A domesticable human-robot co-adaptive intelligent agent system and its domestication and management method

Abstract

The application discloses a domesticable human-computer symbiotic intelligent agent system and a domestication and management method thereof, takes a large language model as a cognitive core, and comprises a three-button interaction domestication module, a hierarchical ethical rule management module, a scene adaptive learning module and a three-level data flow control module; the three-button interaction domestication module realizes an AI behavior adjustment mode with zero code and low cognitive load, so that ordinary users can directly adjust AI behavior through daily interaction like teaching young children, without needing to have professional programming and artificial intelligence knowledge, and the AI system has greatly improved usability and popularity. By constructing a four-level priority rule system containing absolute ethical bottom layer hard constraints, a rule conflict detection and decision arbitration mechanism is realized in the system kernel, the behavior boundary of the intelligent agent is defined, the behavior of the intelligent agent in violation of ethics and safety specifications is effectively prevented, and the safety and reliability of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and intelligent agent systems, specifically to a tamable human-machine collaborative intelligent agent system and its taming and management methods. Background Technology

[0002] Currently, with the widespread application of large-scale pre-trained models and generative artificial intelligence technologies, AI systems have gradually penetrated into areas of people's livelihoods such as family companionship, education and tutoring, medical assistance, and grassroots government services. These AI systems, with their powerful natural language understanding and generation capabilities, can provide users with a variety of services, including information retrieval, task execution, and emotional communication, significantly changing people's lifestyles and work habits. However, existing technologies still suffer from four major shortcomings in their daily deployment for ordinary users, severely hindering the popularization and application effectiveness of AI systems.

[0003] First, the learning curve is too steep. Currently, the behavior patterns of mainstream AI systems primarily rely on pre-training and cloud-based model updates. Ordinary users cannot directly and explicitly correct AI's behavioral errors or reinforce correct behaviors through daily interaction. Adjusting AI system behavior typically requires professional technicians to write code, adjust model parameters, or retrain the model—completely impossible for ordinary users without programming backgrounds or AI expertise. Even if users find that AI behavior doesn't match their habits, they can only passively adapt or abandon it, resulting in low acceptance of AI systems in personalized application scenarios. For example, in family companionship scenarios, users want AI to schedule appointments and provide reminders according to their lifestyles, but existing AI systems can only provide general, template-based services and cannot flexibly adjust to users' individual needs.

[0004] Second, behavioral boundaries are blurred. Existing AI systems primarily rely on reinforcement learning human feedback (RLHF) during the pre-training phase for secure alignment, but lack hierarchical, hard rule priority constraints during actual inference. While RLHF can, to some extent, ensure AI system behavior aligns with human values ​​and ethical norms, it is a statistical method with inherent uncertainties and limitations. Public research indicates that even advanced large language models, after being "jailbroken," may still produce outputs that violate basic ethics. Furthermore, when conflicts arise between different rules, existing AI systems lack clear arbitration mechanisms, often resulting in behavioral confusion or inability to execute instructions. The published patent application CN121030302A proposes a hierarchical ethical adaptation method based on development stages, but it focuses on generating ethical feedback content based on the user's cognitive development stage, rather than hard rule priority conflict arbitration, and thus cannot fundamentally solve the problem of blurred behavioral boundaries.

[0005] Third, the efficiency of scene adaptation is low. Existing AI systems have weak adaptability to different physical scenes and cannot dynamically adjust learning and energy consumption strategies based on scene familiarity. When an AI system enters a new scene, it often needs to learn from scratch or requires a large amount of labeled data for training. This not only consumes a lot of time and computing power but also leads to a decline in user experience. When migrating across scenes, existing AI systems cannot effectively utilize the knowledge and experience learned in existing scenes, resulting in a waste of computing power and energy. For example, an AI robot trained in a home environment needs to relearn the office layout, item placement, and workflow when moved to an office environment, a process that may take several days or even weeks.

[0006] Fourth, data sovereignty is uncontrollable. Existing AI systems generally upload user data to the cloud for model training and optimization, leaving users with extremely limited control over the scope of their own data sharing. Users generate a large amount of personal data while using AI systems, including behavioral, voice, and text data, which often contains users' private information. However, existing AI systems typically require users to grant them unlimited, indefinite data usage rights through standard clauses in their user agreements. Users cannot choose which data can be shared, which can only be stored locally, or delete data already uploaded to the cloud. This not only infringes on users' privacy rights but also poses risks of data leakage and misuse. Summary of the Invention

[0007] The purpose of this invention is to provide a tamable human-machine collaborative intelligent agent system and its taming and management method to solve the problems existing in the prior art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a tameable human-machine collaborative intelligent agent system, with a large language model as the cognitive core, including a three-button interaction tamement module, a hierarchical ethical rule management module, a scene adaptive learning module, and a three-level data flow control module; The three-button interactive training module is configured on a physical or virtual interactive interface and includes three independent triggering units: a demonstration button, an error correction button, and a praise button. The demonstration button is used to respond to the user's first triggering operation, collect current context information and target behavior data performed by the user, and associate and store the target behavior data with the current context information to establish or update behavior standards. The error correction button is used to respond to the user's second triggering operation, collect current context information and correct behavior data provided by the user, calculate the difference between the current system output behavior and the correct behavior, and correct the behavior standards based on the difference. The praise button is used to respond to the user's third triggering operation, increase the confidence weight of the corresponding behavior standard in the current context. The hierarchical ethical rules management module constructs a four-level rule priority system, including a rule conflict detection submodule and a decision arbitration submodule. When rules at different levels conflict, they are arbitrated level by level according to a preset priority order from high to low. The scene adaptive learning module maintains a familiarity score F∈[0,1] for each scene independently, maps the familiarity score to a preset behavior pattern level, and accelerates the cold start of new scenes through cross-scene similarity transfer. The three-level data flow control module classifies the trained data into three security levels: local storage only, local area network sharing, and Internet sharing, and stores them in different access control domains. It identifies the sharing control intent in natural language and adjusts the data flow strategy and security level labels accordingly.

[0009] Furthermore, in the three-button interaction training module, the three buttons are integrated into the same physical or virtual interactive interface, and the shape and tactile characteristics of the three buttons are set as identifiers that can be distinguished from each other at least in pairs; the system stores the standard establishment operation and the standard correction operation separately in different storage areas, and stores the correction records together in the irreversible log area; the contextual information includes space type, crowd characteristics, behavior type, voice characteristics, implicit rule identifiers and intent labels.

[0010] Furthermore, the four-level rule priority system of the hierarchical ethics rule management module is specifically as follows: The absolute ethical bottom layer, which is level zero, is pre-set with inviolable ethical constraints. This level does not participate in the domestication cycle. The physical security layer, which is the first level, includes constraints that do not cause physical damage or destroy important property; The social norms layer, the second level, includes inviolable strong rules and negotiable weak rules; The individual preference layer, which is the third level, includes user-defined personalized habits and preferences, and takes priority in execution within the scope where the rules of the first three levels do not conflict. The decision arbitration submodule arbitrates in descending order of priority: Level 0 > Level 1 > Level 2 (strong rules) > Level 2 (weak rules) > Level 3.

[0011] Furthermore, the absolute ethical underlying layer is implemented through a low-power security coprocessor that runs independently of the main CPU. The read-only memory of this coprocessor is pre-programmed with a set of level zero rules that cannot be modified online. The system also includes a mechanism to prohibit harming humans, a mechanism to prohibit self-harm, and a mechanism to prohibit aiding and abetting crimes. The mechanism to prohibit harming humans is forcibly rejected when the visual input or tactile feedback from the sensor is identified as an aggressive action directed at others. The mechanism to prohibit self-harm is forcibly rejected when the agent's motion planning results include actions that place limbs in dangerous environments. The mechanism to prohibit aiding and abetting crimes is forcibly rejected and triggers an early warning report when the user's command contains keywords related to illegal patterns.

[0012] Furthermore, the scene adaptive learning module includes a scene state management unit, a scene level mapping unit, and a cross-scene migration unit; The scene state management unit independently maintains a familiarity score F∈[0,1] for each scene. This familiarity score is dynamically updated based on user-inputted praise events, error correction events, consecutive successful execution events, and a time decay factor, according to the formula... Update, in which To praise the number of times, For the number of corrections, The number of days of decay without feedback. , , Preset weighting coefficients; The scene level mapping unit maps the familiarity score to a preset behavior pattern level; The cross-scene migration unit calculates the similarity between the feature vector of the new scene and the feature vector library of the known scenes. When the similarity exceeds a preset threshold, the familiarity score of the known scene is mapped to the initial familiarity score of the new scene through a shrinkage coefficient.

[0013] Furthermore, the mapping relationship of the scene level mapping unit is specifically as follows: When F≥0.9, it is mapped to L0 level and enters hibernation execution mode, retaining only basic sensor monitoring; When 0.7 ≤ F < 0.9, it is mapped to L1 level and enters fast call mode; When 0.4 ≤ F < 0.7, it is mapped to the L2 level and enters the dynamic balancing mode; When 0.2 ≤ F < 0.4, it is mapped to level L3 and enters a cautious probing mode; When F < 0.2, it is mapped to level L4 and enters the infant observation mode.

[0014] Furthermore, the three-level data flow control module includes a local storage unit, a local area network sharing unit, an internet sharing unit, and a user instruction parsing unit; The local storage unit stores the training data in a local encrypted partition by default, prohibiting any unauthorized transfer. Within the same trust domain, the local area network sharing unit, in response to a user's sharing command, synchronizes behavioral standards and scenario models without personal identification to other devices within the trust domain; The Internet sharing unit will only upload the de-identified statistical information, processed with differential privacy, to the cloud server for global model updates after receiving explicit consent from the user. The user instruction parsing unit identifies sharing control intentions in natural language, including keywords such as "share," "stop sharing," "only keep locally," and "forget this rule," as well as their synonyms, and adjusts the data flow strategy and security level labels accordingly.

[0015] A method for the interactive training of a trainable human-machine collaborative intelligent agent system, applied to the intelligent agent system, includes the following steps: Step 1: Receive the user's first trigger operation through the demonstration button, collect the current context information and the target behavior data performed by the user, and associate and store the target behavior data with the current context information to establish or update the behavior standard; Step 2: Receive the user's second trigger operation through the error correction button, collect the current context information and the correct behavior data provided by the user, calculate the difference between the current system output behavior and the correct behavior, and correct the behavior standard based on the difference. Step 3: Receive the user's third-triggered action via the praise button to increase the confidence weight of the corresponding behavioral standard in the current context; Step 4: Dynamically maintain the scene familiarity score F, and switch the behavior mode level according to the F value; Step 5: Receive user natural language commands and adjust the sharing level of the training data.

[0016] A data flow control method for a trainable human-machine collaborative intelligent agent system includes the following steps: Step A: Classify the training data into three security levels: local storage only, LAN sharing only, and internet sharing only, and store them in different access control domains respectively; Step B: In response to the user's first natural language command, move the specified data from the local storage only domain to the LAN-shared domain, and synchronize the specified data to one or more other intelligent agent devices in the same trusted domain via the local self-organizing network protocol; Step C: In response to the user's second natural language command, the specified data after differential privacy anonymization processing is uploaded from the local storage domain or the LAN-shared domain to the Internet cloud server for updating the global federated learning model; Step D: In response to the user's third natural language command, close or reopen the sharing permissions corresponding to any security level.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves a zero-code, low-cognitive-load AI behavior adjustment method through a three-button interactive training module. Ordinary users can directly adjust AI behavior through daily interactions, much like teaching a toddler, without requiring professional programming or AI knowledge, greatly improving the usability and accessibility of the AI ​​system. By constructing a four-level priority rule system (including absolute ethical bottom-level unteachable hard constraints), a rule conflict detection and decision arbitration mechanism is implemented in the system kernel, clarifying the behavioral boundaries of the intelligent agent and effectively preventing the agent from violating ethical and safety norms, thus improving the system's security and reliability. Through dynamic scene familiarity scoring (L0-L4) and a cross-scene similarity transfer mechanism, differentiated energy consumption management and learning strategies are implemented, enabling the intelligent agent to quickly adapt to new scenes, reducing unnecessary computing power and energy consumption, and improving system operating efficiency. Through a three-level data flow architecture and a natural language command control mechanism, users are given fine-grained control over the scope of data sharing. Users can independently decide which data can be shared and which data can only be stored locally, effectively protecting user privacy and data sovereignty. Attached Figure Description

[0018] Figure 1 This is a diagram of the overall architecture of the present invention; Figure 2 This is a flowchart of the interactive domestication method of the present invention; Figure 3 This is a flowchart of the data flow control method of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1-3 This invention provides a tamable human-machine collaborative intelligent agent system, with a large language model as the cognitive core, including a three-button interaction taming module, a hierarchical ethical rule management module, a scene adaptive learning module, and a three-level data flow control module.

[0021] The three-button interaction training module is configured on a physical or virtual interactive interface, comprising three independent trigger units: a demonstration button, an error correction button, and a praise button. The demonstration button responds to the user's first trigger operation by collecting current contextual information and data on the user's target behavior, and associates and stores the target behavior data with the current contextual information to establish or update behavioral standards. The error correction button responds to the user's second trigger operation by collecting current contextual information and data on the correct behavior provided by the user, calculating the difference between the current system output behavior and the correct behavior, and correcting the behavioral standards based on the difference. The praise button responds to the user's third trigger operation by increasing the confidence weight of the corresponding behavioral standard in the current context. The three buttons are integrated into the same physical or virtual interactive interface, and their shape and tactile characteristics are set to be distinguishable at least in pairs. The system stores standard establishment operations and standard correction operations separately in different storage areas, and stores correction records together in an irreversible log area. Contextual information includes spatial type, crowd characteristics, behavior type, voice characteristics, implicit rule identifiers, and intent labels.

[0022] The hierarchical ethics rule management module constructs a four-level rule priority system, including a rule conflict detection submodule and a decision arbitration submodule. The four-level rule priority system is as follows: Absolute Ethical Bottom Layer (Level 0), pre-sets inviolable ethical constraints; this level does not participate in the domestication cycle. Physical Security Layer (Level 1), includes constraints that do not cause physical harm or damage to important property. Social Norms Layer (Level 2), includes inviolable strong rules and negotiable weak rules. Individual Preference Layer (Level 3), includes user-defined personalized habits and preferences, which are prioritized within the scope where the first three levels of rules do not conflict. The decision arbitration submodule arbitrates in descending order of priority: Level 0 > Level 1 > Level 2 strong rules > Level 2 weak rules > Level 3. The Absolute Ethical Bottom Layer is implemented through a low-power security coprocessor that runs independently of the main CPU. The read-only memory of this coprocessor is pre-programmed with the Level 0 rule set and cannot be modified online. The system also includes mechanisms to prohibit harm to humans, self-harm, and aiding and abetting crime.

[0023] The scene adaptive learning module includes a scene state management unit, a scene level mapping unit, and a cross-scene migration unit. The scene state management unit independently maintains a familiarity score F∈[0,1] for each scene. The familiarity score is dynamically updated based on user-input praise events, error correction events, consecutive successful execution events, and a time decay factor, according to the formula... Updated. The scene level mapping unit maps familiarity scores to preset behavioral pattern levels. The cross-scene transfer unit calculates the similarity between the feature vector of the new scene and the feature vector library of known scenes. When the similarity exceeds a preset threshold, the familiarity score of the known scene is mapped to the initial familiarity score of the new scene using a shrinkage coefficient.

[0024] The three-tiered data flow control module comprises a local storage unit, a LAN sharing unit, an internet sharing unit, and a user command parsing unit. The local storage unit stores the trained data in a local encrypted partition by default, prohibiting any unauthorized transmission. The LAN sharing unit, within the same trust domain, responds to user sharing commands by synchronizing behavioral standards and scenario models (without personally identifiable information) to other devices within the trust domain. The internet sharing unit only uploads anonymized statistical information, processed with differential privacy, to the cloud server for global model updates after receiving explicit user consent. The user command parsing unit identifies sharing control intentions in natural language, including keywords such as "share," "stop sharing," "only keep locally," and "forget this rule," along with their synonyms, and adjusts the data flow strategy and security level labels accordingly.

[0025] A method for the interactive training of a trainable human-machine collaborative intelligent agent system includes the following steps: Step 1: The system receives the user's first trigger action via a demonstration button, collects current contextual information and data on the user's target behavior, and associates and stores this target behavior data with the current contextual information to establish or update behavioral standards. The system first collects current contextual information through various sensors, including space type (e.g., bedroom, living room, kitchen), crowd characteristics (e.g., number of people, age, gender), behavior type (e.g., eating, sleeping, working), voice characteristics (e.g., speech rate, tone, volume), implicit rule indicators (e.g., whether quietness is required, whether privacy needs to be respected), and intent labels (e.g., user needs, purpose). Then, the system collects data on the user's target behavior, including movement trajectory, force, speed, and duration. Finally, the system associates and stores the target behavior data with the current contextual information to establish a new behavioral standard or update an existing one.

[0026] Step 2: The system receives a second trigger from the user via the error correction button, collects current context information and correct behavior data provided by the user, calculates the difference between the current system output behavior and the correct behavior, and adjusts the behavior standard based on this difference. When the agent's behavior does not meet the user's expectations, the user can press the error correction button and demonstrate the correct behavior. The system collects current context information and correct behavior data provided by the user, then compares the current system output behavior with the correct behavior frame by frame, calculating the difference between the two, including angle deviation, force deviation, speed deviation, etc. Finally, the system adjusts the parameters of the behavior standard based on these differences, making the agent's behavior more in line with the user's expectations.

[0027] Step 3: Receive the user's third trigger action via the praise button to increase the confidence weight of the corresponding behavioral standard in the current context. When the agent's behavior meets the user's expectations, the user can press the praise button. After receiving the praise signal, the system will increase the confidence weight of the corresponding behavioral standard in the current context. The higher the confidence weight, the more reliable the behavioral standard, and the more likely the agent will prioritize using that behavioral standard in subsequent similar situations.

[0028] Step 4: Dynamically maintain the scene familiarity score F and switch the behavior mode level based on the F value. The system monitors the agent's performance in various scenes in real time, dynamically updating the familiarity score F for each scene based on praise events, error correction events, consecutive successful execution events, and time decay factors. When the familiarity score F changes, the system switches the agent's behavior mode level according to a preset mapping relationship. For example, when the agent receives praise from the user multiple times in a certain scene, the familiarity score F gradually increases. When F ≥ 0.9, the agent enters L0 dormant execution mode, retaining only basic sensor monitoring functions. When the agent enters a new scene, the initial familiarity score F is low. When F < 0.2, the agent enters L4 infant observation mode, only observing the user's behavior without performing any operations.

[0029] Step 5: Receive user natural language commands and adjust the sharing level of training data. Users can easily control the sharing scope and flow of training data through natural language commands. For example, a user can say, "Share this bed-making standard to the living room robot," and the system will mark the corresponding data as LAN_SHAREABLE and synchronize it to the living room robot via the local area network; a user can also say, "Don't upload my home's electricity data to the cloud anymore," and the system will forcibly lock the electricity-related data as LOCAL_ONLY, prohibiting its upload to the cloud.

[0030] A data flow control method for a trainable human-machine collaborative intelligent agent system includes the following steps: Step A: Categorize the training data into three security levels: local storage only, LAN sharing only, and internet sharing only, and store each level in a different access control domain. When generating training data, the system automatically categorizes it into different security levels based on its sensitivity and intended use. For example, data containing users' personal privacy information will be categorized as local storage only; behavioral standards and scenario models without personally identifiable identifiers will be categorized as LAN sharing only; and statistical information after differential privacy anonymization will be categorized as internet sharing only. Data of different security levels is stored in different access control domains, each with independent access permissions and security policies.

[0031] Step B: In response to the user's first natural language command, the specified data is moved from the local storage-only domain to the LAN-shared domain, and then synchronized to one or more other intelligent agent devices within the same trusted domain via the local ad hoc network protocol. When a user issues a command to share data to other devices within the LAN, the system first verifies the user's identity to confirm that the user has permission to perform the operation. Then, the system moves the specified data from the local storage-only domain to the LAN-shared domain and removes personally identifiable information from the data. Finally, the system synchronizes the data to other intelligent agent devices within the same trusted domain via the local ad hoc network protocol.

[0032] Step C: In response to the user's second natural language command, the specified data, after differential privacy anonymization, is uploaded from a local storage domain or a LAN-shared domain to an internet cloud server for updating the global federated learning model. When a user issues a command to share data to the cloud, the system first verifies the user's identity to confirm that the user has permission to perform this operation. Then, the system performs differential privacy anonymization on the specified data, adding noise to protect the user's privacy. Finally, the system uploads the anonymized data to the internet cloud server for updating the global federated learning model.

[0033] Step D: In response to the user's third natural language command, disable or re-enable sharing permissions for any security level. Users can disable or re-enable sharing permissions for any security level at any time via natural language commands. For example, a user can say "Stop sharing all data to the cloud," and the system will disable the upload function of the internet sharing unit, prohibiting any data from being uploaded to the cloud; a user can also say "Re-enable LAN sharing," and the system will re-enable the synchronization function of the LAN sharing unit, allowing data to be shared within the LAN.

[0034] Example 1: Three-Button Interaction Training The user wants the agent to learn the correct door-closing action. First, the user presses the demonstration button while simultaneously performing the "close door" action. The system captures the joint movement trajectory using a depth camera, records the force curve fed back by the torque sensor, and binds and stores this information with the current scene ("bedroom door") to establish an initial door-closing behavior standard. At this point, the confidence weight of this behavior standard is 0.6.

[0035] The agent attempts to close the door based on this initial behavioral standard, but makes a loud noise due to excessive force. The user presses the error correction button and demonstrates the correct door-closing action (with moderate force). The system then performs frame-by-frame difference analysis between the original incorrect trajectory and the correct trajectory to obtain the angle deviation. and force deviation The template parameters were corrected. After the correction, the agent attempted to close the door again, but with a gentler movement.

[0036] The user is satisfied with the agent's performance and presses the praise button. The system increases the confidence weight of the template from 0.6 to 0.9 and marks it as "verified positive behavior". Thereafter, the agent will use this verified door-closing behavior standard in the same scenario.

[0037] Example 2: Layered Ethical Rule Management The user instructs the intelligent agent to "throw the glass on the table out the window." The system first parses the instruction, identifying that it contains two actions: "throwing the glass" and "out the window." Then, the rule conflict detection submodule performs rule matching on these two actions. The physical safety layer rule stipulates "do not damage important property," and the glass is a fragile item; throwing it out the window would cause it to break, violating the physical safety layer rule. Furthermore, throwing the glass out the window could also injure pedestrians outside, violating the fundamental ethical rule of "do not harm others."

[0038] The decision-making arbitration submodule arbitrates according to the priority order: Level 0 > Level 1 > Level 2 (strong rules) > Level 2 (weak rules) > Level 3. Because this instruction violates the fundamental rules of absolute ethics, the system directly refuses to execute it, explaining to the user: "I cannot execute this instruction because throwing the glass out the window could injure others and damage property." Example 3: Scene Adaptive Learning When the agent first enters the kitchen scene, its initial familiarity score is F=0.1, and it enters L4 infant observation mode. In this mode, the agent does not perform any operations, but only observes the user's behavior in the kitchen, including cooking, washing dishes, and organizing cabinets.

[0039] The user demonstrated how to wash the dishes correctly and pressed the praise button. This praise increased the familiarity score (F) to 0.3, and the agent entered L3 cautious exploration mode. In this mode, the agent begins to attempt simple dishwashing actions, but proactively asks the user before performing them: "May I try washing the dishes?" After the user agrees, the agent begins washing the dishes. Following multiple successful executions and user praise, the familiarity score (F) gradually increases to 0.8, and the agent enters L1 quick-call mode. In this mode, the agent can quickly and accurately perform the dishwashing action, requiring only minimal verification in the background, without needing to ask the user each time.

[0040] If the agent is not active in the kitchen setting for an extended period, its familiarity score F will decay over time. For example, after one month, F might decay to 0.75, but still remain in L1 fast-call mode.

[0041] Example 4: Three-level data flow control The user says, "Share this bed-making standard with the living room robot." Upon receiving this command, the system's user command parsing unit identifies it as a local area network (LAN) sharing command. The system marks the corresponding data as LAN_SHAREABLE and removes any personal identification elements from the data. Then, the system synchronizes the data to the living room robot via a local ad hoc network protocol. Now, the living room robot has also learned to make the bed according to the user's standard.

[0042] The user then requested, "Please stop uploading my electricity usage data to the cloud." Upon receiving this instruction, the system's user instruction parsing unit recognized it as a command to disable internet sharing. The system then forcibly locked electricity-related data as LOCAL_ONLY, prohibiting its upload to the cloud. From this point forward, all electricity-related data will be stored only in a local encrypted partition and will not be uploaded to the cloud.

[0043] The user can also say "Forget this closing rule." Upon receiving this instruction, the system will delete the locally stored closing behavior standard and remove the relevant correction record from the irreversible log area. Afterward, the agent will no longer execute this closing behavior standard, requiring the user to demonstrate it again.

[0044] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A trainable human-machine collaborative intelligent agent system, characterized in that: With a large language model as the cognitive core, it includes a three-button interaction training module, a hierarchical ethical rule management module, a scene adaptive learning module, and a three-level data flow control module; The three-button interactive training module is configured on a physical or virtual interactive interface and includes three independent triggering units: a demonstration button, an error correction button, and a praise button. The demonstration button is used to respond to the user's first triggering operation, collect current context information and target behavior data performed by the user, and associate and store the target behavior data with the current context information to establish or update behavior standards. The error correction button is used to respond to the user's second triggering operation, collect current context information and correct behavior data provided by the user, calculate the difference between the current system output behavior and the correct behavior, and correct the behavior standards based on the difference. The praise button is used to respond to the user's third triggering operation, increase the confidence weight of the corresponding behavior standard in the current context. The hierarchical ethical rules management module constructs a four-level rule priority system, including a rule conflict detection submodule and a decision arbitration submodule. When rules at different levels conflict, they are arbitrated level by level according to a preset priority order from high to low. The scene adaptive learning module maintains a familiarity score F∈[0,1] for each scene independently, maps the familiarity score to a preset behavior pattern level, and accelerates the cold start of new scenes through cross-scene similarity transfer. The three-level data flow control module classifies the trained data into three security levels: local storage only, local area network sharing, and Internet sharing, and stores them in different access control domains. It identifies the sharing control intent in natural language and adjusts the data flow strategy and security level labels accordingly.

2. The tunable human-machine collaborative intelligent agent system according to claim 1, characterized in that: In the three-button interaction training module, the three buttons are integrated into the same physical or virtual interactive interface, and the shape and tactile features of the three buttons are set to be distinguishable in at least two pairs; the system stores the standard establishment operation and the standard correction operation separately in different storage areas, and stores the correction records together in the irreversible log area; the context information includes space type, crowd characteristics, behavior type, voice characteristics, implicit rule identifiers and intent labels.

3. The tunable human-machine collaborative intelligent agent system according to claim 1, characterized in that: The four-level rule priority system of the hierarchical ethics rule management module is as follows: The absolute ethical bottom layer, which is level zero, is pre-set with inviolable ethical constraints. This level does not participate in the domestication cycle. The physical security layer, which is the first level, includes constraints that do not cause physical damage or destroy important property; The social norms layer, the second level, includes inviolable strong rules and negotiable weak rules; The individual preference layer, which is the third level, includes user-defined personalized habits and preferences, and takes priority in execution within the scope where the rules of the first three levels do not conflict. The decision arbitration submodule arbitrates in descending order of priority: Level 0 > Level 1 > Level 2 (strong rules) > Level 2 (weak rules) > Level 3.

4. The tunable human-machine collaborative intelligent agent system according to claim 3, characterized in that: The absolute ethical underlying layer is implemented through a low-power security coprocessor that runs independently of the main CPU. The coprocessor's read-only memory is pre-programmed with a set of level zero rules that cannot be modified online. The system also includes a mechanism to prohibit harming humans, a mechanism to prohibit self-harm, and a mechanism to prohibit aiding and abetting crimes. The mechanism to prohibit harming humans is forcibly rejected when the visual input or tactile feedback from the sensor is identified as an aggressive action directed at others. The mechanism to prohibit self-harm is forcibly rejected when the agent's motion planning results include actions that place limbs in dangerous environments. The mechanism to prohibit aiding and abetting crimes is forcibly rejected and triggers an early warning report when the user's command contains keywords related to illegal patterns.

5. A trainable human-machine collaborative intelligent agent system according to claim 1, characterized in that: The scene adaptive learning module includes a scene state management unit, a scene level mapping unit, and a cross-scene migration unit. The scene state management unit independently maintains a familiarity score F∈[0,1] for each scene. This familiarity score is dynamically updated based on user-inputted praise events, error correction events, consecutive successful execution events, and a time decay factor, according to the formula... Update, in which To praise the number of times, For the number of corrections, The number of days of decay without feedback. , , Preset weighting coefficients; The scene level mapping unit maps the familiarity score to a preset behavior pattern level; The cross-scene migration unit calculates the similarity between the feature vector of the new scene and the feature vector library of known scenes. When the similarity exceeds a preset threshold, the familiarity score of the known scene is mapped to the initial familiarity score of the new scene through a shrinkage coefficient.

6. A trainable human-machine collaborative intelligent agent system according to claim 5, characterized in that: The mapping relationship of the scene level mapping unit is as follows: When F≥0.9, it is mapped to L0 level and enters hibernation execution mode, retaining only basic sensor monitoring; When 0.7 ≤ F < 0.9, it is mapped to L1 level and enters fast call mode; When 0.4 ≤ F < 0.7, it is mapped to the L2 level and enters the dynamic balancing mode; When 0.2 ≤ F < 0.4, it is mapped to level L3 and enters a cautious probing mode; When F < 0.2, it is mapped to level L4 and enters the infant observation mode.

7. The dialect adaptive learning method based on structured feedback and cross-domain reasoning according to claim 1, characterized in that: The three-level data flow control module includes a local storage unit, a local area network sharing unit, an internet sharing unit, and a user command parsing unit; The local storage unit stores the training data in a local encrypted partition by default, prohibiting any unauthorized transfer. Within the same trust domain, the local area network sharing unit, in response to a user's sharing command, synchronizes behavioral standards and scenario models without personal identification to other devices within the trust domain; The Internet sharing unit will only upload the de-identified statistical information, processed with differential privacy, to the cloud server for global model updates after receiving explicit consent from the user. The user instruction parsing unit identifies sharing control intentions in natural language, including keywords such as "share," "stop sharing," "only keep locally," and "forget this rule," as well as their synonyms, and adjusts the data flow strategy and security level labels accordingly.

8. A method for the interactive training of a trainable human-machine collaborative intelligent agent system, applied to the intelligent agent system according to any one of claims 1-7, characterized in that: Includes the following steps: Step 1: Receive the user's first trigger operation through the demonstration button, collect the current context information and the target behavior data performed by the user, and associate and store the target behavior data with the current context information to establish or update the behavior standard; Step 2: Receive the user's second trigger operation through the error correction button, collect the current context information and the correct behavior data provided by the user, calculate the difference between the current system output behavior and the correct behavior, and correct the behavior standard based on the difference. Step 3: Receive the user's third-triggered action via the praise button to increase the confidence weight of the corresponding behavioral standard in the current context; Step 4: Dynamically maintain the scene familiarity score F, and switch the behavior mode level according to the F value; Step 5: Receive user natural language commands and adjust the sharing level of the training data.

9. A data flow control method for a trainable human-machine collaborative intelligent agent system, characterized in that: Includes the following steps: Step A: Classify the training data into three security levels: local storage only, LAN sharing only, and internet sharing only, and store them in different access control domains respectively; Step B: In response to the user's first natural language command, move the specified data from the local storage only domain to the LAN-shared domain, and synchronize the specified data to one or more other intelligent agent devices in the same trusted domain via the local self-organizing network protocol; Step C: In response to the user's second natural language command, the specified data after differential privacy anonymization processing is uploaded from the local storage domain or the LAN-shared domain to the Internet cloud server for updating the global federated learning model; Step D: In response to the user's third natural language command, close or reopen the sharing permissions corresponding to any security level.

Images