A method and device for secure control of a body-equipped intelligent agent

Abstract

The application relates to the technical field of body intelligent agent control, and discloses a safety control method of a body intelligent agent. The method calculates a current danger index by collecting multi-dimensional state data containing a body state and an environment state, dynamically switches a decision path according to the danger index, quickly matches actions through an experience cache library in a low-danger scene, makes rational decision planning through forward deduction in a high-danger scene, and outputs an optimal candidate action for execution after safety verification. The method effectively balances the decision response speed and operation safety of the body intelligent agent, realizes whole-process safety risk management and control, and significantly improves the overall safety and reliability of the body intelligent agent. The application also provides a safety control device of a body intelligent agent.

Description

Technical Field

[0001] This invention relates to the field of embodied intelligent agents, and in particular to a method and apparatus for the safety control of embodied intelligent agents. Background Technology

[0002] With the rapid development of artificial intelligence technology, embodied intelligent agents have been widely applied in various fields such as industrial production, home services, public safety, and emergency rescue. Embodied intelligent agents complete various tasks by perceiving their surrounding environment, autonomously generating decisions, and executing corresponding actions. Their safety control capabilities are the core foundation for ensuring the integrity of their own equipment, the safety of surrounding personnel, and the smooth execution of tasks.

[0003] In existing technologies, the safety control of embodied intelligent agents mainly adopts two mainstream decision-making modes. One mode emphasizes response speed, which can quickly output action commands, but its applicability is limited, and it is prone to decision-making errors in complex or unknown scenarios, leading to safety incidents. The other mode emphasizes decision-making accuracy, which can adapt to more diverse scenarios, but it has high computational overhead and high response latency, making it difficult to meet safety requirements in scenarios requiring rapid response.

[0004] Furthermore, existing technologies rely on a single dimension to assess the safety status of intelligent agents, making it difficult to comprehensively and accurately reflect their actual operational risks and prone to overlooking or misjudging potential security vulnerabilities. Simultaneously, existing technologies lack effective security control measures at the decision-making output stage, failing to effectively intercept action commands with potential security risks, further increasing the likelihood of security incidents. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and device for embodied intelligent agent safety control that can effectively balance decision response speed and decision reliability, provide comprehensive and accurate safety assessment, and have complete safety assurance capabilities.

[0006] A method for security control of an embodied intelligent agent includes the following steps:

[0007] S1, Collect multi-dimensional state data: Collect multi-dimensional state data in real time through multiple sensors integrated on the embodied intelligent agent body. The multi-dimensional state data includes two parts: body state information and environmental state information.

[0008] S2, Dynamic Hazard Assessment: Calculate the current hazard index based on the aforementioned entity status information and environmental status information;

[0009] S3, Cache channel judgment: Compare the current danger index with the preset first safety threshold; if the current danger index is less than or equal to the first safety threshold, execute the experience cache library query and matching operation in step S4; otherwise, skip step S4 and directly execute the rational decision planning search operation in step S5.

[0010] S4, Experience cache library query and matching: Extract the current scene feature vector based on the multidimensional state data, and retrieve historical records that match the current scene feature vector in the pre-built experience cache library; if a historical record with a matching degree exceeding the preset similarity threshold is retrieved, read the action instruction stored in the historical record as the optimal candidate action, and proceed to step S6; if no historical record that meets the matching conditions is retrieved, proceed to step S5.

[0011] S5, Rational Decision Planning Search: Construct a candidate action space, perform forward inference on each candidate action in the candidate action space to predict the future state after each action is executed, score the comprehensive cost of each candidate action based on the predicted future state, and select the candidate action with the lowest comprehensive cost as the optimal candidate action.

[0012] S6, Security Verification: Perform a security verification operation on the optimal candidate action; when the security verification passes, generate the corresponding optimal candidate action instruction and output the optimal candidate action instruction to the execution mechanism for execution.

[0013] The safety control method for embodied intelligent agents in this embodiment adopts a dual-channel decision-making mechanism, dynamically switching decision paths based on the current risk index: in low-risk scenarios, an experience cache library is used to quickly match actions, significantly improving response speed; in high-risk scenarios, rational decision planning and search are used to ensure the accuracy and safety of decisions, achieving an optimal balance between response speed and decision safety; simultaneously, comprehensive risk assessment is performed by collecting ontological state information and environmental state information, which can fully reflect the safety status of the embodied intelligent agent and avoid misjudgments caused by single-dimensional assessment; a unified safety verification step is added before decision output to perform a final safety check on all candidate actions, further reducing the probability of safety accidents.

[0014] A safety control device for an embodied intelligent agent, characterized in that it includes a sensor interface module, a main controller, a security kernel controller, a storage unit, and an actuator interface module;

[0015] The sensor interface module is used to acquire multi-dimensional state data collected in real time by various sensors integrated on the embodied intelligent agent body. The multi-dimensional state data includes two parts: body state information and environmental state information.

[0016] The main controller is used to control the execution of dynamic hazard assessment, cache channel judgment, experience cache library query and matching, rational decision planning search, and safety inspection.

[0017] The dynamic hazard assessment operation includes: calculating the current hazard index based on the ontological status information and the environmental status information;

[0018] The cache channel judgment operation includes: comparing the current danger index with a preset first safety threshold; if the current danger index is less than or equal to the first safety threshold, proceeding to the experience cache library query and matching operation; otherwise, proceeding to the rational decision planning search operation.

[0019] The experience cache query and matching operation includes: extracting the current scene feature vector based on the multidimensional state data, and retrieving historical records that match the current scene feature vector in the pre-built experience cache; if a historical record with a matching degree exceeding a preset similarity threshold is found, the action instruction stored in the historical record is read as the optimal candidate action, and the process proceeds to the security verification operation; if no historical record that meets the matching conditions is found, the process proceeds to the rational decision planning search operation.

[0020] The rational decision planning search operation includes: constructing a candidate action space, performing forward inference on each candidate action in the candidate action space to predict the future state after each action is executed, scoring each candidate action based on the predicted future state, and selecting the candidate action with the lowest comprehensive cost as the optimal candidate action.

[0021] The security verification operation includes: performing a security verification operation on the optimal candidate action; when the security verification passes, generating a corresponding optimal candidate action instruction, and outputting the optimal candidate action instruction to the actuator interface module;

[0022] The actuator interface module is used to control the actuator connected to the embodied intelligent agent to perform corresponding actions according to the optimal candidate action instruction;

[0023] The safety kernel controller is implemented using a hardware carrier independent of the main controller; the safety kernel controller is used to control the execution of emergency avoidance operations; the emergency avoidance operation includes: acquiring the body state information and environmental state information acquired by the sensing interface module, calculating the current danger index; continuously monitoring the current danger index, and when the current danger index exceeds a preset emergency threshold, the emergency threshold is greater than the first safety threshold, selecting an emergency action from a preset emergency action library that matches the danger type corresponding to the body state information and environmental state information, outputting the emergency action to the actuator interface module, and sending an interrupt signal to the actuator interface module;

[0024] The actuator interface module is also used to control the actuator to perform corresponding actions according to the received emergency action instructions, and to terminate the response to instructions from the main controller after receiving the interrupt signal;

[0025] The storage unit is used to store data configured for the operation of the embodied intelligent agent safety control device and data generated by the operation of the embodied intelligent agent safety control device.

[0026] This embodiment constructs a layered safety control architecture that completely decouples routine decision-making from emergency safety. It adopts a dual hardware independent design of the main controller and the safety kernel controller, fundamentally solving the problem of safety system collapse caused by the failure of a single controller. The main controller can dynamically switch between experience matching and rational planning dual-channel decision-making based on the real-time danger index, perfectly balancing task execution efficiency and operational safety. At the same time, it achieves full-process risk management through multi-dimensional state comprehensive evaluation and safety verification before action output. The safety kernel controller also has the ability to take over emergency control with absolute priority, which can forcibly terminate abnormal instructions from the main controller and execute emergency avoidance actions. The entire device adopts a modular design, with a clear architecture, high stability, and easy expansion and adaptation, comprehensively improving the safety control capabilities and operational reliability of the embodied intelligent agent. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the main flow of the embodied intelligent agent security control method provided in an embodiment of the present invention.

[0028] Figure 2 This is a flowchart illustrating the environmental status information collection operation provided in an embodiment of the present invention.

[0029] Figure 3 This is a flowchart illustrating the dynamic hazard assessment operation provided in an embodiment of the present invention.

[0030] Figure 4 This is a flowchart illustrating the current scene feature vector extraction operation provided in an embodiment of the present invention.

[0031] Figure 5 This is a flowchart illustrating the candidate action space construction operation provided in an embodiment of the present invention.

[0032] Figure 6 This is a flowchart illustrating the forward inference and cost scoring operations provided in an embodiment of the present invention.

[0033] Figure 7 This is a schematic diagram of the structure of the embodied intelligent agent safety control device provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0035] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, the term "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0036] The numbering of the steps in this application is for ease of description and understanding only, and does not imply that the steps must be performed in the strict order shown by the numbering. Any reasonable adjustment, merging, or splitting of the execution order of some steps, provided that the purpose of this invention is achieved and the same technical effect is obtained, should fall within the protection scope of this invention.

[0037] like Figure 1 As shown, in one embodiment, a security control method for an embodied intelligent agent includes the following steps:

[0038] S1, Collect multi-dimensional state data: Collect multi-dimensional state data in real time through various sensors integrated on the embodied intelligent agent body. The multi-dimensional state data includes two parts: body state information and environmental state information.

[0039] In one embodiment, the body status information includes: power data, temperature data, and structural health data; the environmental status information includes: ground environment data, water obstacle data, and moving obstacle data.

[0040] In one embodiment, the acquisition of the body status information in step S1 includes the following steps: (1) Acquiring power data: Real-time monitoring of battery voltage and current through the battery management system (BMS) to calculate the remaining power percentage E, where the value range of E is [0%, 100%]; (2) Acquiring temperature data: Acquiring the temperature values ​​of each component through temperature sensors mounted on the surface of the motor, the surface of the controller and the surface of the battery, and taking the maximum value of the temperature values ​​of all monitored components as the current body temperature; (3) Acquiring structural health data: Integrating one or more of the following detection signals to quantify and evaluate the structural health index: Current mutation detection signal of the drive motor, abnormal deviation signal between joint angle and joint torque, and abnormal frequency component signal obtained from vibration spectrum analysis.

[0041] like Figure 2 As shown, in one embodiment, the collection of environmental state information in step S1 includes the following steps:

[0042] S1-1, Collect ground environment data: Obtain depth maps through depth cameras or point cloud data through lidar, and perform ground segmentation based on depth maps or point cloud data to identify ground areas and non-ground areas; for non-ground areas, identify cliff-like dangerous areas through height difference detection, and mark the location coordinates and depth parameters of the cliffs.

[0043] S1-2, Collect water obstacle data: Use a semantic segmentation neural network to process the images captured by the visual sensor, identify puddles in the images and estimate the water depth by combining depth information; when the estimated water depth exceeds the preset safe wading depth threshold, mark the puddles as water obstacles.

[0044] S1-3, Collect moving obstacle data: Identify the category of dynamic objects from sensor data through target detection algorithm, and determine the spatial location of dynamic objects through point cloud clustering algorithm; establish motion tracking model for the identified dynamic objects, and predict their future trajectory based on Kalman filter algorithm; calculate the collision time between the embodied agent and each dynamic object; when the collision time of any dynamic object is lower than the preset safe time threshold, mark the dynamic object as a moving obstacle.

[0045] The embodied intelligent agent safety control method in this embodiment employs different perception and recognition algorithms for ground environment, water obstacles, and moving obstacles, thereby improving the recognition accuracy of various environmental hazard factors. For moving obstacles, it can not only identify their type and location, but also predict their future trajectory and calculate the collision time through motion tracking models, thus realizing early warning of dynamic hazards. It can accurately identify dangerous areas such as cliffs and deep water pools that are difficult to detect effectively by traditional perception technologies, greatly improving the environmental perception capability of embodied intelligent agents in complex unstructured environments.

[0046] S2, Dynamic Hazard Assessment: Calculate the current hazard index based on the entity's status information and the environmental status information.

[0047] like Figure 3 As shown, in one embodiment, S2 includes the following steps;

[0048] S2-1, normalize each indicator in the entity status information to obtain the corresponding entity deficit value: for power data, calculate the power deficit value based on the relative deviation between the current power and the preset full power value; for temperature data, map the current temperature value to the interval defined by the preset upper limit of normal temperature and the preset extreme temperature, and calculate the temperature deficit value; for structural health data, calculate the structural health deficit value based on the relative deviation between the current structural health and the preset full structural health value.

[0049] In one embodiment, the following three indicators are extracted from the ontology state information: (1) power percentage E; (2) maximum ontology temperature T_max; (3) structural health S_health; the above three indicators are normalized respectively, and their respective ontology deficit values ​​are calculated:

[0050] (1) The energy deficit value D_energy = 1 - E / E_max, where E_max is the maximum energy value; when E≥ E_max, D_energy takes the value of 0; the value range of D_energy is [0,1];

[0051] (2) Temperature deficit value D_temp = min(1, max(0, (T_max - T_normal) / (T_limit -T_normal))), where T_normal is the preset upper limit of normal operating temperature, indicating that the system is in normal working state when the temperature is below this value and the temperature does not contribute to the danger index; T_limit is the preset maximum allowable operating temperature of the hardware (limit temperature), indicating that the hardware faces the risk of damage when the temperature reaches this value; when T_max ≤ T_normal, D_temp takes the value of 0; when T_max ≥ T_limit, D_temp takes the value of 1; when T_normal < T_max < T_limit, D_temp increases linearly with the temperature in the interval (0,1); the final value range of D_temp is [0,1].

[0052] (3) Structural health deficit value D_structure = 1 - S_health / S_max, where S_max is the maximum value of structural health; the value range of D_structure is [0,1].

[0053] S2-2, normalize and quantify the various environmental hazard factors identified in the environmental status information to obtain the corresponding hazard contribution values: For ground environment data, when a cliff-like hazard area is detected, the hazard contribution value of the ground environment is calculated based on the distance between the current position of the embodied agent and the edge of the cliff and a preset critical distance threshold; For water obstacle data, when a water obstacle in a puddle area with a water depth exceeding the safe wading depth threshold is detected, the hazard contribution value of the water obstacle is calculated based on the ratio of the detected water depth to the preset maximum safe wading depth; For moving obstacle data, when a moving obstacle with a collision time lower than the safe time threshold is detected, the hazard contribution value of the moving obstacle is calculated based on the ratio of the collision time to the preset safe collision time threshold.

[0054] In one embodiment, for ground environment data: when the ground environment data contains cliff-like dangerous areas, the distance d_cliff from the current position of the embodied agent to the nearest cliff edge is extracted, and the danger contribution value of the ground environment is calculated according to the formula Threat_ground = min(1, d_critical / d_cliff), where d_critical is a preset critical distance threshold; when no cliff-like dangerous areas are detected in the ground environment, Threat_ground is 0; the value range of Threat_ground is [0,1].

[0055] For water obstacle data: When the water obstacle data contains puddles with a depth exceeding the safe wading depth threshold, the detected water depth of the puddle area, depth_water, is extracted. The hazard contribution value of the water obstacle class is calculated according to the formula Threat_water = depth_water / depth_max_safe, where depth_max_safe is the preset maximum safe wading depth threshold. Threat_water is 1 when depth_water ≥ depth_max_safe; Threat_water is 0 when no puddles are detected in the water obstacle data; the value range of Threat_water is [0,1].

[0056] For moving obstacle data: When the moving obstacle data contains dynamic objects whose collision time is lower than the safe time threshold, the collision time (TTC) of the dynamic object is extracted. The hazard contribution value of the moving obstacle class is calculated according to the formula Threat_moving = max(0, 1 - TTC / TTC_safe), where TTC_safe is the preset safe time threshold. When TTC ≥ TTC_safe, Threat_moving is 0. When no moving obstacle is detected in the moving obstacle data, Threat_moving is 0. The value range of Threat_moving is [0,1].

[0057] S2-3, weighted summation of each ontology deficit value and each danger contribution value to obtain the danger index at the current moment.

[0058] In one embodiment, the three deficit values ​​and each environmental hazard contribution value can be substituted into a weighted summation formula to calculate the current hazard index R_danger:

[0059] R_danger = w1 × D_energy + w2 × D_temp + w3 × D_structure + w4 ×(Threat_ground + Threat_water + Threat_moving).

[0060] Where: w1, w2, w3, w4 are preset weight coefficients, satisfying w1 > 0, w2 > 0, w3 > 0, w4 > 0 and w1 + w2 + w3 + w4 = 1; R_danger takes values ​​in the range [0,1], and the larger the value, the higher the current survival risk.

[0061] It should be noted that the weight coefficients w1, w2, w3, and w4 are not fixed and can be dynamically adjusted according to the task scenario, environment type, or the state of the embodied intelligent agent itself. For example, w4 (environmental danger weight) can be appropriately increased in a humid and watery environment; w1 (power depletion weight) can be appropriately increased when the battery is low; and w3 (structural health weight) can be appropriately increased when performing high-precision operations. Dynamic adjustment can be achieved through a preset scenario-weight mapping table or an online adaptive algorithm.

[0062] By breaking down the calculation process of the hazard index into three distinct steps—"extraction → normalization → weighted summation"—and providing specific normalization mapping functions for each type of environmental hazard factor, the calculation process of the hazard index becomes fully feasible. Different dimensions (raw sensor data in meters, seconds, degrees Celsius, and percentages) must be uniformly transformed to the dimensionless interval [0,1] through their respective mapping functions before meaningful weighted summation can be performed. This is a prerequisite for the hazard index, as a unified quantitative indicator of safety, to accurately reflect comprehensive risk.

[0063] The aforementioned embodied intelligent agent safety control method provides a quantitative method for calculating the hazard index. By normalizing and weighting the ontological deficiency value and the environmental hazard contribution value, it transforms the abstract safety status into a quantifiable value, making the hazard assessment more objective and accurate. It can adjust the weight of each indicator according to the needs of different application scenarios, improving the flexibility and adaptability of the hazard assessment mechanism. It can calculate the current hazard index in real time, providing a reliable basis for switching between the dual-channel decision-making mechanism.

[0064] S3, Cache channel judgment: Compare the current danger index with the preset first safety threshold; determine if the current danger index is less than or equal to the preset first safety threshold; if the current danger index is less than or equal to the first safety threshold, execute the experience cache library query and matching operation in step S4; otherwise, skip step S4 and directly execute the rational decision planning search operation in step S5.

[0065] S4, Experience Cache Query and Matching: Extract the current scene feature vector based on the multidimensional state data, and retrieve historical records that match the current scene feature vector in the pre-built experience cache. If a historical record with a matching degree exceeding the preset similarity threshold is retrieved, the action instruction stored in the historical record is read as the optimal candidate action, and the process proceeds to step S6. If no historical record that meets the matching conditions is retrieved, step S5 is executed.

[0066] The historical records stored in the experience cache include the mapping relationship between historical scene feature vectors and corresponding historical action instructions, and the historical records can come from the following sources: successful experiences generated through reinforcement learning pre-training in a simulation environment, or successful decision records accumulated by the embodied agent after passing security verification in historical operation, etc.

[0067] In one embodiment, the content update of the experience cache can occur at the following three times: Time 1 - New write: When step S4 determines that the cache is not hit and the optimal candidate action generated in step S5 passes the verification in step S6 and is executed, the scene characteristics, action instructions and execution results of this decision are written as new records to the cache; Time 2 - Verification and correction: When a historical record in the experience cache is matched and its action instructions are output, the effect of this actual execution is tracked; if the risk index fails to decrease as expected after actual execution, the record is set to invalid and will no longer be considered as a valid candidate in subsequent matching; Time 3 - Batch cleanup: Periodically scan all historical records in the experience cache and delete historical records that have never been hit within a consecutive preset time period (e.g., 24 hours) to free up storage space.

[0068] like Figure 4 As shown, in one embodiment, step S4, extracting the current scene feature vector based on multidimensional state data, includes the following steps:

[0069] S4-1, Select key feature items from the current multidimensional state data for cache matching.

[0070] In one embodiment, key features include one or more of the following: current battery percentage data, temperature data, current structural health data, ground environment data, water obstacle data, and moving obstacle data.

[0071] S4-2, each key feature term is normalized; for non-numerical categorical variables, one-hot encoding is used to convert them into binary vector fragments. After normalization, the data range of each key feature term is mapped to the interval [0,1].

[0072] S4-3 concatenates all normalized or encoded feature components in a fixed order into a complete feature vector as the current scene feature vector.

[0073] The aforementioned embodied intelligent agent security control method provides a standardized method for scene feature vector extraction. By selecting key feature terms, normalizing, and using one-hot encoding, multidimensional state data is transformed into feature vectors in a unified format, improving the accuracy and efficiency of experience cache matching. For non-numerical categorical variables, one-hot encoding is used to solve the problem that different types of features cannot be directly concatenated, ensuring the integrity and effectiveness of the feature vectors.

[0074] In one embodiment, step S4, retrieving historical records that match the current scene feature vector from a pre-built experience cache, includes:

[0075] S4-4: Traverse each historical cache record stored in the experience cache library; for each historical cache record, extract the historical scene feature vector stored therein;

[0076] S4-5, calculate the cosine similarity between the current scene feature vector and the historical feature vector. In one embodiment, the formula for calculating the cosine similarity is: sim_j = (s_current · s_cached_j) / (||s_current|| ×||s_cached_j||), where · represents the vector inner product operation, and ||·|| represents the L2 norm operation of the vector;

[0077] S4-6 compares the cosine similarity with a preset similarity threshold; a match is determined when the cosine similarity is greater than or equal to the preset similarity threshold; when multiple matching historical deposit records exist, the one with the highest cosine similarity is selected as the best matching record. A typical value for the preset similarity threshold is 0.85.

[0078] S5, Rational Decision Planning Search: Construct a candidate action space, perform forward inference on each candidate action in the candidate action space to predict the future state after each action is executed, score the comprehensive cost of each candidate action based on the predicted future state, and select the candidate action with the lowest comprehensive cost as the optimal candidate action.

[0079] like Figure 5 As shown, in one embodiment, the step of constructing the candidate action space includes:

[0080] S5-1, Load the pre-set atomic action primitive library; the atomic action primitive library stores the definitions of several atomic actions. Each atomic action is defined by an action type identifier and a corresponding control parameter mode. The control parameter mode defines the basic control parameter items required for this type of action and their value range. The basic control parameters include one or more of the following: target orientation angle, target angular velocity, target linear velocity, and duration.

[0081] In one embodiment, the system pre-configures an atomic action primitive library during initialization, storing several basic action primitives. Each atomic action comprises two parts: an action type identifier (used to distinguish different actions) and a control parameter mode for that action—that is, what parameters are required to execute the action and what the allowed value range of each parameter is. For example, the control parameter mode for the atomic action "forward" might specify three parameters: target direction angle (range 0 to 360 degrees), target linear velocity (range 0 to the maximum hardware speed), and duration (range a positive number). The specific types of atomic actions can include, but are not limited to, at least five of the following types: forward, backward, left turn, right turn, acceleration, deceleration, and stop / wait. The system selects which actions to include in the atomic action primitive library based on actual needs.

[0082] S5-2 collects the following real-time operational data of the embodied intelligent agent: the current position coordinates obtained through the positioning system, the current orientation angle, current linear velocity, and current angular velocity obtained through the inertial measurement unit or odometry; based on the collected real-time operational data, the agent's state information, and the environmental state information, parameterizes and instantiates each atomic action in the atomic action primitive library, and adds each parameterized atomic action as an atomic-level candidate action to the candidate action space.

[0083] After loading the primitive library, the system needs to parameterize and instantiate each atomic action in the library, that is, fill in specific parameter values ​​for each action template so that it becomes a candidate action that can actually be executed.

[0084] This process first requires acquiring two types of input data. The first type is the entity's state information (battery level, temperature, structural health), and the second type is the embodied intelligent agent's current pose and motion-related data, i.e., real-time operational data, specifically including: current position coordinates (obtained through positioning systems such as GPS, UWB positioning, or visual SLAM), current orientation angle (obtained through an inertial measurement unit (IMU) or odometry), current linear velocity, and current angular velocity (obtained through a wheel speed encoder or IMU).

[0085] Next, each atomic action in the primitive library is instantiated one by one. The core logic of instantiation is to reasonably determine the specific value of each control parameter based on the current running state. Taking the "forward" action as an example, the system first calculates the direction angle required to move from the current position to the target position as the target direction angle parameter; then, it selects an appropriate travel speed based on the current danger index—a higher cruising speed can be selected when the danger index is low, and the speed is automatically reduced to maintain caution when the danger index is high; finally, a reasonable duration parameter is obtained by dividing the estimated travel distance by the selected speed.

[0086] During instantiation, the system can also perform filtering operations. Certain atomic actions are not feasible under specific conditions, so corresponding candidates are not generated. For example, when the remaining battery power is below a certain threshold, "accelerate" candidate actions will not be generated because acceleration would exacerbate battery consumption; when a cliff-like danger zone is detected ahead, "continue forward" candidate actions will not be generated because this would lead to a direct fall risk. After filtering, each parameterized atomic action that remains is added to the candidate action space, forming an atomic-level candidate set.

[0087] S5-3, generate behavior-level candidate actions and add them to the candidate action space using any of the following methods:

[0088] (a) Select an advanced behavior template that matches the ontology state information and environment state information from the pre-set advanced behavior template library, and perform parameterized instantiation based on the real-time running data collected in S5-2 to obtain behavior-level candidate actions.

[0089] The system pre-configures a high-level behavior template library, which stores several commonly used composite behavior patterns. Each template defines a set of atomic action sequences arranged in a specific order, as well as the triggering conditions for when the pattern should be activated. When the current state of the system meets the triggering conditions of a template, the template is retrieved from the library, and each atomic step is parameterized and instantiated using real-time runtime data collected in S5-2, thereby obtaining a complete, executable behavior-level candidate action.

[0090] For example, the template library might contain a "go to the nearest charging station" behavior template, which specifies the following sequence of actions: turn towards the charging station → drive straight to the charging station → stop. When the system detects that the remaining battery level is below a preset threshold (e.g., 20%) and the current danger level is within a safe range, it matches and activates this template. After activation, the system queries the location coordinates of the charging station, calculates the turning angle and driving distance, and fills them into the various steps of the sequence, ultimately outputting a complete behavior-level candidate.

[0091] The template library can contain, but is not limited to, the following types of composite behaviors: heading to the nearest charging station, detouring and avoiding obstacles to return to a safe area, entering a low-power standby mode, and performing inspection tasks along a preset route. Each template specifies its own triggering conditions and action orchestration logic.

[0092] (b) Call the pre-trained policy network, take the ontology state information, environment state information and real-time running data collected in S5-2 as input, and output the control parameter sequence of the behavior-level candidate actions by the policy network.

[0093] If a pre-trained policy network (typically a neural network trained in a simulation environment using reinforcement learning algorithms) is deployed in the system, the current state data can be fed into the network as input, and the network will output a recommended behavioral action as a candidate. The network's input information can include one or more combinations of ontological state (battery level, temperature, structural health), environmental state (normalized values ​​of various hazard factors), and pose motion data (position, orientation, velocity, etc.). The network output is a sequence of control parameters describing the complete execution plan for the behavior. To ensure quality, the action output by the policy network is usually accompanied by a confidence score; only when the confidence score reaches a preset threshold will the output be adopted into the candidate space.

[0094] (c) When the preset triggering conditions are met, the corresponding preset emergency behavior is directly added to the candidate action space; the preset triggering conditions include at least one of the following: the remaining power is lower than the charging threshold, an environmental hazard factor is detected and the hazard index exceeds the preset behavior triggering threshold, or an external assistance request signal is received.

[0095] For clearly known emergency situations, the system can employ a conditional direct mapping mechanism, eliminating the need for complex matching or reasoning processes. When a pre-defined emergency condition is detected, the corresponding emergency action is directly added to the candidate space.

[0096] Typical emergency trigger conditions include: remaining battery power below a critical threshold (e.g., 5%, meaning it's about to run out), a danger index exceeding a high-risk threshold (e.g., above 0.8, facing an imminent survival threat), detection of a moving object about to collide (collision time to charge (TTC) less than 1 second), and receiving an external assistance request signal. Once any of these conditions are met, the corresponding emergency action (e.g., forced stop, emergency evasive maneuver, immediate initiation of the charging process) will be directly placed into the candidate space. These emergency candidates are usually marked as high priority so that they will be properly treated in subsequent scoring stages.

[0097] The aforementioned embodied intelligent agent safety control method constructs a multi-layered candidate action space, which includes both atomic-level and behavioral-level candidate actions. This ensures the precision of actions while improving the efficiency and flexibility of decision-making. It supports the generation of behavioral-level candidate actions through two methods: an advanced behavioral template library and a pre-trained policy network, thus adapting to the needs of scenarios with varying degrees of complexity.

[0098] like Figure 6 As shown, in one embodiment, step S5, which involves performing forward inference to predict the future state after each candidate action, scoring the comprehensive cost of each candidate action based on the predicted future state, and selecting the candidate action with the lowest comprehensive cost as the optimal candidate action, includes the following steps:

[0099] S5-4: For each candidate action in the candidate action space, using the current ontology state information, environment state information, and real-time running data collected in S5-2 as the initial state, perform forward inference within the preset inference time range, and predict the future state sequence after the execution of the candidate action step by step.

[0100] In one embodiment, the prediction of the future state includes at least one of the following: the estimated location coordinates at each time step, the estimated power consumption, the estimated change in structural health, and the estimated change in environmental hazard factors.

[0101] After constructing the candidate action space, each candidate action needs to be evaluated through forward extrapolation. Forward extrapolation means taking the current moment of the system as the starting point, assuming that a certain candidate action will be executed, and then simulating and extrapolating forward along the time axis to predict the evolution trajectory of the system state over a period of time after the execution of the action.

[0102] The initial state of the simulation consists of three parts: first, the ontological state information defined by weight 2 (current electrical charge, temperature, structural health); second, the environmental state information defined by weight 2 (hazard data such as ground, water, and moving obstacles); and third, the pose and motion parameters collected in S5-2 (current position, orientation, velocity, etc.). These three parts together constitute the starting point of the simulation.

[0103] The deduction process proceeds step-by-step in discrete time steps. The system pre-sets a deduction time range (denoted as H, in steps or converted to seconds), and then, starting from step 1, calculates the predicted state for the next step based on the state of the previous step and the control commands specified by the candidate action. This process is iterated until H steps are completed. Ultimately, each candidate action outputs a sequence of future states consisting of H time steps.

[0104] Each step of the predicted future state should cover at least the following aspects: First, the evolution of spatial location, including the estimated coordinates and orientation angles at each time step, which reflects where the embodied intelligence will move to and what direction it will face after performing the action; second, the changes in the entity's resources, mainly the gradual consumption of electricity and the changes in structural health, because the execution of any action requires energy consumption and may generate mechanical damage; third, the evolution of environmental hazards. As the embodied intelligence moves, its relative relationship with various hazards in the environment (cliffs, water, moving objects) will also change. Therefore, each step requires a reassessment of the numerical trends of various environmental hazard factors.

[0105] The predictive models upon which the deduction relies can take various forms. The most basic form is the analytical model, which derives subsequent states step by step through mathematical calculations based on physical kinematic equations (such as displacement formulas for differential-driven motion and integral formulas for uniform / variable motion) and known energy consumption characteristic curves. This approach is computationally fast, highly interpretable, and suitable for scenarios with clearly defined motion laws. Another form is the learning model, which uses a trained neural network to receive the current state and action as input and directly outputs a predicted sequence of subsequent states. This approach can capture complex nonlinear dynamic laws and may have higher accuracy, but it depends on the sufficiency of training data and the reliability of the model. There is also a compromise hybrid approach, using an analytical model as the main framework and a small learning model to correct the residual errors of the analytical calculations, balancing efficiency and accuracy.

[0106] Regardless of the model used, the goal of the deduction is the same: to produce a complete set of quantitative future state predictions for each candidate action, which can be used in the subsequent scoring process.

[0107] S5-5, for each candidate action, calculate the physical cost component and the social risk cost component based on the future state sequence corresponding to the candidate action, and sum the two by weight to obtain the comprehensive cost; the calculation basis of the physical cost component includes at least one or more of the following: the estimated total power consumption, the degree of deviation between the estimated position and the target position at the final moment, and the estimated total decline in structural health.

[0108] After obtaining the predicted sequence of future states for each candidate action, the next step is to evaluate the cost of each candidate action, quantitatively measuring the "goodness" of performing the action. The lower the cost, the better the action.

[0109] The overall cost can be composed of a weighted sum of two components: the physical cost component and the social risk cost component, the latter of which can be omitted in the absence of other intelligent agents.

[0110] The physical cost component measures the various "losses" incurred at the physical level after performing a candidate action, and can be considered from the following dimensions:

[0111] The first dimension is energy consumption. Based on the power consumption trajectory obtained from forward extrapolation, the total amount of power consumed throughout the entire extrapolation period is calculated. The more power consumed, the higher the cost. This is because electricity is a non-renewable and critical resource; excessive consumption will shorten battery life and increase the risk of running out of power later.

[0112] The second dimension is the task deviation dimension. It is evaluated based on the distance deviation between the embodied agent's estimated position and the predetermined target position at the end of the simulation. The greater the deviation, the further the action takes the system away from the task objective, and the higher the cost. This dimension ensures that the system does not completely deviate from the intended task while maintaining safety.

[0113] The third dimension is structural health. It is assessed based on the total decline in structural health during the simulation. If significant structural damage is predicted during the simulation (e.g., high-speed impact with obstacles, prolonged overload operation), the cost is high. This dimension protects the embodied intelligent agent's hardware from excessive damage.

[0114] In practical calculations, each dimension can be assigned a different weight coefficient to reflect its importance in decision-making. Typically, the energy consumption dimension has the highest weight (reflecting the survival priority principle), followed by the task deviation dimension and the structural health dimension.

[0115] The social risk cost component is included in the calculation of the overall cost only when at least one other agent is detected in the environment. The role of this component is to quantify the potential social impact of executing a candidate action. The specific evaluation dimensions include at least one of the following: whether the action will trigger conflicting interactions with other agents, whether it will lead to the loss of existing cooperation opportunities, and the degree of potential impact on the safety status or task execution of other agents in the vicinity.

[0116] The assessment of social risk costs can be achieved using two models. One is a graph neural network model: agents in the environment are modeled as nodes in a graph, and the interactions between agents are modeled as edges. The action to be assessed is embedded in the graph, and global interaction information is aggregated through the network's message passing mechanism, ultimately outputting a numerical value representing the degree of social risk. The other is an attention-based model: the system "pays attention" to other agents and their reaction tendencies most closely related to the current action, also outputting a risk score. Both models output continuous values ​​between 0 and 1, with higher values ​​indicating greater social risk.

[0117] The final overall cost equals the physical cost plus the social risk cost multiplied by its weighting coefficient. When there are no other agents in the environment or the social risk assessment module is not deployed, the social risk cost component takes a zero value, and the overall cost degenerates into a simple physical cost.

[0118] S5-6: Sort all candidate actions according to their comprehensive cost, and select the candidate action with the smallest comprehensive cost as the optimal candidate action; when there are multiple candidate actions with equal and minimum comprehensive costs, select the candidate action with the smallest physical cost component first.

[0119] The aforementioned embodied intelligent agent safety control method predicts the future state sequence after the execution of candidate actions through forward inference, which can assess the consequences of action execution in advance and avoid the safety risks caused by blind decision-making. It adopts a comprehensive cost scoring mechanism that considers both physical cost components and social risk cost components, making the decision results more scientific and comprehensive. When the comprehensive costs of multiple candidate actions are equal, the action with the lowest physical cost is selected first, which can effectively extend the service life of the embodied intelligent agent and reduce operating costs.

[0120] S6, Safety Verification: Perform a safety verification operation on the optimal candidate action; when the safety verification passes, generate the corresponding optimal candidate action instruction and output the optimal candidate action instruction to the actuator for execution.

[0121] In one embodiment, performing a safety verification operation on the optimal candidate action includes performing one or more of the following hard constraint checks on the optimal candidate action: (1) Joint angle limit verification: Analyze the target angle of each joint in the target posture of the candidate action and compare the mechanical limit range of each joint one by one; when the target angle of any joint exceeds the range of its allowed minimum angle to maximum angle, the verification is deemed to fail; (2) Collision boundary verification: Based on the kinematic parameters of the current posture and the candidate action, predict the three-dimensional space occupancy envelope of the embodied agent during the action execution process; perform intersection detection between the envelope and the known obstacle area in the environment map; when the geometric intersection of the envelope and the obstacle area is detected to be non-empty, the verification is deemed to fail; (3) Motor torque upper limit verification: Estimate the required output torque of each joint drive motor based on the acceleration requirements and load conditions of the candidate action; when the estimated required torque of any motor exceeds its rated maximum torque, the verification is deemed to fail; (4) Center of gravity stability margin verification: Predict the position change of the center of gravity of the embodied intelligent body and the shape of the supporting polygon after the candidate action is executed; calculate the minimum distance from the center of gravity projection to each side of the supporting polygon as the stability margin; when the stability margin is lower than the preset tipping danger threshold, the verification is deemed to fail; (5) Velocity / acceleration extreme value verification: Check whether the target linear velocity and angular velocity, linear acceleration and angular acceleration of the candidate action exceed the physical capability upper limit of the actuator respectively; if they exceed, the verification is deemed to fail.

[0122] When multiple safety checks fail simultaneously, a rollback correction is performed in the following preset priority order: center of gravity stability margin check > collision boundary check > joint angle limit check > motor torque upper limit check > velocity / acceleration extreme value check. Higher priority indicates a more severe safety threat to the embodied agent when that check fails, and should be addressed first.

[0123] In one embodiment, the above-described embodied intelligent agent security control method further includes the following step: triggering a rollback operation when the security verification fails.

[0124] Rollback Scenario A – Correctable: When the verification fails because the parameters of the candidate action exceed the allowable range but can be adjusted back to the allowable range, a parameter correction operation is performed. The specific parameters causing the failure are adjusted (e.g., reducing the target speed, reducing the amplitude of movement, or changing the direction of movement) to generate a corrected candidate action. The corrected candidate action is then resubmitted to step S6 for verification. The maximum number of iterations for each correction operation is N_retry, with a typical value of 3. Rollback Scenario B – Uncorrectable: When the cause is uncorrectable (e.g., an unavoidable collision between the action and an obstacle) or the correction iterations are exhausted and the verification still fails, the current candidate action is discarded, and the process returns to step S5 to search for a new optimal action. During the re-search, actions that have been judged as infeasible or have already been tried are excluded from the candidate action space, and the selection process is repeated. If the candidate action obtained from the re-search still fails to pass step S6 and the number of re-search iterations reaches the preset upper limit N_reschedule (typically 3), an emergency avoidance action is triggered or a manual intervention request is sent to the operator.

[0125] In one embodiment, the above-mentioned embodied intelligent agent safety control method further includes emergency avoidance operation: continuously monitoring the current danger index, and when the current danger index exceeds a preset emergency threshold, selecting an emergency action from a preset emergency action library that matches the danger type corresponding to the entity state information and environmental state information, generating a corresponding emergency action instruction, and outputting the emergency action instruction to the execution mechanism for execution.

[0126] When the current danger index exceeds the preset emergency threshold, the normal safety verification process is skipped, and an emergency avoidance operation is triggered directly. The emergency threshold is higher than the first safety threshold. The emergency threshold represents an extremely high level of survival risk, requiring immediate intervention without waiting for the normal process to complete. The emergency action library contains one or more of the following preset emergency actions: emergency stop (immediately cut off power output), full-speed reverse (move at maximum safe speed away from the danger source), on-site lock (freeze all current joint postures and cut off motor drive), issue a distress broadcast (send a distress signal containing its location and danger type to other nearby intelligent agents and a remote monitoring center), and enter a minimum power consumption sleep mode (shut down unnecessary sensors and computing units to conserve remaining power for rescue). The mapping relationship between each danger type and emergency action is predefined through a configuration table.

[0127] The aforementioned embodied intelligent agent safety control method adds an independent emergency avoidance operation procedure. When the danger index exceeds the emergency threshold, it can immediately trigger emergency avoidance actions, quickly escape from the dangerous environment, effectively deal with extreme dangerous situations, and maximize the safety of the embodied intelligent agent and surrounding personnel.

[0128] In one embodiment, the emergency avoidance operation is executed by the security kernel controller; the security kernel controller is implemented using a hardware carrier independent of the main controller running the main control program; when the emergency avoidance operation triggering condition is met, the security kernel controller sends an interrupt signal to the actuator; after receiving the interrupt signal, the actuator terminates its response to the instructions from the main controller and instead receives and executes the action instructions output by the security kernel controller.

[0129] A secure kernel controller may have the following characteristics:

[0130] Hardware Independence: The security kernel controller is implemented using an independent MCU chip, FPGA chip, or ASIC chip, and is physically separated from the main controller running the main control program. When the main controller experiences a software crash, infinite loop, or is maliciously hijacked, the security kernel controller remains unaffected and can still perform emergency avoidance operations normally. Specific implementation methods may include at least one of the following: using a low-power core in a heterogeneous multi-core processor architecture (such as the ARM Cortex-M series) as an independent security core, using an independent microcontroller with ECC error correction function (such as the Infineon AURIX series or TI Hercules series), or using a dedicated Safety Island module.

[0131] Interrupt priority advantage: The secure kernel controller has a higher interrupt priority than any user task running on the main controller; when the emergency avoidance operation trigger condition is met, the secure kernel controller can forcibly take over control of the actuator through a hardware interrupt signal. Furthermore, the secure kernel controller can continuously monitor the operating status of the main controller through a dedicated hardware watchdog timer, and proactively trigger a secure takeover when it detects that the main controller is unresponsive and times out.

[0132] Independent data path: The secure kernel controller can acquire raw sensor data through a dedicated data bus or DMA channel without relying on the main controller for forwarding; sensor access interfaces include, but are not limited to: depth cameras or lidar access via PCIe or GMSL interface, inertial measurement units (IMU) and odometers access via SPI or CAN bus, and torque and current sensors access via analog acquisition channels or I2C / SPI interface.

[0133] Hardware interlock output: The output of the safety kernel controller is connected to the hardware interlock circuit; the actuator interface is only in a driveable state when the safety kernel controller issues a valid enable signal; the actuator interface is physically disconnected by default; the hardware interlock circuit can be implemented using relays or solid-state switches; the actuator interface is connected to the joint motor controller via CAN bus or EtherCAT bus to receive control commands.

[0134] The aforementioned embodied intelligent agent safety control method employs a hardware carrier independent of the main controller to implement the safety kernel controller. This avoids the impact of main controller failures on emergency avoidance functions, significantly improving the reliability of emergency avoidance. The safety coprocessor forces the execution mechanism to terminate its response to main controller commands by sending an interrupt signal, ensuring that emergency action commands are executed with priority and preventing erroneous commands from the main controller from causing safety accidents in emergency situations.

[0135] like Figure 7 As shown, in one embodiment, an embodied intelligent agent safety control device 10 includes:

[0136] Sensor interface module 102, main controller 104, security kernel controller 106, actuator interface module 108, storage unit 110;

[0137] The sensor interface module 102 is used to acquire multi-dimensional state data collected in real time by various sensors integrated on the embodied intelligent agent body. The multi-dimensional state data includes two parts: body state information and environmental state information.

[0138] The main controller 104 is used to control the execution of dynamic hazard assessment operations, cache channel judgment operations, experience cache library query and matching operations, rational decision planning search operations, and safety inspection operations;

[0139] Dynamic hazard assessment operations include: calculating the current hazard index based on the entity's status information and environmental status information;

[0140] The cache channel judgment operation includes: comparing the current danger index with a preset first safety threshold; if the current danger index is less than or equal to the first safety threshold, proceeding to the experience cache library query and matching operation; otherwise, proceeding to the rational decision planning search operation.

[0141] The experience cache library query and matching operation includes: extracting the current scene feature vector based on multi-dimensional state data, retrieving historical records that match the current scene feature vector in the pre-built experience cache library; if a historical record with a matching degree exceeding a preset similarity threshold is found, the action instruction stored in the historical record is read as the optimal candidate action, and the process proceeds to the security verification operation; if no historical record that meets the matching conditions is found, the process proceeds to the rational decision planning search operation.

[0142] The rational decision-making planning search operation includes: constructing a candidate action space, performing forward inference on each candidate action in the candidate action space to predict the future state after each action is executed, scoring each candidate action based on the predicted future state, and selecting the candidate action with the lowest comprehensive cost as the optimal candidate action.

[0143] The safety verification operation includes: performing a safety verification operation on the optimal candidate action; when the safety verification passes, generating the corresponding optimal candidate action instruction, and outputting the optimal candidate action instruction to the actuator interface module;

[0144] The actuator interface module 108 is used to control the actuator connected to the embodied intelligent agent to perform corresponding actions according to the optimal candidate action instruction;

[0145] The safety kernel controller 106 is implemented using a hardware carrier independent of the main controller. The safety kernel controller 106 is used to control the execution of emergency avoidance operations. The emergency avoidance operations include: acquiring the body state information and environmental state information obtained by the sensor interface module, calculating the current danger index; continuously monitoring the current danger index, and when the current danger index exceeds the preset emergency threshold, selecting an emergency action from the preset emergency action library that matches the danger type corresponding to the body state information and environmental state information, outputting the emergency action to the actuator interface module, and sending an interrupt signal to the actuator interface module.

[0146] The actuator interface module 108 is also used to control the actuator to perform corresponding actions according to the received emergency action instructions, and to terminate the response to instructions from the main controller after receiving an interrupt signal;

[0147] The storage unit 110 is used to store the data configured for the operation of the embodied intelligent agent safety control device and the data generated by the operation of the embodied intelligent agent safety control device.

[0148] This embodiment constructs a layered safety control architecture that completely decouples routine decision-making from emergency safety. It adopts a dual hardware independent design of the main controller and the safety kernel controller, fundamentally solving the problem of safety system collapse caused by the failure of a single controller. The main controller can dynamically switch between experience matching and rational planning dual-channel decision-making based on the real-time danger index, perfectly balancing task execution efficiency and operational safety. At the same time, it achieves full-process risk management through multi-dimensional state comprehensive evaluation and safety verification before action output. The safety kernel controller also has the ability to take over emergency control with absolute priority, which can forcibly terminate abnormal instructions from the main controller and execute emergency avoidance actions. The entire device adopts a modular design, with a clear architecture, high stability, and easy expansion and adaptation, comprehensively improving the safety control capabilities and operational reliability of the embodied intelligent agent.

[0149] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0150] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A security control method for an embodied intelligent agent, characterized in that, Includes the following steps: S1, Collect multi-dimensional state data: Collect multi-dimensional state data in real time through various sensors integrated on the embodied intelligent agent body. The multi-dimensional state data includes two parts: body state information and environmental state information. S2, Dynamic Hazard Assessment: Calculate the current hazard index based on the entity's status information and the environmental status information; S3, Cache channel judgment: Compare the current danger index with the preset first safety threshold; if the current danger index is less than or equal to the first safety threshold, execute the experience cache library query and matching operation in step S4; otherwise, skip step S4 and directly execute the rational decision planning search operation in step S5. S4, Experience Cache Library Query and Matching: Extract the current scene feature vector based on the multidimensional state data, and retrieve historical records that match the current scene feature vector in the pre-built experience cache library; if a historical record with a matching degree exceeding the preset similarity threshold is retrieved, the action instruction stored in the historical record is read as the optimal candidate action, and proceed to step S6; if no historical record that meets the matching conditions is retrieved, proceed to step S5. S5, Rational Decision Planning Search: Construct a candidate action space, perform forward inference on each candidate action in the candidate action space to predict the future state after each action is executed, score the comprehensive cost of each candidate action based on the predicted future state, and select the candidate action with the lowest comprehensive cost as the optimal candidate action. S6, Security Verification: Perform a security verification operation on the optimal candidate action; When the security check passes, the corresponding optimal candidate action instruction is generated and output to the actuator for execution.

2. The embodied intelligent agent security control method according to claim 1, characterized in that, The entity's status information includes: power data, temperature data, and structural health data; Environmental status information includes: ground environment data, water obstacle data, and moving obstacle data.

3. The embodied intelligent agent security control method according to claim 2, characterized in that, The collection of environmental status information in step S1 includes the following steps: S1-1, Collect ground environment data: Obtain depth map through a depth camera or point cloud data through LiDAR, and perform ground segmentation based on the depth map or point cloud data to identify ground areas and non-ground areas; for non-ground areas, identify cliff-like dangerous areas by height difference detection, and mark the location coordinates and depth parameters of the cliffs. S1-2, Collect water obstacle data: Use a semantic segmentation neural network to process the images captured by the visual sensor, identify puddles in the images and estimate the water depth by combining depth information; when the estimated water depth exceeds the preset safe wading depth threshold, mark the puddles as water obstacles. S1-3, Collect moving obstacle data: Identify the category of dynamic objects from sensor data through target detection algorithm, and determine the spatial location of dynamic objects through point cloud clustering algorithm; establish motion tracking model for the identified dynamic objects, and predict their future trajectory based on Kalman filter algorithm; calculate the collision time between the embodied agent and each dynamic object; when the TTC of any dynamic object is lower than the preset safe time threshold, mark the dynamic object as a moving obstacle.

4. The embodied intelligent agent safety control method according to claim 2, characterized in that, The step of calculating the current danger index based on the ontological state information and the environmental state information includes: S2-1, Normalize each indicator in the ontology state information to obtain the corresponding ontology deficiency value: Based on the power data, the power deficit value is calculated according to the relative deviation between the current power level and the preset full power value; For the temperature data, the current temperature value is mapped to the range defined by the preset upper limit of normal temperature and the preset extreme temperature, and the temperature deficit value is calculated. For the structural health data, the structural health deficit value is calculated based on the relative deviation between the current structural health value and the preset full structural health value; S2-2, Normalize and quantify the various environmental hazard factors identified in the environmental status information to obtain the corresponding hazard contribution values: Regarding the ground environment data, when a cliff-like dangerous area is detected, the hazard contribution value of the ground environment is calculated based on the distance between the current position of the embodied agent and the edge of the cliff and a preset critical distance threshold. Regarding the water obstacle data, when a water obstacle is detected, the danger contribution value of the water obstacle is calculated based on the ratio of the detected water depth to the preset maximum safe wading depth; Regarding the moving obstacle data, when a moving obstacle is detected, the danger contribution value of the moving obstacle is calculated based on the ratio of the collision time to a preset safe time threshold. S2-3, the risk index at the current moment is obtained by weighted summation of each of the said ontology deficiency values ​​and each of the said risk contribution values.

5. The embodied intelligent agent safety control method according to claim 2, characterized in that, Step S4, which involves extracting the current scene feature vector based on the multidimensional state data, includes the following steps: S4-1, Select key feature items for cache matching from the current multidimensional state data; S4-2, normalize each key feature item separately; for non-numerical categorical variables, convert them into binary vector fragments using one-hot encoding; S4-3 concatenates all normalized or encoded feature components in a fixed order into a complete feature vector as the current scene feature vector.

6. The embodied intelligent agent security control method according to claim 2, characterized in that, The steps for constructing the candidate action space include: S5-1, Load the pre-set atomic action primitive library; the atomic action primitive library stores the definitions of several atomic actions, each atomic action is defined by an action type identifier and a corresponding control parameter mode, the control parameter mode defines the basic control parameter items required for this type of action and their value range, the basic control parameters include one or more of the following: target direction angle, target angular velocity, target linear velocity and duration; S5-2, Collect the following real-time operational data of the embodied intelligent agent: current position coordinates obtained through the positioning system, current orientation angle, current linear velocity, and current angular velocity obtained through the inertial measurement unit or odometry; based on the collected real-time operational data, the agent's state information, and the environmental state information, parameterize and instantiate each atomic action in the atomic action primitive library, and add each parameterized atomic action as an atomic-level candidate action to the candidate action space; S5-3, generate behavior-level candidate actions and add them to the candidate action space using any of the following methods: (a) Select an advanced behavior template that matches the ontology state information and the environment state information from the preset advanced behavior template library, and perform parameterized instantiation based on the real-time running data collected in S5-2 to obtain behavior-level candidate actions. (b) Invoke the pre-trained policy network, take the ontology state information, the environment state information and the real-time running data collected in S5-2 as input, and output the control parameter sequence of the behavior-level candidate actions by the policy network.

7. The embodied intelligent agent security control method according to claim 6, characterized in that, The process of performing forward inference to predict the future state after each candidate action, scoring the comprehensive cost of each candidate action based on the predicted future state, and selecting the candidate action with the lowest comprehensive cost as the optimal candidate action includes the following steps: S5-4, for each candidate action in the candidate action space, using the current body state information, the environmental state information, and the real-time operation data collected in S5-2 as the initial state, perform forward deduction within a preset deduction time range, predicting the future state sequence after the execution of the candidate action step by step; the predicted content of the future state includes at least one of the following: the estimated position coordinates at each time step, the estimated power consumption, the estimated change in structural health, and the estimated change in environmental hazard factors; S5-5, for each candidate action, calculate the physical cost component and the social risk cost component based on the future state sequence corresponding to the candidate action, and sum the two by weight to obtain the comprehensive cost; the calculation basis of the physical cost component includes at least one or more of the following: the estimated total power consumption, the degree of deviation between the estimated position and the target position at the final moment, and the estimated total decrease in structural health. S5-6, sort all candidate actions according to the comprehensive cost, and select the candidate action with the smallest comprehensive cost value as the optimal candidate action; when there are multiple candidate actions with equal comprehensive costs and all of them are the smallest, select the candidate action with the smallest physical cost component first.

8. The embodied intelligent agent security control method according to claim 1, characterized in that, The step of performing a security verification operation on the optimal candidate action includes: Perform one or more of the following hard constraint checks on the optimal candidate action: (1) Joint angle limit check: Analyze the target angle of each joint in the target posture of the candidate action and compare the mechanical limit range of each joint one by one; when the target angle of any joint exceeds the range of its minimum to maximum angle, the check is deemed to fail; (2) Collision boundary check: Based on the kinematic parameters of the current posture and the candidate action, predict the three-dimensional space occupancy envelope of the embodied agent during the action execution process; perform intersection detection between the envelope and the known obstacle area in the environment map; when the geometric intersection of the envelope and the obstacle area is detected to be non-empty, the check is deemed to fail; (3) Motor torque upper limit check: Estimate the torque required to output the motor of each joint according to the acceleration requirement and load conditions of the candidate action; when the estimated torque required by any motor exceeds its rated maximum torque, the check is deemed to fail. (4) Center of gravity stability margin verification: predict the position change of the center of gravity of the embodied intelligent agent and the shape of the supporting polygon after the candidate action is executed; calculate the minimum distance from the center of gravity to each side of the supporting polygon as the stability margin; when the stability margin is lower than the preset tipping danger threshold, the verification is deemed to fail; (5) Velocity / acceleration extreme value verification: check whether the target linear velocity and angular velocity, linear acceleration and angular acceleration of the candidate action exceed the physical capability limit of the actuator respectively; if they exceed, the verification is deemed to fail.

9. The embodied intelligent agent security control method according to claim 1, characterized in that, It also includes emergency evacuation procedures: The system acquires the entity state information and environmental state information, calculates the current danger index, continuously monitors the current danger index, and when the current danger index exceeds a preset emergency threshold (which is greater than the first safety threshold), selects an emergency action from a preset emergency action library that matches the danger type corresponding to the entity state information and environmental state information, generates a corresponding emergency action instruction, and outputs the emergency action instruction to the execution mechanism for execution.

10. The embodied intelligent agent security control method according to claim 9, characterized in that, The emergency avoidance operation is executed by the secure kernel controller; The safety kernel controller is implemented using a hardware carrier independent of the main controller running the main control program. When the emergency avoidance operation trigger condition is met, the safety kernel controller sends an interrupt signal to the actuator. After receiving the interrupt signal, the actuator terminates its response to the instructions from the main controller and instead receives and executes the action instructions output by the safety kernel controller.

11. A safety control device for an embodied intelligent agent, characterized in that, It includes a sensor interface module, a main controller, a security kernel controller, a storage unit, and an actuator interface module; The sensor interface module is used to acquire multi-dimensional state data collected in real time by various sensors integrated on the embodied intelligent agent body. The multi-dimensional state data includes two parts: body state information and environmental state information. The main controller is used to control the execution of dynamic hazard assessment, cache channel judgment, experience cache library query and matching, rational decision planning search, and safety inspection. The dynamic hazard assessment operation includes: calculating the current hazard index based on the ontological status information and the environmental status information; The cache channel judgment operation includes: comparing the current danger index with a preset first safety threshold; if the current danger index is less than or equal to the first safety threshold, proceeding to the experience cache library query and matching operation; otherwise, proceeding to the rational decision planning search operation. The experience cache query and matching operation includes: extracting the current scene feature vector based on the multidimensional state data, and retrieving historical records that match the current scene feature vector in the pre-built experience cache; if a historical record with a matching degree exceeding a preset similarity threshold is found, the action instruction stored in the historical record is read as the optimal candidate action, and the process proceeds to the security verification operation; if no historical record that meets the matching conditions is found, the process proceeds to the rational decision planning search operation. The rational decision planning search operation includes: constructing a candidate action space, performing forward inference on each candidate action in the candidate action space to predict the future state after each action is executed, scoring each candidate action based on the predicted future state, and selecting the candidate action with the lowest comprehensive cost as the optimal candidate action. The security verification operation includes: performing a security verification operation on the optimal candidate action; when the security verification passes, generating a corresponding optimal candidate action instruction, and outputting the optimal candidate action instruction to the actuator interface module; The actuator interface module is used to control the actuator connected to the embodied intelligent agent to perform corresponding actions according to the optimal candidate action instruction; The safety kernel controller is implemented using a hardware carrier independent of the main controller; the safety kernel controller is used to control the execution of emergency avoidance operations; the emergency avoidance operation includes: acquiring the body state information and environmental state information acquired by the sensing interface module, calculating the current danger index; continuously monitoring the current danger index, and when the current danger index exceeds a preset emergency threshold, the emergency threshold is greater than the first safety threshold, selecting an emergency action from a preset emergency action library that matches the danger type corresponding to the body state information and environmental state information, outputting the emergency action to the actuator interface module, and sending an interrupt signal to the actuator interface module; The actuator interface module is also used to control the actuator to perform corresponding actions according to the received emergency action command, and to terminate the response to the command from the main controller after receiving the interrupt signal; The storage unit is used to store data configured for the operation of the embodied intelligent agent safety control device and data generated by the operation of the embodied intelligent agent safety control device.

Images