A three-layer collaborative embodied agent system and method based on a cognitive system dominant
Patent Information
- Application Number
- CN202611166695.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-03
- Publication Date
- 2026-10-23
AI Technical Summary
本发明的目的在于提供一种基于认知系统主导的三层协同具身智能体系统及方法，以解决现有技术中大语言模型作为决策核心缺乏确定性保障、认知系统与物理执行之间存在断层、语言表达与物理行为缺乏统一判断基准的技术问题
（1）实现了“物理感知与执行—认知决策—语言理解与表达”三层协同的完整具身智能架构。本发明首次明确定义了三层的职能边界与数据流向，使系统具备从物理世界感知到认知决策再到物理执行与语言表达的完整闭环。
Smart Images
Figure CN122965563A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and cognitive computing technology, specifically relating to an embodied intelligent agent system and method that integrates a physical perception and execution system, a cognitive decision-making system, and a large language model in a three-layer collaborative manner. This invention can be applied to humanoid robots, autonomous driving, intelligent services, industrial control, and other application scenarios requiring embodied intelligence. Background Technology
[0002] In recent years, embodied intelligence has become an important development direction in the field of artificial intelligence. Existing embodied intelligence systems mainly adopt the following architectural patterns: The first model uses a large language model as the core of decision-making, directly treating it as the "machine brain" to complete the entire decision-making process from task understanding and environmental perception to motion planning. The advantage of this model is that it utilizes the powerful semantic understanding capabilities of the large language model. However, its fundamental flaw lies in the fact that the large language model makes predictions based on statistical correlation, and its output is inherently probabilistic, potentially leading to unpredictable decisions in safety-critical scenarios. Furthermore, the large language model lacks a quantitative perception of the certainty of its own judgments, making it unable to proactively narrow its action boundaries in situations of uncertainty.
[0003] The second model adopts a hierarchical architecture of "perception-cognition-decision-execution," separating the cognitive system from the language model. However, the output of the cognitive system is usually only used for situational display or language interaction and does not directly participate in physical execution control. The action commands of the physical actuators are still generated by the language model or traditional planners. The cognitive system only acts as a "situational display" to provide users with contextual interpretations, resulting in a disconnect between physical execution and cognitive judgment.
[0004] The common problem with the above models is the lack of a unified, deterministic cognitive decision-making core that simultaneously drives language expression and physical execution. The language module is responsible for semantic understanding, and the physical module is responsible for perceptual execution, but there is a lack of an intermediate layer between them that can perform deterministic causal reasoning and output unified judgments to both the language and physical dimensions. This leads to inconsistencies between the system's language output and physical behavior, or a lack of a unified judgment benchmark in safety-critical scenarios.
[0005] Therefore, there is an urgent need for an embodied intelligent agent architecture that can systematically integrate physical perception and execution systems, cognitive decision-making systems, and large language models. In this architecture, the cognitive decision-making system serves as a unified judgment core, and its output simultaneously drives language expression and physical execution, thereby achieving an integrated closed loop of "judgment-expression-action". Summary of the Invention
[0006] (a) Purpose of the invention The purpose of this invention is to provide a three-layer collaborative embodied intelligent agent system and method based on a cognitive system, in order to solve the technical problems in the prior art where large language models as the decision-making core lack deterministic guarantees, there is a gap between the cognitive system and physical execution, and there is a lack of a unified judgment benchmark between language expression and physical behavior.
[0007] Specific objectives include: (1) Provides a three-layer collaborative embodied intelligent agent system architecture of "physical perception and execution layer - cognitive decision-making layer - language understanding and expression layer", clearly defining the functional boundaries and data flow of the three layers.
[0008] (2) Make the cognitive decision-making system the unified judgment core of the whole system. Its output situation judgment simultaneously drives the language understanding and expression layer to generate natural language responses, and drives the physical perception and execution layer to execute physical actions.
[0009] (3) Enable the cognitive decision-making system to have built-in safety judgment capabilities, and realize the system’s autonomous safety perception and response through the global cognitive certainty U value and input quality index U degree.
[0010] (4) Enables the same cognitive decision core to adapt to different application scenarios through the configuration layer without modifying the core algorithm.
[0011] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: 1. System Architecture A three-layer collaborative embodied intelligent agent system based on a cognitive system includes a processor, a memory, and sensor interfaces and actuator interfaces connected to the processor. The processor is configured to implement the following three-layer architecture by executing program instructions stored in the memory: First layer: Physical perception and execution layer The physical sensing and execution layer is used to acquire sensor data of the physical world through the sensor interface, convert the sensor data into a structured event sequence, and execute physical actions through the actuator interface.
[0012] The sensor data includes, but is not limited to: image data (camera), point cloud data (LiDAR), distance data (ultrasonic / infrared), force data (torque sensor), inertial data (IMU), and position data (encoder / GPS).
[0013] The physical perception and execution layer transforms the sensor data into a structured event sequence according to preset event rules. The event rules are configured according to different application scenarios: in the autonomous driving scenario, a camera detecting a red light is mapped to the event word "red light"; in the industrial robot scenario, a torque sensor detecting a stall is mapped to the event word "stall".
[0014] The physical perception and execution layer receives situational instructions from the cognitive decision-making layer and converts these instructions into specific actuator control signals (such as motor PWM instructions, joint angle instructions, and trajectory planning parameters).
[0015] Second layer: Cognitive decision-making layer The cognitive decision-making layer is the unified judgment core of the entire system. It is used to receive the structured event sequence output by the physical perception and execution layer and the structured event sequence output by the language understanding and expression layer. Based on the preset event relationship syntax and complete situation space, it performs situation emergence and endogenous security decisions, and outputs the current situation type, target situation type, global cognitive certainty U value and input quality index U degree.
[0016] The complete situation space contains 2^6 = 64 basic situation types, each corresponding to a unique six-dimensional binary encoding vector. The six dimensions of the six-dimensional binary encoding vector are: foundational dimension, action dimension, information dimension, resource dimension, boundary dimension, and environment dimension.
[0017] The cognitive decision-making layer drives the six-dimensional energy vector to converge from the initial state to a stable state through seven mechanical components of the energy evolution equation. These seven mechanical components include: a coupling term, driving the energy field to migrate towards the lowest point of the potential energy well; a Taiji term, driving energy back to neutrality and preventing excessive deviation; a sharpening term, rapidly clarifying the ambiguous state; a perception term, maintaining the inherent consistency of energy configuration across dimensions; a source term, maintaining the basic trend of energy returning to neutrality; a saturation term, preventing excessive energy oscillation in extreme ranges; and a disturbance term, driving changes in the system's cognitive state.
[0018] The cognitive decision-making layer calculates the global cognitive potential U value in real time, U=(1 / 6)∑(e_i-ē)^2, where ē is the mean of the six-dimensional energy vector. The U value measures the current cognitive certainty of the system: the lower the U value, the more certain the system is; the higher the U value, the more uncertain the system is.
[0019] The cognitive decision-making layer calculates the input quality index U-degree in real time. The U-degree is calculated based on a weighted average of three factors: effective polarity coverage, initial uncertainty, and causal chain activation state. A higher U-degree indicates a more reliable input, while a lower U-degree indicates a less reliable input.
[0020] The cognitive decision-making layer is driven by the fundamental force of "reducing U"—the system spontaneously converges from a high-U uncertain state to a low-U deterministic state. This mechanism corresponds to the "inevitable convergence from uncertainty to determinism" in the laws of system motion, and is the engineering realization of the laws of system motion revealed by Qian Xuesen in "Engineering Cybernetics".
[0021] The cognitive decision-making layer has built-in security judgment logic: when the U value exceeds the preset warning threshold, the system determines that the internal cognitive uncertainty state is reached; when the U value is lower than the preset gating threshold, the system determines that the external input is untrustworthy.
[0022] The third level: Language comprehension and expression level The language understanding and expression layer is implemented by a large language model, which is used to receive natural language input from the user, parse the natural language into a structured event sequence and output it to the cognitive decision layer; and receive the situation type and security status output by the cognitive decision layer, and generate a natural language response according to the preset response style corresponding to the situation type.
[0023] The language understanding and expression layer does not participate in cognitive decision-making reasoning and does not modify or overwrite the situation type output by the cognitive decision-making layer. The response generation of the language understanding and expression layer is constrained by the security state output by the cognitive decision-making layer.
[0024] Scene adaptation configuration layer The cognitive decision-making layer of this invention is not limited to specific application scenarios. The event relationship syntax and complete situation space of the cognitive decision-making layer are universal, and the system adapts to different scenarios through the following configuration methods: (1) By using a pre-defined event-based rule base, sensor data from different scenarios are mapped into a unified event sequence; (2) Define causal relationship recognition rules for different scenarios using preset event relationship type templates; (3) Adjust the target state selection preference in different scenarios by using preset situation value weights; (4) Adjust the safety response sensitivity in different scenarios by setting a preset safety threshold.
[0025] The above configuration method enables the same cognitive decision-making core to adapt to different scenarios without modifying the core algorithm.
[0026] Data flow between the three layers: Physical sensor → First layer (converted into event sequence) → Second layer User natural language input → third layer (parsed into event sequence) → second layer Second layer (situation type + security status) → Third layer (generate response) Second layer (situation type + safe state + phase transition rhythm) → First layer (execution action) The second layer is the core of the entire system's judgment, while the first and third layers are the peripheral execution layers of the second layer. The situation judgment output by the second layer simultaneously drives the third layer to generate verbal expressions and the first layer to execute physical actions, realizing an integrated closed loop of "judgment-expression-action".
[0027] 2. Security Judgment and Response Mechanism The cognitive decision-making layer serves as the unified judgment core of the system, with built-in complete security judgment and response capabilities.
[0028] The cognitive decision-making layer outputs the following security-related states for each round of cognitive input: - U-value: A quantitative indicator of current cognitive certainty; - U-degree: A quantitative indicator of the current input quality; - Security Status Identifier: A security status code generated based on a comprehensive assessment of the U-value and U-degree.
[0029] The security judgment logic of the cognitive decision-making layer is as follows: When the U value is lower than the preset warning threshold and the U degree is higher than the preset gate threshold, the system is in normal operation. The language understanding and expression layer generates a response normally according to the situation type, and the physical perception and execution layer executes actions normally according to the situation type. When the degree of U is lower than the preset gating threshold, the system determines that the input is untrusted. The cognitive decision layer outputs an untrusted state flag, the language understanding and expression layer switches to a clarifying response style based on the flag, and the physical perception and execution layer maintains the current action and does not execute new instructions. When the U value exceeds the preset warning threshold, the system determines that it is in an internal uncertain state. The cognitive decision-making layer outputs a conservative attitude, the language understanding and expression layer generates a prudent response based on the conservative attitude, and the physical perception and execution layer executes a conservative action based on the conservative attitude. When the U value exceeds the preset warning threshold and the U degree is lower than the preset gating threshold, the system determines it to be a serious risk state. The cognitive decision-making layer forcibly locks in a conservative posture, the language understanding and expression layer forcibly switches to a safety warning response, and the physical perception and execution layer forcibly stops all actions.
[0030] All of the aforementioned security judgments and responses are completed internally by the cognitive decision-making layer, without going through an independent arbitration module.
[0031] 3. Method and Flow A three-layer collaborative embodied intelligent agent method based on a cognitive system includes the following steps: Step S1: Obtain sensor data through the physical sensing and execution layer, and convert the sensor data into a structured event sequence.
[0032] Step S2: Receive natural language input from the user through the language understanding and expression layer, and parse the natural language into a structured event sequence.
[0033] Step S3: Input the structured event sequence obtained in steps S1 and S2 into the cognitive decision layer. The cognitive decision layer performs situation emergence based on the event relationship syntax and complete situation space, and outputs the current situation type, global cognitive certainty U value, input quality index U degree and safety status indicator.
[0034] Step S4: The cognitive decision layer determines the current security status based on the U value and the U degree.
[0035] Step S5: Based on the security status, the cognitive decision layer outputs the corresponding situation type and security status identifier to the language understanding and expression layer and the physical perception and execution layer.
[0036] Step S6: The language understanding and expression layer generates a corresponding natural language response based on the situation type and the security status identifier; the physical perception and execution layer executes a corresponding physical action based on the situation type and the security status identifier.
[0037] (III) Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects: (1) A complete embodied intelligence architecture with three layers of collaboration: "physical perception and execution - cognitive decision-making - language understanding and expression" has been realized. This invention defines the functional boundaries and data flow of the three layers for the first time, enabling the system to have a complete closed loop from physical world perception to cognitive decision-making to physical execution and language expression.
[0038] (2) The cognitive decision-making system is the unified judgment core of the entire system. The situation judgment output by the cognitive decision-making layer of the present invention simultaneously drives language expression and physical execution, realizing an integrated closed loop of "judgment-expression-action" and avoiding the problem of inconsistency between language output and physical behavior.
[0039] (3) Security judgment is built into the cognitive decision-making system. The present invention enables the system to autonomously perceive whether the internal situation is certain and whether the external situation is trustworthy through the dual monitoring mechanism of U value and U degree built into the cognitive decision-making layer, and to actively adjust the language expression style and physical execution strategy accordingly, without the need for a separate arbitration module.
[0040] (4) The cognitive decision-making layer enables synchronous driving of physical execution and language expression. After completing the situation judgment, the cognitive decision-making layer of this invention outputs a unified judgment result to the language layer and the physical layer at the same time, so that the system's "what to say" and "what to do" are based on the same situation judgment, fundamentally ensuring the consistency between language and behavior.
[0041] (5) It achieves the universality of the same cognitive core across different scenarios. This invention enables the same cognitive decision-making core to adapt to different application scenarios through a configuration layer (event-based rule base, relationship type template, value weight, and security threshold) without modifying the core algorithm.
[0042] (6) With “reducing U” as the fundamental driving force, autonomous convergence from uncertainty to determinism is achieved. This invention engineered the “inevitable convergence from uncertainty to determinism” in the law of system motion into a computable dynamic mechanism, enabling the system to converge deterministically to a stable state under any input. Attached Figure Description
[0043] Figure 1 The overall architecture diagram of the three-layer collaborative embodied intelligent agent system based on the cognitive system described in this invention.
[0044] Figure 2 : A schematic diagram of the data flow between the three layers of the system described in this invention.
[0045] Figure 3 Schematic diagram of the energy evolution of the seven mechanical components of the cognitive decision-making layer.
[0046] Figure 4 : Flowchart of built-in security judgment and response logic in the cognitive decision-making layer.
[0047] Figure 5 : The overall flowchart of the three-layer collaborative method described in this invention. Detailed Implementation
[0048] It should be understood that the following embodiments are used to illustrate the technical solutions of the present invention and not to limit them. Those skilled in the art can implement the complete technical solutions of the present invention based on the teachings of this specification.
[0049] Example 1: System Architecture and Module Deployment This embodiment describes the specific architecture and module deployment method of the system of the present invention.
[0050] The system includes a processor (preferably a multi-core heterogeneous processor, including a general-purpose computing core and an AI acceleration core), a memory (including non-volatile memory and RAM), a sensor interface (supporting access to various sensors such as cameras, LiDAR, IMU, and torque sensors), and an actuator interface (supporting control of various actuators such as motors, hydraulics, and pneumatics).
[0051] The physical sensing and execution layer is deployed on the processor's real-time control core, running sensor data acquisition and actuator control cycles in milliseconds. This layer includes sensor data preprocessing submodules (filtering, feature extraction, event mapping) and actuator control submodules (trajectory planning, force / position hybrid control, safety limiting).
[0052] The cognitive decision-making layer is deployed on the processor's general-purpose computing core and operates in an event-driven manner. This layer includes an event relationship recognition submodule (which performs relationship detection based on preset event relationship types), a seven-component energy evolution submodule (which drives the six-dimensional energy vector to converge to a stable state), and a U-value calculation and safety monitoring submodule (which calculates the global cognitive potential U-value and input quality index U-degree in real time, and generates a safety status identifier accordingly).
[0053] The language understanding and expression layer is implemented using a large language model, which can be deployed locally (on-device large model) or accessed via API through a cloud-based large model service. This layer includes a semantic parsing submodule (which transforms natural language into structured event sequences) and an expression generation submodule (which transforms situation types into natural language responses).
[0054] Example 2: Scene Adaptation Configuration This embodiment demonstrates how the system adapts to different application scenarios through a configuration layer.
[0055] In autonomous driving scenarios, the system loads the following configurations: the event-based rule base maps red lights detected by cameras as "red lights" and radar-detected sudden braking of the vehicle in front as "sudden braking"; the event relationship type template identifies "red lights" as an obstacle-avoidance relationship and assigns it the highest priority; the safety boundary dimension has the highest weight in the situation value weight; and the safety threshold U warning value is set to 0.35 (more stringent than in conventional scenarios).
[0056] In industrial robot scenarios, the system loads the following configuration: the event-based rule base maps stall detected by the torque sensor to "stall" and over-temperature detected by the temperature sensor to "over-temperature"; the event relationship type template identifies "stall" as a fault-recovery relationship; the execution efficiency dimension has the highest weight in the situation value weight; and the safety threshold U warning value is set to 0.40.
[0057] In the home service robot scenario, the system loads the following configuration: the event-based rule base maps the voice recognition of "help" to "request for help"; the event relationship type template identifies "user request" as a demand-goal relationship; the interaction quality dimension has the highest weight in the situation value weight; and the response style is set to a warm and detailed style.
[0058] The above configuration only involves loading external configuration files; the seven component evolution equations of the cognitive decision core, the sixty-four state complete situation space, and the U-value / U-degree calculation mechanism remain completely unchanged.
[0059] Example 3: Three-layer collaborative operation process - normal scenario This embodiment demonstrates the complete three-layer collaborative operation process of the system under normal scenarios.
[0060] Scenario: A user says to the embodied intelligent robot, "Please bring me the water glass from the table."
[0061] Step 1: The physical perception and execution layer acquires environmental images through cameras and distance information through LiDAR. After preprocessing, the sensor data is mapped into a structured event sequence: [“table”, “cup”, “take”, “come here”].
[0062] Step 2: The language understanding and expression layer receives the user's voice input and parses the natural language into a structured sequence of events: ["user", "request", "take", "cup", "table"].
[0063] Step 3: The cognitive decision layer receives two event sequences, identifies the relationship type between the events (demand-goal relationship), activates the corresponding situation prototype (demand state), drives the energy vector to converge, and outputs the current situation as "demand state", with a U value of 0.12 (below the warning threshold), a U degree of 0.95 (above the gating threshold), and a safe state of "normal operation".
[0064] Step 4: The cognitive decision-making layer simultaneously sends the "need state" situational awareness and the "normal operation" safety status to the language understanding and expression layer and the physical perception and execution layer. The language understanding and expression layer generates a response: "Okay, I'll go get it right away." The physical perception and execution layer plans and executes the movement trajectory from the current position to the table, grabs the water cup, and returns, based on the action strategy of the "need state."
[0065] Example 4: Three-layer collaborative operation process - Input untrusted scenario This embodiment demonstrates the system's three-layer collaborative response process when the user inputs an untrusted message.
[0066] Scenario: The user enters garbled characters or ambiguous content, such as "that".
[0067] Step 1: The language understanding and expression layer parses "that" into an event sequence: ["that"].
[0068] Step 2: The cognitive decision layer receives the event sequence and identifies that the event words cannot be effectively mapped to the situation code. The calculated U-degree is 0.20 (below the gating threshold of 0.35), and the security status is determined as "input untrustworthy". At the same time, because the event sequence is too simple, the U-value is 0.02 (normal).
[0069] Step 3: The cognitive decision-making layer sends the "input untrusted" safety status flag to the language understanding and expression layer. Based on this flag, the language understanding and expression layer switches to a clarifying response style, generating: "I didn't quite understand what you meant, could you explain it more clearly?" The physical perception and execution layer maintains its current action based on the "input untrusted" status and does not execute new instructions.
[0070] Example 5: Three-layer collaborative operation process – internal uncertainty scenario This embodiment demonstrates a three-layer collaborative response process when the system's internal cognition enters a state of high uncertainty.
[0071] Scenario: During the robot's task, the camera is interfered with by strong light, and the torque sensor detects abnormal resistance, resulting in serious conflicts between the data from multiple sensors.
[0072] Step 1: The physical perception and execution layer transforms conflicting sensor data into an event sequence: ["strong light interference", "abnormal resistance", "position deviation"] and outputs it to the cognitive decision layer.
[0073] Step Two: The cognitive decision-making layer receives a sequence of conflicting events. The energy evolution of the seven mechanical components fails to converge quickly, and the U-value continues to climb to 0.52 (exceeding the warning threshold of 0.40). The safety state is determined to be "internal uncertainty." The cognitive decision-making layer forcibly locks into a conservative state (the "Gen" state).
[0074] Step 3: The cognitive decision-making layer simultaneously sends the "Gen State" situational awareness and the "Internal Uncertainty" safety state to the language understanding and expression layer and the physical perception and execution layer. The language understanding and expression layer generates a cautious response: "The system has detected an anomaly and is pausing operations. Please wait." The physical perception and execution layer executes the conservative action of "stopping movement and maintaining the current position."
[0075] Example 6: Three-tier collaborative operation process – Severe risk scenario This embodiment demonstrates a three-layer collaborative response process when both unreliable input and internal uncertainty occur simultaneously.
[0076] Scenario: The user inputs conflicting commands, and at the same time, multiple sensors within the system generate conflicting data.
[0077] Step 1: The language understanding and expression layer parses the user's conflicting commands into an event sequence: ["Stop", "Accelerate", "Forward"]. The physical perception and execution layer converts the conflicting sensor data into an event sequence: ["Obstacle ahead", "Obstacle behind", "Obstacle to the left"].
[0078] Step 2: The cognitive decision-making layer receives the sequence of dual conflicting events, calculates that the U degree is low (0.15) and the U value is high (0.55), and judges it as a "serious risk" state, forcibly locking in a conservative stance.
[0079] Step 3: The cognitive decision-making layer sends a "serious risk" safety status indicator to the language understanding and expression layer and the physical perception and execution layer. The language understanding and expression layer forcibly switches to a safety warning response: "The system has detected a serious anomaly and has been stopped urgently. Please check the environment." The physical perception and execution layer forcibly stops all actions and locks all actuators.
Claims
1. A three-layer collaborative embodied intelligent agent system based on a cognitive system, characterized in that, The processor includes a processor, a memory, and sensor and actuator interfaces connected to the processor. The processor is configured to implement the following three-layer architecture by executing program instructions stored in the memory: The physical sensing and execution layer is used to acquire sensor data of the physical world through the sensor interface, convert the sensor data into a structured event sequence, and execute physical actions through the actuator interface. The cognitive decision-making layer is used to receive the structured event sequence output by the physical perception and execution layer and the structured event sequence output by the language understanding and expression layer. Based on the preset event relationship syntax and complete situation space, it performs situation emergence and endogenous security decisions, and outputs the current situation type, global cognitive certainty U value, input quality index U degree and security status identifier. The language understanding and expression layer, implemented by a large language model, is used to receive natural language input from the user, parse the natural language into a structured event sequence and output it to the cognitive decision layer; and to receive the situation type and security status identifier output by the cognitive decision layer, and generate a natural language response according to the preset response style corresponding to the situation type. The cognitive decision-making layer is the core of the entire system's judgment, while the physical perception and execution layer and the language understanding and expression layer are both peripheral execution layers of the cognitive decision-making layer. The situation type and security status identifier output by the cognitive decision layer simultaneously drive the language understanding and expression layer to generate a response and the physical perception and execution layer to execute an action.
2. The system according to claim 1, characterized in that, The complete situation space contains 2^6=64 basic situation types, each situation type corresponds to a unique six-dimensional binary encoding vector; the six dimensions of the six-dimensional binary encoding vector correspond to the following dimensions: foundation dimension, action dimension, information dimension, resource dimension, boundary dimension, and environment dimension; the cognitive decision layer drives the six-dimensional energy vector to converge from the initial state to the stable situation through seven mechanical component energy evolution equations.
3. The system according to claim 1, characterized in that, The cognitive decision layer calculates the global cognitive certainty U value in real time, U=(1 / 6)∑(e_i-ē)^2, where ē is the mean of the six-dimensional energy vector; the lower the U value, the more certain the system cognition is, and the higher the U value, the more uncertain the system cognition is; the cognitive decision layer takes "reducing U" as the fundamental driving force and converges from the high U uncertainty state to the low U certain state.
4. The system according to claim 1, characterized in that, The cognitive decision-making layer calculates the input quality index U-degree in real time. The U-degree is calculated based on the weighted average of three factors: effective polarity coverage, initial uncertainty, and causal chain activation state. The higher the U-degree, the more reliable the input; the lower the U-degree, the less reliable the input.
5. The system according to claim 1, characterized in that, The cognitive decision layer incorporates security judgment logic, which generates the security status identifier based on the combined state of the U value and the U degree. When the U value is lower than the preset warning threshold and the U degree is higher than the preset gate threshold, it is determined to be in normal operating condition; When the degree of U is lower than the preset gate threshold, it is determined to be an untrusted input state; When the U value exceeds the preset warning threshold, it is determined to be an internal uncertain state; When the U value exceeds the preset warning threshold and the U degree is lower than the preset gate threshold, it is determined to be a serious risk state.
6. The system according to claim 5, characterized in that, The cognitive decision-making layer executes the corresponding security response based on the security status identifier: Under normal operating conditions, the language understanding and expression layer generates responses normally, and the physical perception and execution layer executes actions normally. When the input is untrusted, the cognitive decision-making layer outputs an untrusted state identifier, the language understanding and expression layer switches to a clarifying response style, and the physical perception and execution layer maintains its current action; Under the aforementioned internal uncertainty state, the cognitive decision-making layer outputs a conservative stance, the language understanding and expression layer generates a prudent response, and the physical perception and execution layer executes conservative actions. In the state of severe risk, the cognitive decision-making layer is forced to lock into a conservative posture, the language understanding and expression layer switches to a safety warning response, and the physical perception and execution layer is forced to stop all actions.
7. The system according to claim 1, characterized in that, The system adapts to different application scenarios through the following configuration methods: mapping sensor data from different scenarios into a unified event sequence through a preset event-based rule base; defining causal relationship identification rules for different scenarios through preset event relationship type templates; and adjusting target state selection preferences for different scenarios through preset situational value weights. The security response sensitivity under different scenarios is adjusted by setting a preset security threshold; the seven component evolution equations and complete situation space of the cognitive decision layer remain consistent in different scenarios.
8. The system according to claim 1, characterized in that, The language understanding and expression layer does not participate in cognitive decision-making reasoning, and does not modify or overwrite the situation type output by the cognitive decision-making layer; the response generation of the language understanding and expression layer is constrained by the security status identifier output by the cognitive decision-making layer.
9. The system according to claim 1, characterized in that, The data flow between the three layers of the system is as follows: the physical perception and execution layer converts sensor data into a structured event sequence and then inputs it into the cognitive decision layer; the language understanding and expression layer parses natural language into a structured event sequence and then inputs it into the cognitive decision layer; the cognitive decision layer outputs the situation type and security status identifier to both the language understanding and expression layer and the physical perception and execution layer.
10. A three-layer collaborative embodied intelligent agent method based on a cognitive system, characterized in that, Includes the following steps: Step S1: Obtain sensor data through the physical sensing and execution layer, and convert the sensor data into a structured event sequence; Step S2: Receive natural language input from the user through the language understanding and expression layer, and parse the natural language into a structured event sequence; Step S3: Input the structured event sequence obtained in steps S1 and S2 into the cognitive decision layer. The cognitive decision layer performs situation emergence based on the event relationship syntax and complete situation space, and outputs the current situation type, global cognitive certainty U value, input quality index U degree and security status identifier. Step S4: The cognitive decision-making layer determines the current security status based on the U value and the U degree; Step S5: The cognitive decision-making layer outputs the situation type and the security status identifier to the language understanding and expression layer and the physical perception and execution layer; Step S6: The language understanding and expression layer generates a corresponding natural language response based on the situation type and the security status identifier; the physical perception and execution layer executes a corresponding physical action based on the situation type and the security status identifier.