backtracking and matching of historical data is based on a clear absolute spatial and temporal starting point. After obtaining the current time and spatial location parameters, the extracted target timestamp and target three-dimensional global topological coordinates, together with the time and spatial parameters attached to the candidate historical tags to be screened in the memory pool, are used as input items and input into the pre-built spatial trajectory continuity determination model. The above-mentioned multi-dimensional parameter input mechanism transforms the feature matching verification logic from simply relying on account identifiers or visual facial feature comparison to a cross-validation mechanism based on three-dimensional physical space and time dimensions, providing underlying verification data support for determining whether discrete multimodal data belong to the same physical entity. Upon receiving the above input, the spatial trajectory continuity determination model executes physical verification logic to calculate the physical movement trajectory of the target user between the target timestamp and the corresponding time parameters of the candidate historical tags. This determines whether the trajectory meets the preset spatial continuity distribution characteristics. Specifically, the preset spatial continuity distribution characteristics require the target user to estimate their movement speed between two time points, ensuring it falls within a preset threshold range for reasonable human movement speed. The following mathematical formula can be used to estimate the physical movement speed and determine the conditional constraints: ; ; in, This represents the estimated movement speed of the target user between two modal data collection sessions. , , The target's three-dimensional global topological coordinates represent the newly written basic feature labels. , , Represents the three-dimensional global topological coordinates corresponding to the candidate historical labels. and These represent the time parameters for the target timestamp and the candidate historical tags, respectively. This represents the preset upper limit threshold for reasonable human movement speed. When the calculated speed... If the above threshold constraints are met, the candidate historical tag is determined to belong to the target user and is extracted as a historical basic feature tag. In a specific application scenario, suppose a robot companion in the bedroom generates a candidate historical tag indicating it is reading at 10:00:00, along with the bedroom's specific 3D coordinates. Then, at 10:00:03, a robot vacuum cleaner in the living room generates a newly written basic feature tag indicating it is drinking water, along with the living room's 3D coordinates. The coordinates and timestamps of both are input into a spatial trajectory continuity determination model. The model then uses a formula to calculate the estimated movement speed if the physical straight-line distance between the two coordinate points in the bedroom and living room is 15 meters and the time difference is 3 seconds. The speed is 5 meters per second, assuming a preset threshold for reasonable human movement speed. The speed was set to 8 meters per second. Since the calculated speed of 5 meters per second did not exceed the preset threshold for reasonable human movement speed, the judgment model confirmed that the two actions conformed to the spatial continuity distribution characteristics. Therefore, it was determined that the reading status data obtained in the bedroom and the drinking status data in the living room belonged to the same physical user, and the bedroom data was extracted as the historical basic feature label. The above mathematical quantification calculation mechanism based on biophysical limits objectively eliminated the data misassociation caused by different people appearing in physically adjacent areas within a very short period of time.
[0025] Furthermore, the baseline constraint profile is extracted in S3, including: The frequency and duration of historical effective interactions between each of the multiple physical robots within the target scenario and the target user are statistically analyzed. Based on the frequency and duration of historical effective interactions, the master node robot with the highest historical confidence is selected from multiple physical robots. The profile data, which contains the long-term deterministic state preferences of the target user and is constructed locally by the master node entity robot based on long-term interaction data, is extracted as the benchmark constraint profile.
[0026] Specifically, after completing the tracing and extraction of historical basic feature tags, the next step is to extract the benchmark constraint archive. First, the frequency and duration of historical effective interactions between the multiple physical robots in the target scene and the target user are statistically analyzed. Specifically, for physical robots of different forms deployed in the scene, their underlying historical interaction logs are retrieved, and the total number of effective command interactions between each physical robot and the target user in the set target scene is accumulated and calculated, as well as the total cumulative duration of state following, voice dialogue or physical contact in each interaction. The above statistical process aims to quantitatively evaluate the familiarity of each physical robot in the scene with the target user's behavioral habits and physiological state through objective historical interaction behavior data, and provide a quantitative basis for establishing the priority of data confidence in the future. Subsequently, based on the statistically obtained historical effective interaction frequency and duration, the master node robot with the highest historical confidence score is selected from multiple entity robots. In the specific implementation, by configuring corresponding weight coefficients for interaction frequency and interaction duration, a comprehensive historical confidence score for each entity robot is calculated. The specific calculation uses the following numerical conversion formula: ; in, Represents historical confidence score. Represents the frequency of effective interactions in history. Represents the duration of the interaction. and These represent the preset frequency weighting coefficient and duration weighting coefficient, respectively. After calculation, the physical robot with the highest score in the scene is determined as the main node physical robot. For example, if the companion robot that stays in the bedroom for a long time to accompany the user has a higher comprehensive score in terms of command response frequency and eye-following duration than the sweeping robot that only performs periodic cleaning tasks in the living room, then the companion robot is determined as the main node. This step clarifies the data trust level in the multi-machine collaborative environment from the perspective of physical devices, ensuring that the comparison benchmark used to constrain and verify new data has the highest data reliability. Finally, the profile data of the main node robot, which is constructed based on long-term interaction data and contains the long-term deterministic state preferences of the target user, is extracted as the benchmark constraint profile. Specifically, the main node robot has constructed a highly deterministic local state profile node based on the high-frequency behavior patterns and stable preferences of the target user during its long-term local operation and interaction. This local state profile data is directly extracted and elevated to the benchmark constraint data source for global verification. This transforms the target user behavior characteristics in the scene into single-machine local data and into a reference boundary in the global dynamic update process. This provides an objective and highly confident comparison benchmark for the subsequent calculation of the deviation of newly written feature labels and the generation of constraint penalty terms.
[0027] S4. Calculate the attenuation weight based on the time difference between the new and old basic feature labels, generate a constraint penalty term based on the deviation between the newly written basic feature labels and the benchmark constraint file, and calculate the cross-modal fusion confidence based on the modal type. Furthermore, in S4, the attenuation weight is calculated based on the time difference between the old and new basic feature labels, including: Extract the content attribute categories represented by historical basic feature labels, and configure corresponding state variability indices for different content attribute categories; The time difference and the state variability index are input into the decay calculation model as common inputs, and the decay weight is output. Among them, the decay weight has an inverse mapping relationship with both the time difference and the state variability index.
[0028] Specifically, in the step of calculating the decay weight based on the time difference between the old and new basic feature labels, the content attribute categories of the historical basic feature label table are first extracted, and corresponding state variability indices are configured for different content attribute categories. Specifically, the business semantics contained in the traced historical basic feature labels are parsed and classified into specific content attribute categories such as instantaneous physiological signs, short-term behavioral states, or long-term biological characteristics. Subsequently, a predefined state variability mapping relationship table is queried to assign corresponding numerical state variability indices to the extracted content attribute categories. The predefined state variability mapping table is a numerical lookup table pre-embedded in the underlying storage medium. This mapping table is pre-constructed based on the objective evolution law of human physiological characteristics and behavioral habits. It divides various feature data into different change frequency levels and binds a specific decay constant value to each level. For example, in the mapping table, the high variability index value corresponding to the category of "heart rate / blood pressure" which represents instantaneous changes is set to 0.9, the medium variability index value corresponding to the category of "sleep posture" which represents short-term movements is set to 0.4, and the low variability index value corresponding to the category of "skeletal gait" which represents long-term stability is set to only 0.05. After extracting the content attribute category, the mapping table can be directly compared to convert the qualitative feature category into the quantitative constant required for subsequent index calculation. The operation of configuring the state variability index objectively reflects the natural loss law of different physiological or physical states in the time dimension. After configuring the index, the calculated time difference and state variability index are input into the decay calculation model as common inputs. The model outputs decay weights. In the specific calculation process, the decay calculation model uses an exponential decay function with the state variability index as the adjustment coefficient to quantify the degree of loss of timeliness of historical data. The decay weights are calculated using the following mathematical formula: ; in, The attenuation weight represents the output. Represents the natural constant. This represents the state volatility index configured after querying the mapping relationship table for the corresponding content attribute category. The time parameter represents the time at which new basic feature labels are written. The time parameter representing the historical basic feature label. By characterizing the time difference, this mathematical formula is used to algebraically fuse the time span with the volatility characteristics of the state itself, providing a quantitative decay benchmark for subsequent confidence fusion of cross-modal data. The attenuation calculation model's attenuation weights exhibit an inverse mapping relationship with both the time difference and the state variability index. Specifically, the greater the time difference between the old and new basic feature labels, the larger the absolute value of the exponential term in the calculation formula, resulting in a smaller output attenuation weight. Similarly, the higher the state variability index corresponding to the extracted content attribute category, the larger the product term becomes, and the smaller the output attenuation weight. Continuing with the aforementioned specific numerical scenario, under the condition that the time difference is two hours, after substituting into the above formula for calculation... The decay weight of the instantaneous heart rate label output with a value of 0.9 will drop sharply, approaching zero, while Even a skeletal gait label with a value of 0.05 can still maintain a very high decay weight. The rules for establishing a reverse mapping relationship can accelerate the expiration process of high-frequency variable data and ensure the long-term retention of low-frequency stable data, preventing expired and volatile features from interfering with the cross-modal determination of the target user's current real state.
[0029] Furthermore, constraint penalty terms are generated based on the deviation between the newly written basic feature labels and the baseline constraint file, and cross-modal fusion confidence is calculated in conjunction with the modality type, including: The state values represented by the newly written basic feature labels are compared with the values of the deterministic archive nodes of the corresponding dimensions in the baseline constraint archive to calculate the state deviation. If the deviation from the state is greater than the preset constraint threshold, a negative adjustment value is generated as a constraint penalty term. Determine whether the modal type of the new and old basic feature labels belongs to positive interactive complementarity modality. If so, generate positive complementary gain values. The cross-modal fusion confidence score is calculated by weighting and summing the attenuation weights, constraint penalty terms, and complementary gain values.
[0030] Specifically, in the implementation process, it is first necessary to generate constraint penalty terms based on the deviation between the newly written basic feature labels and the baseline constraint file. Specifically, the state values represented by the newly written basic feature labels are extracted and compared with the values of the deterministic file nodes of the corresponding dimensions in the baseline constraint file to calculate the state deviation. If the state deviation exceeds the preset constraint threshold, it indicates that there is a significant conflict between the new data and the target user's long-term habit file. At this time, a corresponding negative adjustment value is generated as a constraint penalty term, which quantifies the inconsistency between the newly collected data and the baseline file and objectively reflects the risk indicators of data flow. The relevant calculation formula is as follows: ; ; in, This represents the degree of deviation from the state. This represents the state value of the newly written basic feature label. This represents the file node value for the corresponding dimension in the baseline constraint file. Represents constraint and penalty items. This represents the preset constraint threshold. This represents the penalty amplification factor; While calculating the constraint penalty term, it is determined whether the modal type of the new and old basic feature labels belongs to positive interactive complementarity modality. If the determination result is yes, a positive complementary gain value is generated. Positive interactive complementarity modality refers to the significant verification or enhancement effect of two different modal data (such as physiological pulse and EEG signal) in representing the same physiological event. That is, through the complementary characteristics between modalities, the accuracy of user behavior judgment is improved. If it is determined to belong to complementary modality, a positive gain value is obtained by looking up a table or calculating the model to offset part of the confidence loss caused by occasional data noise. Finally, the attenuation weights, constraint penalty terms, and complementary gain values are weighted and summed to calculate the cross-modal fusion confidence score. The weighted summation logic ensures that the confidence score calculation comprehensively considers the attenuation timeliness of historical data, logical consistency with the baseline archive, and the collaborative effectiveness between modalities. The specific calculation formula is as follows: ; in, This represents the final calculated cross-modal fusion confidence level. This represents the previously calculated attenuation weight. Represents constraint and penalty items. Represents the positive complementary gain value. , , These are the weighting coefficients for the corresponding factors; In a specific application example, suppose the target user's baseline constraint profile records a "resting heart rate" of 60 beats / min, and the newly written "auditory data" represents a heart rate of 120 beats / min, with a state deviation of... The value is 60, which exceeds the preset threshold. (Assuming the value is 20), a negative constraint penalty term is generated, significantly reducing the confidence level. However, if "physiological pulse data" is simultaneously collected and determined to be a positive complementary modality with the auditory data, a positive complementary gain value is generated. Ultimately, even if the heart rate deviates from the baseline, the intervention of the gain value will maintain a moderately high confidence level, thus enabling the subsequent update process to be initiated cautiously, rather than discarding data directly due to deviation in a single dimension. This ensures that the cross-modal fusion logic can maintain a dynamic balance when facing sudden changes in user status, taking into account both the stability of historical baselines and the sensitivity to new sudden features.
[0031] S5. If the cross-modal fusion confidence is within the detection trigger range, then the candidate entity robot is scheduled to perform physical interaction actions based on the active detection strategy, and supplementary feature labels are extracted to update the cross-modal fusion confidence. Furthermore, in S5, candidate entity robots are scheduled to perform physical interaction actions based on an active detection strategy, including: The physical movement energy consumption data of each candidate physical robot among multiple physical robots, as well as the disturbance risk index corresponding to the execution of detection actions, are obtained, and the physical movement energy consumption data and the disturbance risk index are used as penalty parameters. Calculate the expected probability of obtaining the missing supplementary modal data, and use the expected probability as a parameter of the revenue term; The penalty and benefit parameters are input into the reinforcement learning decision model, which then outputs physical interaction action instructions assigned to specific candidate robot entities. Among them, the reward function of the reinforcement learning decision model is positively correlated with the expected probability and negatively correlated with the physical movement energy consumption data and the disturbance risk index.
[0032] Specifically, if the cross-modal fusion confidence is within the detection trigger range, it indicates that the current multimodal data fusion result has logical ambiguity or insufficient confidence. At this time, an active detection strategy is initiated, and candidate entity robots are scheduled to perform physical interaction actions and obtain supplementary feature labels to update the cross-modal fusion confidence. The detection trigger range is preset to the critical range of the confidence value. The setting of this range is intended to avoid blindly performing physical interactions due to poor data quality, and to ensure that the physical spatial mobility of entity robots is used only when necessary to obtain key supplementary information, thereby realizing the transformation from passive data fusion to active perception and decision-making. In the process of scheduling candidate physical robots, the cost and benefit of physical interaction actions are first quantified. Physical movement energy consumption data of multiple physical robots in the target scene when performing specific physical actions (such as moving to the side of the user, adjusting the sensor angle, and making prompting sounds) are obtained, as well as the disturbance risk index of the action to the target user. The two are integrated into a penalty parameter. At the same time, based on the currently acquired feature data, the expected probability that each candidate physical robot can obtain missing supplementary modal data through physical interaction is predicted, and this expected probability is used as a benefit parameter. The acquisition of the above parameters makes the action decision-making process change from a single distance-first logic to a comprehensive optimization decision that takes into account energy consumption, user experience, and perception gain. The penalty and reward parameters mentioned above are used as inputs to a pre-trained reinforcement learning decision model. The model calculates and outputs the optimal physical interaction action command assigned to a specific candidate robot. The reward function of this reinforcement learning decision model is defined as follows: ; in, This represents the reward value output by the model. This represents the expected probability of obtaining supplemental modal data. Represents energy consumption data for physical movement. Represents the risk index of disturbing the peace. The weighting coefficient representing the expected probability of return. The penalty weight coefficient represents the energy consumption of physical movement. The penalty weighting coefficient representing the risk of disturbing the peace is used to maximize the reward value. Reinforcement learning decision models can output interactive action commands that maximize perceived benefits while minimizing energy consumption and user interference. In a specific application example, suppose the target user's confidence level is within the detection trigger range. At this time, there are two candidate physical robots in the living room: a companion robot that is closer and a sweeping robot that is farther away. If the companion robot only requires very low energy to move ( (Low risk) and no risk of disturbing residents ( (lower), while the movement of a robotic vacuum cleaner involves a large-scale path planning ( (Higher), and it is expected that the companion robot will be able to clearly obtain the missing visual features after adjusting its position ( If the value is relatively high, the reinforcement learning decision model will issue a physical interaction command to the escort robot to "move to the side of the user and adjust the camera angle" by calculating the reward function. After the escort robot executes the command, it successfully obtains the user's facial supplementary feature label, updates the cross-modal fusion confidence based on this, and removes it from the detection trigger range, thereby completing the confirmation of the user's true state.
[0033] Please see the appendix Figure 3 Furthermore, the reinforcement learning decision model outputs physical interaction action instructions assigned to specific candidate entity robots, including: Extract the target user's current physiological or environmental state data recorded in the baseline constraint file and input it as boundary constraint conditions into the reinforcement learning decision model; The reinforcement learning decision model, based on boundary constraints, removes intrusive interaction actions with state conflicts from the preset set of physical interaction instructions, and outputs physical interaction action instructions from the remaining interaction instructions.
[0034] Specifically, before the reinforcement learning decision model outputs physical interaction action instructions, the current physiological or environmental state data of the target user recorded in the baseline constraint file is first extracted and used as boundary constraint conditions to input into the reinforcement learning decision model. This data covers multiple dimensions of parameters such as the target user's current real-time heart rate, sleep quality, environmental noise decibels, or activity level. These data are used as environmental constraint boundaries in the reinforcement learning process to ensure that the subsequently generated interaction actions are logically consistent with the target user's current physiological and behavioral taboos, thereby ensuring the safety and suitability of the physical robot performing physical interaction actions from the source. Based on the aforementioned boundary constraints, the reinforcement learning decision model removes intrusive interaction actions with state conflicts from the preset set of physical interaction instructions, and outputs physical interaction action instructions from the remaining instructions. Specifically, the reinforcement learning decision model evaluates the state consistency of each interaction action in the instruction set. If the physical feedback caused by the action conflicts with the current user state (such as disturbing rest or causing fright), the instruction is directly marked as invalid and removed from the candidate set. This filtering logic is defined by the following mathematical expression: ; in, This represents the set of valid interactive instructions remaining after conflict removal. This represents the complete set of preset physical interaction instructions. Represents specific interactive actions, This represents data indicating the current physiological or environmental state of the user. This represents a compatibility matching function between interactive actions and user states. This represents a preset state safety threshold. Through this screening process, physically reasonable and psychologically non-intrusive action instructions can be accurately filtered from the candidate strategy space. In specific application scenarios, when the target user is in a deep sleep state, the state data in the baseline constraint file is "deep sleep". If the candidate interaction instruction set includes two actions, "loud voice inquiry" and "soft light projection", the reinforcement learning decision model will use "deep sleep" as the boundary constraint input when performing the selection. Through the aforementioned compatibility matching function, it is found that the compatibility value of the "loud voice inquiry" action is... Below the threshold (i.e., those judged to have serious state conflicts) are defined as intrusive interaction actions and eliminated, while the compatibility value of "gentle light projection" meets the requirements and is finally output as an execution command. This process ensures that the physical robot can automatically select the lightest and least disruptive interaction method within the time window when the user's physiological state is highly sensitive, significantly improving the comfort and safety of the human-machine co-existence experience.
[0035] S6. When the cross-modal fusion confidence meets the file update conditions, determine the dynamic overwrite weight and write the data after fusing the old and new basic feature labels into the target user's global user file.
[0036] Please see the appendix Figure 3 Furthermore, in S6, dynamic overwrite weights are determined based on cross-modal fusion confidence. These weights are then used to fuse the old and new basic feature labels and write them into the target user's global user profile, including: The cross-modal fusion confidence score is converted into a learning rate parameter calculated using a moving average, and the learning rate parameter is set as a dynamic overwrite weight. By employing dynamic overwrite weights, a moving average calculation is performed on the feature values represented by the new basic feature labels and the historical basic feature labels to generate fused feature data with smooth transition characteristics. By using fused feature data, a non-mutational numerical overwrite update is performed on the corresponding node in the global user profile of the target user.
[0037] Specifically, when the cross-modal fusion confidence meets the file update conditions, the dynamic overwrite weight is determined. The cross-modal fusion confidence value is converted into a learning rate parameter calculated by moving average, and this learning rate parameter is set as the dynamic overwrite weight. The value of the dynamic overwrite weight represents the correction strength of the newly acquired feature data to the existing values in the global file. The larger the value, the higher the contribution weight of the new data in this feature dimension, thereby ensuring the dynamic adaptability and sensitivity of the file update. Dynamic overwrite weights are used to perform a moving average calculation on the feature values represented by the new basic feature labels and the historical basic feature labels, generating fused feature data with smooth transition characteristics. This moving average calculation logic eliminates the interference caused by sudden noise by weighted fusion of historical data and current data. The specific mathematical calculation formula is as follows: ; in, Represents the generated fused feature data, Represents dynamic overwrite weights. The feature values represented by the historical basic feature labels The feature values represented by the new basic feature labels are weighted and fused to achieve a smooth correction of the existing archive values, avoiding drastic numerical changes in the archives caused by a single abnormal data. By using fused feature data, a non-mutational numerical overwrite update is performed on the corresponding nodes in the global user profile of the target user. The calculated fused feature data is directly written back to the corresponding data field of the global user profile to complete the rolling update of the data nodes. This non-mutational numerical overwrite mechanism ensures the stability of the user profile in the long-term evolution process, reflecting both the real-time dynamic changes of user behavior preferences and maintaining the continuity of numerical evolution. For example, if the target user's global user profile records a "resting heart rate" of 70 beats / min, and the newly collected baseline feature label value is 80 beats / min, and the cross-modal fusion confidence level is determined to be 0.8, then the dynamic overwrite weight is set accordingly. The fused feature data is obtained by calculating using the above formula: At this point, the merged value of 78 times / minute is written into the global user profile, replacing the original 70 times / minute. This update method does not directly adopt the sudden change of 80 times / minute, but rather approximates the new data appropriately based on reliability. This allows the profile to maintain numerical updates while effectively filtering out the impact of collection errors or instantaneous fluctuations, ensuring the consistency of the user profile logic.
[0038] Furthermore, when the cross-modal fusion confidence level fails to meet the archive update conditions, the following fault-tolerant correction process is executed: Extract the newly written basic feature labels that fail to meet the target due to the negative adjustment of the constraint penalty term, and store them in the temporary queue; The cumulative truncation frequency of basic feature labels representing the same feature values in the temporary queue is counted within a sliding time window of fixed length. When the cumulative truncation frequency reaches the preset wall breaking threshold, the constraint penalty term is removed from the downward adjustment restriction of the cross-modal fusion confidence to trigger the numerical overwrite update, and a reverse feature correction instruction is generated simultaneously and sent to the master node entity robot that provides the benchmark constraint file to overwrite the corresponding expired file node. Among them, multiple physical robots adopt a federated learning architecture. The original stream of local modal data acquired by each physical robot is kept locally, and only the extracted basic feature labels and their corresponding spatiotemporal parameters are uploaded to the memory pool.
[0039] Specifically, when the cross-modal fusion confidence level fails to meet the preset file update conditions, a fault-tolerant correction process is executed to address the problem of normal feature updates being hindered due to expired baseline constraint files or preference shifts. Specifically, this involves extracting newly written basic feature labels that fail to meet the criteria due to negative adjustments caused by constraint penalties and storing them in a preset temporary queue. This step, by caching "rejected" data nodes, prevents the system from missing the evolution trend of user behavior due to excessive initial deviations, providing necessary data caching space for subsequent data backtracking and behavior pattern mining. Within a fixed-length sliding time window, the cumulative truncation frequency of basic feature labels representing the same value in the temporary queue is statistically analyzed. This statistical step distinguishes between instantaneous data noise and persistent behavioral state changes. The specific mathematical logic is as follows: ; in, This represents the cumulative truncation frequency within the sliding time window. This represents the preset fixed sliding time length. Representative at Rejected labels with similar target feature values are stored in the temporary queue at all times. This is an indicator function that takes the value 1 when the label value meets the preset clustering range, and 0 otherwise; When the cumulative truncation frequency reaches the preset threshold, the constraint penalty term's restriction on the reduction of cross-modal fusion confidence is lifted to trigger numerical overwrite update. Simultaneously, a reverse feature correction instruction is generated and sent to the master node entity robot that provides the baseline constraint file to overwrite the corresponding expired file node. This operation marks the confirmation of the user behavior pattern change, that is, the lag of the old file node is confirmed through long-term verification. Under the federated learning architecture, each entity robot always keeps the original stream of acquired local modal data locally, and only uploads the extracted basic feature labels and their corresponding spatiotemporal parameters to the memory pool. This ensures that when performing the reverse feature correction process, both the dynamic update of the global file is realized, and the user's local privacy data security is protected to the greatest extent. In a specific application example, if a user's original habit was to wake up at 8 am every day, and the wake-up status tag recorded in the global user profile was also "8 am", due to a recent adjustment in their schedule, the user changed to waking up at 6 am every day. As a result, the wake-up data at 6 am for several consecutive days deviated from the 8 am profile benchmark, was identified as abnormal, and was continuously blocked by the constraint penalty item and stored in a temporary queue. As time went on, the cumulative truncation frequency of the 6 am wake-up tag within the sliding time window reached the threshold. The method then determined that this was not a false trigger or noise, but a substantial change in the user's schedule. It automatically lifted the penalty restriction for that period, triggered the value overwrite, and pushed the "6 am" correction instruction to the main node entity robot that recorded the original 8 am profile, updating it to the new wake-up preference, thereby completing the accurate synchronization and correction of the user's long-term behavioral preferences.
[0040] Although embodiments of the invention 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 to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for dynamically updating user profiles of physical robots based on AI, characterized in that, Applicable to target scenarios involving multiple physical robots, including: S1. Obtain the local modal data of the target user and extract the corresponding basic feature labels. The local modal data includes: visual data, auditory data, and distance perception data. S2. Configure time and space parameters for the basic feature labels and store them in the memory pool; S3. In response to the newly written basic feature tags in the memory pool, based on the time parameters and spatial parameters, trace and extract the historical basic feature tags belonging to the target user, and extract the benchmark constraint file; S4. Calculate the attenuation weight based on the time difference between the new and old basic feature labels, generate a constraint penalty term based on the deviation between the newly written basic feature labels and the benchmark constraint file, and calculate the cross-modal fusion confidence level in combination with the modality type. S5. If the cross-modal fusion confidence is within the detection trigger range, then the candidate entity robot is scheduled to perform physical interaction actions based on the active detection strategy, and supplementary feature labels are extracted to update the cross-modal fusion confidence. S6. When the cross-modal fusion confidence meets the file update condition, determine the dynamic overwrite weight and write the data after fusing the new and old basic feature tags into the global user file of the target user. When the cross-modal fusion confidence fails to meet the file update conditions, the following fault-tolerant correction process is performed: Extract the newly written basic feature labels that fail to meet the target due to the negative adjustment of the constraint penalty term, and store them in a temporary queue; The cumulative truncation frequency of basic feature labels representing the same feature value in the temporary queue is counted within a sliding time window of fixed length. When it is determined that the cumulative truncation frequency reaches the preset wall breaking threshold, the restriction on the reduction of the cross-modal fusion confidence by the constraint penalty term is lifted to trigger the numerical overwrite update, and a reverse feature correction instruction is generated simultaneously and sent to the master node entity robot that provides the benchmark constraint file to overwrite the corresponding expired file node. The multiple entity robots adopt a federated learning architecture. The original stream of local modal data acquired by each entity robot is kept locally, and only the extracted basic feature labels and their corresponding spatiotemporal parameters are uploaded to the memory pool.
2. The method for dynamically updating user profiles of physical robots based on AI as described in claim 1, characterized in that, In step S2, configuring time and spatial parameters for the basic feature labels and storing them in a memory pool includes: Record the timestamp of the moment when the corresponding physical robot acquires the local modal data, and use the timestamp as the time parameter; Extract the three-dimensional global topological coordinates when the corresponding physical robot collects the local modal data, and the field of view direction vector of the corresponding sensing sensor on the physical robot, and combine the three-dimensional global topological coordinates and the field of view direction vector to form the spatial parameters; The basic feature labels are bound to the corresponding time parameters and spatial parameters using data structures to generate a multidimensional feature data package, and the multidimensional feature data package is stored in the memory pool. The timestamp is based on the Network Time Protocol to achieve clock synchronization among the multiple physical robots, and the three-dimensional global topological coordinates are obtained based on a fusion system of ultra-wideband positioning, real-time positioning and map building.
3. The method for dynamically updating user profiles of physical robots based on AI as described in claim 1, characterized in that, In step S3, based on the time and space parameters, historical basic feature tags belonging to the same target user are extracted, including: Extract the target timestamp and the target's three-dimensional global topological coordinates corresponding to the newly written basic feature labels; The target timestamp, the target three-dimensional global topological coordinates, and the time and spatial parameters of the candidate historical tags in the memory pool are used as input items and input into the spatial trajectory continuity determination model. If the spatial trajectory continuity determination model calculates that the physical motion trajectory of the target user between the target timestamp and the time parameters corresponding to the candidate historical tags satisfies the preset spatial continuity distribution characteristics, then the candidate historical tags are determined to belong to the target user and are extracted as the historical basic feature tags. The preset spatial continuity distribution feature includes: the estimated movement speed of the target user between the target timestamp and the corresponding time parameter of the candidate historical tag is within the preset reasonable human movement speed threshold range.
4. The method for dynamically updating user profiles of an AI-driven physical robot according to claim 3, characterized in that, The baseline constraint file extracted in S3 includes: The frequency and duration of each of the multiple physical robots within the target scene interacting with the target user in the past are statistically analyzed. Based on the frequency of historical valid interactions and the duration of the interactions, the master node entity robot with the highest historical confidence is selected from the plurality of entity robots. The profile data, which contains the long-term deterministic state preferences of the target user and is constructed locally by the master node entity robot based on long-term interaction data, is extracted as the benchmark constraint profile.
5. The method for dynamically updating user profiles of physical robots based on AI as described in claim 1, characterized in that, In step S4, the attenuation weight is calculated based on the time difference between the old and new basic feature labels, including: Extract the content attribute categories represented by the historical basic feature tags, and configure corresponding state volatility indices for different content attribute categories; The time difference and the state variability index are input into the decay calculation model as common inputs, and the decay weight is output. The attenuation weight is inversely mapped to both the time difference and the state variability index.
6. The method for dynamically updating user profiles of an AI-driven physical robot according to claim 5, characterized in that, Constraint penalty terms are generated based on the deviation between the newly written basic feature labels and the baseline constraint profile, and cross-modal fusion confidence is calculated in conjunction with the modality type, including: The state value represented by the newly written basic feature label is compared with the value of the deterministic archive node of the corresponding dimension in the benchmark constraint archive to calculate the state deviation. If the deviation of the state is greater than the preset constraint threshold, a negative adjustment value is generated as the constraint penalty item; Determine whether the modal type of the new and old basic feature labels belongs to positive interactive complementarity modality. If so, generate a positive complementary gain value. The cross-modal fusion confidence score is calculated by weighting and summing the attenuation weight, the constraint penalty term, and the complementary gain value.
7. The method for dynamically updating user profiles of an AI-driven physical robot according to claim 1, characterized in that, In step S5, scheduling candidate entity robots to perform physical interaction actions based on the active detection strategy includes: Obtain the physical movement energy consumption data of each candidate physical robot among the multiple physical robots, as well as the disturbance risk index corresponding to the execution of the detection action, and use the physical movement energy consumption data and the disturbance risk index as penalty parameters; Calculate the expected probability of obtaining the missing supplementary modal data, and use the expected probability as a revenue parameter; The penalty term parameter and the benefit term parameter are input into the reinforcement learning decision model, and the reinforcement learning decision model outputs physical interaction action instructions assigned to specific candidate entity robots. The reward function of the reinforcement learning decision model is positively correlated with the expected probability and negatively correlated with the physical mobility energy consumption data and the disturbance risk index.
8. The method for dynamically updating user profiles of an AI-driven physical robot according to claim 7, characterized in that, The physical interaction action instructions assigned to specific candidate entity robots by the reinforcement learning decision model include: Extract the current physiological or environmental state data of the target user recorded in the baseline constraint file and input it as boundary constraint conditions into the reinforcement learning decision model; Based on the boundary constraints, the reinforcement learning decision model removes intrusive interaction actions with state conflicts from the preset set of physical interaction instructions, and outputs the physical interaction action instructions from the remaining interaction instructions.
9. The method for dynamically updating user profiles of an AI-driven physical robot according to claim 1, characterized in that, In step S6, a dynamic overwrite weight is determined based on the cross-modal fusion confidence level. The new and old basic feature labels are then fused using the dynamic overwrite weight and written into the target user's global user profile. This includes: The cross-modal fusion confidence score is converted into a learning rate parameter calculated by moving average, and the learning rate parameter is set as the dynamic overwrite weight. Using the dynamic overwrite weight, a moving average calculation is performed on the feature values represented by the newly written basic feature labels and the historical basic feature labels to generate fused feature data with smooth transition characteristics. Using the fused feature data, a non-mutational numerical overwrite update is performed on the corresponding node in the global user profile of the target user.
Citation Information
Patent Citations
Accompanying robot decision-making method, device and equipment based on cross-modal knowledge graph and medium
CN120941406A
Intelligent interactive accompanying system supporting multi-module cooperation and cross-device memory undertaking
CN121302344A
