Old-age robot interaction decision-making method and system based on timing intention and reliability weighting

Abstract

The application belongs to the field of service robots and artificial intelligence interaction decision, and discloses an old-age care robot interaction decision method and system based on time sequence intention and reliability weighting, which comprises the following steps: S1. Collecting interaction observation information in an interaction scene within a preset time window, and pre-processing to obtain an implicit interaction intention scalar; S2. Calculating the variance of the implicit interaction intention scalar within the preset time window based on S1, thereby calculating a continuously differentiable reliability weight, weighting the implicit interaction intention scalar to generate a reliability weighted intention scalar; S3. Splicing a user behavior state, the reliability weighted intention scalar, the reliability weight and a service feasibility constraint mask into a robot response strategy state; S4. Inputting the robot response strategy state into a hybrid response strategy network based on proximal policy optimization, and outputting a discrete response type and a continuous response parameter to generate an old-age care robot interaction response instruction. The application reduces misjudgment caused by instantaneous observation through time sequence behavior feature coding.

Description

Technical Field

[0001] This invention belongs to the field of service robot and artificial intelligence interactive decision-making, specifically involving an interactive decision-making method and system for elderly care robots based on temporal intent and reliability weighting. Background Technology

[0002] As intelligent service devices designed for home-based, community-based, and institutional elderly care environments, elderly care robots typically undertake tasks such as daily companionship, status reminders, assistance with items, risk alerts, and emergency calls. Compared to ordinary service robots, elderly care robots interact with users exhibiting more pronounced non-standard interaction characteristics. In real-life scenarios, elderly users may experience unclear speech, delayed reactions, slow movements, tremors, unstable gaze, and unclear intentions to interact. Therefore, elderly care robots need to understand the implicit interaction intentions of users based on continuous behavioral changes over a period of time.

[0003] Current elderly care robots typically rely on voice recognition, touch buttons, preset gesture recognition, or simple anomaly detection for interaction decisions. Voice recognition requires users to clearly express their needs, touch buttons require users to actively complete operations, preset gesture recognition requires user actions to conform to established patterns, and simple anomaly detection usually only detects falls, prolonged stillness, or abnormal postures. These methods are applicable when users can clearly express their needs, but they are prone to missed responses, false responses, and response delays in non-explicit interaction scenarios.

[0004] The key challenge of indirect elderly care interaction lies in the fact that user needs are usually not directly expressed by a single action, but rather formed by a combination of gaze direction, head posture, hand movements, limb displacement, distance to the target object, and the sustained trend of the action. For example, when a user continuously looks at a target object and makes a stable approaching movement, they may have a need for proactive assistance; when a user briefly glances at a target object but the movement is not sustained, it may not constitute a proactive service need; and when a user makes high-frequency changes in movement due to limb tremors, it should not be directly judged as a service request. If the robot judges the current action only based on a fixed threshold, it is easy to identify sporadic actions or observation noise as real needs, leading to false triggers; if the robot is too conservative in its handling of non-standard actions, it is easy to ignore real needs, resulting in response delays.

[0005] Based on this, the interactive decision-making of elderly care companion robots faces the following shortcomings: First, instantaneous observation information is difficult to characterize the continuity and trend of user behavior; second, non-explicit user needs are difficult to transform into calculable intention variables; third, fixed threshold decision-making is prone to frequent switching of response strategies under conditions of jitter, noise, or short-term occlusion; fourth, when the robot performs proactive assistance, it also needs to consider whether the service recipient is accessible, whether the movement is safe, whether the task is feasible, and whether the user's current state allows the service. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an interaction decision-making method and system for elderly care robots based on temporal intent and reliability weighting. The aim is to reduce the impact of short-term jitter, sporadic actions, and observation noise on intent judgment in non-explicit elderly care interaction scenarios, and to improve the response stability and safety of elderly care robots.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The interaction decision-making method for elderly care robots based on temporal intent and reliability weighting includes the following steps: S1. Collect interaction observation information in the interaction scenario within a preset time window, and obtain implicit interaction intention quantity after preprocessing, wherein: the interaction observation information includes the user behavior status of elderly users, the relative relationship information between users and interaction objects, and the service feasibility constraint mask; S2. Calculate the variance of the implicit interaction meaning quantity within a preset time window based on the implicit interaction meaning quantity, and then calculate a continuously differentiable reliability weight based on the variance of the implicit interaction meaning quantity. Softly weight the implicit interaction meaning quantity using the reliability weight to generate a reliability-weighted meaning quantity. S3. Concatenate the user behavior state, the reliability-weighted intention quantity, the reliability weight, and the service feasibility constraint mask to construct the robot response strategy state; S4. Input the robot response strategy state into a hybrid response strategy network based on proximal strategy optimization, output discrete response types and continuous response parameters, thereby generating interactive response commands for the elderly care robot.

[0008] Preferably, in step S1, the user behavior state in the interaction scenario within the preset time window satisfies the following relationship: ; in, For user behavior state vectors, Indicates head posture characteristics. Indicates the characteristics of the direction of the line of sight. Indicates the location features of key points on the hand. Indicates limb displacement characteristics. Indicates the speed characteristics of the action. Indicates the rate of change of action, superscript The transpose operator indicates that the vector is arranged in columns; The relative relationship information sequence includes the relative distance between the user and the target object, the relative distance between the user and the elderly care robot, and the relative distance between the user and the interactive interface; The service feasibility constraint mask is constructed as follows: ; in, This indicates the reachability constraint of the service object. Indicates safety constraints for robot movement. Indicates the execution constraints of interactive tasks. This represents the user's current state constraints. This indicates an emergency response constraint.

[0009] Preferably, in step S1, the interactive observation information is preprocessed to obtain implicit interactive intent quantities in the following manner: S1.1 Construct a user's historical behavior state sequence, relative relationship information sequence, and action change trend sequence based on a preset time window, and splice the user's historical behavior state sequence, relative relationship information sequence, and action change trend sequence to form a user interaction input sequence; S1.2 Input the user interaction input sequence into a unidirectional long short-term memory network and output the user's temporal behavior features; S1.3 Based on the user's temporal behavior characteristics, the implicit interaction intent quantity is inferred through a regression mapping layer and a Sigmoid function.

[0010] Preferably, in step S1.1, the user interaction input sequence is formed in the following manner: ; in, Represents a sequence of user interaction inputs. Represents a sequence of historical user behavior states. Represents a sequence of relative information. Indicates a sequence of trends in action changes. Indicates the preset time window length. Indicates the current time; This is a concatenation function used to concatenate multiple vectors / sequences into a single tensor along their dimensions.

[0011] Preferably, in step S1.2, the user's temporal behavior characteristics are output in the following manner: User interaction input sequence By inputting the data into a unidirectional long short-term memory network, the user's temporal behavioral characteristics are obtained, i.e. ; in: Represents a sequence of user interaction inputs. Represents user temporal behavior characteristics. It is a unidirectional long short-term memory network function that encodes time series sequences and outputs time series feature vectors.

[0012] Preferably, in step S1.3, the implicit interaction intent quantity is obtained in the following manner: User time-series behavior characteristics The input regression mapping layer is used, and the implicit interaction intention quantity is output through the Sigmoid function; that is... in, This represents the Sigmoid function. Represents the regression mapping parameters. Indicates bias. For implicit interaction, icon quantity This represents the user's temporal behavioral characteristics.

[0013] Preferably, in step S2, the continuously differentiable reliability weights and reliability-weighted schematic values ​​are obtained in the following manner: that is, S2.1 Calculate the average number of implicit interaction intents within a preset time window. Variance of implicit interaction intent within the preset time window ; ; ; in, The time window length, for Implicit interaction icon quantity at all times The mean of implicit intent within the window; Var( ) It is a variance function used to calculate the degree of dispersion of implicit intentional quantities within a window; for τ Implicit symbolic quantity at any moment. This represents the mean of implicit intents within the window.

[0014] S2.2 Based on the variance of implicit interaction intentions within a preset time window Calculate the reliability weights for differentiable connections: ; in, It is a natural exponential function; This is the sensitivity coefficient. This is used to adjust the degree of influence of variance on reliability weights; The implicit intent variance within the window. Reliability weight; S2.3 Reliability weight ( ) and the current implicit interaction icon quantity ( Multiplying these together yields the reliability-weighted semantic value, i.e.: ; in, As a reliability weight, The current implicit interaction icon quantity, Weighted indicators for reliability.

[0015] Preferably, in step S3, the robot response strategy state is constructed in the following manner: ; in, For user behavior status, Weighted icons for reliability As a reliability weight, Service feasibility constraint mask; The concatenation function concatenates the four types of information into a policy state vector. This step enables the robot to consider user behavior, intent strength, intent reliability, and service executability simultaneously when making decisions.

[0016] Preferably, in step S4, the discrete response type and continuous response parameters are output in the following manner: ; in, It is a discrete response type. These are continuous response parameters.

[0017] Furthermore, discrete response types include five categories: silent companionship, confirmation inquiry, proactive assistance, risk warning, and emergency call; continuous response parameters include robot movement speed, interaction distance, voice volume, prompt frequency, and movement amplitude.

[0018] Preferably, in step S4, the interactive response commands for the elderly care robot are generated in the following manner: S4.1 Training a hybrid response policy network based on proximal policy optimization, with robot response policy states as input. ; in, ; The objective function of the PPO algorithm is used to optimize the policy network parameters so that the policy is updated towards higher rewards. Learnable parameters of a hybrid response policy network based on proximal policy optimization; The time-step mathematical expectation operator represents the average of samples over all time steps t, used to calculate the expectation of the objective function; This represents the clipping function, which takes the input value as the clipping value. Limited to the interval [ ], that is: Used to limit importance weights The update range should be adjusted to prevent excessive fluctuations in the strategy. For the sake of advantage estimation, For trimming parameters, As the current strategy, This is the old strategy; For the sake of advantage estimation, For trimming parameters, As the current strategy, For the old strategy, The policy state vector; S4.2 Preset the response reward signal for the elderly care robot, that is, ; in, Indicates a matching reward for the response. This indicates that the penalty was triggered accidentally. Indicates penalty for missing response. Indicates penalty for switching actions. This indicates a penalty for violating service constraints. This indicates a penalty for delayed response.

[0019] A decision-making system for elderly care robots based on temporal intent and reliability weighting is proposed, implementing the aforementioned decision-making method steps; the system includes, The interactive observation information acquisition module is used to collect information on the user behavior status of elderly users, the relative relationship between users and interactive objects, and the service feasibility constraint mask. The temporal behavior feature extraction module is used to construct user interaction input sequences and output user temporal behavior features through a unidirectional long short-term memory network. The implicit intent inference module is used to output implicit interaction intent quantities based on user temporal behavior characteristics; The reliability soft-weighting module is used to calculate the implicit interaction intention quantity variance, reliability weight, and reliability-weighted intention quantity within a preset time window; The response strategy state construction module is used to concatenate user behavior state, reliability-weighted intention icon quantity, reliability weight, and service feasibility constraint mask. The hybrid response decision module is used to output discrete response types and continuous response parameters; The response execution module is used to generate and execute interactive response instructions based on discrete response types and continuous response parameters.

[0020] A computing device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it performs the steps of the method described above.

[0021] A storage medium storing a program / instructions that, when executed by a processor, implement the steps of the method described above.

[0022] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0023] Compared with the prior art, the beneficial effects of the present invention are: This invention reduces misjudgments caused by instantaneous observations through temporal behavioral feature encoding; enhances the computability of non-explicit user needs through implicit interaction intent metrics; reduces the impact of short-term jitter, sporadic actions, and observation noise on response decisions through a reliability soft weighting mechanism; reduces the risk of proactive responses being triggered under unexecutable or unsafe conditions through service feasibility constraint masks; and enables the elderly care robot to continuously adjust execution intensity while clearly defining the response category by simultaneously outputting response type and response parameters through a hybrid response strategy. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a flowchart of the preprocessing process in step S1 of the present invention; Figure 3 This is a detailed flowchart of the reliability weights and reliability weighted symbol quantities in step S2 of the present invention; Figure 4 This is a detailed flowchart of generating interactive response instructions in step S3 of the present invention; Figure 5 This is a comparison chart of recognition response performance in a simulated elderly care interaction scenario in Embodiment 2 of the present invention, where (a) is the accuracy rate (%); (b) is the active false trigger rate (%); (c) is the missed response rate (%); and (d) is the insecure response rate (%). Figure 6 This is a comparison chart of the safety and stability performance of the simulated elderly care interaction scenario in Embodiment 2 of the present invention; where (a) is the average response delay and (b) is the action switching frequency. Figure 7 This is a graph showing the relationship between the variance of intent and the reliability weight in this invention. Figure 8 This is a flowchart of each module in Embodiment 3 of the present invention. Detailed Implementation

[0025] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments are described in detail with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand this invention, but does not constitute a limitation of this invention.

[0026] Example 1 like Figure 1-7 As shown, the interaction decision-making method for elderly care robots based on temporal intent and reliability weighting includes the following steps: S1. Collect interaction observation information in the interaction scene within a preset time window, and preprocess it to obtain implicit interaction intent quantities, wherein: the interaction observation information includes the user behavior state of elderly users, the relative relationship information between users and interaction objects (the relative relationship information refers to the relative relationship information at a single moment), and the service feasibility constraint mask; specifically: obtain the user behavior state in the interaction scene within the preset time window through cameras, depth sensors, microphones, distance sensors, or robot body state sensors, which satisfies the following relationship: ; in, For user behavior state vectors, Indicates head posture characteristics. Indicates the characteristics of the direction of the line of sight. Indicates the location features of key points on the hand. Indicates limb displacement characteristics. Indicates the speed characteristics of the action. Indicates the rate of change of action; The relative relationship information sequence includes the relative distance between the user and the target object, the relative distance between the user and the elderly care robot, and the relative distance between the user and the interactive interface. This step solves the problem that single sensor information is insufficient to express the user's interaction state, enabling subsequent models to extract temporal intent information from multiple behavioral features. The service feasibility constraint mask is constructed as follows: ; in, This indicates the reachability constraint of the service object. Indicates safety constraints for robot movement. Indicates the execution constraints of interactive tasks. This represents the user's current state constraints. This represents an emergency response constraint; the service feasibility constraint mask can be represented by binary or continuous variables. For binary variables, a value of 0 or 1 indicates that the corresponding service condition is met, and a value of 0 indicates that the corresponding service condition is not met. For continuous variables, a value between 0 and 1 indicates that the larger the value, the more sufficient the corresponding service condition. In addition, the interactive observation information is preprocessed in the following manner to obtain the implicit interactive intent quantity: S1.1 Construct a user historical behavior state sequence, a relative relationship information sequence, and an action change trend sequence based on a preset time window. Then, concatenate the user historical behavior state sequence, the relative relationship information sequence (which refers to the time-series result of splicing together information of this type from multiple moments within the preset time window in chronological order), and the action change trend sequence to form a user interaction input sequence; specifically, the user interaction input sequence is formed in the following manner: ; in, Represents a sequence of historical user behavior states. Represents a sequence of relative information. Indicates a sequence of trends in action changes. Indicates the preset time window length. This indicates the current moment; this step saves the evolution of user behavior through a time window, reducing the risk of the robot response being directly triggered by a single moment of observation. S1.2 Input the user interaction input sequence into a unidirectional long short-term memory network and output the user temporal behavior features; specifically, the user temporal behavior features are output in the following manner: User interaction input sequence By inputting the data into a unidirectional long short-term memory network, the user's temporal behavioral characteristics are obtained, i.e. ; in: Represents a sequence of user interaction inputs. The network represents the user's temporal behavioral characteristics; the unidirectional long short-term memory network encodes changes in user behavior along the time direction, retaining behavioral evolution information prior to the current moment; preferably, the unidirectional long short-term memory network adopts a single-layer structure with a hidden unit dimension of 128; the input dimension corresponds to the user interaction input sequence. Consistent dimensions; preset time window length The value ranges from 10 to 30 frames, preferably 20 frames; S1.3 Based on the aforementioned user temporal behavior characteristics, implicit interaction intent values ​​are inferred through a regression mapping layer and a sigmoid function; specifically, implicit interaction intent values ​​are obtained in the following manner: User time-series behavior characteristics The input regression mapping layer is used, and the implicit interaction intention quantity is output through the Sigmoid function; that is... in, This represents the Sigmoid function. Represents the regression mapping parameters. Indicates bias. This implicit interaction intention quantity is used to characterize the strength of a user's tendency to initiate service interactions or accept robot assistance within the current interaction window. S2. Calculate the variance of the implicit interaction meaning quantity within a preset time window based on the implicit interaction meaning quantity, then calculate continuously differentiable reliability weights based on the variance of the implicit interaction meaning quantity, and softly weight the implicit interaction meaning quantity using the reliability weights to generate a reliability-weighted meaning quantity; specifically, obtain the continuously differentiable reliability weights and reliability-weighted meaning quantity in the following manner; that is, S2.1 Calculate the average number of implicit interaction intents within a preset time window. Variance of implicit interaction intent within the preset time window ; ; ; in, It is a variance function used to calculate the degree of dispersion of implicit intentional quantities within a window; Let τ be the implicit symbolic quantity at time τ. The mean of implicit intent within the window; S2.2 Based on the variance of implicit interaction intentions within a preset time window Calculate the reliability weights for differentiable connections: ; in, It is a natural exponential function; This is the sensitivity coefficient. This is used to adjust the degree of influence of variance on reliability weights; The implicit intent variance within the window. Reliability weight; Preferably, The value ranges from 0.5 to 2.0, preferably 1.0. When the implicit interaction intention value remains stable within the window, Smaller, reliability weight High; when the amount of implicit interaction icons fluctuates significantly within the window, Larger, reliability weight reduce; Additionally, it should be noted that the reliability weighting logic's algorithm can be described as follows: Input the implicit intent sequence, window length, and sensitivity coefficient within the time window; calculate the average value of the implicit intent sequence within the time window; calculate the variance of the implicit intent sequence within the time window based on the average value; calculate the continuously differentiable reliability weight based on the variance; multiply the reliability weight by the implicit intent at the current moment to obtain the reliability-weighted intent; output the reliability-weighted intent and reliability weight. The above process is merely an exemplary description of the algorithm logic, used to clearly illustrate the technical solution of this invention, and is not intended to limit the specific software implementation. S2.3 Reliability weight ( ) and the current implicit interaction icon quantity ( Multiplying these together yields the reliability-weighted semantic value, i.e.: ; in, As a reliability weight, The current implicit interaction icon quantity, Weighted indicators for reliability.

[0027] This step reduces the impact of unstable intentions on the policy state through continuous soft weighting, thereby avoiding sudden response changes caused by hard threshold switching; S3. Concatenate the user behavior state, the reliability-weighted intention index, the reliability weight, and the service feasibility constraint mask to construct the robot response strategy state; specifically, the robot response strategy state is constructed in the following manner: ; in, For user behavior status, Weighted icons for reliability As a reliability weight, Service feasibility constraint mask; The concatenation function concatenates the four types of information into a policy state vector. This step enables the robot to consider user behavior, intent strength, intent reliability, and service feasibility simultaneously when making decisions; S4. Input the robot response strategy state into a hybrid response strategy network based on proximal strategy optimization, and output discrete response types and continuous response parameters to generate interactive response instructions for the elderly care robot; specifically, output discrete response types and continuous response parameters in the following manner: ; in, It is a discrete response type. The parameters are for continuous response; furthermore, the discrete response types include five categories: silent companionship, confirmation inquiry, proactive assistance, risk warning, and emergency call; the continuous response parameters include robot movement speed, interaction distance, voice volume, prompt frequency, and movement amplitude; preferably, the robot movement speed, interaction distance, voice volume, prompt frequency, and movement amplitude are all normalized to the 0 to 1 range, and then mapped to the actual execution range according to the specific robot hardware parameters.

[0028] In step S4, the interactive response commands for the elderly care robot are generated in the following manner: S4.1 Training a hybrid response policy network based on proximal policy optimization, with robot response policy states as input. ; in, ; The objective function of the PPO algorithm is used to optimize the policy network parameters so that the policy is updated towards higher rewards. Learnable parameters of a hybrid response policy network based on proximal policy optimization; For the sake of advantage estimation, For trimming parameters, As the current strategy, This is the old strategy; PPO training uses a learning rate. Discount factor Cutting parameters Each round of updates occurs 10 times; For the sake of advantage estimation, For trimming parameters, As the current strategy, This is the old strategy; S4.2 Preset the response reward signal for the elderly care robot, that is, ; in, 1 These are weighting coefficients, representing the importance of each reward / penalty item, and their values ​​are non-negative real numbers. Indicates a matching reward for the response. This indicates that the penalty was triggered accidentally. Indicates penalty for missing response. Indicates penalty for switching actions. This indicates a penalty for violating service constraints. Indicates a penalty for response delay; preferably, , The weights of each penalty item can be adjusted between 0.1 and 0.5. This reward signal enables the hybrid response strategy network to simultaneously consider response correctness, response stability, service security, and interaction timeliness during the optimization process.

[0029] In addition, to address the problem of insufficient instantaneous observation information, this invention employs a one-way long short-term memory network to encode the continuous evolution sequence of user behavior, capturing the characteristics of behavior persistence, proximity trends, and short-term fluctuations, thus avoiding reliance solely on observations at the current moment for response judgment.

[0030] To address the difficulty in quantifying non-explicit needs, this invention maps user temporal behavior characteristics to implicit interaction intention quantities between 0 and 1, enabling fuzzy user needs to be directly used by the subsequent response strategy network.

[0031] To address the issue of unstable response strategies, this invention constructs continuously differentiable reliability weights based on the variance of implicit interaction intent quantities within a time window, and uses these reliability weights to apply soft weighting to the current intent quantity. When intent changes are stable, the reliability weights are higher; when intent fluctuates significantly due to jitter, noise, occlusion, or sporadic actions, the reliability weights are lowered, thereby reducing the impact of unstable intents on the response strategy.

[0032] To address the execution risks of proactive services, this invention introduces a service feasibility constraint mask to characterize the reachability of the service object, robot motion safety, interactive task executability, user's current state, and emergency response conditions. This constraint mask, along with a reliability-weighted intent, is input into a hybrid response strategy network, enabling the robot to simultaneously consider user needs, intent credibility, and service executability when generating response instructions.

[0033] Example 2 To illustrate the technical effectiveness of the method, a simulation experiment was conducted to build a sequence of interactive behaviors for elderly care. This simulation example does not use real elderly user data, nor does it represent the results of real robot hardware deployment. The simulation script generates a synthetic behavior sequence according to a fixed random seed and preset behavior rules, and incorporates motion jitter, observation noise, short-term occlusion, label perturbation, and service constraint failure to illustrate the relative effects of the reliability weighting mechanism and the service feasibility constraint mask under simulated perturbation conditions.

[0034] The simulation scenarios include normal silence, brief glance with no need, hand tremor with no need, uncertain request, water assistance, medicine assistance, risk of slow standing up, and abnormal stillness call. Multiple user behavior sequences are generated for each scenario, with sequence features including gaze direction, hand proximity, tremor intensity, posture risk, degree of stillness, observation noise, short-term occlusion, service object accessibility, robot movement safety, and task executability.

[0035] The comparison methods include the threshold rule method, the temporal intent recognition method alone, the method without reliability weighting, the method without service constraint masking, and the method of this invention. Specifically, the threshold rule method does not employ temporal coding, reliability weighting, or service constraint masking; the temporal intent recognition method uses a unidirectional long short-term memory network and intent metrics, but does not employ reliability weighting or service constraint masking; the method without reliability weighting retains temporal coding and service constraint masking, but does not apply reliability weighting to the intent metrics; the method without service constraint masking retains temporal coding and reliability weighting, but does not input a service feasibility constraint mask; and the method of this invention employs temporal coding, reliability weighting, and service feasibility constraint masking simultaneously.

[0036] Evaluation metrics include accuracy (%), active false trigger rate (%), and missed response rate (%). Table 1 is established based on these metrics to simulate the recognition and response performance comparison in elderly care interaction scenarios. Table 2 is established based on the unsafe response rate (%), average response latency, and action switching frequency to simulate the safety and stability performance comparison in elderly care interaction scenarios. Specifically: accuracy measures the proportion of robot output responses that match the user's actual needs; average response latency measures the average time from user behavior sequence input to robot output response instructions; active false trigger rate measures the proportion of the robot generating confirmation inquiries or proactive assistance in scenarios without service demand; missed response rate measures the proportion of genuine assistance, reminder, or call requests that are judged as silent or confirmation inquiries; unsafe response rate measures the proportion of the robot still outputting physical execution responses such as movement, approach, or assistance under conditions where service is infeasible or safety constraints are not met; and action switching frequency measures the stability of the response strategy.

[0037] Table 1 Comparison of recognition response performance in simulated elderly care interaction scenarios Table 2 Comparison of safety and stability performance in simulated elderly care interaction scenarios As shown in Tables 1 and 2, in simulated elderly care interaction scenarios incorporating motion jitter, observation noise, short-term occlusion, and service constraint failure, the accuracy of the method presented in this invention is 94.5%, the active false trigger rate is 3.1%, the missed response rate is 4.9%, the unsafe response rate is 0.7%, the average response latency is 0.23 seconds, and the action switching frequency is 9.3 times per minute. The above data are synthetic simulation results and do not represent actual measured results from real elderly users or real robot hardware. These results illustrate that the reliability-weighted intention icon can mitigate the impact of highly fluctuating intentions on response strategies, and that the service feasibility constraint mask can reduce the risk of active responses under service infeasibility conditions.

[0038] The plot is based on the accuracy, active false trigger rate, missed response rate, and unsafe response rate in Tables 1 and 2. Figure 5This is to visually illustrate the differences between different methods in terms of response accuracy, false trigger control, missed response control, and safety constraint control.

[0039] like Figure 5 As shown, compared to threshold rule methods, temporal intent methods, unreliability weighted methods, and service-constraint-free methods, the method of this invention achieves the highest accuracy and the lowest rates of active false triggering, missed responses, and insecure responses. This result demonstrates that unidirectional temporal behavioral feature extraction, reliability soft weighting, and service feasibility constraint masking can synergistically reduce the interference of occasional actions, short-term jitter, and service infeasibility conditions on the response strategy.

[0040] Furthermore, plotting based on the average response delay and action switching frequency in Table 2. Figure 6 This is to illustrate the impact of the method of the present invention on response timeliness and strategy stability.

[0041] like Figure 6 As shown, the average response latency of the method of this invention is 0.23s, lower than that of the comparative methods; the action switching frequency is 9.3 times per minute, lower than the threshold rule method's 18.9 times per minute, the temporal intent method's 14.9 times per minute, the unreliability-weighted method's 13.5 times per minute, and the unservice-constraint method's 12.6 times per minute. This result indicates that the reliability-weighted intent scale can suppress the direct impact of highly fluctuating intents on the response strategy, thereby reducing frequent response switching of the robot in non-explicit interaction scenarios.

[0042] The sensitivity coefficient of the reliability weight in the simulation is taken as Reliability weights are based on Calculation. To avoid inconsistencies in the data, the reliability weights in Table 3 maintain a monotonic correspondence with the standard deviation of the intention measure; that is, under the same conditions, the higher the standard deviation of the intention measure, the lower the reliability weight.

[0043] Table 3 Reliability weighted diagnostic results in simulated scenarios As shown in Table 3, the standard deviation of the intention quantity in the scenario with no hand tremor is higher than that in most stable behavior scenarios, and its reliability weight is correspondingly reduced. This indicates that the reliability weight can reduce the impact of the original intention quantity on the subsequent response strategy under simulated high-fluctuation input conditions.

[0044] To further illustrate the continuous adjustment relationship of the reliability weight's fluctuation degree in the arbitrary graph, a plot was drawn based on the reliability weight calculation formula. Figure 7 .

[0045] like Figure 7As shown, with a fixed sensitivity coefficient α, the larger the window variance of the implicit interaction intent quantity, the lower the reliability weight; conversely, the smaller the window variance of the implicit interaction intent quantity, the higher the reliability weight. This continuously decreasing relationship enables the robot's response strategy to assign higher response weights to stable intents and suppress highly fluctuating intents caused by short-term jitter, sporadic actions, or observation noise, thereby avoiding sudden response changes caused by hard threshold judgments.

[0046] Example 3 This embodiment also provides an interactive decision-making system for elderly care robots based on temporal intent and reliability weighting, implementing the method steps described in Embodiment 1.

[0047] like Figure 8 As shown, the system includes, The interactive observation information acquisition module 201 is used to collect information on the user behavior status of elderly users, the relative relationship between users and interactive objects, and the service feasibility constraint mask. The temporal behavior feature extraction module 202 is used to construct the user interaction input sequence and output the user temporal behavior features through a unidirectional long short-term memory network; The implicit intent inference module 203 is used to output implicit interaction intent quantities based on user temporal behavior characteristics; The reliability soft-weighting module 204 is used to calculate the implicit interaction intention quantity variance, reliability weight, and reliability-weighted intention quantity within a preset time window; The response strategy state construction module 205 is used to concatenate user behavior state, reliability weighted intention icon quantity, reliability weight, and service feasibility constraint mask; The hybrid response decision module 206 is used to output discrete response type and continuous response parameters; The response execution module 207 is used to generate and execute interactive response instructions based on the discrete response type and continuous response parameters.

[0048] Those skilled in the art will understand that the above-mentioned functional modules can be implemented through software programs, hardware circuits, firmware, or a combination of software and hardware. Any implementation that can accomplish the above-mentioned functions of information acquisition, temporal feature extraction, intent inference, reliability weighting, strategy state construction, and response decision-making falls within the scope of this invention's technical solution.

[0049] Example 4 A computing device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it performs the steps of the method described above.

[0050] A storage medium storing a program / instructions that, when executed by a processor, implement the steps of the method described above.

[0051] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0052] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications or variations that can be made by those skilled in the art without creative effort within the scope of the appended claims are still within the scope of protection of this patent.

Claims

1. A decision-making method for elderly care robots based on temporal intent and reliability weighting, characterized in that, Includes the following steps: S1. Collect interaction observation information in the interaction scenario within a preset time window, and obtain implicit interaction intention quantity after preprocessing, wherein: the interaction observation information includes the user behavior status of elderly users, the relative relationship information between users and interaction objects, and the service feasibility constraint mask; S2. Calculate the variance of the implicit interaction meaning quantity within a preset time window based on the implicit interaction meaning quantity, and then calculate a continuously differentiable reliability weight based on the variance of the implicit interaction meaning quantity. Softly weight the implicit interaction meaning quantity using the reliability weight to generate a reliability-weighted meaning quantity. S3. Concatenate the user behavior state, the reliability-weighted intention quantity, the reliability weight, and the service feasibility constraint mask to construct the robot response strategy state; S4. Input the robot response strategy state into a hybrid response strategy network based on proximal strategy optimization, output discrete response types and continuous response parameters, thereby generating interactive response commands for the elderly care robot.

2. The elderly care robot interaction decision-making method based on temporal intent and reliability weighting according to claim 1, characterized in that: In step S1, the user behavior states in the interaction scenario within the preset time window satisfy the following relationship: ； in, For user behavior state vectors, Indicates head posture characteristics. Indicates the characteristics of the direction of the line of sight. Indicates the location features of key points on the hand. Indicates limb displacement characteristics. Indicates the speed characteristics of the action. Indicates the rate of change of action, superscript The transpose operator indicates that the vector is arranged in columns; The relative relationship information sequence includes the relative distance between the user and the target object, the relative distance between the user and the elderly care robot, and the relative distance between the user and the interactive interface; The service feasibility constraint mask is constructed as follows: ； in, This indicates the reachability constraint of the service object. Indicates safety constraints for robot movement. Indicates the execution constraints of interactive tasks. This represents the user's current state constraints. Indicates emergency response constraints; In step S1, the interactive observation information is preprocessed to obtain implicit interactive intent quantities in the following manner: S1.1 Construct a user's historical behavior state sequence, relative relationship information sequence, and action change trend sequence based on a preset time window, and splice the user's historical behavior state sequence, relative relationship information sequence, and action change trend sequence to form a user interaction input sequence; S1.2 Input the user interaction input sequence into a unidirectional long short-term memory network and output the user's temporal behavior features; S1.3 Based on the user's temporal behavior characteristics, the implicit interaction intent quantity is inferred through a regression mapping layer and a Sigmoid function.

3. The elderly care robot interaction decision-making method based on temporal intent and reliability weighting according to claim 2, characterized in that: In step S2, the continuously differentiable reliability weights and reliability-weighted schematic values ​​are obtained in the following manner; that is, S2.1 Calculate the average number of implicit interaction intents within a preset time window. Variance of implicit interaction intent within the preset time window ; ； ； in, The time window length, for Implicit interaction icon quantity at all times The mean of implicit intent within the window. Var( ) It is a variance function used to calculate the degree of dispersion of implicit intentional quantities within a window; S2.2 Variance of Implicit Interaction Intention Quantity within a Preset Time Window Calculate the reliability weights for differentiable connections: ； in, It is a natural exponential function; This is the sensitivity coefficient. This is used to adjust the degree of influence of variance on reliability weights; The implicit intent variance within the window. Reliability weight; S2.3 Reliability weight ( ) and the current implicit interaction icon quantity ( Multiplying these together yields the reliability-weighted semantic value, i.e.: ； in, As a reliability weight, The current implicit interaction icon quantity, Weighted indicators for reliability.

4. The elderly care robot interaction decision-making method based on temporal intent and reliability weighting according to claim 3, characterized in that: In step S3, the robot response strategy state is constructed as follows: ； in, For user behavior state vectors, Weighted icons for reliability As a reliability weight, Service feasibility constraint mask; The concatenation function concatenates the four types of information into a policy state vector. This step enables the robot to consider user behavior, intent strength, intent reliability, and service executability simultaneously when making decisions.

5. The elderly care robot interaction decision-making method based on temporal intent and reliability weighting according to claim 4, characterized in that: In step S4, the discrete response type and continuous response parameters are output in the following manner: ； in, It is a discrete response type. These are continuous response parameters.

6. The elderly care robot interaction decision-making method based on temporal intent and reliability weighting according to claim 5, characterized in that: In step S4, the interactive response commands for the elderly care robot are generated in the following manner: S4.1 Training a hybrid response policy network based on proximal policy optimization, with robot response policy states as input. ； in, ； The objective function of the PPO algorithm is used to optimize the policy network parameters so that the policy is updated towards higher rewards. Learnable parameters of a hybrid response policy network based on proximal policy optimization; For the sake of advantage estimation, For trimming parameters, As the current strategy, For the old strategy, The policy state vector; S4.2 Preset the response reward signal for the elderly care robot, that is, ； in, 1 These are weighting coefficients, representing the importance of each reward / penalty item, and their values ​​are non-negative real numbers. Indicates a matching reward for the response. This indicates that the penalty was triggered accidentally. Indicates penalty for missing response. Indicates penalty for switching actions. This indicates a penalty for violating service constraints. This indicates a penalty for delayed response.

7. A decision-making system for elderly care robots based on temporal intent and reliability weighting, implemented by the decision-making method described in any one of claims 1-6; characterized in that, The system includes, The interactive observation information acquisition module is used to collect information on the user behavior status of elderly users, the relative relationship between users and interactive objects, and the service feasibility constraint mask. The temporal behavior feature extraction module is used to construct user interaction input sequences and output user temporal behavior features through a unidirectional long short-term memory network. The implicit intent inference module is used to output implicit interaction intent quantities based on user temporal behavior characteristics; The reliability soft-weighting module is used to calculate the implicit interaction intention quantity variance, reliability weight, and reliability-weighted intention quantity within a preset time window; The response strategy state construction module is used to concatenate user behavior state, reliability-weighted intention icon quantity, reliability weight, and service feasibility constraint mask. The hybrid response decision module is used to output discrete response types and continuous response parameters; The response execution module is used to generate and execute interactive response instructions based on discrete response types and continuous response parameters.

8. A computing device, comprising a processor and a memory for storing a processor-executable program, characterized in that: When the processor executes the program stored in the memory, it implements the steps of the method as described in any one of claims 1-6.

9. A storage medium storing a program / instruction, characterized in that: When the program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1-6.

Images