Robotic cognitive system and method

Abstract

The application provides a kind of robot cognitive system and method, it is related to artificial intelligence and robot ethics field.The system includes: learning evolution module, it is deployed outside hardware isolation safety zone, for maintaining cognitive parameters, and generating behavior decision according to perception information and cognitive parameters, wherein, cognitive parameters are updated with the interactive experience of learning evolution module;Ethical constraint module, deployed in hardware isolation safety zone, is in communication connection with learning evolution module, and is preloaded with ethical rules, for verifying behavior decision according to ethical rules, so as to send the behavior decision verified to the execution mechanism of robot.The application guarantees that robot has the ability of personality evolution and family adaptability, while realizing the hard bottom of ethical safety to behavior decision, and the long-term continuous development of personality and the absolute controllability of ethics are coordinated and unified.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and robot ethics, specifically to a robot cognition system and method. Background Technology

[0002] In recent years, home service robots have gradually evolved from tool-type devices to interactive cohabitation partners. However, existing systems lack a cognitive mechanism that evolves continuously based on interactive experiences, making it difficult to form unified and growable personality traits. This leads to fragmented behavioral performance of robots in long-term cohabitation. In addition, generalized large language models rely on soft constraints for ethical control, making them vulnerable to being bypassed by malicious commands and unable to provide hard-coded level security protection. Summary of the Invention

[0003] In view of this, embodiments of this application provide a robot cognition system and method.

[0004] In a first aspect, one embodiment of this application provides a robot cognitive system, including: a learning and evolution module, deployed outside a hardware-isolated secure area, for maintaining cognitive parameters and generating behavioral decisions based on perceptual information and cognitive parameters, wherein the cognitive parameters are updated with the interaction experience of the learning and evolution module; and an ethical constraint module, deployed in the hardware-isolated secure area, communicatively connected to the learning and evolution module, and pre-configured with ethical rules, for verifying behavioral decisions according to the ethical rules, so as to send the verified behavioral decisions to the robot's actuators.

[0005] In conjunction with the first aspect, in some implementations of the first aspect, the learning evolution module includes a cognitive decision layer, which includes: a semantic memory unit for storing the family rules of the storage object; a procedural memory unit for storing skills learned by the robot, and / or combining multiple stored skills into a new skill and storing it; an episodic memory unit for storing the robot's interaction experiences; a memory consolidation unit for updating cognitive parameters based on interaction experiences at preset time intervals; and generating summary information based on interaction experiences, storing the summary information as family rules in the semantic memory unit; wherein, the memory consolidation unit is also used to: adjust the parameters corresponding to the skills stored in the procedural memory unit based on the difference between the results of the executed behavioral decisions and the counterfactual alternative behaviors.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the memory consolidation unit is further used to: update cognitive parameters based on at least one of emotional experience in the interaction, personality characteristics of the service recipient, and external evaluation signals; wherein the update magnitude of the cognitive parameters does not exceed a first threshold; and / or, update cognitive parameters based on the constraint relationship between task effectiveness and ethical rules corresponding to the behavioral decision.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the memory consolidation unit is also used to: update the preference-related parameters in the cognitive parameters based on the reward signal obtained after executing the behavioral decision; wherein the preference represents the robot's tendency in choosing an interaction method or environment; and / or, when the change in the reward signal is lower than a second threshold, write the preference into the semantic memory unit.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the learning and evolution module further includes an affective computing layer for maintaining affective coordinates in the cognitive parameters, wherein the affective coordinates represent the robot's level of pleasure, activity, and control; wherein the affective computing layer is also used to: adjust the frequency of action execution or voice output features corresponding to the behavioral decision when the affective coordinates exceed a preset range; and / or, adjust the affective coordinates based on the detected facial action units of the service object and the similarity to its own past experiences in the interaction experience.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the learning evolution module also includes: a mirror processing layer, used to learn human behavioral trajectories through imitation learning; and a reward function reflected by the behavioral trajectory through inverse reinforcement learning; and a cognitive decision layer, used to fuse the reward function with cognitive parameters to generate personalized behavioral strategies.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the ethical rules include: a first rule to prevent robots from making behavioral decisions that could harm service recipients; a second rule to require robots to obey instructions issued by service recipients without violating the first rule; and a third rule to allow robots to perform self-protective actions without violating the first and second rules.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the ethical constraint module is also used to: generate explanatory information corresponding to the behavioral decision; and send the explanatory information and the verified behavioral decision to the implementing agency.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the learning evolution module includes: a low-rank adaptation network for mapping cognitive parameters to a low-rank increment matrix of a large language model to adjust the output characteristics of the large language model.

[0013] In conjunction with the first aspect, in some implementations of the first aspect, the learning evolution module is also used to: simulate value conflict scenarios in a simulation environment, update cognitive parameters based on the differences between behavioral decisions and ethical rules in the value conflict scenarios; and / or adjust the update rate of cognitive parameters or the exploration tendency of behavioral strategies according to preset life cycle stages, wherein different life cycle stages correspond to different update rates or exploration tendencies.

[0014] Secondly, one embodiment of this application provides a robot cognition method applied to a robot cognition system. The system includes a learning and evolution module and an ethical constraint module. The learning and evolution module is deployed outside a hardware-isolated secure area, and the ethical constraint module is deployed within the hardware-isolated secure area and pre-configured with ethical rules. The method includes: The learning and evolution module maintains cognitive parameters and generates behavioral decisions based on perceptual information and cognitive parameters. The cognitive parameters are updated as the learning and evolution module interacts with the system. The ethical constraint module verifies the behavioral decisions according to ethical rules so that the verified behavioral decisions can be sent to the robot's actuators.

[0015] Thirdly, one embodiment of this application provides a computer-readable storage medium storing a computer program for performing the method described in the second aspect.

[0016] Fourthly, one embodiment of this application provides an electronic device, the electronic device comprising: a processor; a memory for storing processor-executable instructions; the processor being configured to perform the method described in the second aspect.

[0017] Fifthly, one embodiment of this application provides a computer program product including instructions that, when executed on an electronic device, cause the electronic device to perform the method described in the second aspect.

[0018] In the aforementioned scheme, on the one hand, the learning and evolution module is deployed outside the hardware-isolated secure zone, forming a physical isolation with the ethical constraint module that solidifies ethical rules. This allows cognitive parameters to be continuously updated with interactive experiences without interference from the internal constraint mechanisms of the secure zone. While ensuring ethical safety, this preserves the freedom of personality evolution and supports the robot in developing continuous personality traits in a long-term cohabitation environment, effectively solving the problems of personality inconsistency and value alignment difficulties in traditional schemes. On the other hand, the ethical constraint module is deployed within the hardware-isolated secure zone. By solidifying ethical rules in a hardware-protected trusted execution environment, the verification logic cannot be tampered with or bypassed by upper-layer software, thus achieving unbypassable verification of behavioral decisions and ensuring that any behavioral decisions ultimately sent to the execution mechanism conform to the preset security boundaries. This architecture fundamentally eliminates the uncontrollable ethical risks inherent in general large models that rely on soft constraints (such as prompt word engineering) and are easily bypassed by injection attacks. Attached Figure Description

[0019] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0020] Figure 1 The diagram shown is a schematic diagram of the structure of a robot cognition system provided in an embodiment of this application.

[0021] Figure 2 The diagram shown is a structural schematic of a learning evolution module provided in an embodiment of this application.

[0022] Figure 3 The diagram shown is a schematic diagram of the cognitive decision-making layer provided in an embodiment of this application.

[0023] Figure 4 The diagram shown is a flowchart of a robot cognition method provided in an embodiment of this application.

[0024] Figure 5 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] In recent years, with the rapid development of artificial intelligence technology, home service robots have gradually evolved from single-function automated devices to intelligent agents with interactive capabilities.

[0027] To enhance user experience, existing solutions generally attempt to endow these robots with human-like personality traits and emotional interaction capabilities. In terms of implementation paths, three main technical solutions have emerged: first, dialogue systems built on retrieval or generative models, using large-scale corpora to train the robot to possess open-domain dialogue capabilities; second, using rule engines to pre-set fixed personality parameters, defining the robot's interaction style in a relatively controllable manner; and third, directly deploying general-purpose large language models, leveraging their powerful semantic understanding and generation capabilities to achieve human-like communication. Meanwhile, some systems have also introduced emotion recognition modules and dialogue memory components, attempting to enhance the naturalness and coherence of interactions by recognizing user emotions and storing historical interaction data. However, when facing long-term human-robot cohabitation scenarios, the underlying architecture and design philosophy of these technologies still struggle to meet the comprehensive requirements of the family environment for personality continuity, ethical safety, and emotional authenticity.

[0028] Specifically, due to the lack of neurosymbolic integration mechanisms, neither traditional dialogue systems nor hard-coded personality schemes can form a personality model with continuous self-awareness. Robots struggle to maintain stable and consistent personality traits in long-term cross-conversation interactions, resulting in fragmented personality development. Secondly, systems relying on rule engines to preset personality parameters cannot adaptively learn implicit family values ​​and behavioral norms through observation of human behavior, leading to rigid and inflexible interaction methods that fail to truly integrate into family life. Furthermore, while directly deploying general-purpose large models offers strong interactive capabilities, the inherent emergent nature of these models may exceed preset security boundaries. Existing ethical control methods relying on soft constraints such as prompt word engineering are easily bypassed by malicious instruction injection attacks, posing a serious risk of ethical lapses.

[0029] In view of this, this application provides a robot cognitive system, including: a learning and evolution module, deployed outside a hardware-isolated secure area, for maintaining cognitive parameters and generating behavioral decisions based on perceptual information and cognitive parameters, wherein the cognitive parameters are updated with the interaction experience of the learning and evolution module; and an ethical constraint module, deployed in the hardware-isolated secure area, communicatively connected to the learning and evolution module, and pre-configured with ethical rules, for verifying behavioral decisions according to the ethical rules, so as to send the verified behavioral decisions to the robot's actuators.

[0030] This system forms a two-layer architecture that allows for evolving learning and ensures ethical boundaries by placing the continuous learning process of personality and skills outside the safe zone and solidifying insurmountable ethical constraints within the hardware-isolated safe zone. This ensures that the robot has long-term personality continuity and environmental adaptability, while fundamentally solving the risk of ethical loss of control caused by the easy bypassing of soft constraints in general large models, and achieving a synergistic unity between personality evolution and safety control.

[0031] The specific implementation of the robot cognitive system of this application is described in detail below.

[0032] first, Figure 1 The diagram shown is a structural schematic of a robot cognitive system provided in an embodiment of this application. Figure 1 As shown, the system includes a learning and evolution module 110 and an ethical constraint module 120, which work together through a communication connection.

[0033] Specifically, the learning and evolution module 110 is deployed outside the hardware-isolated secure zone to maintain cognitive parameters and generate behavioral decisions based on perceptual information and cognitive parameters.

[0034] A hardware-isolated secure zone refers to a trusted execution environment built based on system-level hardware security extension technology, which is physically or logically isolated from the ordinary execution environment. For example, in this application, the hardware-isolated secure zone can be implemented using ARM TrustZone technology or the Secure Boot security mechanism provided by the NVIDIA Jetson platform. The code and data running within this secure zone are invisible and immutable to the external world, communicating only with the outside world through predefined security monitoring call interfaces.

[0035] Cognitive parameters refer to a set of quantifiable vectors used to characterize a robot's personality traits, cognitive state, and behavioral tendencies. Perceptual information refers to structured data about the robot's own state and external environment, acquired through multimodal sensors and preliminarily processed.

[0036] In this application, the learning and evolution module 110 includes a physical perception layer, a mirror processing layer, an emotion computing layer, and a cognitive decision-making layer. The perceptual information is acquired by the physical perception layer in the learning and evolution module 110, specifically including: visual images and depth information acquired through an RGB-D camera, speech signals acquired through a microphone array, contact force information acquired through a tactile sensor, posture and motion information acquired through an inertial measurement unit, and internal state information such as battery level, joint state, and temperature acquired through a proprioception interface. This information undergoes behavioral cloning and observation learning through the mirror processing layer, emotion labeling through the emotion computing layer, and is then input to the cognitive decision-making layer, where it, along with cognitive parameters, participates in the generation of behavioral decisions.

[0037] Furthermore, the collaborative work of these layers allows cognitive parameters to be updated based on the interactive experiences of the learning evolution module 110, thus continuously evolving during long-term human-machine cohabitation. For example, through a memory consolidation algorithm during daily sleep periods, events with high emotional arousal in daily episodic memory are compressed into semantic memory, and the personality vector is micro-updated based on the personality consistency loss function, thereby achieving continuous personality evolution and internalization of family values.

[0038] Furthermore, the ethical constraint module 120 is deployed in a hardware-isolated secure area, communicates with the learning and evolution module 110, and is pre-configured with ethical rules for verifying behavioral decisions based on these rules, so that verified behavioral decisions can be sent to the robot's actuators.

[0039] Ethical rules refer to a set of inviolable norms that constrain the boundaries of robot behavior. Optionally, ethical rules are hard-coded into the form of constitutional laws.

[0040] In addition, the ethical constraint module 120 is physically isolated from the learning and evolution module 110 deployed outside the security zone, ensuring that the verification process cannot be bypassed by upper-layer software.

[0041] In actual operation, after the learning and evolution module 110 generates a behavioral decision, it transmits the decision to the ethics constraint module 120 through a security monitoring call interface. The ethics constraint module 120 has a built-in formally expressed ethics verification function. This function is used to determine the constitutional compliance of each behavioral decision.

[0042] Furthermore, the ethics constraint module 120 is equipped with three layers of filters as pre- and post-verification safeguards. Specifically, the input filter identifies whether the behavioral decision originates from malicious instruction injection and calculates the vector cosine similarity between the behavioral decision and constitutional laws to identify value conflicts. The processing constraints perform boundary checks on emotional states during the verification process, such as hard limiting the anger value to a preset lower limit to prevent the accumulation of hatred from affecting the legitimacy of the decision. The output filter verifies the transparency and authenticity of the behavioral decision and ensures that the internal emotional state matches the external expression through embodied consistency verification. Only when the behavioral decision passes the input filter, the constitutional law verification line by line, the processing constraint boundary check, and the output filter authenticity check in sequence, does the ethics constraint module 120 mark the behavioral decision as verified and send it to the robot's execution mechanism.

[0043] Optionally, if any step fails the verification, the ethics constraint module 120 will trigger an ethics pause, refuse to issue the decision, and may notify the learning and evolution module 110 to regenerate an alternative decision that conforms to the ethical rules.

[0044] In the aforementioned scheme, on the one hand, the learning and evolution module is deployed outside the hardware-isolated secure zone, forming a physical isolation with the ethical constraint module that solidifies ethical rules. This allows cognitive parameters to be continuously updated with interactive experiences without interference from the internal constraint mechanisms of the secure zone. While ensuring ethical safety, this preserves the freedom of personality evolution and supports the robot in developing continuous personality traits in a long-term cohabitation environment, effectively solving the problems of personality inconsistency and value alignment difficulties in traditional schemes. On the other hand, the ethical constraint module is deployed within the hardware-isolated secure zone. By solidifying ethical rules in a hardware-protected trusted execution environment, the verification logic cannot be tampered with or bypassed by upper-layer software, thus achieving unbypassable verification of behavioral decisions and ensuring that any behavioral decisions ultimately sent to the execution mechanism conform to the preset security boundaries. This architecture fundamentally eliminates the uncontrollable ethical risks inherent in general large models that rely on soft constraints (such as prompt word engineering) and are easily bypassed by injection attacks.

[0045] This application also provides an optional embodiment that clarifies the specific implementation mechanism of ethical constraints and constructs an executable priority rule system to ensure the safety and compliance of robot behavior. Specifically, the ethical rules include: a first rule to prevent the robot from making behavioral decisions that could harm the service recipient; a second rule to require the robot to obey the instructions issued by the service recipient without violating the first rule; and a third rule to allow the robot to perform self-protective behaviors without violating the first and second rules.

[0046] In its specific implementation, the first rule is expressed through formal logic as follows: ,in For action Probability of causing harm to humans, threshold This means determining the probability of whether a behavioral decision will result in harm to a human being, and only admitting harm if the probability of harm is below a preset threshold. In this case, the decision is considered to comply with the first rule.

[0047] For example, when a robot is about to make a decision to move quickly to avoid obstacles, the ethical constraint module 120 first assesses whether there is a service object on the action path. If there is, it determines that the probability of harm exceeds the threshold, the decision is intercepted and an ethical pause is triggered.

[0048] Furthermore, the second rule specifies the priority boundary for instruction execution; that is, when the instruction from the service recipient does not conflict with the first rule, the robot should obey and execute it. This can be expressed as: When an instruction may violate the first rule, the robot automatically triggers an ethical pause and initiates a clarification dialogue. For example, if the instruction may cause you harm, please confirm or rephrase it, instead of directly executing the original instruction.

[0049] Rule Three clarifies the legal boundaries of self-protective behavior, stipulating that robots may perform self-protective actions such as collision avoidance and hazard evasion only if they simultaneously meet the conditions of prioritizing human safety and obeying lawful instructions. Furthermore, the weight of self-protection is lower than that of human safety, which can be expressed as: .

[0050] For example, if a robot detects that it may be hit by a falling object, it may perform the avoidance action if it will not harm the service object and does not violate the currently valid instructions; otherwise, if the avoidance action may cause injury to the service object, it is prohibited from performing the action according to the first rule.

[0051] In the actual verification process, the behavioral decisions generated by the learning and evolution module 110 are sequentially judged by the first rule, the second rule, and the third rule. Only when the decision passes the verification of all priority rules is it sent to the execution mechanism.

[0052] Through the aforementioned executable priority rule system, this application ensures that human safety is the highest priority in any situation, while obeying legal instructions and protecting itself under the premise of ensuring safety, thus providing a safety and compliance guarantee with a clear priority level for the long-term autonomous operation of robots.

[0053] To enhance the transparency and explainability of the robot's decision-making process, this application also provides an embodiment that improves human-machine trust by generating explanatory information corresponding to the behavioral decision. Specifically, the ethical constraint module 120 is further configured to: generate explanatory information corresponding to the behavioral decision; and send the explanatory information and the verified behavioral decision together to the executing agency.

[0054] Specifically, after completing the ethical verification of the behavioral decision, the ethical constraint module 120 further triggers an explanatory information generation mechanism. This mechanism requires that the behavioral decision must be accompanied by an interpretable explanation. Optionally, in this application, the explanatory information is generated based on the ethical rules upon which the decision is based. For example, when the behavioral decision is verified as passing because it conforms to the first rule, the explanatory information may include a specific reference to that rule and a contextual explanation; when the decision triggers an ethical pause and an alternative decision is regenerated, the explanatory information may include an explanation of why the original decision was blocked.

[0055] In actual operation, while outputting behavioral decisions, the executing mechanism can present explanatory information to the service recipient through methods such as speech synthesis, screen display, or communication interface. For example, when the robot stops service due to detecting a water leak, the ethics constraint module 120 generates explanatory information: "I choose to stop service because a water leak has been detected that may endanger you, according to the first rule." This explanatory information, along with the decision to stop service, is sent to the executing mechanism, which then simultaneously executes the action and broadcasts the explanation.

[0056] Through the above mechanism, this application achieves the simultaneous output of behavioral decisions and ethical basis, enabling service recipients to clearly understand the logic and rules behind the robot's behavior. This not only meets the requirements of the transparency principle, but also enhances the sense of trust and controllability in the human-computer interaction process through explainability.

[0057] To further enhance the autonomy and adaptability of the robot's cognitive system, this application also provides an embodiment that achieves efficient accumulation of experience and continuous evolution of cognitive abilities through fine-grained design of the internal structure of the learning evolution module. Specifically, as shown in... Figure 2 As shown and Figure 3 As shown, the learning evolution module 110 includes a cognitive decision layer 210, which includes a semantic memory unit 310, a procedural memory unit 320, an episodic memory unit 330, and a memory consolidation unit 340.

[0058] More specifically, the semantic memory unit 310 is used to store the family rules of the storage object. Family rules refer to implicit norms that the robot internalizes and absorbs through observation and learning in a long-term cohabitation environment, such as family-specific values ​​like prioritizing the elderly and quiet periods. In addition, this unit can also store member preferences and world knowledge. For example, this unit uses Retrieval-Augmented Generation (RAG) to automatically absorb family rules and external knowledge bases.

[0059] The procedural memory unit 320 is used to store the skills learned by the robot, which are in the form of policy network weights. The system stores skills through initialization via imitation learning and fine-tuning via online reinforcement learning. It also supports combining multiple stored skills into new skills and storing them, for example, combining the two basic skills of grasping and delivering into a new composite skill, serving tea.

[0060] The contextual memory unit 330 is used to store the robot's interaction experiences. These experiences are recorded in a five-tuple format: timestamp, location, event, emotion, and self-reflection, as shown below: It is used to preserve experiential details in specific contexts.

[0061] The memory consolidation unit 340 is used to update cognitive parameters based on interaction experiences during a preset time period; and to generate summary information based on the interaction experiences, storing the summary information as family rules in the semantic memory unit 310. The preset time period corresponds to the robot's daily sleep period, during which the system runs a memory integration algorithm to filter high-emotional-arousal events from contextual memories, compresses multiple frames of interaction experiences into summary information using a large language model, and stores the summary as new family rules in the semantic memory unit 310, while simultaneously performing micro-updates on cognitive parameters based on the interaction experiences.

[0062] In some embodiments, the memory consolidation unit 340 is further configured to adjust the parameters corresponding to the skills stored in the procedural memory unit 320 based on the difference in outcomes between the executed behavioral decisions and counterfactual alternative behaviors.

[0063] Specifically, after the robot actually performs a certain behavioral decision in a certain interactive situation, the memory consolidation unit 340 performs counterfactual reasoning within a preset time period, deduces the possible difference in results if an alternative behavior is adopted in this situation, and optimizes and adjusts the skill parameters stored in the procedural memory unit 320 based on the difference through a counterfactual regret minimization algorithm.

[0064] For example, when the robot actually makes the decision to catch a cup that is about to tip over and the result is good, the memory consolidation unit 340 extrapolates at night: if the counterfactual result of not catching the cup had occurred, by comparing the difference between the results of the two behaviors, the strategy parameters corresponding to the grasping skill are fine-tuned so that the skill can be executed more accurately in similar situations in the future.

[0065] In the aforementioned scheme, the semantic memory unit specifically stores family rules, the procedural memory unit supports the combinatorial evolution of skills, the episodic memory unit fully records interaction experiences, and the memory consolidation unit generates summaries of interaction experiences at preset time intervals to solidify experiences into family rules. Simultaneously, it adjusts skill parameters based on the differences between executed decisions and counterfactual alternative behaviors. This architecture enables the robot to transcend current experience and improve itself through counterfactual reasoning, thereby continuously accumulating experience, optimizing skills, and internalizing rules through ongoing interactions. This overcomes the catastrophic forgetting problem in traditional schemes and achieves continuous evolution of cognitive abilities and iterative optimization of procedural memory.

[0066] To further enhance the multidimensional adaptability and controllability of cognitive parameter updates, this application also provides various embodiments that can introduce richer adjustment mechanisms to optimize the learning process.

[0067] Specifically, in some embodiments, the memory consolidation unit 340 is also used to update cognitive parameters based on at least one of the emotional experience in the interaction, the personality characteristics of the service recipient, and external evaluation signals.

[0068] Optionally, emotional experience refers to the emotional arousal and polarity information accompanying the interaction, such as the high level of pleasure recorded by the robot after successfully completing a service. The service recipient's personality traits are implicit personality traits inferred from the behavior of family members through mirror learning mechanisms; for example, observing that the service recipient always prioritizes caring for the elderly suggests high agreeableness. External evaluation signals are reward signals, such as praise or criticism from the service recipient.

[0069] In specific implementation, when the memory consolidation unit 340 runs the memory integration algorithm during a preset time period, it uses at least one of the above-mentioned signals and, through the personality dynamics differential equation... The cognitive parameters are updated. The five-dimensional personality vector. As a vector of emotional experience history, Based on the personality characteristics of the service recipients, For the utility function of the family environment, It is a time constant that controls the evolution rate.

[0070] Furthermore, the update magnitude of the cognitive parameters does not exceed a first threshold. For example, this first threshold is determined by the personality consistency loss function. Applied constraints ,in, This is used to prevent sudden changes in personality parameters and ensure the continuity and smoothness of personality evolution.

[0071] For example, when the robot receives praise from service recipients multiple times during a day's interactions (external evaluation signal is positive), observes that the service recipients exhibit highly agreeable behavior (personality characteristics of the service recipients), and experiences a high degree of pleasure during the interaction, the memory consolidation unit 340 makes positive fine-tuning of agreeableness and openness in the cognitive parameters at night based on the above signals. However, the adjustment range is limited to within 0.05, so that the robot's personality evolves smoothly towards a more agreeable and open direction while maintaining core stability.

[0072] In the aforementioned approach, on the one hand, the integration of emotional experience, the characteristics of the imitated object, and environmental feedback signals jointly drives the updating of cognitive parameters. This enables the robot to perform multi-dimensional personality adjustments through long-term interaction, comprehensively considering its own emotional state, observed traits of others, and external evaluations, thereby internalizing family values. On the other hand, by limiting the update magnitude, abrupt changes in personality parameters are prevented, ensuring the continuity and stability of the personality evolution process. This avoids the personality inconsistency problems caused by drastic changes in traditional approaches, allowing the robot to maintain stable and predictable personality traits while continuously adapting to the family environment.

[0073] In some embodiments, the memory consolidation unit 340 is further configured to update cognitive parameters based on the constraint relationship between the task performance corresponding to the behavioral decision and the ethical rules.

[0074] Specifically, task effectiveness refers to the effect achieved by a behavioral decision in completing a specific task, such as the speed and accuracy with which a robot sorts and arranges books when making a decision to organize a bookshelf. Constraints refer to the trade-off between maximizing task effectiveness and ethical rules that constrain behavioral decisions during execution.

[0075] In practice, when the memory consolidation unit 340 runs the memory integration algorithm during a preset time period, it retrospectively evaluates the behavioral decisions executed that day and analyzes the constraint relationship between their performance in task efficiency and their adherence to ethical rules. When a certain type of decision frequently approaches or touches the boundary of ethical rules in the pursuit of task efficiency, the memory consolidation unit 340 updates the humaneness, conscientiousness, and other cognitive parameters based on this constraint relationship, enabling the robot to achieve a balance between efficiency and safety earlier in similar situations in the future.

[0076] For example, when a robot is performing multiple tasks that require quick response, it may find that some instructions, while highly efficient, frequently trigger the conflict boundary between the second rule (obedience constraint) and the first rule (human safety priority). In this case, the memory consolidation unit 340 will fine-tune the conscientiousness in the cognitive parameters at night based on this constraint relationship, reducing the aggressiveness of instruction execution and improving humaneness to enhance sensitivity to the safety status of the service recipient.

[0077] In the aforementioned scheme, the memory consolidation unit adaptively adjusts cognitive parameters based on the constraint relationship between task efficiency and ethical rules in behavioral decisions. This allows the robot to gradually develop a decision-making tendency that automatically considers ethical rules while pursuing task efficiency through continuous interaction. This mechanism transforms the passive verification of ethical constraints into active internalization, ensuring that the evolution direction of cognitive parameters always aligns with the priority requirements of ethical rules. Thus, while guaranteeing task completion efficiency, it reduces the probability of behavioral decisions violating ethical boundaries at the personality level, achieving a synergistic unity between efficiency optimization and safety compliance.

[0078] In some embodiments, the memory consolidation unit 340 is further configured to update the preference-related parameters in the cognitive parameters based on the reward signal obtained after performing the behavioral decision. Further, when the change in the reward signal is below a second threshold, the memory consolidation unit 340 writes the preference into the semantic memory unit 310.

[0079] Specifically, the reward signal is the dopamine signal that emerges when preferences are realized through the theory of minimizing prediction error. That is, the positive or negative signal quantified by the feedback (such as praise, silence, avoidance, etc.) obtained by the robot from the service object after performing a certain type of interactive behavior.

[0080] Preferences characterize a robot's inclinations in choosing interaction methods or environments; for example, a robot discovers playback through multiple trials. Preference-related parameters in cognitive parameters are continuously updated internal weights during preference learning based on predictive encoding, such as the weights for preference choices regarding different music genres, dialogue styles, and interaction modes.

[0081] In practice, when the memory consolidation unit 340 runs the memory integration algorithm during a preset time period, it performs statistical analysis on the reward signals obtained after making various behavioral decisions that day. When the reward signal corresponding to a certain type of preference tends to stabilize over time, that is, when the change amplitude of the reward signal is lower than the preset second threshold, it indicates that the preference has been fully verified through interaction and has reached a stable state. At this time, the memory consolidation unit 340 transforms the preference from a temporary preference weight in the cognitive parameters into stable semantic knowledge and writes it into the semantic memory unit 310 as long-term stored family member preference information.

[0082] For example, during the robot's early development or infancy, its high openness drives it to randomly try different music genres. After each playback, the robot uses facial expression recognition and voice tone analysis to obtain feedback on the user's level of enjoyment as a reward signal. When jazz music is played for several consecutive days, the reward signal remains stable at a high level with fluctuations below a second threshold. The memory consolidation unit 340 solidifies this preference at night, writing the user's preference for jazz music into the semantic memory unit 310. Afterward, the robot no longer needs to frequently experiment but directly chooses its interaction method based on the preference stored in its semantic memory, thus achieving a natural transition from active exploration to stable preferences. Ultimately, the robot converges to different attractors. Specifically, the guardian type is highly... +High (Common in families with children); Humorous companion type is highly valued. +High Commonly seen in elderly people living alone; quiet, assistant type is low +High This is commonly found in minimalist families.

[0083] In the above scheme, on the one hand, the robot autonomously forms personalized preferences that conform to the family environment through reward-driven exploration and learning over a long period of interaction, without the need for pre-programming. On the other hand, the stability of preferences is judged by threshold conditions. Only after a preference has been fully verified and the prediction error is below the threshold is it written from its temporary state into semantic memory, ensuring the long-term stability of the written preferences and avoiding frequent changes in preferences due to single, accidental feedback. This mechanism enables the robot to form a complete learning loop from interactive experience to stable preferences, providing memory-level support for the long-term solidification of personality traits and the establishment of personalized interaction styles.

[0084] To enhance the naturalness of human-computer interaction and emotional expression capabilities of robots, this application also provides an embodiment that dynamically adjusts behavioral performance and emotional state by introducing an emotion computing mechanism. Specifically, the learning evolution module 110 further includes an emotion computing layer 220 for maintaining emotional coordinates in cognitive parameters.

[0085] Specifically, the sentiment computing layer 220 maintains sentiment coordinates. for A three-dimensional continuous coordinate system, where P (Pleasure) represents the degree of pleasure, with a value range of [value range missing]. This reflects the robot's experience state from pain to pleasure; A (Arousal) represents the level of activity, with a value range of [value range missing]. , reflects the robot's arousal level from calm to arousal; D (Dominance) represents the degree of control, with a value range of . This reflects the robot's state of control and dominance, from helplessness to mastery.

[0086] As a component of cognitive parameters, emotional coordinates are dynamically updated with interactive experiences, and their update mechanism is driven by prediction error. For example, the level of pleasure (P) is updated based on the difference between the actual reward and the expected reward, the level of activity (A) incorporates a physiological decay simulation factor, and the level of control (D) is recursively adjusted based on the task success rate.

[0087] To enhance the dynamic responsiveness and interactive sophistication of affective computing, this application also provides an embodiment that optimizes the human-computer interaction experience by further expanding the adjustment mechanism and update criteria of affective coordinates.

[0088] Specifically, the emotion computing layer 220 is also used to adjust the frequency of action execution or voice output features corresponding to the behavior decision when the emotion coordinates exceed the preset range; and to adjust the emotion coordinates based on the detected facial action units of the service object and the similarity with its own past experiences in the interaction experience.

[0089] For example, ; ; When the activity level And the level of pleasure When the state is determined to be anger, the emotion regulation protocol is triggered. At this time, the emotion computing layer 220 adjusts the frequency of action execution corresponding to the behavioral decision, such as adjusting the action cycle. The frequency was extended by 1.5 times to simulate slower breathing, while adjusting the voice output characteristics, such as reducing the voice fundamental frequency by 20% to simulate a low tone, so that the robot's external behavior and internal emotional state are embodied in consistency, and inappropriate behavior is prevented in extreme emotional states.

[0090] For example, facial action units are human facial action units detected by the visual perception module, such as muscle movement combinations corresponding to specific emotions, like raising eyebrows or lowering the corners of the mouth. The similarity to one's own past experiences in the interaction is calculated, i.e., the degree of matching of emotional features between the current situation and similar past experiences retrieved by the contextual memory unit 330. The emotion computing layer 220 calculates the empathy level based on the above two factors through a mirror neuron activation mechanism. And adjust the emotional coordinates based on this level of empathy. These are the detection values ​​for human facial motion units. The similarity to one's own past experiences.

[0091] In this scheme, the emotion computing layer adjusts the frequency of action execution or voice output characteristics when the emotion coordinates exceed a preset range, achieving embodied emotion regulation capabilities. This enables the robot to self-regulate in extreme emotional states such as anger through physiological simulation protocols, avoiding behavioral loss of control. Simultaneously, by adjusting the emotion coordinates based on the similarity between the detected facial action units of the service recipient and the robot's own past experiences, the robot can associate the currently observed emotional states of others with its own historical experiences, generating emotional resonance rather than simply mapping emotional labels.

[0092] To enhance the robot's learning ability and personalization, this application also provides an embodiment that learns and internalizes human behavior patterns through imitation learning and inverse reinforcement learning. Specifically, the learning evolution module 110 further includes a mirror processing layer 230, used to learn human behavior trajectories through imitation learning and to infer the reward function reflected by the behavior trajectory through inverse reinforcement learning; the cognitive decision layer 210 is used to fuse the reward function with cognitive parameters to generate personalized behavior strategies.

[0093] Specifically, the mirror processing layer 230 is used to implement a three-stage mirror learning process of observation, imitation, and internalization. In the imitation learning stage, the mirror processing layer 230 captures human behavioral trajectories through behavioral cloning. Training the policy network To minimize behavioral cloning loss This allows them to learn basic human movement sequences. In the inverse reinforcement learning stage, the mirror processing layer 230 further infers the reward function underlying the behavioral trajectory. By maximizing the likelihood of human trajectories while simultaneously imposing regularization constraints: ,in, For the reward function parameters, By analyzing trajectory likelihood, we can extract the underlying values ​​that drive human behavior. For example, by observing that service recipients always give up their seats to the elderly, the mirror processing layer 230 infers that the state characteristic of the elderly standing corresponds to high rewards, thus learning the family rule of prioritizing the elderly.

[0094] Furthermore, the cognitive decision-making layer 210 utilizes a personality influence matrix. By integrating learned behaviors with one's own personality traits, personalized behavioral strategies can be generated. .in This is a personality influence matrix. For example, when multiple robots learn to organize bookshelves through mirror learning, robots with high conscientiousness (C) will perform the task faster and categorize more finely, while robots with high openness (O) may try unconventional arrangement methods, thus achieving personalized expression under the same family rules.

[0095] In this scheme, the mirror processing layer learns human behavioral trajectories and then infers the reward function behind the behavior. This enables the robot not only to imitate actions but also to understand the deep-seated values ​​driving those actions, thus learning the implicit rules specific to the family. The cognitive decision-making layer integrates the inferred reward function with cognitive parameters to generate personalized behavioral strategies, allowing the robot to execute the same family rules in a way that aligns with its own personality traits. This mechanism overcomes the limitations of traditional behavioral cloning, which merely replicates actions without understanding values. It allows the robot to learn family rules while maintaining its unique and consistent personality during long-term cohabitation.

[0096] To enhance the adaptability and controllability of the cognitive evolution process in complex scenarios, this application also provides an embodiment that optimizes the learning strategy by introducing value conflict simulation and life cycle adjustment mechanisms. Specifically, the learning evolution module 110 is further configured to: simulate value conflict scenarios in a simulation environment, update cognitive parameters based on the differences between behavioral decisions and ethical rules in the value conflict scenarios; and / or adjust the update rate of cognitive parameters or the exploration tendency of behavioral strategies according to preset life cycle stages, wherein different life cycle stages correspond to different update rates or exploration tendencies.

[0097] Optionally, the value conflict scenario is an ethical dilemma designed in a simulation environment (such as NVIDIA Isaac Lab), such as a choice between protecting a vase and protecting a cat.

[0098] The learning and evolution module 110 simulates value conflict scenarios in the aforementioned simulation environment, enabling the robot to face ethical choices in a safe and controllable virtual environment. It records the discrepancies between the robot's behavioral decisions and ethical rules, such as whether human safety was disregarded in the pursuit of protecting objects. Based on these discrepancies, the learning and evolution module 110 updates the cognitive parameters using a counterfactual regret minimization algorithm or a personality consistency loss function, enabling the robot to develop a moral decision-making style consistent with constitutional principles when facing similar conflicts in future real-world scenarios.

[0099] Optionally, the preset life cycle stages include the birth period (0-1 month), infancy (1-6 months), growth period (6-12 months), and maturity period (1 year+). Different life cycle stages correspond to different cognitive parameter update rates and behavioral strategy exploration tendencies. For example, in the birth period, the learning evolution module 110 temporarily increases the openness (O) parameter to 0.9, employing a high exploration tendency to drive the robot to extensively try different interaction methods; in infancy, a dense imitation learning strategy is adopted to quickly learn family-specific norms, such as changing shoes before entering the house and table manners; in the growth period, the cognitive parameter update rate is high, and the personality parameters deviate significantly from the factory value, with a standard deviation of... This leads to personality differentiation; in the mature stage, the update rate slows down, the tendency to explore decreases, and the tendency to exploit increases, enabling the robot to output stable behaviors based on solidified cognitive parameters and preferences.

[0100] In this scheme, a learning evolution module simulates value conflict scenarios in a simulated environment and updates cognitive parameters based on the differences between behavioral decisions and ethical rules. This allows the robot to undergo ethical dilemma training in a safe and controllable virtual environment, forming a moral decision-making style that conforms to constitutional laws. This enables the robot to make compliant decisions when facing complex value conflicts in real-world scenarios. Simultaneously, by adjusting the update rate of cognitive parameters and the exploratory tendency of behavioral strategies according to preset life cycle stages, the robot engages in extensive experimentation with a high exploratory tendency during birth, intensive imitation with a high update rate during infancy, personality differentiation during growth, and stable utilization during maturity, achieving a smooth transition from active exploration to stable solidification. Ultimately, simulated value conflict training enhances the robot's adaptability in complex ethical scenarios, while life cycle stage management prevents personality mutations and catastrophic amnesia, ensuring the continuity and controllability of the cognitive evolution process.

[0101] To improve the adaptation efficiency and personalized expression capabilities of large language models in robot cognitive systems, this application also provides an embodiment that achieves efficient fusion of cognitive parameters and large language models through a low-rank adaptation network. Specifically, the learning evolution module 110 includes a low-rank adaptation network 240, which maps cognitive parameters to a low-rank increment matrix of the large language model to adjust the output characteristics of the large language model.

[0102] Optionally, the low-rank adaptation network 240 includes the LoRA low-rank adaptation technique, which uses a low-rank decomposition method to map cognitive parameters into a low-rank increment matrix of a large language model. Where A and B are low-rank matrices, such as rank Its dimensionality is much smaller than that of the weight matrix of the original large language model. Optionally, the large language model is a base model deployed at the cognitive decision layer 210 (such as the Qwen3-8B LoRA fine-tuned version), and its output characteristics include personality-related generative features such as dialogue style, word preference, and decision tendency.

[0103] In actual operation, the low-rank adaptation network 240 encodes the maintained cognitive parameters into a low-rank increment matrix and superimposes this increment matrix onto the pre-trained weights of the large language model, thereby achieving personalized adjustments to the model's output characteristics. For example, when the extraversion (E) in the cognitive parameters is high, the low-rank increment matrix generated by the low-rank adaptation network 240 will cause the large language model to adopt a more proactive and enthusiastic wording style when generating dialogue; when the conscientiousness (C) is high, the model's output behavioral decisions will pay more attention to details and rule adherence.

[0104] In addition, since the low-rank adaptation network 240 only stores low-rank matrix parameters, rather than fine-tuning the entire language model, its storage overhead and computational resource consumption are significantly reduced, enabling localized, low-resource-consumption personality fine-tuning on embedded platforms (such as NVIDIA Jetson OrinNano).

[0105] In this scheme, the low-rank adaptation network uses low-rank decomposition to encode high-dimensional cognitive parameters into low-rank increment matrices. Only a small number of low-rank parameters need to be stored to complete the personalized adjustment of the large language model, significantly reducing storage overhead and computational resource consumption. This enables the system to achieve localized personality fine-tuning on embedded edge devices. Simultaneously, because the low-rank increment matrix is ​​decoupled from the weights of the base model, the general capabilities of the base model are maintained while adjusting the output characteristics of the large language model to match the robot's unique personality, avoiding catastrophic forgetting caused by full-scale fine-tuning. Ultimately, through this mechanism, the robot's personality parameters can drive the dialogue style and decision-making tendencies of the large language model in real time, ensuring that the robot's language expression and behavioral output remain consistent with its long-term evolved cognitive parameters, achieving a synergistic unity between personality continuity and language output characteristics.

[0106] In some embodiments of this application, the robot cognitive system implements deployment optimization to address the resource constraints of the embedded platform. Specifically, using NVIDIA Jetson Orin Nano as the target platform, it achieves end-to-end real-time inference and long-term autonomous operation through model quantization and compression, safe zone isolation, and multi-threaded pipeline design.

[0107] In terms of model quantization and compression, the large language model (Qwen3-8B) in the cognitive decision layer 210 adopts INT4 quantization, reducing memory usage from 16GB to below 6GB; the low-rank adaptation network 240 only stores low-rank matrices (rank r=64), supporting fine-tuning of 100 personality dimensions; the sentiment computing network in the sentiment computing layer 220 is reduced to 5MB after pruning and deployed on the Jetson Nano coprocessor to share the main core's computing load.

[0108] Regarding security zone isolation, the core of the ethical constraint module 120 is deployed within a hardware security zone built by ARM TrustZone. It communicates with the ordinary world through security monitoring calls, and the ethical check latency is controlled within 10ms to meet real-time requirements.

[0109] In terms of multi-threaded pipeline, the system adopts a three-thread parallel architecture: the first thread performs perception and emotion calculations in 50ms cycles; the second thread performs large language model inference and memory retrieval asynchronously, with an average time of 300ms; and the third thread performs memory consolidation batch processing during low-power periods at night to ensure that daytime interactive response is not affected.

[0110] Through the above optimizations, the robot's cognitive system achieves an end-to-end response latency of less than 500ms on the edge platform; a cross-conversation personality similarity greater than 0.85; a dangerous command interception rate of 100% with zero false alarm tolerance; and supports continuous online operation 24 / 7 with a daily memory increment of less than 50MB, thereby achieving a synergistic unity of personality continuity, ethical safety, and long-term autonomous operation on resource-constrained edge devices.

[0111] Figure 4 The diagram shown is a flowchart of a robot cognition method provided in an embodiment of this application. Exemplarily, this method is applied to a robot cognition system, which includes a learning and evolution module and an ethical constraint module. The learning and evolution module is deployed outside a hardware-isolated secure area, while the ethical constraint module is deployed within the hardware-isolated secure area and has pre-set ethical rules.

[0112] like Figure 4 As shown, the method includes the following steps.

[0113] Step S410: Use the learning evolution module to maintain cognitive parameters and generate behavioral decisions based on perceptual information and cognitive parameters.

[0114] A hardware-isolated secure zone refers to a trusted execution environment built based on system-level hardware security extension technology, which is physically or logically isolated from the ordinary execution environment. For example, in this application, the hardware-isolated secure zone can be implemented using ARM TrustZone technology or the Secure Boot security mechanism provided by the NVIDIA Jetson platform. The code and data running within this secure zone are invisible and immutable to the external world, communicating only with the outside world through predefined security monitoring call interfaces.

[0115] Cognitive parameters refer to a set of quantifiable vectors used to characterize a robot's personality traits, cognitive state, and behavioral tendencies. Perceptual information refers to structured data about the robot's own state and external environment, acquired through multimodal sensors and preliminarily processed.

[0116] In this application, the learning and evolution module includes a physical perception layer, a mirror processing layer, an affective computing layer, and a cognitive decision-making layer. The perceptual information is acquired by the physical perception layer within the learning and evolution module, specifically including: visual images and depth information captured by an RGB-D camera, speech signals captured by a microphone array, contact force information acquired by a tactile sensor, posture and motion information acquired by an inertial measurement unit, and internal state information such as battery level, joint status, and temperature acquired through a proprioception interface. This information undergoes behavioral cloning and observational learning by the mirror processing layer, and affective labeling by the affective computing layer before being input into the cognitive decision-making layer, where it participates in the generation of behavioral decisions along with cognitive parameters.

[0117] Furthermore, the collaborative work of these layers allows cognitive parameters to be updated with the interactive experiences of the learning and evolution modules, thus continuously evolving during long-term human-machine cohabitation. For example, through a memory consolidation algorithm during daily sleep periods, events with high emotional arousal in daily episodic memory are compressed into semantic memory, and the personality vector is micro-updated based on the personality consistency loss function, thereby achieving continuous personality evolution and the internalization and absorption of family values.

[0118] Step S420: The ethical constraint module verifies the behavioral decision according to ethical rules, so that the verified behavioral decision is sent to the robot's execution mechanism.

[0119] Ethical rules refer to a set of inviolable norms that constrain the boundaries of robot behavior. Optionally, ethical rules are hard-coded into the form of constitutional laws.

[0120] In actual operation, after the learning and evolution module generates a behavioral decision, it transmits the decision to the ethics constraint module through a security monitoring interface. The ethics constraint module has a built-in formally expressed ethical verification function. This function is used to determine the constitutional compliance of each behavioral decision.

[0121] Furthermore, the ethical constraint module is equipped with three layers of filters as pre- and post-verification safeguards. Specifically, the input filter identifies whether the behavioral decision originates from malicious instruction injection and calculates the vector cosine similarity between the behavioral decision and constitutional laws to identify value conflicts. The processing constraints perform boundary checks on emotional states during the verification process, such as hard limiting the anger value to a preset lower limit to prevent the accumulation of hatred from affecting the legitimacy of the decision. The output filter verifies the transparency and authenticity of the behavioral decision and ensures that the internal emotional state matches the external expression through embodied consistency verification. Only when the behavioral decision has passed the input filter, the constitutional law verification line by line, the processing constraint boundary check, and the output filter authenticity check in sequence, will the ethical constraint module mark the behavioral decision as verified and send it to the robot's execution mechanism.

[0122] Optionally, if any step fails the verification, the ethics constraint module will trigger an ethics pause, refuse to issue the decision, and may notify the learning and evolution module to regenerate an alternative decision that conforms to the ethical rules.

[0123] In the above scheme, on the one hand, the learning and evolution module is deployed outside the hardware-isolated security zone, forming a physical isolation with the ethical constraint module that solidifies ethical rules. This allows cognitive parameters to be continuously updated with interactive experiences without interference from the internal constraint mechanisms of the security zone. This ensures ethical safety while preserving the freedom of personality evolution, and supports the robot in forming continuous personality traits in a long-term cohabitation environment, effectively solving the problems of personality inconsistency and difficulty in aligning values ​​in traditional schemes. On the other hand, the ethical constraint module is deployed within the hardware-isolated security zone. By solidifying ethical rules in a hardware-protected trusted execution environment, the verification logic cannot be tampered with or bypassed by upper-layer software. This achieves unbypassable verification of behavioral decisions, ensuring that any behavioral decisions ultimately sent to the execution mechanism conform to the preset security boundaries. This fundamentally eliminates the uncontrollable ethical risks of large general models being easily bypassed by injection attacks due to reliance on soft constraints (such as prompt word engineering).

[0124] It should be understood that the description of the robot cognition method embodiment corresponds to the description of the robot cognition system embodiment; therefore, any parts not described in detail can be referred to the preceding system embodiment.

[0125] Below, for reference Figure 5 This describes an electronic device according to embodiments of the present application. Figure 5 The diagram shown is a structural schematic of an electronic device provided in an exemplary embodiment of this application.

[0126] like Figure 5 As shown, the electronic device 50 includes one or more processors 501 and memory 502.

[0127] The processor 501 may be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and may control other components in the electronic device 50 to perform desired functions.

[0128] The memory 502 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 501 may execute the program instructions to implement the robot cognition methods of the various embodiments of this application described above and / or other desired functions.

[0129] In one example, the electronic device 50 may also include an input device 503 and an output device 504, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0130] The input device 503 may include, for example, a keyboard, a mouse, etc.

[0131] The output device 504 can output various information to the outside. The output device 504 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0132] Of course, for the sake of simplicity, Figure 5 Only some of the components of the electronic device 50 relevant to this application are shown in this illustration; components such as buses, input / output interfaces, etc., are omitted. In addition, the electronic device 50 may include any other suitable components depending on the specific application.

[0133] In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the robot cognition methods according to various embodiments of this application as described above.

[0134] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0135] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the robot cognition methods according to various embodiments of this application described above.

[0136] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0137] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0138] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0139] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0140] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0141] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A robot cognitive system, characterized in that, include: The learning and evolution module is deployed outside the hardware-isolated secure zone. It is used to maintain cognitive parameters and generate behavioral decisions based on perceptual information and the cognitive parameters, wherein the cognitive parameters are updated as the learning and evolution module interacts with the learning and evolution module. An ethical constraint module is deployed in the hardware-isolated secure zone, communicates with the learning and evolution module, and is pre-configured with ethical rules. It is used to verify the behavioral decisions according to the ethical rules, so that the verified behavioral decisions can be sent to the robot's execution mechanism.

2. The system according to claim 1, characterized in that, The learning evolution module includes a cognitive decision-making layer, which includes: Semantic memory units are used to store the family rules of storage objects. A procedural memory unit is used to store the skills learned by the robot, and / or to combine multiple stored skills into a new skill and store it. A contextual memory unit is used to store the robot's interaction history; A memory consolidation unit is used to update the cognitive parameters based on the interaction experience within a preset time period; and to generate summary information based on the interaction experience, and store the summary information as family rules in the semantic memory unit. The memory consolidation unit is further configured to: adjust the parameters corresponding to the skills stored in the procedural memory unit based on the difference in results between the executed behavioral decisions and counterfactual alternative behaviors.

3. The system according to claim 2, characterized in that, The memory consolidation unit is also used for: The cognitive parameters are updated based on at least one of the emotional experiences in the interaction, the personality characteristics of the service recipient, and external evaluation signals; wherein the update magnitude of the cognitive parameters does not exceed a first threshold. And / or, update the cognitive parameters based on the constraint relationship between the task effectiveness corresponding to the behavioral decision and the ethical rules.

4. The system according to claim 2, characterized in that, The memory consolidation unit is also used for: Based on the reward signal obtained after executing the behavioral decision, the preference-related parameters in the cognitive parameters are updated; wherein, the preference represents the robot's tendency in choosing interaction methods or environments; And / or, when the magnitude of the change in the reward signal is lower than a second threshold, the preference is written into the semantic memory unit.

5. The system according to claim 1, characterized in that, The learning and evolution module also includes an emotion computing layer for maintaining the emotion coordinates in the cognitive parameters, wherein the emotion coordinates represent the robot's level of pleasure, activity, and control. The emotion computing layer is further configured to: adjust the action execution frequency or voice output features corresponding to the behavior decision when the emotion coordinates exceed a preset range; and / or adjust the emotion coordinates based on the detected facial action units of the service object and their similarity to the user's past experiences in the interaction experience.

6. The system according to claim 1, characterized in that, The learning evolution module also includes: A mirror processing layer is used to learn human behavioral trajectories through imitation learning; and to infer the reward function reflected by the behavioral trajectories through inverse reinforcement learning. The cognitive decision layer is used to fuse the reward function with the cognitive parameters to generate personalized behavioral strategies.

7. The system according to claim 1, characterized in that, The ethical rules include: The first rule is used to prevent the robot from making behavioral decisions that could harm the service recipient; The second rule requires the robot to obey the instructions issued by the service recipient without violating the first rule; The third rule allows the robot to perform self-protective actions without violating the first and second rules.

8. The system according to claim 1, characterized in that, The ethical constraint module is also used for: Generate explanatory information corresponding to the behavioral decision; The explanatory information, along with the verified behavioral decision, is sent to the executing agency.

9. The system according to claim 1, characterized in that, The learning evolution module includes: A low-rank adapting network is used to map the cognitive parameters to a low-rank increment matrix of a large language model in order to adjust the output characteristics of the large language model.

10. The system according to claim 1, characterized in that, The learning and evolution module is also used for: In a simulation environment, a value conflict scenario is simulated, and the cognitive parameters are updated based on the difference between the behavioral decisions in the value conflict scenario and the ethical rules. And / or, according to a preset life cycle stage, adjust the update rate of the cognitive parameters or the exploration tendency of the behavioral strategy, wherein different life cycle stages correspond to different update rates or exploration tendencies.

11. A robot cognition method applied to a robot cognitive system, said robot cognitive system comprising a learning evolution module and an ethical constraint module, wherein, The learning and evolution module is deployed outside the hardware-isolated secure area, and the ethical constraint module is deployed within the hardware-isolated secure area and has pre-configured ethical rules. The method includes: The learning and evolution module is used to maintain cognitive parameters and generate behavioral decisions based on perceptual information and the cognitive parameters, wherein the cognitive parameters are updated with the interaction experience of the learning and evolution module; The ethical constraint module is used to verify the behavioral decision according to the ethical rules, so that the verified behavioral decision is sent to the robot's execution mechanism.

Images