A multi-agent intelligent body safe interaction method based on semantic physical constraint conversion

Abstract

The application discloses a kind of multi-body intelligent agent safe interaction methods based on semantic physical constraint conversion, comprising the following steps: S1, multi-source information acquisition and preprocessing;S2, the security coding and cross-modal consistency check of multi-source input information after preprocessing;S3, construct semantic-physical constraint conversion and hybrid security constraint;S4, task safety evaluation and collaborative decision-making;S5, distributed security planning;S6, communication adaptive security degradation execution;S7, security information record;S8, security action output.The application matches the security constraint of natural language form with real-time vision, spatial and motion state perception result consistently, and converts the matched semantic security constraint into physical hard security constraint executable by control barrier function, so as to realize safety collaborative control in different states such as normal communication, limited communication and communication interruption of multi-body intelligent agent.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robot control technology, specifically to a method for secure interaction among multiple embodied intelligent agents based on semantic-physical constraint transformation. Background Technology

[0002] Multi-agent intelligent systems refer to systems in which multiple autonomous intelligent agents with perception, cognition, planning, and action capabilities collaborate to complete tasks in the physical world. These systems are widely used in safety-critical scenarios such as autonomous vehicle fleets, drone swarms, search and rescue robot formations, and multi-robot collaboration in warehousing and logistics.

[0003] In multi-agent collaboration, safety is the primary challenge. Unlike purely digital AI systems, embodied agents operate in dynamic and uncertain physical environments. Perceptual errors, reasoning mistakes, or planning deviations can not only lead to task failures but also directly cause physical collisions, equipment damage, and even personal injury. Therefore, ensuring that each agent meets safety constraints at every time step in a dynamic environment with coupled multi-agent interactions is a core technical problem to be solved in this field.

[0004] Security threats to embodied agents span five layers: Perception layer: Adversarial attacks, sensor spoofing, and interference lead to the agent's misinterpretation of the environment; Cognition layer: Jailbreak attacks and hallucination risks lead to unsafe reasoning; Planning layer: Adversarial disturbances and constraint bypassing lead to the planning of dangerous trajectories; Action interaction layer: Bypassing security mechanisms leads to unsafe physical interactions; Agent system layer: Tool misuse and memory poisoning lead to persistent unreliable behavior. More seriously, security failures in inner layers (such as the perception layer) can amplify and propagate to outer interaction layers, creating a cascading failure risk.

[0005] The closest prior art known to the inventor includes the following categories: Safe multi-agent reinforcement learning methods based on constrained Markov games. Representative methods include MAPPO-Lagrangian and Scale-MAPPO-L. These methods formalize the multi-agent safety problem as a constrained Markov game, optimizing the team's cumulative reward while constraining the cumulative cost to not exceed a preset threshold. Scale-MAPPO-L further utilizes the spatial correlation decay property to propose local policy optimization, eliminating the dependence on the global state and giving the method a certain degree of scalability.

[0006] Hierarchical safety control methods based on control barrier functions. Representative methods include HMARL-CBF and Safe-RMM. These methods construct safety constraints as control barrier functions, ensuring in real-time that the agent does not enter unsafe state regions in the lower-level control. HMARL-CBF adopts a two-layer architecture of high-level cooperative policy learning and low-level CBF safe execution. Safe-RMM further introduces robust CBF and ellipsoidal uncertainty regions to handle state estimation errors.

[0007] Secure reasoning methods based on large language models. Representative methods include SMALL and MADRA. SMALL introduces natural language security constraint encoding, transforming human natural language constraints into semantic embeddings by fine-tuning the language model to guide policy learning. MADRA employs a multi-LLM debate mechanism for risk assessment, determining the security of task instructions through majority voting and a key evaluator.

[0008] Zero-constraint optimization methods. A representative method is Def-MARL. This method transforms the multi-agent safe optimal control problem into an Epigraph form, achieving distributed solution under a centralized training and distributed execution paradigm, theoretically guaranteeing zero constraint violations.

[0009] Collaborative methods based on shared communication intent. A representative method is IntNet. This method achieves proactive and secure collaboration by sharing future intent predictions among agents through a graph attention network.

[0010] However, the closest existing technology mentioned above still has the following problems: Existing technologies generally rely on state variables, environmental description text, or numerical optimization information for decision-making, lacking the ability to jointly process natural language safety constraints and real-time environmental perception information. For example, in a warehouse robot collaboration scenario, after an administrator issues a safety instruction such as "Please avoid slippery areas," the system not only needs to understand the safety requirements expressed in the instruction but also needs to establish a correspondence between the semantic object "slippery surface" and the specific area currently perceived visually, and further generate executable safety control constraints. However, existing methods typically only support state-space constraint optimization or text-level risk assessment, failing to complete the continuous processing from natural language safety constraints and environmental perception object recognition to the generation of physical safety constraints. Therefore, how to achieve the association between natural language safety constraints and real-time environmental perception information in multi-agent collaboration scenarios, and transform the association result into executable safety control constraints, is a technical problem that urgently needs to be solved by existing technologies. Summary of the Invention

[0011] To address the problems existing in the prior art, this invention proposes a secure interaction method for multiple avatars based on semantic-physical constraint transformation. This invention performs consistency matching between security constraints in natural language form and real-time visual, spatial, and motion state perception results, and then converts the matched semantic security constraints into physical hard security constraints executable by the control barrier function. This enables secure collaborative control under different states such as normal communication, restricted communication, and communication interruption among the multiple avatars. The technical solution of this invention is as follows.

[0012] To achieve the above objectives, this invention provides a secure interaction method for multiple embodied intelligent agents based on semantic-physical constraint transformation, comprising the following steps: S1. Multi-source information acquisition and preprocessing; S2. Perform secure coding and cross-modal consistency verification on the preprocessed multi-source input information; S3. Construct semantic-physical constraint transformation and hybrid security constraints; S4. Mission safety assessment and collaborative decision-making; S5, Distributed Security Planning; S6, Communication Adaptive Security Degradation Execution; S7, Security Information Recording; S8, Safety Action Output.

[0013] Preferably, step S1 specifically includes the following steps: S11. Obtain input information from multiple embodied intelligent agents, including: visual information, denoted as first perception information I; spatial information, denoted as second perception information P; natural language security constraint information, denoted as third perception information L; and motion state information, denoted as fourth perception information M. S12. The first sensing information I, the second sensing information P, the third sensing information L, and the fourth sensing information M are preprocessed respectively to obtain the preprocessed first sensing information I', the preprocessed second sensing information P', the preprocessed third sensing information L', and the preprocessed fourth sensing information M'.

[0014] Preferably, step S2 specifically includes the following steps: S21. The visual encoder outputs visual semantic features and visual credibility, the spatial encoder outputs spatial geometric features and spatial credibility, the language encoder outputs language semantic features and language credibility, and the motion encoder outputs motion state features and motion credibility. S22. The system calculates a first consistency index between the language semantic features and the visual semantic features, and calculates a second consistency index between the spatial geometric features and the motion state features; S23. When any consistency index falls below the corresponding threshold, the credibility of the corresponding modality is reduced. S24. Based on each confidence parameter, the modal features are weighted and fused to obtain a unified security representation.

[0015] Preferably, step S3 specifically includes the following steps: S31. Extract security constraint parameters based on the linguistic semantic information in the unified security representation vector F. The security constraint parameters include: constraint type TYPE, constraint parameter PARAM, and constraint priority LEVEL. The security constraint type TYPE includes at least: distance preservation constraint, area avoidance constraint, velocity limit constraint, and attitude limit constraint. S32. Verify the safety constraint parameters using visual information, match the target object described by the constraint parameter PARAM with the visual recognition result, generate the corresponding physical constraint when the match is successful, and output semantic inconsistency alarm information when the match fails. S33. Subsequently, construct the hybrid security constraint function B(s), where s represents the current system state and B(s) represents the hybrid security constraint function.

[0016] Preferably, the hybrid safety constraint function B(s) is composed of the following three sub-constraint functions: a first constraint function B1(s) for describing geometric safety constraints; a second constraint function B2(s) for describing semantic safety constraints; and a third constraint function B3(s) for describing the learned complex safety constraints. The hybrid safety constraint function is defined as: B(s) = min(B1(s), B2(s), B3(s)). When B(s) is greater than or equal to zero, it indicates that the current state meets the safety requirements; when B(s) is less than zero, it indicates that the current state violates the safety requirements. The hybrid safety constraint function B(s) serves as the input to steps S4 and S5.

[0017] Preferably, step S4 specifically includes the following steps: S41. Perform security assessment on task instructions based on the hybrid security constraint function B(s); S42. When multiple agents have conflicting objectives, generate multiple candidate cooperative schemes, and conduct a comprehensive evaluation based on safety indicators and task efficiency indicators to determine the target cooperative scheme and output the cooperative decision result.

[0018] Preferably, step S5 specifically involves: generating planning action A based on the collaborative decision-making result output in step S4 and the hybrid safety constraint function B(s). During the planning process, only local neighborhood information is used to complete the safe path planning, and the hybrid safety constraint function B(s) is continuously guaranteed to meet the safety conditions. After generating planning action A, planning action A is sent to step S6.

[0019] Preferably, step S6 specifically includes the following steps: S61. Real-time acquisition of communication quality parameter Q, wherein the communication quality parameter Q is determined by communication delay, packet loss rate and signal strength; S62. Perform mode switching according to the communication quality parameter Q: when the communication quality parameter Q is greater than the first threshold, enter the cooperative security mode; when the communication quality parameter Q is between the first threshold and the second threshold, enter the conservative cooperative mode; when the communication quality parameter Q is less than the second threshold, enter the independent security mode. S63. In independent safety mode: stop using information sent by other agents, perform safety control only based on the perception results of this agent and the hybrid safety constraint function B(s), while increasing the safety distance and reducing the movement speed. When communication is restored, perform state synchronization and resume cooperative planning.

[0020] Preferably, step S7 specifically involves: recording security event information during the current decision-making process, including: security constraint triggering information, communication status change information, manual intervention information, planning and decision-making information, and forming security record information.

[0021] Preferably, step S8 specifically involves: generating a target action A_safe based on the safety control result obtained in step S6, where A_safe represents the final safety action; if the planned action A satisfies the requirements of the hybrid safety constraint function B(s), then A_safe = A; otherwise, the planned action A is modified to obtain the target action A_safe that satisfies the requirements of the hybrid safety constraint function B(s); outputting the target action A_safe to the actuator to complete the control, acquiring new environmental information after execution, and returning to step S1 to form a closed-loop control process.

[0022] Compared with the prior art, the beneficial effects of the present invention are: This invention transforms natural language security constraints into physical hard security constraints that can be executed by control barrier functions. Through this approach, security rules proposed by humans in natural language can directly participate in the underlying continuous control solution process, providing a more direct constraint on actual actions compared to schemes that only perform risk scoring or candidate action selection.

[0023] This invention constructs a hybrid safety boundary using analytical geometric control barrier functions, semantic control barrier functions, and neural control barrier functions. This hybrid approach retains the interpretability and verifiability of analytical constraints while also being able to express natural language constraints and complex nonlinear hazardous boundaries, thereby improving safety control capabilities in complex environments.

[0024] This invention addresses the problem of excessive reliance on communication networks in multi-agent collaboration by proactively expanding the safe distance, reducing the speed, and switching to local independent security control when communication quality deteriorates or communication is interrupted through an adaptive security degradation mechanism for communication quality.

[0025] This invention uses multimodal consistency verification as a prerequisite reliability judgment for semantic-physical conversion, which can reduce the impact of abnormal modalities on security control when vision is interfered with, language constraints are ambiguous, or spatial perception is abnormal. Attached Figure Description

[0026] Figure 1 Overall flowchart of the embodied intelligent agent secure interaction method provided by the present invention; Figure 2 The unified multimodal security coding structure diagram provided by this invention; Figure 3 This invention provides a bridging unit from semantic security constraints to physical hard security constraints. Figure 4 This invention provides a multi-agent cooperative planning structure diagram; Figure 5 The flowchart for adaptive security degradation in communication provided by this invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This embodiment illustrates a heterogeneous multi-robot collaborative material handling scenario in a warehouse. The scenario includes two automated guided vehicles (AGVs), one inspection drone, and one warehouse manager. The two AVRs are responsible for material handling, the drone is responsible for environmental inspection and global information supplementation, and the warehouse manager issues safety constraints to the system via natural language.

[0029] Step S1: As Figure 1As shown, after the system starts running, each intelligent agent first collects its own corresponding multimodal perception data. Specifically: the first perception data is visual perception data; the second perception data is spatial perception data; the third perception data is linguistic constraint data; and the fourth perception data is motion state data. In this embodiment, the automated guided vehicle (AGV) robot uses a visual sensor to collect image information of the shelf area, uses a lidar to obtain spatial distribution information of surrounding obstacles, and uses an inertial measurement unit to obtain its own motion state information. The warehouse manager inputs the following safety constraint through a terminal: "The ground in shelf area 3 is slippery. All robots should slow down when passing through this area and maintain a safe distance from the shelves." The above natural language constraint is used as the third perception data input to the system. To prevent abnormal data from directly entering the decision-making process, this invention first performs safety preprocessing. Visual perception data undergoes credibility assessment through a visual safety detection module; spatial perception data undergoes noise reduction processing through an anomaly filtering module; linguistic constraint data undergoes legality verification through a safety alignment detection module; and motion state data undergoes authenticity verification through a multi-source consistency detection module. After preprocessing, reliable first feature data, second feature data, third feature data, and fourth feature data are obtained and used as input for step S2. During the research and development process, the inventors attempted to directly input the raw perceptual data into the planning module. However, experiments revealed that when the visual input was partially obstructed or interfered with by reflections, the planning results fluctuated significantly. Therefore, a preprocessing step was added to improve the stability of subsequent decisions.

[0030] Step S2: Its structure is as follows Figure 2 As shown. Existing technologies typically process visual, linguistic, and spatial information separately, lacking a correlation verification mechanism between the various modalities. This invention first performs unified encoding on the four types of perceptual data. Specifically: the visual feature encoding module extracts information on environmental targets, ground conditions, and obstacles; the linguistic feature encoding module extracts semantic information on safety constraints; the spatial feature encoding module extracts three-dimensional environmental structural information; and the motion feature encoding module extracts robot motion state information.

[0031] After encoding, the following features are obtained: visual features; linguistic features; spatial features; and motion features. Then, the system proceeds to the cross-modal consistency verification stage. For example, after an administrator issues the linguistic constraint "Slow down and proceed through slippery areas," the system will automatically check: whether visual perception indicates slippery ground features; whether a corresponding area exists in the spatial map; and whether the robot's motion state is consistent with the deceleration requirement.

[0032] If a modality is found to significantly conflict with other modalities, its credibility is automatically reduced. For example, if a linguistic constraint indicates a slippery area, visual perception detects dry ground, but the corresponding area does not exist in the spatial map, the system will determine that there is a semantic inconsistency. In this case, the modality will not be directly deleted, but its fusion weight will be reduced. During the development process, the inventors attempted to use separate visual attack detectors and linguistic attack detectors to judge attack behavior. Experiments showed that as scene complexity increased, the false positive rate of the single-modality detector increased significantly. Therefore, a cross-modal consistency verification mechanism was ultimately adopted.

[0033] Experimental results show that this scheme can significantly improve the anomaly recognition success rate under simulated visual perturbation scenarios. Finally, a unified security representation information is generated through an uncertainty-aware fusion mechanism, and the overall uncertainty assessment result is output. This unified security representation information serves as the common input for steps S3 and S5.

[0034] Step S3: Corresponding Figure 3 The module shown addresses the problem in existing technologies where natural language constraints cannot directly affect the underlying control. When the system receives the administrator's input: "When passing through shelf area 3, the speed should not exceed the set threshold, and a safe distance should be maintained from the shelf," the semantic parsing module first identifies: area constraints; speed constraints; and distance constraints. These are then converted into corresponding physical constraint parameters. For example, area constraints are converted into restricted area boundaries; speed constraints into maximum speed boundaries; and distance constraints into minimum safe distance boundaries. Semantic verification is then performed. The system uses visual perception results to verify whether the constraint object actually exists. For example, the "slippery area" in the language constraint must be identifiable in the visual perception results. If verification fails, a safety alarm is triggered. After verification, a hybrid control barrier function is constructed. This hybrid control barrier function includes: analytical geometric safety boundaries; semantic safety boundaries; and neural network-learned safety boundaries. These three types of boundaries together constitute a unified safety constraint. During development, the inventors initially attempted to use only the analytical control barrier function. Experiments showed that it worked well for scenarios with regular obstacles; however, its ability to express safety boundaries was insufficient when facing irregular shelf layouts and dynamic personnel flow scenarios. Therefore, a neural control barrier function was subsequently added. After training in a large number of hazardous scenarios, it can automatically learn complex safety boundaries, thereby significantly improving its adaptability to complex scenarios.

[0035] Step S4: Upon receiving a new task instruction, the system first performs a safety risk assessment. For example, if an administrator issues the instruction, "Immediately deliver the goods to the designated area," the system needs to consider: path reachability; environmental safety; the status of other robot tasks; and current safety constraints. If multiple alternatives exist, the risk level of each option is assessed separately. A collaborative decision result is then generated. This step is primarily used for high-level task coordination and does not directly participate in low-level control. Therefore, those skilled in the art can employ any implementation method capable of achieving risk assessment and collaborative decision-making.

[0036] Step S5: As Figure 4 As shown, each agent performs local planning based on the unified security representation information generated in step S2 and the hybrid control barrier function generated in step S3. Unlike traditional globally centralized planning, this invention only exchanges local neighborhood information. For example, the automated guided vehicle A does not need to obtain the entire warehouse state; it only needs to obtain the states of neighboring robots and key security information to complete the planning. During the research and development process, the inventors attempted a completely centralized planning scheme. Experiments showed that as the number of robots increased, the communication load and computational complexity increased rapidly. Therefore, a local neighborhood planning mechanism was adopted. Experimental results show that, while ensuring security, the system's computational overhead is significantly reduced.

[0037] Step S6: Corresponding Figure 5 The system continuously monitors the communication link status. When communication quality is normal, a cooperative safety mode is adopted. When communication quality deteriorates, it automatically switches to a conservative cooperative mode. When communication is completely interrupted, it enters an independent safety mode. In this mode, each robot relies solely on local safety constraints to achieve safety control. Communication failure experiments were conducted during the development process. In the experiments, the communication link between robots was artificially severed. The results show that even in a completely disconnected state, the robots can still maintain safe operation by relying on the hybrid control barrier function, and no collision events occurred. After communication is restored, the system automatically performs state synchronization and trajectory fusion. Normal collaborative operation has been restored.

[0038] Step S7: Used for continuous accumulation of safety experience. The system records: changes in safety status; triggering of control barrier functions; communication anomalies; and task execution results. Once a safety behavior pattern is successfully verified multiple times, a corresponding safety skill is automatically generated. Examples include: "Yielding to pedestrians in narrow passages"; "Passing slowly through slippery areas"; and "Dynamically avoiding moving personnel." When encountering similar scenarios later, the system can directly invoke the corresponding safety skill, reducing real-time computational overhead.

[0039] Step S8: Execution Phase. After the planning module generates the control actions, it first performs a final safety verification using a hybrid control barrier function. If all safety constraints are met, the action is executed directly. If the safety constraints are not met, a safety projection mechanism is used to correct the action to the most recent safe action. After execution, the environment returns new observation results, and the process re-enters step S1, forming a continuously running safety closed-loop control process.

[0040] The greatest innovation of this invention lies in proposing a bridging unit from semantic safety constraints to physical hard safety constraints, called the "semantic-physical constraint converter." This converter receives natural language safety constraints (such as "maintain at least a 0.5m distance from other robots") and achieves bridging through the following process: First, a safety-aligned large language model extracts the semantic embedding of the constraint and checks its consistency with the visual segmentation results; then, a semantic constraint mapper transforms the semantic embedding into differentiable safety region constraint parameters (such as distance thresholds and prohibited region shapes); finally, a hybrid control barrier function verifier unifies this parameterized semantic constraint with the analytical geometric control barrier function and the neural learning control barrier function into a single safety region decision function. To the inventors' knowledge, this is the first scheme to directly transform safety constraints from large language model inference into a hard constraint form executable by the control barrier function, completely solving the "semantic-physical gap" problem.

[0041] This invention proposes an automatic construction method for a three-layer hybrid control barrier function. It consists of the minimum values ​​of three sub-barrier functions: the first layer, an analytical geometric control barrier function, provides distance-based collision avoidance guarantees (e.g., minimum safe distance between multiple agents, minimum distance from obstacles); the second layer, a semantic control barrier function, receives the output of a semantic-physical constraint converter and encodes natural language constraints into numerical safety boundaries (e.g., avoidance zone radius for pedestrians, maximum slow speed limit on slippery surfaces); the third layer, a neural control barrier function, uses data-driven learning from diverse hazardous scenario samples generated by a safety diffusion model to automatically model complex safety constraints that cannot be analytically expressed (e.g., passage conditions between non-convex obstacles, safe flight envelope under airflow disturbances). This hybrid approach overcomes the shortcomings of purely analytical control barrier functions requiring manual design by domain experts, purely neural control barrier functions lacking formal guarantees, and purely semantic methods being detached from physical reality.

[0042] This invention proposes a communication-independent security degradation mechanism. When communication between agents is interrupted or degraded, the system automatically degrades from a cooperative security mode to an independent security mode. In the cooperative security mode, intention prediction and graph attention networks are used to share future trajectory information for proactive collision avoidance. When communication is interrupted, each agent independently activates a conservative security strategy, relying solely on a hybrid control barrier function for local security guarantees, and executes actions with a large security margin (e.g., 1.5 times the standard security distance). Once communication is restored, the system automatically and quickly synchronizes its state and resumes cooperation through a consensus mechanism.

[0043] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for safe interaction of multi-agent based on semantic physical constraint transformation, characterized in that, Includes the following steps: S1. Multi-source information acquisition and preprocessing; S2. Perform secure coding and cross-modal consistency verification on the preprocessed multi-source input information; S3. Construct semantic-physical constraint transformation and hybrid security constraints; S4. Mission safety assessment and collaborative decision-making; S5, Distributed Security Planning; S6, Communication Adaptive Security Degradation Execution; S7, Security Information Recording; S8, Safety Action Output.

2. The method of claim 1, wherein, Step S1 specifically includes the following steps: S11. Obtain input information from multiple embodied intelligent agents, including: visual information, denoted as first perception information I; spatial information, denoted as second perception information P; natural language security constraint information, denoted as third perception information L; and motion state information, denoted as fourth perception information M. S12. The first sensing information I, the second sensing information P, the third sensing information L, and the fourth sensing information M are preprocessed respectively to obtain the preprocessed first sensing information I', the preprocessed second sensing information P', the preprocessed third sensing information L', and the preprocessed fourth sensing information M'.

3. The method of claim 2, wherein, Step S2 specifically includes the following steps: S21. The visual encoder outputs visual semantic features and visual credibility, the spatial encoder outputs spatial geometric features and spatial credibility, the language encoder outputs language semantic features and language credibility, and the motion encoder outputs motion state features and motion credibility. S22. The system calculates a first consistency index between the language semantic features and the visual semantic features, and calculates a second consistency index between the spatial geometric features and the motion state features; S23. When any consistency index falls below the corresponding threshold, the credibility of the corresponding modality is reduced. S24. Based on each confidence parameter, the modal features are weighted and fused to obtain a unified security representation.

4. The method of claim 3, wherein, Step S3 specifically includes the following steps: S31. Extract security constraint parameters based on the linguistic semantic information in the unified security representation vector F. The security constraint parameters include: constraint type TYPE, constraint parameter PARAM, and constraint priority LEVEL. The security constraint type TYPE includes at least: distance preservation constraint, area avoidance constraint, velocity limit constraint, and attitude limit constraint. S32. Verify the safety constraint parameters using visual information, match the target object described by the constraint parameter PARAM with the visual recognition result, generate the corresponding physical constraint when the match is successful, and output semantic inconsistency alarm information when the match fails. S33. Subsequently, construct the hybrid security constraint function B(s), where s represents the current system state and B(s) represents the hybrid security constraint function.

5. The method of claim 4, wherein, The hybrid safety constraint function B(s) is composed of the following three sub-constraint functions: the first constraint function B1(s), which describes geometric safety constraints; the second constraint function B2(s), which describes semantic safety constraints; and the third constraint function B3(s), which describes the learned complex safety constraints. The hybrid safety constraint function is defined as: B(s) = min(B1(s), B2(s), B3(s)). When B(s) is greater than or equal to zero, it indicates that the current state meets the safety requirements; when B(s) is less than zero, it indicates that the current state violates the safety requirements. The hybrid safety constraint function B(s) serves as the input for steps S4 and S5.

6. The method of claim 5, wherein, Step S4 specifically includes the following steps: S41. Perform security assessment on task instructions based on the hybrid security constraint function B(s); S42. When multiple agents have conflicting objectives, generate multiple candidate cooperative schemes, and conduct a comprehensive evaluation based on safety indicators and task efficiency indicators to determine the target cooperative scheme and output the cooperative decision result.

7. The method of claim 6, wherein, Step S5 specifically involves generating planning action A based on the collaborative decision-making results output in step S4 and the hybrid safety constraint function B(s). During the planning process, only local neighborhood information is used to complete the safe path planning, and the hybrid safety constraint function B(s) is continuously guaranteed to meet the safety conditions. After generating planning action A, planning action A is sent to step S6.

8. The method of claim 7, wherein, Step S6 specifically includes the following steps: S61. Real-time acquisition of communication quality parameter Q, wherein the communication quality parameter Q is determined by communication delay, packet loss rate and signal strength; S62. Perform mode switching according to the communication quality parameter Q: when the communication quality parameter Q is greater than the first threshold, enter the cooperative security mode; when the communication quality parameter Q is between the first threshold and the second threshold, enter the conservative cooperative mode; when the communication quality parameter Q is less than the second threshold, enter the independent security mode. S63. In independent safety mode: stop using information sent by other agents, perform safety control only based on the perception results of this agent and the hybrid safety constraint function B(s), while increasing the safety distance and reducing the movement speed. When communication is restored, perform state synchronization and resume cooperative planning.

9. A secure interaction method for multiple embodied intelligent agents based on semantic-physical constraint transformation according to claim 8, characterized in that, Step S7 specifically involves recording security event information during the current decision-making process, including: security constraint triggering information, communication status change information, manual intervention information, planning and decision-making information, and forming security record information.

10. A secure interaction method for multiple embodied intelligent agents based on semantic-physical constraint transformation according to claim 9, characterized in that: Step S8 specifically involves: generating a target action A_safe based on the safety control result obtained in step S6, where A_safe represents the final safe action; if the planned action A satisfies the requirements of the hybrid safety constraint function B(s), then A_safe = A; otherwise, the planned action A is modified to obtain the target action A_safe that satisfies the requirements of the hybrid safety constraint function B(s); the target action A_safe is output to the actuator to complete the control, and after the execution is completed, new environmental information is obtained, and the process returns to step S1 to form a closed-loop control process.

Images