Intelligent pet toy action sequence control method and system based on deep learning

Abstract

The application provides a deep learning-based intelligent pet toy action sequence control method and system. The method comprises the following steps: obtaining voice instruction data of a user, parsing the voice instruction data into a task target for intelligent pet toy action control through a pre-constructed deep learning model; obtaining current execution state data of the intelligent pet toy, wherein the current execution state data comprises a distance state between the user and the pet toy and a posture state of the pet toy body, and generating sequence constraint information; adjusting the weight and correcting the parameters of an action unit in an action template library according to the sequence constraint information to form a target action sequence; and driving the intelligent pet toy to perform continuous actions according to the target action sequence and the execution sequence. The application effectively improves the problems in the prior art, such as mechanical action response, excessive rigidity, uncontrollable random behavior, and difficulty in balancing the main task and the performance of the pet toy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent pet toy motion sequence control, and particularly relates to an intelligent pet toy motion sequence control method and system based on deep learning. Background Technology

[0002] Existing products typically receive user voice commands and trigger preset actions such as sitting, raising a hand, walking, shaking the head, and sticking out the tongue, demonstrating a certain level of technological accumulation in basic interaction and pet-like performance. Existing technologies mainly include solutions that directly map voice recognition results to fixed actions and execute them in a preset order, as well as solutions that superimpose several random actions or facial expressions on top of the fixed actions to enhance fun. While these solutions offer a relatively straightforward implementation path and can meet basic command-based action requirements, they still have significant limitations in practical use. Most existing technologies remain at the level of direct mapping from voice to single actions or fixed action chains, lacking the ability to organize continuous actions around a task objective. This results in toys being able to perform a single action but struggling to form a natural, flowing response process. Meanwhile, some solutions improve the mechanical feel by adding random actions, but the insertion of random actions usually relies on simple selection, polling, or static rules, lacking constraints between the main task and the current execution state. This easily leads to problems such as abrupt action transitions, inappropriate insertion timing, interruptions to the main task's progress, and imbalances in the execution rhythm. Furthermore, intelligent pet toys differ fundamentally from industrial equipment in their control objectives. Their motion control doesn't simply pursue the shortest path and highest repeatability; rather, it requires, within limited structural and actuation conditions, accurately reflecting the user's voice-expressed task intent while dynamically organizing the initiation, pace, and transition of actions based on the relative positions of the user and the toy, as well as the toy's current posture. This ensures the final action performance balances task orientation, natural flow, and a sense of pet-like interaction. The reason existing technologies struggle to achieve this is that they haven't integrated voice task understanding, current state constraints, and action sequence construction into a single continuous control process, nor have they smoothly converted the generated action sequence into control commands directly usable by the actuator.

[0003] Therefore, existing technologies urgently need a motion sequence control method for smart pet toy scenarios, enabling the system to start from voice commands and complete the task objective determination, state constraint generation, target motion sequence construction, and execution control implementation layer by layer, thereby outputting a more coherent, natural, and pet-like interaction response process. Summary of the Invention

[0004] The purpose of this invention is to propose a method and system for controlling the action sequence of intelligent pet toys based on deep learning, thereby solving the above-mentioned problems.

[0005] To achieve the above objectives, a method for controlling the action sequence of an intelligent pet toy based on deep learning is provided in a first aspect of the present invention, the method comprising the following steps: Acquire user voice command data, and parse the voice command data into task objectives for the motion control of smart pet toys through a pre-built deep learning model; The system acquires the current execution status data of the smart pet toy, which includes the distance between the toy and the user and the toy's posture. Based on the task objective and the current execution status data, it generates sequence constraint information, which includes the priority of transition actions and the progress of the main task, to constrain the starting order and pace of the action sequence. The corresponding action template library is invoked according to the task objective, and the action units in the action template library are weighted and their parameters are corrected according to the sequence constraint information to form a target action sequence; wherein, the target action sequence includes action type, execution order, duration parameter and driving parameter; According to the target action sequence, the continuous parameters and drive parameters of each action unit are converted into control instructions for the actuator, and the smart pet toy is driven to perform continuous actions in the execution order.

[0006] Furthermore, the pre-built deep learning model is a semantic parsing model, which consists of a continuously connected feature embedding structure, a temporal coding structure, and a task mapping structure.

[0007] Furthermore, the step of parsing the voice command data into task objectives for the motion control of intelligent pet toys using a pre-built deep learning model specifically involves: The collected voice command data is framed and features are extracted to obtain a voice feature matrix; The speech feature matrix is ​​input into a pre-constructed deep learning model, wherein the temporal coding structure uses gated recurrent units to output semantic vectors, and the task mapping structure maps the semantic vectors to a predefined action task space and selects the task category with the highest probability as the task target; wherein each dimension of the action task space corresponds to a type of executable action task of the smart pet toy.

[0008] Furthermore, the current execution state data is collected in real time by the state sensing unit on the toy body; The distance between users is obtained by continuous sampling by an infrared rangefinder or an ultrasonic rangefinder within a short time window. The controller reads multiple distance values ​​within the window, removes abnormal values ​​with obvious jumps in time sequence, and then takes the intermediate stable value as the current distance measurement result. The attitude state of the toy body is composed of the orientation, forward and backward tilt and left and right tilt output by the inertial measurement unit. The controller also reads the attitude data multiple times within a short time window, and obtains the current attitude result after removing instantaneous spikes. The controller maps the distance between the users and the posture of the toy body into distance deviation and posture deviation, respectively.

[0009] Furthermore, the priority of the transition action is obtained by weighting the attitude deviation amount and the product of the distance deviation amount and the attitude deviation amount; The distance deviation and attitude adaptation correction terms of the main task progress are obtained by weighted combination. The task weight parameters and coupling parameters used in the weighted combination are obtained from the task parameter table based on the task objective.

[0010] Furthermore, the action template library has a preset sequence skeleton for each task target, and each action unit in the sequence skeleton is accompanied by a basic weight, a main task marker, a transition marker, and a micro-interaction marker; Specifically, adjusting the weights and correcting the parameters of the action units in the action template library based on the sequence constraint information involves: The weight adjustment: The base weight of each action unit is combined with a correction coefficient composed of the main task progress and transition action priority in the sequence constraint information and the tags of the corresponding action unit to obtain the actual sequence weight of the corresponding action unit in the current task and current state, and the action units to be retained are selected according to the sequence weight. The parameter correction is as follows: For the retained action units, the persistence parameters of action units with transition markers or micro-interaction markers are extended according to the transition action priority; the driving parameters of action units with main task markers are enhanced according to the main task progress, while the driving parameters of action units with transition markers are weakened according to the transition action priority.

[0011] Further, the step of selecting and retaining action units based on the sequence weights specifically involves: Based on the actual sequence weights of each action unit, all action units are scanned according to the pre-set sequential relationship of their respective action templates. Action units that reach the template threshold are retained, while maintaining their original relative order.

[0012] Furthermore, the step of converting the driving parameters of each action unit into control instructions for the actuator based on the target action sequence specifically involves: Based on the basic control amplitude of the action unit in the drive parameters and execution mapping table, and combined with the actual control amplitude of the previous action unit, a smooth transition calculation is performed to obtain the actual control amplitude of the current action unit. The smooth transition calculation uses a preset smoothing adjustment coefficient to weight and correct the difference between the actual control amplitude of the previous action unit and the basic control amplitude of the current action unit.

[0013] Furthermore, the step of converting the continuous parameters of each action unit into control instructions for the actuator based on the target action sequence specifically includes: The continuous parameter is divided into several fixed control cycles. In each control cycle, time interpolation is performed according to the standard execution curve of the corresponding action in the execution mapping table to generate a continuously changing control signal.

[0014] A second aspect of the invention provides a deep learning-based intelligent pet toy motion sequence control system, the system comprising: The voice parsing module is used to acquire the user's voice command data and parse the voice command data into task objectives for the motion control of smart pet toys through a pre-built deep learning model. The state constraint module is used to acquire the current execution state data of the smart pet toy. The current execution state data includes the distance state between the toy and the user and the posture state of the toy itself. Based on the task objective and the current execution state data, sequence constraint information is generated. The sequence constraint information includes the priority of transition actions and the progress of the main task, which are used to constrain the starting order and progress rhythm of the action sequence. The sequence generation module is used to call the corresponding action template library according to the task objective, and adjust the weights and correct the parameters of the action units in the action template library according to the sequence constraint information to form a target action sequence; wherein, the target action sequence includes action type, execution order, duration parameters and driving parameters; The execution control module is used to convert the continuous parameters and drive parameters of each action unit into control instructions for the actuator according to the target action sequence, and drive the smart pet toy to perform continuous actions in the order of execution.

[0015] The beneficial technical effects of the present invention are at least as follows: This invention reconstructs the action sequence control process of intelligent pet toys in voice interaction scenarios. Its core is not to simply increase the number of actions or add independent functions, but to establish a continuous control link from understanding the voice task to executing the action. Specifically, this invention first parses user voice commands into task objectives for toy motion control, giving the system a clear task orientation rather than remaining at the level of general text or isolated action labels. Then, combining the current distance between the user and the toy, as well as the current posture of the toy itself, it generates sequence constraint information that reflects the initiation method of the action and the rhythm of the main task. This allows the subsequent action organization process to no longer rely on fixed templates or simple random insertions, but to dynamically determine which transitional actions to arrange first, when to enter the main task action, and at what rhythm, based on the current interaction conditions. On this basis, it calls the action template corresponding to the task objective, and selects, rearranges, and adjusts the parameters of the action units according to the sequence constraint information, thereby forming a target action sequence that maintains task consistency and has natural connection and pet-like performance characteristics. Finally, it converts the target action sequence into control commands for the actuators and introduces continuous adjustment during action switching, so that the toy can maintain a smooth transition when performing multiple action units, reducing abrupt changes and abruptness. Through the above technical approach, this invention unifies voice understanding, state adaptation, sequence organization, and execution control into a single solution, enabling smart pet toys to not only complete corresponding tasks when responding to user commands, but also to exhibit a more natural interactive rhythm in terms of action sequence, action duration, and changes in action intensity. This effectively improves the problems of mechanical action response, excessive stiffness, random behavior out of control, and difficulty in balancing the main task with the performance of a simulated pet in existing technologies. Attached Figure Description

[0016] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0017] Figure 1 This is a flowchart of the intelligent pet toy motion sequence control method based on deep learning according to the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] like Figure 1As shown in the figure, the method for controlling the action sequence of an intelligent pet toy based on deep learning provided by an embodiment of the present invention includes: S1. Obtain the voice command data of the user, and parse the voice command data into a task target for controlling the actions of the intelligent pet toy through a pre-constructed deep learning model. Specifically, the core of S1 is to directly convert the voice command data into a task target that can be used for subsequent action sequence control, so that the voice input completes the mapping from the sound signal to the action control semantics after entering the system. The input is voice command data

[0020] collected by a microphone built into the intelligent pet toy. The microphone is usually arranged at the front side or above the head of the toy to cover the voice interaction area directly in front of the user. The collected analog sound wave is converted into a continuous digital signal sequence through an analog-to-digital conversion module. This sequence records the change of the voice amplitude over time according to a fixed sampling frequency. For example, when the user says "come here" within a distance of about one to two meters, a continuous waveform data lasting about one second will be formed. For the convenience of subsequent processing, the continuous voice signal is segmented into frames according to a fixed time window. For example, it is segmented into frames with about 25 milliseconds per frame, and a certain proportion of overlap is reserved between adjacent frames so that the voice transition part can be completely captured. Each frame of voice obtains a spectrum representation through a short-time frequency domain transformation and further undergoes logarithmic compression processing, thereby forming a stable voice feature matrix. This matrix is arranged in a sequence of time frames, and each frame corresponds to a set of frequency feature values, which are used to characterize the energy distribution of the voice in different frequency bands. Further, after obtaining the voice feature matrix, it is input into a semantic parsing model. This model is a deep learning model and consists of a continuously connected feature embedding structure, a time series encoding structure, and a task mapping structure. The feature embedding structure uses a one-dimensional convolutional unit that slides along the time direction to jointly process adjacent voice frames, thereby extracting the local patterns of the voice in the continuous time change. For example, the spectrum change trend between the two syllables "come" will be encoded into a more stable feature representation; after processing, each frame will be mapped into a vector sequence of a fixed length. This vector sequence then enters the time series encoding structure. This structure uses a gated recurrent unit to process the input sequence frame by frame and maintains an internal state vector that is updated over time, forming an overall representation of the semantics of the entire sentence by accumulating the front and back voice information. For example, when the voice gradually inputs "come", "here", and "a moment", the internal state of the model will gradually transition from a single syllable to the expression of a complete action intention, and finally output a semantic vector containing the time series semantic information of the entire segment of voice.

[0021] Further, in the task mapping structure, the semantic vector Mapping to a predefined action target space, where each dimension corresponds to a type of executable action task of the smart pet toy, such as "approaching the user," "raising front paws," "turning head left or right," or "interacting in place." The mapping process is completed through the following calculations: ; in, The semantic vector output by the temporal coding structure is obtained by gradually accumulating the frame-level features of the entire speech by the gated recurrent unit. This is the task mapping weight matrix, whose specific values ​​come from the parameter table that is stored in the device or controller memory after the model is trained, and each row corresponds to a specific action task category. This is a bias term, the value of which also comes from the model parameters after training, and is used to adjust the basic response level for different task categories; This is a normalization mapping operation used to convert the linear mapping result into probability values ​​corresponding to each task category; The specific value of the target value is determined by the category index corresponding to the maximum probability among all categories. Taking a real-world interaction as an example, when a user issues the voice command "Come here," the semantic vector obtained after the above processing... After performing an inner product operation on the weighted rows corresponding to the "proximity to user" category, the result is significantly higher than that of other categories. After bias adjustment and normalization, this category has the highest probability value and is therefore selected as the task target. The objective of this mission It directly corresponds to the main task in the subsequent action sequence control, such as "moving towards and approaching the user", without the need for an additional semantic-to-action conversion process.

[0022] S2. Obtain the current execution status data of the smart pet toy, which includes the distance status between the toy and the user and the posture status of the toy itself. Based on the task objective and the current execution status data, generate sequence constraint information. The sequence constraint information includes the priority of transition actions and the progress of the main task, which are used to constrain the starting order and progress rhythm of the action sequence. Specifically, S2 does not simply process "task categories", but rather "how the current task should be carried out in the current state".

[0023] Furthermore, in the smart pet toy scenario, after the voice parsing determines the task objective of "approaching the user," the controller does not directly write this objective as a fixed forward movement. Instead, it first reads the current execution state data. , and then and This is collectively transformed into a constraint expression of the starting order and progression rhythm of the action sequence. The reason for this is that the action output of a pet toy is not a single-path response like an industrial actuator, but rather requires a balance between task completion, smooth posture, and a pet-like transition. When the user says "come here," if the toy is already facing the user and in a stable position, the action sequence can directly enter the approaching process; if the toy is turned to the side or has just finished a head-turning action, it is more suitable to first arrange a turning, stabilizing posture, or a short pause before entering the approaching action. The sequence constraint information generated in step two... It means representing this "what to do first, what to do next, and to what extent to advance the main action" in a structured way.

[0024] Furthermore, the current execution status data The data is collected in real time by a state sensing unit on the toy itself. Firstly, the forward distance state is obtained by continuous sampling within a short time window using an infrared or ultrasonic rangefinder. The controller reads multiple distance values ​​within this window, discards outliers with significant jumps in time, and then takes the intermediate stable value as the current distance measurement result. Secondly, the body attitude state is composed of the orientation, forward / backward tilt, and left / right tilt output by the inertial measurement unit. The controller also reads multiple attitude data within a short time window, discarding instantaneous spikes to obtain the current attitude result. To ensure that different state variables can be incorporated into a unified constraint calculation process, the controller maps the distance measurement result to a distance deviation. Map the attitude result to attitude deviation. .in, Describe the actual distance between the person and the toy relative to the task objective. The degree of deviation from the corresponding ideal interaction distance; Describe the current orientation and stable state of the machine relative to the mission objective. The required deviation from the initial action posture. The "ideal interaction distance" and "initial action posture" here are not calculated temporarily, but are control parameters pre-configured for each type of task when the action task set is established. For example, the ideal interaction distance range for "approaching the user" is farther than that for "raising forelimbs for interaction," while the initial action posture for "raising forelimbs for interaction" requires a more stable center of gravity for the machine. The controller determines the deviation based on the task objective output in step one. By querying the task parameter table, the ideal distance configuration and initial attitude configuration corresponding to the task can be obtained, and the sensor data can be converted into distance deviation accordingly. and attitude deviation After mapping, and All fall within the same normalization range, which facilitates subsequent combinations.

[0025] Furthermore, it is used to generate sequence constraint information. The fundamental relationship stems from two classic computational approaches. The first part originates from weighted linear combination, which essentially synthesizes multiple state factors based on task relevance; this approach is widely used in state evaluation and decision scoring in control systems. The second part comes from bivariate coupling terms, which essentially describe the joint effect on the outcome when two state factors increase simultaneously. This form is frequently used in statistical modeling and control correction to express situations where "single-factor deviation is acceptable, but additional correction is needed when both factors deviate simultaneously." Considering the movement organization characteristics of pet toys, step two does not directly address the distance deviation. and attitude deviation Instead of performing a single summation, the result is broken down into two interrelated constraint components: one expressing the "priority of transitional actions," and the other expressing the "progress of the main task." The former corresponds to the priority of transitional actions such as turning, posture stabilization, head probing, and limb adduction in the subsequent action sequence, while the latter corresponds to the progression of the main task actions into the early part of the sequence. Thus, the sequence constraint information... Defined as being by and The resulting two components, in which Indicates the priority of transition actions. Indicates the progress of the main task.

[0026] Considering different mission objectives Because of the different sensitivities to distance deviation and attitude deviation, the controller configures two types of task parameters in the task parameter table for each type of task: task weight parameters. and coupling parameters .in, Used to adjust distance deviation With attitude deviation The relative proportion of influence; This is used to adjust the amplification of transitional action priority when both distance and posture deviations increase simultaneously. Taking "approaching the user" as an example... The value is relatively high because the task first requires resolving proximity relationships; taking "forelimb raising interaction" as an example. The value is set relatively low because this task places greater emphasis on attitude stability. Based on the above design, the two components of the sequence constraint information are calculated according to the following formula: ; in, The priority of transitional actions is derived from the amount of posture deviation. and range attitude coupling term The combination; The origin of this concept is the idea of ​​weighted linear combination, which means that the less the task depends on distance advancement, the greater the impact of attitude deviation on the arrangement of transition actions. The source is the bivariate coupling correction, which indicates that when the distance deviation... With attitude deviation At higher levels, subsequent sequences need to further increase the priority of steering, stabilization, or cushioning actions. The first item is the main task progress. Also derived from weighted linear combination, it indicates that the more the task emphasizes distance correction, the stronger the intention to advance the main task is when the distance deviation is greater; the second term Derived from posture adaptation correction, among which This means that the closer the posture is to the ideal starting state, the easier it is for the main task action to directly enter the beginning of the sequence, and This is used to prevent the main task's progress from being prematurely increased in tasks with strong coupling corrections. The distance deviation is obtained by mapping the distance measurement results to the ideal interaction distance configuration for the task. The attitude deviation is obtained by mapping the current machine attitude result to the initial attitude configuration of the mission. and All are the task objectives output from step one. Found in the task parameter table; Depend on Together they form a sequence used for generating subsequent action sequences.

[0027] The logical relationship between these two formulas is continuous. First, the controller bases its decisions on the task objective. Obtain from the task parameter table and Secondly, the controller bases its decisions on the current execution status data. Obtain distance deviation With attitude deviation Then, obtain the results using the above formulas. and Finally, put Combined into sequence constraint information The reasoning here can be directly derived from the needs of motion organization: when the body posture deviates significantly, the sequence is more suitable for arranging transitional movements first, thus requiring a follow-up sequence. The amount that increases with size leads to a basis... The basic items; when both distance deviation and attitude deviation are large, directly advancing the main task action can easily produce abrupt action transitions, therefore, a joint item is added. Prioritize transition actions; simultaneously, to ensure the main task's progress increases when "distance deviation is significant and posture is basically adapted," the main task's progress needs to receive data simultaneously. Positive contributions and The positive contribution from this leads to the second equation. Because , , , Both are normalized pure numerical quantities; the two calculation results and It also remains within a unified constraint expression range and can be directly used as input for subsequent action sequence generation modules.

[0028] In the context of this invention, a specific computational embodiment can be provided. Assume the task objective output in step one is... The task is configured to be "closer to the user" in the task parameter table. , Current execution status data The distance deviation was obtained after processing. This indicates a significant deviation in distance between the toy and the user; the amount of posture deviation. This indicates that the toy's current orientation is significantly off. Substituting these values ​​into the above calculation formula, we get: ; here, This indicates that a certain proportion of transitional movements need to be retained in the initial stage of the sequence, with priority given to turning or stabilizing posture. This indicates that the main task is progressing well, therefore, after a short transition, the subsequent action sequence should quickly move to the approaching action. In another interaction scenario, with the same task objective... The following measurements were obtained , Then, using the same parameters, we can obtain , The corresponding sequence constraints are characterized by "almost no need for specialized steering and posture stabilization actions, allowing for relatively quick entry into short-amplitude approach or subsequent light interactive actions." For example, when the task objective... When switching to "forelimb lifting interaction", the task parameter table can be configured to a lower value. With a high attitude-related coupling strength, even if the distance deviation is large... Medium, as long as the posture deviation is moderate Too high, in the calculation results This will also significantly improve the sequence, allowing subsequent movement sequences to prioritize posture stabilization and center of gravity adjustment before moving on to the forelimb lifting motion. In this way, sequence constraint information... It is no longer a single threshold judgment, but can provide fine-grained action organization tendencies under different tasks and different state combinations.

[0029] The final output is sequence constraint information. This includes the priority of transition actions. Progress relative to main mission The controller will Write to the input buffer of the action sequence generation module, and subsequent steps will read it. Then, it can be based on The insertion intensity of the transition action is determined based on... Determines the pace at which the main task actions enter the sequence.

[0030] S3. Based on the task objective, call the corresponding action template library, and adjust the weights and correct the parameters of the action units in the action template library according to the sequence constraint information to form a target action sequence; wherein, the target action sequence includes action type, execution order, duration parameters and driving parameters. Specifically, S3 uses the task objectives output by S1. and the sequence constraint information output by S2 To form a common input, a target action sequence is generated. .in, The main task type to be completed in this round of interaction is given. Transition action priority and main mission progress The components reflect the proportion of transitional actions and the pace at which the main task actions should enter the sequence in the current state. In this step, the controller calls a pre-stored action template library. This library is not a simple list of actions, but rather tailored to each task objective. The established sequence skeleton set. Taking "approaching the user" as an example, the sequence skeleton can consist of "turning," "probing," "starting," "continuous approaching," and "decelerating and stopping"; taking "raising forelimbs for interaction" as an example, the sequence skeleton can consist of "stabilizing posture," "shifting the center of gravity backward," "raising," "holding," and "retracting." Each action unit is assigned a basic weight when it is entered into the database. It also has three binary tags: the main task tag. Transition markers and micro-interactive tags .in, This indicates whether the action directly leads to the completion of the task. This indicates whether the action serves to steer, stabilize posture, decelerate, or cushion the impact. These markers are used to indicate whether the action is a light, pet-like movement, such as a short pause, a slight head tilt, or a head peek. These markers are configured based on the mechanical structure and the desired action presentation during the action template creation phase and are stored along with the template.

[0031] Furthermore, the first type of calculation relationship used in this step originates from the classic weighted scoring model. This model was originally used to comprehensively score multiple candidate items based on different conditions, and its basic form is "base score plus several conditional adjustment terms". When applying this idea to pet toy action sequences, a single action unit can be regarded as a candidate item, and the result obtained in step two... and This is considered the current condition, and then the weights are adjusted based on the function type of the action unit in the sequence. Based on this original idea, a product adjustment term closely related to the pet toy scenario is further added. Its function is to allow the insertion of light micro-interactions into the sequence during the transition interval when the priority of the transitional action is high and the main task has not yet reached its maximum progress. This makes the action connection more similar to the way a pet observes, hesitates, and probes before executing the main behavior. Following this line of reasoning, the first... Sequence weights of each action unit We obtain the following formula: ; in, For the first The actual sequence weights of each action unit in the current task and current state; This represents the base weight of the action unit in the template; and Sequence constraint information directly from the output of step two ; , and The first The main task marker, transition marker, and micro-interaction marker for each action unit. Because , , , , , All are normalized pure numerical values, so These are still pure numerical quantities of the same type and can be directly used for subsequent filtering and sorting. The derivation of this formula can be understood in three steps: first, using the basic weights... Characterize the importance of action units in the standard template, and then use Strengthen the priority of the main task action within the current task, then use... Strengthen the priority of transition actions in the current state, and finally use... The accessibility of micro-interactions is only increased within the range where a transition is needed and the main task has not yet fully unfolded. This results in... It can not only distinguish between main action and transitional action, but also insert natural little movements into pet toys at appropriate times, making the sequence more lifelike.

[0032] Furthermore, after obtaining the actual sequence weights of each action unit... Afterwards, the controller does not directly shuffle the template order. Instead, it first scans all action units according to the pre-set sequential relationship of each action template, retains action units that reach the template threshold, and maintains their original relative order to ensure the continuity of action logic. Next, the controller modifies the parameters of the retained action units so that the same action exhibits different rhythms and intensities under different constraints. The second type of calculation relationship used in this step comes from the classic linear proportional correction method, the basic idea of ​​which is to conditionally scale the standard parameters obtained in the action calibration stage. For the pet toy action sequence, the standard parameters are not directly copied and executed, but the continuous parameters and driving parameters are differentiated according to the priority of transition actions and the progress of the main task: transition actions and micro-interaction actions in A higher level is more suitable for extending the duration to create a smoother entry; the main action is... Higher elevations are more suitable for enhancing its propulsion performance. Based on this idea, the first... Actual duration parameters of each action unit and actual driving parameters We obtain the following formula: ; in, and These are the standard continuous parameters and standard drive parameters of the action unit in the template, respectively, both of which come from the action calibration records during the equipment commissioning phase. and These are the persistent and driving parameters that are actually written into the target action sequence in the current state; , , , and This is consistent with the meaning of the previous equation. Since all terms within the parentheses are purely numerical scaling factors, the parameter types before and after the correction remain consistent, and the calculation relationship can be directly used for numerical updates within the controller. There is a clear logical relationship between the two equations: the first equation answers "which actions enter the sequence and what their priority is," and the second equation answers "how each action, after entering the sequence, is executed in terms of continuity and drive." In other words, the first equation completes the sequence structure reconstruction, and the second equation completes the sequence performance reconstruction; together, they determine the target action sequence. The content.

[0033] In light of the scenario described in this invention, a complete computational embodiment can be provided. Assume that the task objective output in step one is... To be "closer to the user", the sequence constraint information output in step two for , The action template sequence for this task is "Steering", "Probe", "Start", "Continuous Approach", and "Decelerate and Stop", with their base weights set as follows: , , , , Main task marker In order Transition marker In order Micro-interaction tags In order Substituting into the first equation, the sequence weight for "turning" is: The sequence weight of the "probe" is The sequence weight of "starting point" is The weight of the "continuously close" sequence is: The sequence weight for "deceleration and pause" is: If the threshold for retaining actions in this template is set to... Then all five action units are retained, and a candidate sequence "steering—probe—starting—continuous approach—deceleration and stop" is formed according to the template order. Further assuming that the standard duration parameters of these five action units are as follows: , , , , The standard drive parameters are as follows: , , , , Substituting into the second equation, the actual duration parameter of "steering" is: The actual driving parameters are The actual continuous parameters of the "probe" are: The actual driving parameters are The actual duration parameter for "starting" is: The actual driving parameters are The actual duration parameter for "continuous approach" is: The actual driving parameters are The actual duration parameter for "deceleration and pause" is: The actual driving parameters are The resulting target action sequence The behavior is as follows: first, a gentle turn is completed, followed by a short probing motion, then a strong acceleration to initiate and continuously approach, and finally, a gentle deceleration to stop in front of the user. This result is highly consistent with the expected interaction in a pet toy scenario; the user perceives not simply "moving forward," but a continuous behavior with a rhythm of observation, confirmation, and approach.

[0034] If in another scenario, the same task objective Next step two to obtain , When calculated using the same template, the sequence weight of "probe" will drop to near the template threshold, while the weights of "starting" and "continuous approaching" will further increase, corresponding to a more direct "starting—continuous approaching—deceleration and stopping" sequence. If the task objective is switched to "raising forelimbs interaction," the "stabilizing posture," "slight head tilt," "shifting center of gravity backward," "raising," "holding," and "retracting" sequences in the template will be reconstructed within the same calculation framework into a sequence of first stabilizing posture, then briefly tilting the head, and finally entering the forelimb raising action. This indicates that this step is not a special case treatment for a specific task, but rather revolves around the task objective given in Step 1. and the sequence constraint information given in step two A unified action sequence generation mechanism was established. Ultimately, the controller will retain the action units and their updated versions sequentially. and Write the target action sequence Each sequence element includes an action type, sequence position, actual duration parameter, and actual drive parameter, which are then directly converted into execution control instructions for the next step.

[0035] S4. Based on the target action sequence, convert the continuous parameters and drive parameters of each action unit into control instructions for the actuator, and drive the smart pet toy to perform continuous actions according to the execution order.

[0036] Specifically, S4 outputs the target action sequence from S3. Based on this, the already determined action types, execution order, and duration parameters in the sequence are... and driving parameters Converted into control commands that can be directly applied to the actuator. This process drives the intelligent pet toy to complete continuous action outputs. Step three has completed the organization of the action level, giving each action unit a clear duration and intensity. However, these parameters remain at the abstract control layer and need to be further mapped to underlying drive signals. To achieve this conversion, an execution mapping table is pre-established within the controller. This table is indexed by action type and stores the execution unit and basic control amplitude corresponding to each type of action. And the standard execution curve. Basic control amplitude. Experimental records from the action calibration phase, such as gradually adjusting the input signals of the motor or servo in the test environment, recording the control amplitude range that can stably complete the action, and storing the intermediate value as the standard value; the standard execution curve is used to describe the trend of the action in the time dimension, such as slow acceleration in the initial stage, stability in the middle stage, and gradual deceleration in the final stage.

[0037] Furthermore, in the action sequence When converting to control instructions, the action units are first read sequentially according to the sequence order. For the first... Each action unit extracts its basic control amplitude from the execution mapping table. And combined with the driving parameters output in step three Calculate the target control amplitude for the action in the current state. This calculation relationship is derived from the classic proportional control principle, which states that the control output is proportional to the target amplitude. Based on this, a transition adjustment term is introduced between adjacent actions to ensure the control signal remains continuously variable during action switching. The specific calculation is as follows: ; in, For the first The actual control amplitude of each action unit is directly used to generate drive signals; From step three, this indicates the intensity adjustment ratio of the action relative to the standard action; The basic control amplitude is derived from the execution mapping table. This is the actual control amplitude of the previous action unit, calculated from the previous round; The smoothing adjustment coefficient, whose value is set during system initialization, controls the degree to which the current action inherits from the previous action. The derivation of this formula is based on two fundamental ideas: one is the proportional control principle, through... The first is to amplify or reduce the amplitude of standard movements; the second is the concept of first-order smoothing, achieved by introducing... This ensures continuity in the transition from the current motion amplitude to the target amplitude, thereby preventing abrupt changes in the actuator. Because... , and Since they are all of the same type of control amplitude, the calculation results are... Maintaining physical consistency allows it to be directly used as execution input.

[0038] Furthermore, in the range of control Then, the controller further determines the parameters based on the continuous parameters. Generate time-series control signals. Duration parameter. From step three, the value indicates the length of time the action should last in the sequence. The controller will... The process is divided into several fixed control cycles. Within each cycle, the control amplitude of the current action is output, and time interpolation is performed according to the standard execution curve of the corresponding action in the execution mapping table to obtain a continuously changing control signal. For example, for a "forward" action lasting several control cycles, the control signal can be generated in a "low amplitude—gradually increasing—stable output—gradually decreasing" manner, rather than a constant output. This avoids mechanical shock and improves the smoothness of the action. The control signals of multiple consecutive action units are connected end-to-end on the time axis. The smoothing calculation relationship described above ensures a continuous amplitude transition between adjacent actions, thus forming a complete sequence of execution control instructions. .

[0039] To illustrate the specific effects of this method, we can use a set of typical data for calculation. Assume a target action sequence. It contains three continuous motion units, whose driving parameters are as follows: , , The corresponding basic control amplitudes are respectively , , Smoothing adjustment coefficient For the first action unit, since there is no previous action, take... Then there is The control amplitude of the second action unit is The control range of the third action unit is It can be seen that although the second action corresponds to a greater driving demand, its amplitude of control does not increase abruptly, but rather gradually transitions from a lower amplitude in the previous action to a higher level; the third action, while maintaining continuity, gradually reduces its amplitude, making the entire sequence exhibit a smooth acceleration and deceleration process during execution. This variation is similar to the behavioral rhythm of a pet when moving or interacting, which can effectively enhance the naturalness of the interaction.

[0040] Ultimately, the controller uses the time-deployed sequence of control signals as the execution control commands. The output is sent to the drive modules of each execution unit. The drive modules control the motors or servos based on the received control amplitude and timing signals, thereby causing the toy to follow the target motion sequence. Complete continuous actions. By combining the action sequence parameters generated in step three with the execution mapping table and the proportional smoothing control method, this step achieves a stable transformation from abstract action descriptions to specific execution signals, making the transitions between actions natural, the execution process smooth, and accurately reflecting the behavioral characteristics defined by the task objectives and sequence constraints.

[0041] This invention also provides a deep learning-based intelligent pet toy motion sequence control system, the system comprising: The voice parsing module is used to acquire the user's voice command data and parse the voice command data into task objectives for the motion control of smart pet toys through a pre-built deep learning model. The state constraint module is used to acquire the current execution state data of the smart pet toy. The current execution state data includes the distance state between the toy and the user and the posture state of the toy itself. Based on the task objective and the current execution state data, sequence constraint information is generated. The sequence constraint information includes the priority of transition actions and the progress of the main task, which are used to constrain the starting order and progress rhythm of the action sequence. The sequence generation module is used to call the corresponding action template library according to the task objective, and adjust the weights and correct the parameters of the action units in the action template library according to the sequence constraint information to form a target action sequence; wherein, the target action sequence includes action type, execution order, duration parameters and driving parameters; The execution control module is used to convert the continuous parameters and drive parameters of each action unit into control instructions for the actuator according to the target action sequence, and drive the smart pet toy to perform continuous actions in the order of execution.

[0042] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0043] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0044] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0045] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A deep learning-based intelligent pet toy motion sequence control method, characterized in that, The method includes: Acquire user voice command data, and parse the voice command data into task objectives for the motion control of smart pet toys through a pre-built deep learning model; The system acquires the current execution status data of the smart pet toy, which includes the distance between the toy and the user and the toy's posture. Based on the task objective and the current execution status data, it generates sequence constraint information, which includes the priority of transition actions and the progress of the main task, to constrain the starting order and pace of the action sequence. The corresponding action template library is invoked according to the task objective, and the action units in the action template library are weighted and their parameters are corrected according to the sequence constraint information to form a target action sequence; wherein, the target action sequence includes action type, execution order, duration parameter and driving parameter; According to the target action sequence, the continuous parameters and drive parameters of each action unit are converted into control instructions for the actuator, and the smart pet toy is driven to perform continuous actions in the execution order.

2. The method for controlling the action sequence of an intelligent pet toy based on deep learning according to claim 1, characterized in that, The pre-built deep learning model is a semantic parsing model, which consists of a continuously connected feature embedding structure, a temporal coding structure, and a task mapping structure.

3. The method for controlling the action sequence of an intelligent pet toy based on deep learning according to claim 2, characterized in that, The process of parsing the voice command data into task objectives for the motion control of intelligent pet toys using a pre-built deep learning model specifically involves: The collected voice command data is framed and features are extracted to obtain a voice feature matrix; The speech feature matrix is ​​input into a pre-constructed deep learning model, wherein the temporal coding structure uses gated recurrent units to output semantic vectors, and the task mapping structure maps the semantic vectors to a predefined action task space and selects the task category with the highest probability as the task target; wherein each dimension of the action task space corresponds to a type of executable action task of the smart pet toy.

4. The method for controlling the action sequence of an intelligent pet toy based on deep learning according to claim 1, characterized in that, The current execution status data is collected in real time by the status sensing unit on the toy body; The distance between users is obtained by continuous sampling by an infrared rangefinder or an ultrasonic rangefinder within a short time window. The controller reads multiple distance values ​​within the window, removes abnormal values ​​with obvious jumps in time sequence, and then takes the intermediate stable value as the current distance measurement result. The attitude state of the toy body is composed of the orientation, forward and backward tilt and left and right tilt output by the inertial measurement unit. The controller also reads the attitude data multiple times within a short time window, and obtains the current attitude result after removing instantaneous spikes. The controller maps the distance between the users and the posture of the toy body into distance deviation and posture deviation, respectively.

5. The method for controlling the action sequence of an intelligent pet toy based on deep learning according to claim 4, characterized in that, The priority of the transition action is obtained by weighting the attitude deviation and the product of the distance deviation and the attitude deviation. The distance deviation and attitude adaptation correction terms of the main task progress are obtained by weighted combination. The task weight parameters and coupling parameters used in the weighted combination are obtained from the task parameter table based on the task objective.

6. The method for controlling the action sequence of an intelligent pet toy based on deep learning according to claim 1, characterized in that, The action template library has a preset sequence skeleton for each task objective. Each action unit in the sequence skeleton is accompanied by a basic weight, a main task marker, a transition marker, and a micro-interaction marker. Specifically, adjusting the weights and correcting the parameters of the action units in the action template library based on the sequence constraint information involves: The weight adjustment: The base weight of each action unit is combined with a correction coefficient composed of the main task progress and transition action priority in the sequence constraint information and the tags of the corresponding action unit to obtain the actual sequence weight of the corresponding action unit in the current task and current state, and the action units to be retained are selected according to the sequence weight. The parameter correction is as follows: For the retained action units, the persistence parameters of action units with transition markers or micro-interaction markers are extended according to the transition action priority; the driving parameters of action units with main task markers are enhanced according to the main task progress, while the driving parameters of action units with transition markers are weakened according to the transition action priority.

7. The method for controlling the action sequence of an intelligent pet toy based on deep learning according to claim 6, characterized in that, The step of selecting and retaining action units based on the sequence weight specifically involves: Based on the actual sequence weights of each action unit, all action units are scanned according to the pre-set sequential relationship of their respective action templates. Action units that reach the template threshold are retained, while maintaining their original relative order.

8. The method for controlling the action sequence of an intelligent pet toy based on deep learning according to claim 1, characterized in that, The step of converting the driving parameters of each action unit into control instructions for the actuator based on the target action sequence specifically includes: Based on the basic control amplitude of the action unit in the drive parameters and execution mapping table, and combined with the actual control amplitude of the previous action unit, a smooth transition calculation is performed to obtain the actual control amplitude of the current action unit. The smooth transition calculation uses a preset smoothing adjustment coefficient to weight and correct the difference between the actual control amplitude of the previous action unit and the basic control amplitude of the current action unit.

9. The method for controlling the action sequence of an intelligent pet toy based on deep learning according to claim 1, characterized in that, The step of converting the continuous parameters of each action unit into control instructions for the actuator based on the target action sequence specifically includes: The continuous parameter is divided into several fixed control cycles. In each control cycle, time interpolation is performed according to the standard execution curve of the corresponding action in the execution mapping table to generate a continuously changing control signal.

10. A deep learning-based intelligent pet toy motion sequence control system, characterized in that, The system includes: The voice parsing module is used to acquire the user's voice command data and parse the voice command data into task objectives for the motion control of smart pet toys through a pre-built deep learning model. The state constraint module is used to acquire the current execution state data of the smart pet toy. The current execution state data includes the distance state between the toy and the user and the posture state of the toy itself. Based on the task objective and the current execution state data, sequence constraint information is generated. The sequence constraint information includes the priority of transition actions and the progress of the main task, which are used to constrain the starting order and progress rhythm of the action sequence. The sequence generation module is used to call the corresponding action template library according to the task objective, and adjust the weights and correct the parameters of the action units in the action template library according to the sequence constraint information to form a target action sequence; wherein, the target action sequence includes action type, execution order, duration parameters and driving parameters; The execution control module is used to convert the continuous parameters and drive parameters of each action unit into control instructions for the actuator according to the target action sequence, and drive the smart pet toy to perform continuous actions in the order of execution.

Images