Robots and methods for predicting their social intentions

By collecting and analyzing the behavioral signals of target users, establishing individual baselines, and using a pre-reflective prediction model for multi-dimensional anomaly analysis, the accuracy problem in predicting the social intentions of robots is solved, and efficient and accurate response decisions are achieved.

CN123115668APending Publication Date: 2026-11-17BEIJING XINGCHI POWER TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611300182.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing robots have poor accuracy in predicting social intentions and cannot effectively distinguish individual differences, leading to false positives and false negatives.

Method used

By collecting behavioral signals from target users, extracting current behavioral features, establishing individual baselines based on historical behavioral features, calculating deviations, and inputting them into a prediction model to generate response levels, the prediction results do not include intermediate expressions in natural language. A pre-reflective prediction model is used for multi-dimensional anomaly analysis to construct a response level classification decision.

Benefits of technology

It improves the accuracy and response speed of social intent prediction, avoids false alarms and false negatives caused by individual differences, achieves sub-second accurate response, and ensures the consistency of user experience and continuous improvement of accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115668A_ABST
    Figure CN123115668A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of robotics technology, specifically providing a robot and a method for predicting its social intentions. The method includes: collecting behavioral signals of a target user; extracting the target user's current behavioral features from the behavioral signals; establishing and updating an individual baseline based on the target user's historical behavioral features; calculating the deviation between the current behavioral features and the individual baseline; inputting the deviation into a prediction model to obtain a prediction result; the prediction result does not include intermediate expressions in natural language form; and determining the robot's response level based on the prediction result. This invention improves the accuracy of predicting the robot's social intentions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of robotics technology, specifically providing a robot and a method for predicting its social intentions. Background Technology

[0002] With the widespread application of service robots in scenarios such as home care, elderly care, disability assistance, and public guidance, robots need to possess the ability to perceive users' social states and adjust their own behavior accordingly. Social intention prediction refers to the robot's ability to infer users' current emotional state, inner needs, and behavioral intentions by collecting users' behavioral signals (including voice features, facial features, and motion features), and then decide how to respond to users based on this information.

[0003] Current service robots generally use a uniform population model to perceive human states. That is, emotion recognition models and behavior recognition models trained on public datasets use the same set of discrimination parameters for all users. This perception method averages out individual differences.

[0004] Because the same behavioral signal can have opposite meanings for different individuals. For example, a speech rate of 120 words per minute might be considered low for a talkative person, but normal for a quiet person. A uniform population model cannot distinguish these differences, leading to continuous false positives for users who deviate from the population mean but are within their normal range, and false negatives for users who are abnormal but still within the normal range. Therefore, existing robots' perception methods suffer from poor accuracy in predicting social intentions. Summary of the Invention

[0005] One objective of this invention is to address the problem of poor accuracy in predicting social intentions in existing robots.

[0006] To achieve the above objectives, the present invention provides, in a first aspect, a method for predicting the social intentions of a robot, comprising: Collect behavioral signals from target users; Extract the current behavioral features of the target user from the behavioral signals; Establish and update individual baselines based on the historical behavioral characteristics of the target users; Calculate the deviation between the current behavioral characteristics and the individual baseline; The bias is input into the prediction model to obtain the prediction result; the prediction result does not include intermediate expression results in natural language form; The robot's response level is determined based on the prediction results.

[0007] Optionally, establishing and updating the individual baseline based on the target user's historical behavioral characteristics includes: Obtain the cumulative sample size of the historical behavioral characteristics of the target user; When the cumulative sample size does not reach a preset threshold, the group statistical reference value and the statistical value of the target user's historical behavioral characteristics are combined according to weights to form the individual baseline; the weight of the group statistical reference value decreases as the cumulative sample size increases; Once the cumulative sample size reaches the preset threshold, the individual baseline is updated based on the historical behavioral characteristics of the target user; the individual baseline uses the statistical value of the historical behavioral characteristics within a preset time window as a benchmark.

[0008] Optionally, the deviation is the difference between the value of the current behavioral feature and the mean of the individual baseline, divided by the standard deviation of the individual baseline.

[0009] Optionally, the prediction model is a pre-reflective prediction model, and the deviation includes deviation values ​​in multiple dimensions; the step of inputting the deviation into the prediction model to obtain the prediction result includes: The deviation is input into the pre-reflective prediction model; The pre-reflective prediction model uses the deviation value of each dimension of the deviation as the anomaly significance index of that dimension. Based on the relative strength of the anomaly significance indices of each dimension and whether the anomaly significance indices of each dimension appear simultaneously, it generates a comprehensive state representation of the target user. The pre-reflective prediction model directly outputs the prediction results based on the comprehensive state representation. The prediction results include the target user's emotional state distribution, demand distribution, and the probability of acceptance of each candidate response level.

[0010] Optionally, the response level is a plurality of progressive response levels that increase sequentially from low to high, and the prediction result includes the acceptance probability corresponding to each of the progressive response levels. Determining the robot's response level based on the prediction result includes: Starting from the highest level among the multiple progressive response levels, the acceptance probability of each level is compared with its corresponding preset threshold in turn. The first level where the acceptance probability reaches or exceeds its corresponding preset threshold is determined as the response level; If none of the acceptance probabilities reach their corresponding preset thresholds, then the lowest level among the plurality of progressive response levels is determined as the response level.

[0011] Optionally, the method further includes: Execute the determined response level; Within a preset time period after execution, the target user's reaction signal is collected; The reaction signal is determined to be a positive or negative sample according to a preset rule. The reaction signals that are determined to be positive or negative samples are used as calibration samples and grouped and stored according to the combination of the target user and the response level; When the number of calibration samples in any group reaches a preset lower limit, the acceptance probability of the corresponding response level is updated according to the proportion of positive samples in that group.

[0012] Optionally, updating the acceptance probability of the corresponding response level based on the proportion of positive samples within the group includes: Obtain the difference between the proportion of positive samples within the group and the current acceptance probability; The update step size is determined according to the magnitude of the difference and a preset mapping relationship; the larger the difference, the larger the update step size. The current acceptance probability is corrected using the update step size to obtain the updated acceptance probability.

[0013] Optionally, the update step size is calculated using the following formula: ; Where η0 is the base step size, α is the preset scaling factor, Δp is the difference between the positive sample ratio and the current acceptance probability, and N is the total number of calibration samples in the group.

[0014] Optionally, the method further includes: Before determining the robot's response level based on the prediction results, the veto bits are queried; the veto bits include physical veto bits and social veto bits. When the physical veto position is set, the robot is forced to execute the lowest response level and the result is recorded. When the social veto position is set, the robot is forced to execute the lowest response level and the result is recorded.

[0015] In a second aspect, the present invention provides a robot including a processor and a memory, wherein the memory stores a computer program, and the processor, when executing the computer program, is capable of implementing the method described in any one of the first aspects.

[0016] Based on the foregoing description, those skilled in the art will understand that in the aforementioned technical solution of this invention, an individual baseline is established and updated based on the historical behavioral characteristics of the target user, enabling this invention to establish a unique individual baseline for each user. Then, the deviation between the current behavioral characteristics and the individual baseline is calculated; and this deviation is input into the prediction model to obtain the prediction result, thus improving the accuracy of social intention prediction. Therefore, compared to the prior art's method of using a unified population model for discrimination, this invention fundamentally eliminates false positives and false negatives caused by individual differences. Furthermore, because the prediction result of this invention does not include intermediate expressions in natural language form, it avoids the time delay and loss of continuous information caused by transcribing the perceived result into natural language and then using a large language model for inference, further improving the robot's response speed and accuracy in predicting social intention.

[0017] Furthermore, by acquiring the cumulative sample size of historical behavioral characteristics, and when the cumulative sample size has not reached a preset threshold, the group statistical reference value and individual statistical values ​​are combined according to weights to form an individual baseline. After the cumulative sample size reaches the preset threshold, the statistical value of historical behavioral characteristics within a preset time window is used as the benchmark for the individual baseline, achieving a smooth transition for new users from a cold start with no data to one with sufficient data. Specifically, during the cold start phase, the group statistical reference value can compensate for the lack of individual data. Then, as the sample accumulates, the group weight gradually decreases while the individual weight gradually increases, with no baseline jumps throughout the process and no need for manual intervention, ensuring the consistency of the target user's experience across time periods.

[0018] Furthermore, by using the deviation value of each dimension of the deviation as the anomaly significance index for that dimension in the pre-reflective prediction model, and generating a comprehensive state representation based on the relative strength and synchronous occurrence of the anomaly significance indices of each dimension, the model directly outputs the emotion distribution, demand distribution, and graded acceptance probability, thus directly capturing the co-occurrence pattern of deviations. In short, the pre-reflective prediction model of this invention can determine the state of the target user based on synchronous anomalies in multiple behavioral dimensions (rather than anomalies in a single dimension), improving the accuracy of the robot's prediction of social intentions and reducing misjudgments caused by random fluctuations or accidental anomalies in a single dimension.

[0019] Furthermore, by using preset rules to determine whether the target user's response signal is a positive or negative sample, the determination result is used as a calibration sample, which is then grouped and stored according to the combination of the target user and the response level. When the number of calibration samples in any group reaches a preset lower limit, the corresponding acceptance probability is updated based on the proportion of positive samples, thus constructing a continuous calibration closed loop that does not disturb the target user and fundamentally avoids interference from the confirmation process. Grouped storage ensures fine-grained calibration—the acceptance level of the same target user for different levels is calibrated independently, and the acceptance level of different target users for the same level is also calibrated independently, thereby enabling the prediction accuracy to continuously improve with the accumulation of interactions.

[0020] Furthermore, by obtaining the difference between the proportion of positive samples within a group and the current acceptance probability, and determining the update step size based on this difference, the current acceptance probability is corrected using the update step size. This achieves adaptive calibration where "the larger the difference, the faster the correction; the more samples, the more stable the correction." Specifically, during the cold start phase when the sample size is small, the update step size can be larger for a rapid response; once the sample size is sufficient, the update step size can be automatically reduced for fine convergence.

[0021] Other beneficial effects of the present invention will be described in detail below with reference to the accompanying drawings, so that those skilled in the art can more clearly understand the improved objectives, features and advantages of the present invention. Attached Figure Description

[0022] To more clearly illustrate the technical solution of the present invention, some embodiments of the present invention will be described below with reference to the accompanying drawings. Those skilled in the art should understand that the same reference numerals may indicate the same or similar parts or components in different drawings; the drawings of the present invention are not necessarily drawn to scale. In the drawings: Figure 1 This is a flowchart of the steps of the social intention prediction method in some embodiments of the present invention; Figure 2 yes Figure 1 Flowchart of the specific steps in step S130; Figure 3 yes Figure 1 Flowchart of the specific steps in step S150; Figure 4 yes Figure 1 Flowchart of the specific steps in step S160; Figure 5 This is a flowchart of some steps of the social intention prediction method in some embodiments of the present invention; Figure 6 yes Figure 5 Flowchart of the specific steps in step S250; Figure 7This is a flowchart of some steps of the social intention prediction method in other embodiments of the present invention; Figure 8 This is a block diagram of a robot provided by the present invention. Detailed Implementation

[0023] Those skilled in the art should understand that the embodiments described below are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. These partial embodiments are intended to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of the present invention.

[0024] It should be noted that in the description of this invention, terms such as "center," "upper," "lower," "top," "bottom," "left," "right," "vertical," "horizontal," "inner," and "outer," which indicate direction or positional relationships, are based on the direction or positional relationships shown in the accompanying drawings. These are used merely for ease of description and do not indicate or imply that the corresponding device or element must have a specific orientation, or be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0025] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can also refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] Furthermore, it should be noted that in this invention, the flowchart is not intended to indicate that the operations of the method will be performed in any particular order, or that all operations of the method are included in all every case. Moreover, the method may include additional operations. Within the scope of the technical concept provided by the method of this embodiment, additional variations can be made to the following method.

[0027] Finally, it should be noted that in the description of this invention, each functional module can be a physical module composed of multiple structures, components, or electronic devices, or a virtual module composed of multiple programs; each functional module can be an independent module or a module divided from a whole module according to its function. Those skilled in the art should understand that, provided the technical solution described in this invention can be implemented, any changes in the configuration, implementation, or positional relationship of the functional modules will not deviate from the technical principles of this invention, and therefore should all fall within the protection scope of this invention.

[0028] like Figure 1 As shown, in some embodiments of the present invention, the method for predicting the social intentions of a robot includes: Step S110: Collect behavioral signals of the target user.

[0029] Specifically, behavioral signals of the target user can be collected through a multimodal perception module installed on the robot. This multimodal perception module may include visual sensors (such as cameras and depth cameras), audio sensors (such as microphone arrays), tactile force sensors (such as tactile sensors and pressure sensors), body sensors (such as joint angle sensors and IMU inertial measurement units), and environmental sensors (such as distance sensors and light sensors). These sensors can be integrated into the robot body or distributed in the environment in which the robot operates (e.g., fixed cameras and microphones in a smart home environment).

[0030] In this invention, behavioral signals refer to multimodal raw signals that can reflect the user's behavioral state, including but not limited to: the user's voice signals, facial image sequences, human posture image sequences, body movement trajectories, and interactive manipulation forces with the robot. These behavioral signals are continuously collected by corresponding sensors (as described in the front end) at a preset sampling frequency.

[0031] In this invention, the target user can be a registered user, specifically a user who has registered in the robot's system.

[0032] Step S120: Extract the current behavioral features of the target user from the behavioral signals.

[0033] Specifically, the multimodal behavioral signals collected in step S110 can be preprocessed and feature extracted by a feature extraction module installed on the robot to obtain the current behavioral features. The current behavioral features refer to the numerical feature vector extracted from the behavioral signals at the current moment, which can quantitatively describe the user's behavioral state.

[0034] In this embodiment, the current behavioral characteristics may include behavioral characteristics of multiple dimensions.

[0035] For example, the current behavioral characteristics include features from eight channels: speech rate (the number of words or syllables spoken by the user per unit time), volume (the amplitude or energy of the speech signal), fundamental frequency (the basic frequency of the speech signal used to reflect pitch), facial expression activity (the intensity of facial movements calculated based on facial key point detection), movement speed (the speed at which the user's body moves in three-dimensional space), interaction frequency (the number of times the user actively interacts with the robot per unit time), response delay (the time delay in the user responding to the interaction request initiated by the robot), and circadian rhythm (the current position relative to the user's daily circadian rhythm, such as the time interval since the last time they fell asleep / woke up). These eight channel features characterize the user's state from four behavioral dimensions: speech, facial expression, movement, and interaction rhythm.

[0036] In this embodiment, feature extraction can be implemented using signal processing methods known in the art. Specifically, as follows.

[0037] For speech signals, speech rate, volume, and fundamental frequency can be calculated after extracting effective speech segments through speech activity detection (VAD).

[0038] For facial images, the activity intensity of facial motion units can be calculated as expression activity through facial keypoint detection (such as using a 68-point or 106-point facial keypoint detection model).

[0039] For human pose images, the positions of human key points and the movement speed can be extracted and calculated through human pose estimation (such as using algorithms like OpenPose and MediaPipe).

[0040] For interactive force signals, the frequency of interaction can be calculated directly from the readings of the contact force sensor.

[0041] The delay in response can be calculated by recording the difference between the timestamp of the robot sending the interaction signal and the timestamp of the user's response.

[0042] Circadian rhythms can be calculated by recording the user's daily active and resting periods.

[0043] Step S130: Establish and update individual baselines based on the historical behavioral characteristics of target users.

[0044] Historical behavioral characteristics refer to all or part of the behavioral characteristics data of the target user collected before the current moment.

[0045] Individual baselines refer to the normal reference values ​​obtained by statistically analyzing various behavioral characteristic channels of a single user, including the mean and standard deviation of historical behavioral characteristics for each channel. Individual baselines differ from the unified population model. Specifically, the unified population model uses the same set of discriminant parameters for all users, while individual baselines are established based on the user's own historical data, forming a unique reference system for that user.

[0046] Specifically, this step can store and continuously update the individual baseline of the target user by using the target user's identifier (such as user ID, facial feature code, or voiceprint feature code) as an index in a preset individual baseline library.

[0047] The individual baseline database is used to maintain the individual baseline of each target user and is continuously updated as new behavioral characteristics are collected. For example, the individual baseline uses a 14-day rolling window and is maintained in buckets according to the four time periods of the day (morning, noon, evening, and night). Historical samples are exponentially forgotten according to a preset half-life (7 days) to take into account both the periodicity and recentity of the work and rest cycle.

[0048] Step S140: Calculate the deviation between the current behavioral characteristics and the individual baseline.

[0049] In this step, for each behavioral feature channel of the current observation, its deviation from the individual baseline is calculated. This deviation reflects the degree to which the current observation deviates from the target user's individual norm.

[0050] Wherein, the deviation is the standardized deviation, and its calculation formula is: In the formula, x i Let μ be the current observation value of the i-th behavioral feature channel. i and σ i Let Δz be the baseline mean and standard deviation of the user in the i-th behavioral feature channel, respectively. i Let be the standardized deviation value of the i-th channel. Standardized deviation is dimensionless and measures the degree to which the current observation deviates from the mean, expressed in standard deviation. Channels can be directly compared.

[0051] For example, suppose a user's baseline speech rate is 100 words per minute (MB / s) with a standard deviation of 15 MB / s. If the current observed speech rate is 73 MB / s, the speech rate deviation is (73-100)÷15=-1.8, meaning the current speech rate is 1.8 standard deviations lower than the user's baseline. If the current observed speech rate is 115 MB / s, the speech rate deviation is (115-100)÷15=1.0, meaning the current speech rate is 1.0 standard deviation higher than the user's baseline.

[0052] The deviation values ​​of each channel constitute the deviation vector z. t =[Δz1,Δz2,…,Δz m ]⊺ , where m is the number of behavioral feature channels. For example, in the eight-channel embodiment mentioned above, m=8.

[0053] Step S150: Input the deviation into the prediction model to obtain the prediction result.

[0054] The prediction results do not include intermediate expressions in natural language form. That is, the prediction model does not convert the bias vector obtained in step S140 into natural language text (such as "the user speaks slowly and has inactive facial expressions"), nor does it output the intermediate inference process in natural language form. Instead, it directly outputs the prediction results in the form of a structured probability distribution. In this way, the time delay and loss of continuous information (bias magnitude, confidence level, etc.) caused by natural language processing can be avoided.

[0055] Step S160: Determine the robot's response level based on the prediction results.

[0056] The response levels are multiple progressively higher levels, arranged from low to high. For example, the response levels include five levels, R1 to R5, arranged from low to high: R1 is silent observation, where the robot does not perform any explicit actions but only maintains environmental awareness; R2 is for fine-tuning the posture, where the robot adjusts the orientation of its head or torso to show attention; R3 is the minimum disturbance preparatory action, in which the robot performs preparatory actions such as approaching at a low speed or amplitude. R4 represents non-verbal service actions, where the robot performs non-verbal service actions such as playing music or adjusting lights. R5 is for language interaction, where the robot actively speaks to the target user.

[0057] In this embodiment, the prediction result includes the acceptance probability corresponding to each progressive response level. That is, the prediction model outputs an independent acceptance probability value R for each level. a (R i ), where R a R represents the probability of acceptance. i Let i represent the response level, i = 1, 2, ..., 5. The acceptance probability of each response level is independent.

[0058] In this embodiment, the acceptance probability of a certain response level represents the probability that the target user will accept the robot's response when the robot executes that particular response level.

[0059] Based on the foregoing description, those skilled in the art will understand that this invention eliminates false positives and false negatives by replacing the traditional unified population model with an individual baseline; it achieves sub-second prediction through a languageless intermediate representation; and it avoids interference in low-acceptance scenarios by replacing the traditional binary decision-making with a tiered response level. Therefore, this invention improves the robot's response speed and accuracy in predicting social intentions.

[0060] It should be noted that the first embodiment described above is merely a basic embodiment of the present invention. In other embodiments of the present invention, those skilled in the art can adjust, optimize, and configure the schemes and steps in the above embodiments as needed to achieve further technical effects. Other embodiments of the present invention, different from the above embodiments, will be described below with reference to the accompanying drawings. Of course, those skilled in the art can also appropriately modify the execution order, operating conditions, and number of steps in the embodiments described below according to actual needs. The modified embodiments will not deviate from the technical concept and / or technical principles of the present invention and should still fall within the protection scope of the present invention.

[0061] like Figure 2 As shown, in some embodiments of the present invention, step S130 may further include: Step S131: Obtain the cumulative sample size of the target user's historical behavioral characteristics.

[0062] Specifically, using the target user's identifier as an index, the total number of historical behavioral features stored for that target user is queried and denoted as n. Each historical behavioral feature is a complete multi-channel feature vector extracted from the behavioral signal within one acquisition period. The cumulative sample size n is a non-negative integer, with n=0 indicating that the user has no historical data yet.

[0063] Step S132: When the cumulative sample size does not reach the preset threshold, the group statistical reference value and the statistical value of the target user's historical behavioral characteristics are combined into an individual baseline according to the weights.

[0064] Among them, the weight of the group statistical reference value decreases as the cumulative sample size increases.

[0065] The preset thresholds are defined as both the number of consecutive observation days and the cumulative number of samples reaching preset values. For example, the preset thresholds are: consecutive observation days ≥ 3 days and cumulative sample count ≥ 200.

[0066] When the cumulative sample size does not reach the preset threshold (i.e., n<200 or the number of consecutive observation days<3 days), the robot system is in the cold start stage. At this time, the individual historical data of the target user is insufficient. The individual baseline library synthesizes the individual baseline by weighting the group statistical reference value and the statistical value of the target user's historical behavioral characteristics.

[0067] The group statistical reference value is a pre-set reference value based on large-scale population statistics, including the mean and standard deviation of each behavioral characteristic channel in the population. The group statistical reference value can be factory-preset or obtained in the cloud based on statistics from a large amount of user data. The purpose of the group statistical reference value is to provide a reasonable initial baseline when individual user data is insufficient, enabling the robot to have a usable reference benchmark from the very first interaction.

[0068] The statistical values ​​of a target user's historical behavioral characteristics refer to the mean and standard deviation calculated based on all or part of the user's accumulated historical behavioral characteristics. When historical data is limited, these statistical values ​​may not be stable enough, and therefore their weight is relatively low.

[0069] In this step, the synthesis method is a weighted summation: In the formula above, w represents the weight of the population statistical reference value, and 0 ≤ w ≤ 1, which monotonically decreases as the cumulative sample size n increases. An example is shown below: Where N0=500 is the preset attenuation constant. When n=0, w=1, the individual baseline is exactly equal to the population statistical reference value; when n=100, w≈0.819, the population statistical reference value dominates; when n=500, w≈0.368, the individual statistical value is more than half; when n=3000, w≈0.002, the population statistical reference value is essentially eliminated.

[0070] For example, suppose a user registers on the second day, with a cumulative sample size of 100 entries. The average speech rate in the group statistical reference is 120 words per minute, and the average speech rate in the individual user statistics is 90 words per minute. Then w = exp(-100 ÷ 500) ≈ 0.819, and the average speech rate in the composite baseline μ = 0.819 × 120 + (1 - 0.819) × 90 = 114.57 words per minute.

[0071] Step S133: When the cumulative sample size reaches the preset threshold, update the individual baseline based on the historical behavioral characteristics of the target user.

[0072] The individual baseline uses the statistical values ​​of historical behavioral characteristics within a preset time window as a benchmark.

[0073] Once the cumulative sample size reaches the preset threshold (i.e., n≥200 and the number of consecutive observation days≥3 days), the robot's system exits the cold start phase and directly uses the statistical values ​​of the target user's own historical behavioral characteristics within the preset time window as the individual baseline. The group statistical reference values ​​are no longer included in the synthesis.

[0074] The preset time window can be a 14-day rolling window, and maintenance is carried out in separate buckets according to the four time periods of the day (morning, noon, evening, and night).

[0075] Furthermore, for each behavioral characteristic channel, the mean and standard deviation of the user's historical data within the current time period and time window coverage area can be calculated separately to serve as an individual baseline.

[0076] Furthermore, historical samples can be exponentially forgotten according to a preset half-life T, that is, the older the sample, the lower its weight, and the weight decay factor is exp(-t÷T). Here, t is the time since the sample was taken; T is the half-life, which can be any feasible duration such as 7 days, 8 days, 10 days, etc.

[0077] Taking a half-life of T=7 days as an example, let the current time be t0 and the timestamp of a certain historical sample be t. i Then the weight of this sample in the baseline statistics is exp[-(t0-t i ()÷7). This means that the sample weight 7 days ago is reduced to about 37%, and the sample weight 14 days ago is reduced to about 14%, thus making the individual baseline more sensitive to the user's recent status changes.

[0078] For example, if a user's speech rate during the evening hours (6:00 PM - 10:00 PM) has a 14-day rolling window, and the half-life-weighted data has a mean of 95 words per minute and a standard deviation of 12 words per minute, then the user's individual baseline mean for that time period is 95 words per minute and the standard deviation is 12 words per minute. If the user's speech rate is collected at 6:00 PM on a certain day and is 68 words per minute, then the deviation is (68-95)÷12=-2.25.

[0079] Based on the foregoing description, those skilled in the art will understand that, through steps S131 to S133, the present invention achieves a smooth transition from the cold start phase to the regular phase. Specifically, during the cold start phase, insufficient individual data can be compensated for by using group statistical reference values. Then, as the sample accumulates, the group weight gradually decreases while the individual weight gradually increases, with no baseline jumps throughout the process and no need for manual intervention, ensuring the consistency of the target user's experience across different time periods.

[0080] Furthermore, in some embodiments of the present invention, the prediction model is a pre-reflective prediction model, and the bias includes bias values ​​in multiple dimensions (such as the bias vector z described above). t Furthermore, step S150 can be as follows: Figure 3 The illustrated further includes: Step S151: Input the deviation into the pre-reflection prediction model.

[0081] Specifically, the deviation vector z calculated in step S140t Pre-input reflection prediction model.

[0082] The pre-reflective prediction model is a computational model that performs forward reasoning based on the current deviation and directly outputs the prediction result before the robot executes a response action. The term "pre-reflective" means that this model completes the prediction before the robot performs the action, without relying on retrospective evaluation of historical interactions or self-correction of already performed actions. This is completely opposite in the time dimension to the existing "reflection" mechanism (which evaluates its own performance after execution and optimizes accordingly).

[0083] The pre-reflective prediction model can be implemented using known neural network architectures in the field (such as multilayer perceptrons or Transformer encoders) and trained through supervised learning: using historical interaction data labeled with emotion categories, demand categories and acceptance probabilities at each level as training samples, using the weighted sum of cross-entropy loss and mean squared error loss as the loss function, and optimizing the model parameters through backpropagation.

[0084] Step S152: The pre-reflective prediction model uses the deviation value of each dimension of the deviation as the anomaly significance index of that dimension. Based on the relative strength of the anomaly significance indices of each dimension and whether the anomaly significance indices of each dimension appear synchronously, a comprehensive state representation of the target user is generated.

[0085] Specifically, the pre-reflective prediction model treats each dimension of the deviation vector as a deviation value Δz. i Let's interpret this. The physical meaning of this deviation value is the degree to which this behavioral characteristic dimension deviates from the user's personal norm. The absolute value of the deviation value |Δz i | This directly reflects the anomalous significance of this dimension, | Δz i The larger the value, the further the dimension deviates from the norm, and the more attention it deserves.

[0086] Building upon this, the pre-reflective prediction model analyzes the relative strength of the anomaly indicators. For example, if the speech rate deviation is -1.8 (significantly decreased), the facial expression activity deviation is -1.2 (significantly decreased), the response delay deviation is +1.1 (significantly increased), and the volume deviation is -0.2 (close to normal), then the pre-reflective prediction model determines that the combination of "decreased speech rate + decreased facial expression + increased response delay" is the most significant anomaly pattern, followed by the volume dimension. This analysis of relative strength enables the pre-reflective prediction model to distinguish between primary and secondary anomalies, thereby gaining a more accurate understanding of the target user's state.

[0087] Simultaneously, the pre-reflective prediction model also determines whether the anomalies in each dimension occur synchronously. If the deviation values ​​of multiple dimensions are simultaneously large (e.g., a decrease in speech rate and a decrease in facial expression activity occur simultaneously), the pre-reflective prediction model determines that there is a co-occurrence relationship between the anomalies in these behavioral feature dimensions. This co-occurrence relationship reflects a holistic change in the user's state. This is because anomalies in a single dimension may be caused by random factors, while synchronous anomalies in multiple dimensions may be an external manifestation of internal state changes. For example, the synchronous occurrence of a decrease in speech rate and an increase in response delay is more likely to represent a user's depressed mood than a decrease in speech rate alone.

[0088] Based on the above analysis, the pre-reflective prediction model generates a comprehensive state representation of the target user. This comprehensive state representation is a hidden vector that integrates the relative strength of anomalies across various dimensions and co-occurrence patterns. This vector encodes the complete state information of the target user's current deviation from the norm.

[0089] Step S153: The pre-reflective prediction model directly outputs the prediction results based on the comprehensive state representation. The prediction results include the target user's emotional state distribution, demand distribution, and the probability of acceptance of each candidate response level.

[0090] Specifically, the pre-reflective prediction model maps the comprehensive state representation generated in step S152 into an emotional state distribution, a demand distribution, and a graded acceptance probability, and outputs it directly in the form of a structured probability distribution without going through natural language.

[0091] Here, the emotional state distribution is a probability vector Pe∈ℝ Ke Ke is the preset number of emotion categories. For example, Ke=8, the emotion categories include: calm, happy, depressed, angry, surprised, fearful, disgusted, and neutral. The j-th element of Pe represents the probability that the target user is currently in the j-th emotion category, and the sum of all elements is 1.

[0092] Wherein, the demand distribution is a probability vector Pn∈ℝ Kn Kn is the preset number of need categories. For example, Kn=12, the need categories include: companionship, help, information, comfort, space, confirmation, encouragement, guidance, interaction, rest, eating, and no specific need. The k-th element of Pn represents the probability that the target user currently has the k-th type of need, and the sum of all elements is 1.

[0093] Here, the acceptance probability for each candidate response level is a vector r. a ∈[0,1] LL represents the number of candidate response levels. For example, L=5 corresponds to five progressive response levels from R1 to R5. The l-th element of ra represents the probability that the target user will accept the robot's response corresponding to the l-th response level. Each probability value is independent and is output by the Sigmoid function.

[0094] The aforementioned emotion distribution, demand distribution, and tiered acceptance probability constitute a structured prediction result, which is directly used for subsequent response tier determination without going through natural language processing.

[0095] Based on the foregoing description, those skilled in the art will understand that the pre-reflective prediction model of the present invention can determine the state of the target user based on synchronous anomalies in multiple behavioral dimensions (rather than anomalies in a single dimension), thereby improving the accuracy of the robot's prediction of social intentions and reducing misjudgments caused by random fluctuations or accidental anomalies in a single dimension.

[0096] like Figure 4 As shown, in some embodiments of the present invention, step S160 may further include: Step S161: Starting from the highest level among multiple progressive response levels, the acceptance probability of each level is compared with its corresponding preset threshold.

[0097] Taking a response level of 5 as an example, first denote the acceptance probability of each progressive response level output in step S150 as R. a (R1),R a (R2),…,R a (R5). Then, starting from the highest level R5, the acceptance probability of each level is compared with its corresponding preset threshold. Each level is configured with a corresponding preset threshold: the threshold for R5 is 0.65, the threshold for R4 is 0.55, the threshold for R3 is 0.50, the threshold for R2 is 0.30, and the threshold for R1 is 0. Furthermore, those skilled in the art can adjust the values ​​of the preset thresholds for each level as needed.

[0098] Specifically, first compare R a If R5 is ≥ 0.65, it is considered an R5 candidate (but R5 still needs to pass language-level careful verification, see below for details). If not, compare R... a If (R4) is ≥ 0.50, then it is determined to be R4. If not, then compare R... a If (R3) is ≥ 0.50, then it is determined to be R3. If not, then compare R... a (R2) is ≥0.30. If it is satisfied, it is determined to be R2; if it is not satisfied, it is determined to be R1.

[0099] Among them, R5 (voice interaction) is the highest level, with the highest degree of social intrusion (once triggered, it means the robot needs to actively speak to the user, which is far more disruptive than non-verbal actions), therefore its admission requirements are the most stringent. Specifically, it not only requires R... a If R5 ≥ 0.6, language-level deliberate verification is required. This language-level deliberate verification is an additional verification step independent of the prediction model, used to comprehensively judge factors such as user state, scenario suitability, and historical interaction records before formally executing language interaction. If R5 ≥ 0.6, it must pass language-level deliberate verification. a If (R5) ≥ 0.65, but fails the language-level careful verification, it will be downgraded to R4.

[0100] Step S162: The first level whose acceptance probability reaches or exceeds its corresponding preset threshold is determined as the response level.

[0101] Specifically, starting from the highest level and traversing downwards, the first level that satisfies "acceptance probability ≥ corresponding threshold" is determined as the response level. This "first-to-satisfy" principle allows the robot to prioritize the currently acceptable highest level response. That is, if the user's acceptance level allows, the robot executes a more invasive (and thus more expressive) response; if the high-level condition is not met, it automatically reverts to a lower level.

[0102] Step S163: If none of the acceptance probabilities reach their corresponding preset thresholds, then the lowest level among the multiple progressive response levels is determined as the response level.

[0103] Specifically, if R a (R²) < 0.30, R a (R3)<0.50, R a (R4)<0.55, R a If (R5) < 0.65, then all non-lowest levels do not meet the admission criteria, and the response level is determined to be R1 (silent observation). As the lowest level, R1's core characteristic is "not performing any explicit actions, but only maintaining environmental awareness." That is, when the probability of user acceptance is extremely low, the robot uses silent observation instead of any form of active response, fundamentally avoiding disturbing the user.

[0104] Furthermore, after executing a designated response level, if execution fails or is rejected during execution, the system rolls back to the next lower level. For example, if the current execution of level R5 fails or is rejected, it rolls back to level R4; if level R4 also fails or is rejected, it rolls back to level R3; and so on until level R1. Each rollback is recorded with the reason (e.g., "R5 language-level prudence check failed," "R4 execution was rejected by the social rejection flag"), and this record of the reason provides a data source for subsequent calibration and statistical analysis.

[0105] Those skilled in the art will understand that, through steps S161 to S163, this invention achieves a transformation from a binary "do / don't do" decision to a multi-level progressive decision of "to what extent to do it." The robot no longer simply chooses between responding and not responding, but rather determines the level of response based on the user's acceptance of each level. In low-acceptance scenarios, it maintains the lowest level to avoid disturbance, while in high-acceptance scenarios, it executes a higher-level response. Therefore, this invention eliminates the disturbance problem in low-acceptance scenarios at the mechanism level and enables the robot's response strategy to gradually converge from "model prediction" to "personalized experience-driven" as interaction accumulates, allowing the robot's response accuracy to monotonically improve.

[0106] like Figure 5 As shown, in some embodiments of the present invention, compared to the method in the foregoing embodiments, steps S210 to S250 are further included after step S160. Specifically: Step S210: Execute the determined response profile.

[0107] Specifically, the robot performs the corresponding action according to the response level determined in step S160. For example, if it is determined to be R3 (minimum disturbance preparatory action), it selects an action with the smallest disturbance coefficient from the preparatory action library and performs it; if it is determined to be R4 (non-verbal service action), it performs non-verbal service actions such as playing soft music or turning on ambient lights; if it is determined to be R5 (verbal interaction), it performs an active voice greeting or dialogue.

[0108] Step S220: Collect the target user's reaction signal within a preset time period after execution.

[0109] The preset duration is the observation window, and its duration is related to the response level being executed, with a value range of 1 to 5 seconds.

[0110] The reaction signal refers to the behavioral feedback signal generated by the target user after receiving the robot's response, including but not limited to: gaze direction and duration, changes in torso orientation angle, body movement trends of moving closer or further away, and verbal response content. The reaction signal and the behavioral signal collected in step S110 may overlap in type, but their collection purposes differ. Specifically, step S110 collects behavioral signals to perceive the user's current state, while step S220 collects reaction signals to assess the user's acceptance of the robot's executed response.

[0111] Step S230: Determine whether the reaction signal is a positive or negative sample according to preset rules.

[0112] Specifically, the reaction signals collected in step S220 are classified as positive or negative samples according to the preset judgment rules.

[0113] The criteria for determining a positive sample include any of the following: eye contact with the robot for ≥1 second, torso turning towards the robot, actively approaching the robot, or giving a clear verbal response. These conditions indicate that the user accepts or approves of the robot's response.

[0114] The criteria for identifying negative samples include any of the following: explicit rejection statements (such as "don't" or "go away"), turning the torso away from the robot's direction, or leaving the area. These conditions indicate that the user has rejected or disagreed with the robot's response.

[0115] If the response signal does not meet any of the above criteria for positive and negative samples (i.e., the user has no obvious response or the response is ambiguous), it is judged as an ambiguous sample and is discarded and will not be used for subsequent calibration.

[0116] Step S240: The reaction signals that are determined to be positive or negative samples are used as calibration samples and grouped and stored according to the combination of target user and response level.

[0117] Specifically, the calibration samples that are determined to be positive or negative samples in step S230 are grouped and stored according to the combination of "target user × response level".

[0118] The "Target User × Response Level" group identifies the source of the calibration sample: the sample was collected after the target user u executed response level R. Even if different users execute the same level, their calibration samples are stored in different groups; similarly, if the same user executes different levels, their calibration samples are also stored in different groups. Those skilled in the art will understand that this grouping method ensures that the calibration of each acceptance probability group is independent and highly targeted.

[0119] For example, calibration samples collected by user A when executing R4 mode are stored in group (A, R4), and calibration samples collected by user B when executing R4 mode are stored in group (B, R4). The two groups are statistically analyzed independently and do not affect each other.

[0120] Step S250: When the number of calibration samples in any group reaches a preset lower limit, update the acceptance probability of the corresponding response level according to the proportion of positive samples in that group.

[0121] The preset lower limit ranges from 20 to 50 values.

[0122] In this step, the update method is as follows: calculate the proportion of positive samples in the group (i.e., the number of positive samples divided by the total number of valid samples in the group), and use this proportion as the actual acceptance rate to update the acceptance probability of the response level corresponding to the group.

[0123] For example, the preset lower limit is 30 samples. In group (User A, R4), 30 calibration samples have accumulated, with 21 positive samples and 9 negative samples. Therefore, the proportion of positive samples is 21 ÷ 30 = 0.70. If the predicted acceptance probability for the current R4 level is 0.58, then the actual acceptance rate of 0.70 is fused with the predicted value of 0.58 (e.g., using isotonic regression or piecewise linear fitting) to obtain the updated acceptance probability for the R4 level. This value, after calibration, becomes the updated calibration value for that level, used for subsequent response level determination.

[0124] like Figure 6 As shown, step S250 may further include: Step S251: Obtain the difference between the proportion of positive samples in the group and the current acceptance probability.

[0125] Assume the proportion of positive samples within a group is r, and the current acceptance probability is p (this value is the acceptance probability output by the prediction model in step S150, after previous calibrations). Calculate the difference Δp = rp. If Δp > 0, it means the actual acceptance rate is higher than the current predicted value, and upward correction is needed. If Δp < 0, it means the actual acceptance rate is lower than the current predicted value, and downward correction is needed. If Δp = 0, it means the predicted value is consistent with the actual value, and no correction is needed.

[0126] Step S252: Determine the update step size according to the magnitude of the difference and the preset mapping relationship. The larger the difference, the larger the update step size.

[0127] Specifically, the update step size η is determined based on the absolute value of Δp. The mapping relationship is: the larger |Δp| ​​is, the larger η is; the smaller |Δp| ​​is, the smaller η is. When Δp approaches zero, η approaches a minimum value (but is not zero) to ensure that the system can continuously fine-tune.

[0128] Step S253: Adjust the current acceptance probability with the update step size to obtain the updated acceptance probability.

[0129] Specifically, the current acceptance probability p is combined with the update step size η to obtain the updated acceptance probability p′.

[0130] p′=p+η×sign(Δp), where sign(Δp) is the sign (positive or negative) of Δp, which determines the direction of the correction.

[0131] Furthermore, the update step size is calculated using the following formula: .

[0132] Here, η0 is the base step size, and η0 > 0. η0 is used to ensure that the system can still maintain its ability to fine-tune when the difference Δp is very small.

[0133] Where α is a preset scaling factor, and 0.1≤α≤1.0. α is used to control the amplification factor of the difference term on the step size.

[0134] Where Δp is the difference between the proportion of positive samples and the current acceptance probability (see step S251).

[0135] Where N is the total number of calibration samples within the group.

[0136] The above formula can be synergistically adjusted using the difference Δp and the total number of calibration samples N. Specifically, the larger the difference Δp between the proportion of positive samples and the current acceptance probability, the further the current estimate deviates from reality, thus requiring a larger correction step size. The larger the total number of calibration samples N within a group, the higher the confidence level of the estimate, thus requiring a smaller correction step size to avoid oversensitivity. The two factors (Δp and N) work together in a coupled manner to update the step size, achieving adaptive calibration where "the larger the difference, the faster the correction; the more samples, the more stable the correction."

[0137] For example, suppose η0 = 0.01 and α = 0.5.

[0138] If N=30 and Δp=0.12 in a certain group, then η=0.01+0.5×0.12×1 / 30≈0.02096.

[0139] If N=100 and Δp=0.12 in a certain group, then η=0.01+0.5×0.12×0.1=0.016.

[0140] It is evident that, for the same difference, the larger the sample size, the smaller the step size.

[0141] Based on the foregoing description, those skilled in the art will understand that the present invention, through steps S210 to S250, can perform continuous calibration without disturbing the user. Specifically, by collecting implicit response signals such as eye contact, torso orientation, and proximity / avoidance, explicit voice confirmation is replaced, fundamentally avoiding disturbance to the user during the confirmation process; the "target user × response level" grouping and storage method ensures fine-grained and targeted calibration; and the two-factor coupled step-size update formula makes the calibration results both fast and stable, and the prediction accuracy can be continuously improved with the accumulation of interactions.

[0142] like Figure 7 As shown, in some other embodiments of the present invention, compared with the methods in some embodiments described above, steps S310 to S330 are further included after step S160. Specifically: Step S310: Before determining the robot's response level based on the prediction results, query the veto bit.

[0143] The veto mechanism includes two independent flags: a physical veto and a social veto. The physical veto represents a hard constraint at the physical safety level, typically triggered by the robot's safety sensors (such as collision detection and overload sensors) or environmental sensors. The social veto represents a hard constraint at the social acceptance or authorization level, determined by the authorization status or the output (acceptance probability) of the predictive model.

[0144] In this embodiment, the veto bit is output by an arbitrator independent of the pre-reflective prediction model. This arbitrator does not participate in the reasoning process of the prediction model and can directly generate the veto bit based on external sensor signals and authorization status signals.

[0145] Step S320: When the physical rejection position is set, force the robot to execute the lowest response level and record it.

[0146] In this step, when the physical veto position is active (i.e., the physical veto position is valid), it indicates that a physical safety hazard has been detected (such as the robot detecting a collision, overload, or tipping risk, or environmental sensors detecting an obstacle). The response level is then set to R1 (the lowest level), and the robot immediately stops all active response actions and enters a safe state. Simultaneously, this physical veto event can be recorded (including a timestamp, reason for veto, etc.).

[0147] Step S330: When the social rejection position is set, force the robot to execute the lowest response level and record it.

[0148] In this step, when the social veto position is active (i.e., the social veto position is in a valid state), it indicates that an unacceptable condition at the social level has been detected (such as insufficient authorization from the target user for the current interaction, or the highest acceptance probability being lower than the minimum admission threshold of 0.30). The response level is then set to R1 (the lowest level), and the robot remains in a silent observation state. Simultaneously, this social veto event can be recorded.

[0149] In this embodiment, the difference between a physical veto and a social veto lies in the fact that a physical veto originates from the robot's own safety detection (a hard constraint at the physical level), while a social veto originates from insufficient authorization or acceptance probability (a hard constraint at the social level). Together, they constitute a two-layer decision-making architecture that prioritizes hard constraints over soft predictions. Regardless of the prediction model's output, as long as a veto is in place, the lowest level is enforced to ensure that the robot will never execute a physically unsafe or socially unacceptable response under any circumstances.

[0150] like Figure 8 As shown, the robot 001 of the present invention includes a processor 100 and a memory 200. The memory 200 stores a computer program 201. When the processor 100 executes the computer program 201, it can implement the method described in any of the preceding embodiments.

[0151] The processor 100 may be adapted to execute stored instructions, and the processor 100 may be a single-core processor, a multi-core processor, a computing cluster, or any number of other configurations.

[0152] The memory 200 can provide temporary storage space for the operation of the above instructions during operation. The memory 200 may include random access memory (RAM), read-only memory, flash memory, or any other suitable storage system.

[0153] The computer program 201 used to perform the operations of this invention may be assembly instructions, Instruction Set Architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages ​​and procedural programming languages. The computer program 201 may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can connect to the user's computer via any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or it can connect to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, to perform aspects of the invention, electronic circuits, including, for example, programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), can execute computer-readable program instructions to personalize the electronic circuits by utilizing state information of computer-readable program instructions.

[0154] The technical solutions of the present invention have been described in conjunction with several embodiments above. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is not limited to these specific embodiments. Without departing from the technical principles of the present invention, those skilled in the art can disassemble and combine the technical solutions in the above embodiments, and can also make equivalent changes or substitutions to related technical features. Any changes, equivalent substitutions, improvements, etc., made within the technical concept and / or technical principles of the present invention will fall within the scope of protection of the present invention.

Claims

1. A method for predicting the social intentions of a robot, comprising: Collect behavioral signals from target users; Extract the current behavioral features of the target user from the behavioral signals; Establish and update individual baselines based on the historical behavioral characteristics of the target users; Calculate the deviation between the current behavioral characteristics and the individual baseline; The bias is input into the prediction model to obtain the prediction result; the prediction result does not include intermediate expression results in natural language form; The robot's response level is determined based on the prediction results.

2. The method according to claim 1, wherein, The process of establishing and updating individual baselines based on the historical behavioral characteristics of the target user includes: Obtain the cumulative sample size of the historical behavioral characteristics of the target user; When the cumulative sample size does not reach a preset threshold, the group statistical reference value and the statistical value of the target user's historical behavioral characteristics are combined according to weights to form the individual baseline; the weight of the group statistical reference value decreases as the cumulative sample size increases; Once the cumulative sample size reaches the preset threshold, the individual baseline is updated based on the historical behavioral characteristics of the target user; the individual baseline uses the statistical value of the historical behavioral characteristics within a preset time window as a benchmark.

3. The method according to claim 1, wherein, The deviation is the difference between the value of the current behavioral characteristic and the mean of the individual baseline, divided by the standard deviation of the individual baseline.

4. The method according to claim 1, wherein, The prediction model is a forward-reflective prediction model, and the deviation includes deviation values ​​in multiple dimensions. The step of inputting the deviation into the prediction model to obtain the prediction result includes: The deviation is input into the pre-reflective prediction model; The pre-reflective prediction model uses the deviation value of each dimension of the deviation as the anomaly significance index of that dimension. Based on the relative strength of the anomaly significance indices of each dimension and whether the anomaly significance indices of each dimension appear simultaneously, it generates a comprehensive state representation of the target user. The pre-reflective prediction model directly outputs the prediction results based on the comprehensive state representation. The prediction results include the target user's emotional state distribution, demand distribution, and the probability of acceptance of each candidate response level.

5. The method according to claim 1, wherein, The response levels are multiple progressive response levels that increase sequentially from low to high, and the prediction results include the acceptance probability corresponding to each of the progressive response levels. Determining the robot's response level based on the prediction result includes: Starting from the highest level among the multiple progressive response levels, the acceptance probability of each level is compared with its corresponding preset threshold in turn. The first level where the acceptance probability reaches or exceeds its corresponding preset threshold is determined as the response level; If none of the acceptance probabilities reach their corresponding preset thresholds, then the lowest level among the plurality of progressive response levels is determined as the response level.

6. The method according to any one of claims 1 to 5, further comprising: Execute the determined response level; Within a preset time period after execution, the target user's reaction signal is collected; The reaction signal is determined to be a positive or negative sample according to a preset rule. The reaction signals that are determined to be positive or negative samples are used as calibration samples and grouped and stored according to the combination of the target user and the response level; When the number of calibration samples in any group reaches a preset lower limit, the acceptance probability of the corresponding response level is updated according to the proportion of positive samples in that group.

7. The method according to claim 6, wherein, The step of updating the acceptance probability of the corresponding response level based on the proportion of positive samples within the group includes: Obtain the difference between the proportion of positive samples within the group and the current acceptance probability; The update step size is determined according to the magnitude of the difference and a preset mapping relationship; the larger the difference, the larger the update step size. The current acceptance probability is corrected using the update step size to obtain the updated acceptance probability.

8. The method according to claim 7, wherein, The update step size is calculated using the following formula: ; Where η0 is the base step size, α is the preset scaling factor, Δp is the difference between the positive sample ratio and the current acceptance probability, and N is the total number of calibration samples in the group.

9. The method according to claim 1, further comprising: Before determining the robot's response level based on the prediction results, the veto bits are queried; the veto bits include physical veto bits and social veto bits. When the physical veto position is set, the robot is forced to execute the lowest response level and the result is recorded. When the social veto position is set, the robot is forced to execute the lowest response level and the result is recorded.

10. A robot comprising a processor and a memory, the memory storing a computer program, wherein the processor, when executing the computer program, is capable of implementing the method of any one of claims 1 to 9.