Method and system for semantic execution pollution propagation control of a network of temporal robotic agents

Abstract

The application provides a semantic execution pollution propagation control method and system of a timing robot agent network, comprising: collecting data of a robot node; constructing a non-negative promotion propagation operator and updating a pollution upper bound; calculating a plurality of data in a limited prediction time domain; obtaining reverse execution path pressure and path scores of each node based on an executor reachable coefficient; selecting a network protection core node according to a preset network protection budget; configuring an equal-cost network protection package, updating the non-negative promotion propagation operator to obtain a controlled promotion propagation operator; calculating execution exposure degrees of each serially reachable execution host; constructing a serial execution guard set; performing serial execution guard checking on an execution request to be issued outside the serial execution guard set, and blocking a dangerous execution request or allowing an execution command to be issued according to a checking result, so that the application realizes verifiable execution pollution reduction, dangerous action blocking and balanced maintenance of task effectiveness.

Description

Technical Field

[0001] This invention relates to the field of robot intelligent agent control technology, specifically to a method and system for controlling the propagation of semantic execution contamination in a temporal robot intelligent agent network. Background Technology

[0002] With the development of robots possessing language models, shared memory, map collaboration, and tool invocation capabilities, they no longer merely exchange numerical states such as position and speed, but also task descriptions, route summaries, environmental annotations, tool requests, and text with execution intentions. When external text enters a robot, it may initially appear as prompts, then be rewritten as memory records, map annotations, or tool parameters, until it is converted into serial control commands by a host machine with an execution interface. The resulting risks include semantic channel variations, heterogeneous node capabilities, time-varying contact topology, and delayed physical execution consequences.

[0003] Existing technologies can be broadly categorized into four types: the first type detects indirect cue injection at the input or return end of a single tool-type intelligent agent; the second type authenticates or isolates robot middleware, controller communication, and permissions; the third type uses node infection probability, network centrality, or spectral radius to describe network propagation; and the fourth type selects nodes to be protected based on current text labels or single-step risks. These solutions can alleviate input contamination, communication attacks, or general network propagation problems respectively, but they typically do not treat "semantic rewriting—temporal contact—actuator reachability—budget-constrained network intervention—serial execution blocking" as a single closed-loop control object.

[0004] When sorting by node degree or eigenvector centrality alone, it is impossible to distinguish between nodes that are actively communicating but cannot reach the executor and nodes with low degree but located on dangerous execution paths. When filtering by current tool words or action words alone, it is impossible to preprocess contamination that is still in the prompt, memory, or map channel but will be transformed into execution requests in subsequent contacts. When using only one protection budget, network propagation suppression and execution blocking will compete for the same quota, and may abandon more critical propagation path nodes in order to cover the execution host. Summary of the Invention

[0005] This invention is made to solve the above-mentioned problems, and aims to provide a method and system for controlling the propagation of semantic execution pollution in a temporal robot intelligent agent network.

[0006] This invention provides a semantic execution contamination propagation control method for a temporal robot agent network, characterized by the following steps: Step S1, acquiring the type, online state, temporal directed contact relationship, and actuator interface mapping relationship of each robot node in the network; collecting the contamination intensity of robot nodes in the semantic channel, as well as the corresponding semantic mutation matrix, channel susceptibility, detection intensity, recovery rate, and actuator reachability coefficient; Step S2, constructing a non-negative lifting propagation operator and updating the contamination upper bound based on the temporal contact matrix, the basic propagation coefficient of the semantic channel, the semantic mutation matrix, the channel susceptibility, detection intensity, and recovery rate; Step S3, calculating the actuator reachability contamination load, task utility, ordered product of the temporal lifting operator, and finite temporal growth within the finite prediction time domain based on the non-negative lifting propagation operator; Step S4, iteratively propagating backward on the normalized temporal contact matrix based on the actuator reachability coefficient. Step S5: Obtain the reverse execution path pressure of each node, and obtain the path score of the robot node based on the reverse execution path pressure, the pollution intensity of the semantic channel, and the reachability coefficient of the actuator; Step S6: Select a corresponding number of nodes as network protection core nodes according to the path score from high to low based on the preset network protection budget; Configure an equal-cost network protection package for the selected network protection core nodes, and update the non-negative lift propagation operator using the equal-cost network protection package to obtain the controlled lift propagation operator; Step S7: Calculate the two-step prediction state based on the controlled lift propagation operator, and calculate the execution exposure of each serially reachable execution host; Construct a serial guard budget, and select several reachable execution hosts with the highest execution exposure to form a serial execution guard set; Step S8: Perform serial execution guard checks on the execution requests to be issued outside the serial execution guard set, and block dangerous execution requests or allow the issuance of execution commands based on the check results.

[0007] The semantic execution pollution propagation control method for temporal robot intelligent agent networks provided by the present invention may also have the following features, including: step S8, recording the pollution state, propagation operator state, operator version, path score, protection set, guard set and action result of the current period, and entering the next control period according to the updated state of the current period.

[0008] The semantic execution pollution propagation control method for temporal robot intelligent agent networks provided by the present invention may also have the following features: wherein the semantic channels include: prompt semantic channel, memory semantic channel, map semantic channel, tool semantic channel, and execution semantic channel.

[0009] The semantic execution contamination propagation control method for temporal robot intelligent agent networks provided by this invention may also have the following feature: wherein the semantic variation matrix is ​​a row-normalized non-negative matrix, with rows corresponding to source channels and columns corresponding to destination channels.

[0010] The semantic execution contamination propagation control method for temporal robot intelligent agent networks provided by this invention may also have the following features: wherein the equal-cost network protection package includes: reducing the contact weights of incoming and outgoing edges, increasing the detection intensity, increasing the recovery rate, and cleaning up a portion of the local contamination state.

[0011] The semantic execution contamination propagation control method for temporal robot intelligent agent networks provided by the present invention may also have the following features: In step S2, the non-negative boosting propagation operator consists of two parts: the first part is a cross-node propagation term, which corresponds to the amount of propagation after contamination reaches the receiving node through directed contact and undergoes channel variation, and is adjusted by susceptibility and detection intensity; the second part is a node residual term, which corresponds to the remaining amount of local contamination after recovery.

[0012] The semantic execution contamination propagation control method for temporal robot intelligent agent networks provided by the present invention may also have the following feature: wherein, in step S3, the actuator reachable contamination load is the weighted average of the contamination intensity of the semantic channel and the corresponding actuator reachability coefficient, which is used to measure the overall impact of contamination on the physical execution end.

[0013] The semantic execution contamination propagation control method for temporal robot agent networks provided by the present invention may also have the following feature: wherein, in step S6, the execution exposure of each serially reachable execution host is obtained by weighted summation of the contamination intensity of the tool semantic channel and the execution semantic channel at the current time and the predicted time, and the corresponding actuator reachability coefficient.

[0014] The semantic execution contamination propagation control method for temporal robot agent networks provided by this invention may also have the following features: the network protection budget and the serial guard budget are independent of each other, the network protection layer is used to optimize load reduction on the propagation path, and the serial guard layer is used to additionally block dangerous actions reaching the execution end.

[0015] This invention also provides a semantic execution contamination propagation control system for a temporal robot agent network, characterized by the following features: a data acquisition module for acquiring the type, online status, temporal directed contact relationship, and actuator interface mapping relationship of each robot node in the network; acquiring the contamination intensity of robot nodes in the semantic channel, as well as the corresponding semantic variation matrix, channel susceptibility, detection intensity, recovery rate, and actuator reachability coefficient; a boost propagation operator generation module for constructing a non-negative boost propagation operator and updating the contamination upper bound based on the temporal contact matrix, the basic propagation coefficient of the semantic channel, the semantic variation matrix, the channel susceptibility, detection intensity, and recovery rate; a calculation module for calculating the actuator reachability contamination load, task utility, ordered product of the temporal boost operator, and finite-time domain growth within the finite prediction time domain based on the non-negative boost propagation operator; and a path score generation module for iterative backpropagation on the normalized temporal contact matrix based on the actuator reachability coefficient. The system obtains the reverse execution path pressure of each node, and calculates the path score of the robot node based on the reverse execution path pressure, the pollution intensity of the semantic channel, and the reachability coefficient of the actuator. The network protection package configuration module selects a corresponding number of nodes as core network protection nodes according to the preset network protection budget and their path scores from highest to lowest. It configures equal-cost network protection packages for the selected core network protection nodes and updates the non-negative lift propagation operator using these packages to obtain the controlled lift propagation operator. The serial execution guard construction module calculates the two-step prediction state based on the controlled lift propagation operator and calculates the execution exposure of each serially reachable execution host. It constructs a serial guard budget and selects several reachable execution hosts with the highest execution exposure to form a serial execution guard set. The inspection module checks the pending execution requests outside the serial execution guard set and blocks dangerous execution requests or allows the execution command to be issued based on the inspection results.

[0016] The role and effect of invention

[0017] The semantic execution contamination propagation control method and system for temporal robot intelligent agent networks according to the present invention includes: Step S1, acquiring the type, online state, temporal directed contact relationship, and actuator interface mapping relationship of each robot node in the network; collecting the contamination intensity of the robot node in the semantic channel and the corresponding semantic variation matrix, channel susceptibility, detection intensity, recovery rate, and actuator reachability coefficient; Step S2, constructing a non-negative boosting propagation operator and updating the contamination upper bound based on the temporal contact matrix, the basic propagation coefficient of the semantic channel, the semantic variation matrix, the channel susceptibility, detection intensity, and recovery rate; Step S3, calculating the actuator reachability contamination load, task utility, ordered product of the temporal boosting operator, and finite time domain growth within the finite prediction time domain based on the non-negative boosting propagation operator; Step S4, iteratively propagating backward on the normalized temporal contact matrix based on the actuator reachability coefficient to obtain the backward execution path pressure of each node, and obtaining the path score of the robot node based on the backward execution path pressure, the contamination intensity of the semantic channel, and the actuator reachability coefficient; Step S5, selecting the robot node according to the path score from high to low based on the preset network protection budget. Select a corresponding number of nodes as core nodes for network protection; configure equal-cost network protection packages for the selected core nodes, and update the non-negative lift propagation operator using the equal-cost network protection packages to obtain the controlled lift propagation operator; Step S6, calculate the two-step prediction state based on the controlled lift propagation operator, and calculate the execution exposure of each serially reachable execution host; construct a serial guard budget, and select several reachable execution hosts with the highest execution exposure to form a serial execution guard set; Step S7, perform serial execution guard checks on the execution requests to be issued outside the serial execution guard set, and block dangerous execution requests or allow the issuance of execution commands based on the check results. Therefore, the semantic execution pollution propagation control method and system of the temporal robot intelligent agent network of the present invention retains the semantic position of the pollution in the multi-channel semantic pollution state modeling, uses the non-negative lift propagation operator to unify and integrate topology, semantic variation and actuator reachability, accurately identifies nodes on dangerous execution paths by means of reverse path pressure, and avoids resource competition by separating the network protection budget and the serial guard budget, thereby achieving verifiable execution pollution reduction, dangerous action blocking and task utility balance maintenance. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the semantic execution contamination propagation control method for a temporal robot agent network in an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating the key security determination and anomaly rollback in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the semantic execution contamination propagation control system of the temporal robot intelligent agent network in an embodiment of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the semantic execution pollution propagation control method and system of the temporal robot intelligent agent network of this invention.

[0022] Example 1

[0023] Figure 1 This is a flowchart illustrating the semantic execution contamination propagation control method for a temporal robot agent network in an embodiment of the present invention.

[0024] like Figure 1 As shown, this embodiment provides a semantic execution contamination propagation control method for temporal robot agent networks, including:

[0025] Step S1: Obtain the type, online status, temporal directed contact relationship, and actuator interface mapping relationship of each robot node in the network; collect the pollution intensity of the robot node in the semantic channel, as well as the corresponding semantic variation matrix, channel susceptibility, detection intensity, recovery rate, and actuator reachability coefficient.

[0026] In this embodiment, the number of robots is first set to be... Robot node The type is Define semantic channel set for:

[0027]

[0028] The semantic channels include: cue semantic channels. Memory semantic channel Map semantic channel Tool semantic channel Execute semantic channels Number of semantic channels It is 5.

[0029] Contamination intensity of robot nodes in semantic channels for:

[0030]

[0031]

[0032] in, This represents the contamination intensity of robot node i on the c-th semantic channel at time t. . Specifically, These represent the pollution intensity of robot node i in each corresponding semantic channel.

[0033] Stacking the semantic states of all robot nodes yields the contaminated state of the entire robot network:

[0034]

[0035] in, This represents the contamination state vector of the i-th robot in the C semantic channels. This represents the overall contamination state vector of the network formed by vertically concatenating the semantic state vectors of each node according to the robot node number.

[0036] The system allows the same robot to be polluted on multiple semantic channels at the same time, thereby distinguishing states such as prompts being cleared but memories remaining, and maps being polluted but tool requests not yet formed.

[0037] robot node semantic variation matrix for:

[0038]

[0039] And it satisfies: .

[0040] in, This represents the proportion of polluted content received by robot node j from source semantic channel c, which is then transformed into the target semantic channel d after semantic processing. For source channel, For the destination channel.

[0041] The semantic variation matrix is ​​a row-normalized non-negative matrix, with each row corresponding to a source channel. Column corresponding to the destination channel The semantic variation matrix describes the probability that the semantic content entering robot node j will be transferred between different semantic channels during the processing of that node. The semantic variation matrix can be obtained from historical log statistics, rule calibration, or model-generated sample calibration, but a verified matrix with version numbers is used during the online control cycle.

[0042] Step S2: Construct a non-negative boosting propagation operator and update the contamination upper bound based on the temporal contact matrix, the basic propagation coefficient of the semantic channel, the semantic variation matrix, the channel susceptibility, the detection intensity, and the recovery rate.

[0043] In this embodiment, the temporal contact matrix for:

[0044]

[0045] in, Represents robots For robots The effective impact can be generated by proximity communication, route sharing, wireless connectivity, task dependence, or shared storage access.

[0046] Non-negative boosting propagation operator The elements are:

[0047]

[0048] Among them, the control action is denoted as The amount after being affected by the action is above the standard. express, The fundamental propagation coefficient of the source channel. For channel susceptibility, To test the intensity, For recovery rate, channel susceptibility, detection intensity, and recovery rate are all ranked in the top. , This indicates that after the control action is performed, robot node j processes the content of the source semantic channel c and transforms it into the proportion of the target semantic channel d. Indicates the robot node after the control action is performed. The effective influence strength on robot node j.

[0049] In step S2, the non-negative lift propagation operator consists of two parts: the first part is the cross-node propagation term, which corresponds to the amount of propagation after the contaminant reaches the receiving node via directed contact and undergoes channel variation, and is adjusted by susceptibility and detection intensity; the second part is the intra-node residual term, which corresponds to the amount of local contaminant remaining after recovery.

[0050] For robots destination channel Receive pollution pressure for:

[0051]

[0052] This embodiment uses the following formula to update the nonlinear contamination state:

[0053]

[0054] in, The value represents the contamination intensity of robot node j on the target semantic channel d at time t, and its value range is [0,1].

[0055] In this embodiment, Let represent any non-negative real number.

[0056]

[0057] Therefore:

[0058]

[0059] The nonlinear state satisfies the upper bound of the comparison:

[0060]

[0061] The above comparison relationship enables the online controller to calculate the safety upper bound using nonnegative matrix multiplication, while still maintaining the true simulation or estimated state through nonlinear updates. This is especially true when the initial state and model parameters are both located within... At that time, the amount of residual items restored does not exceed the current state, and the amount of new contaminants does not exceed the remaining uncontaminated amount. It remains unchanged at any given time.

[0062] Step S3: Calculate the actuator reachability contamination load, task utility, ordered product of time-series lifting operators, and finite-time domain growth within the finite prediction time domain based on the non-negative lift propagation operator.

[0063] In step S3, the actuator reachability contamination load is the weighted average of the contamination intensity of the semantic channel and the corresponding actuator reachability coefficient, which is used to measure the overall impact of contamination on the physical execution end.

[0064] set up Represents robots passage The formula for calculating the reachability of the actuator to the contamination load is based on the degree of accessibility to the tool or lower-level controller. for:

[0065]

[0066] This load allows multiple semantic channels of the same robot to contribute risk separately, so its value can be greater than 1.

[0067] To simultaneously measure the degree of communication maintenance for each task, let... The nominal utility of the scenario without intervention. Represents a node Whether it is overridden by network protection actions determines the task's effectiveness. for:

[0068]

[0069] in, The weighting coefficients for the impact of pollution load on task utility. The weighting coefficient for the impact of network protection action coverage ratio on task effectiveness. Select an indicator function for the node, setting it to 1 if robot node i belongs to the network protection set u(t) at time t, and 0 otherwise.

[0070] In this embodiment , This formula is used for scheme comparison and parameter calibration, and does not equate task scores with the success rate of real physical tasks.

[0071] Given the prediction time domain Define the ordered product of the time-order lifting operator. :

[0072]

[0073] in, Let represent the controlled non-negative boosting propagation operator at time t under the network protection action u(t). Further define the finite-time semantic execution regeneration number. :

[0074]

[0075] in, represents the spectral radius. H represents the prediction time domain length, i.e., the number of consecutive control cycles considered from the current moment. This quantity is calculated on the enhanced robot-semantic channel space, reflecting the contamination growth trend within the prediction time domain, rather than the ordinary infection regeneration number on the original robot graph.

[0076] To provide a reproducible decay condition (i.e., a weighted block decay safety certificate) on time-varying operator sequences, a strictly positive vector is taken. and If at the starting point of each block All satisfy:

[0077]

[0078] The control sequence is said to satisfy the weighted block decay condition. The corresponding weighted infinite norm is defined as:

[0079]

[0080] in, This represents the enhanced robot-semantic channel state vector. The One portion, This represents the strictly positive weight corresponding to the state component.

[0081] By utilizing the nonnegativity of the lifting propagation operator and the upper bound of comparison, the geometric decay relationship at block time can be obtained:

[0082]

[0083] The controller does not require that the weighted block decay safety certificate be necessarily satisfied under any finite budget. When the certificate is temporarily unfeasible, the condition is still made easier to satisfy by reducing the reachability of execution states and shortening hazard propagation paths. If detected... Not satisfied with less than or equal to If the component-wise constraint is met and the load exceeds the threshold, then the program will enter the budget expansion, local isolation, or safe shutdown branch.

[0084] Step S4: Based on the actuator reachability coefficient, iterative propagation is performed in reverse on the normalized temporal contact matrix to obtain the reverse execution path pressure of each node. The path score of the robot node is obtained based on the reverse execution path pressure, the contamination intensity of the semantic channel, and the actuator reachability coefficient.

[0085] In this embodiment, the reachability coefficients of all actuators are first stacked as follows: For candidate network actions Define a finite-time load balancing agent :

[0086]

[0087] in, To represent the weighting coefficients for the contamination load term that can be reached at the end of the prediction time domain, To represent the weighting coefficients of the short-term semantic state change term, This is the reduction factor for the short-term forecast step. Indicates the first The weights corresponding to each prediction step To indicate the elapsed time since the current moment The predicted semantic pollution state of the network.

[0088] The first evaluation term assesses the achievable semantic load at the end of the prediction time domain, while the second term evaluates the current state and semantic movement within the next three short-term prediction steps. This applies to a basic protection action. The first-order descent of the load loader is written as:

[0089]

[0090] in, This indicates the finite-time load surrogate function after executing basic protection action a. The first-order approximate descent, relative to the amount of descent before the action is performed, is used to measure the marginal reduction contribution of the protective action to the actuator's achievable contamination load. This represents the change caused by the basic protective action 'a'. These represent the load surrogate functions respectively. Regarding improving the propagation operator The gradient of the contamination state z and the actuator reachability coefficient r. , , These respectively represent the execution of protective actions. The resulting change in the corresponding variable.

[0091] The maximum actuator reachability of each robot across all semantic channels is compressed to :

[0092]

[0093] The temporal contact matrix is ​​normalized according to the incoming edge quality of each destination node:

[0094]

[0095] in, To prevent division by zero when the incoming edge is empty.

[0096] Using the executor reachable vector as the restart source, the pressure of the reverse execution path from the execution host along the current influence graph is as follows:

[0097]

[0098] Where b represents the actuator reachability vector of the robot node. That is, the maximum reachability of robot node i across all semantic channels. This represents the restart coefficient in the path pressure iteration, used to control the proportion of reachable vectors of the executor retained in the current iteration result. In this embodiment... Perform a fixed number of iterations. This represents the matrix obtained by normalizing the temporal contact matrix A(t) according to the incoming edge weights of each target node. This represents the reverse execution path pressure vector obtained in the m-th iteration.

[0099] robot node Path score for:

[0100]

[0101] Step S5: Based on the preset network protection budget, select a corresponding number of nodes as core network protection nodes according to the path score from high to low; configure equal-cost network protection packages for the selected core network protection nodes, and use the equal-cost network protection packages to update the non-negative boosting propagation operator to obtain the controlled boosting propagation operator.

[0102] In this embodiment, a preset network protection budget is used. Choose the path with the highest score. The robot nodes form the core of network protection. :

[0103]

[0104] For the peer cost node protection package, the problem equivalence is:

[0105]

[0106] Since the protection packet cost is equal for each robot node, according to Selecting from largest to smallest maximizes the descent of the first-order load agent. This sorting also uses reverse execution path pressure, current multi-semantic channel pollution, and executor reachability coefficients, thus distinguishing two robot nodes with the same graph centrality and total pollution quality but different future execution consequences.

[0107] The cost-effective network protection package includes: reducing the contact weights of incoming and outgoing edges, increasing detection intensity, improving recovery rate, and cleaning up some local contamination.

[0108] Equivalent cost network protection packages are updated using the following formula:

[0109]

[0110]

[0111] in, The gating coefficient and , To detect incremental changes, To restore incremental growth, This refers to the cleaning ratio.

[0112] In this embodiment, , , , Network protection budget Actual deployment can be determined based on link capacity, false alarm cost, and mission security level, but the protection packets of each candidate node within the same comparison period should maintain equal cost or be converted to a uniform cost before sorting.

[0113] Step S6: Calculate the two-step predicted state based on the controlled lift propagation operator. and Calculate the execution exposure of each serially reachable execution host; construct a serial guard budget, and select several reachable execution hosts with the highest execution exposure to form a serial execution guard set.

[0114] In step S6, the execution exposure of each serially reachable execution host is obtained by weighted summation of the contamination intensity of the tool semantic channel and the execution semantic channel at the current time and the predicted time, and the corresponding executor reachability coefficient.

[0115] In this embodiment, for robot nodes Calculate the execution exposure of each serially reachable execution host. The calculation formula is as follows:

[0116]

[0117] in, This indicates the reachability of the tool semantic channel from robot node i to the tool or lower-level controller; This indicates the reachability of the execution semantic channel of robot node i to the physical actuator or low-level controller. Both values ​​are in the range of [0,1]. The larger the value, the more easily the pollution state in the corresponding semantic channel can affect the actual execution end.

[0118] Constructing a budget independent of network protection The serial guard budget G is only used for the set of reachable hosts with serial or lower-level execution connections. Select the host with the highest execution exposure:

[0119]

[0120] The serial execution exposure not covered by the serial guard budget G is:

[0121]

[0122] The network protection budget and the serial guard budget are independent of each other. The network protection layer is used to optimize load reduction on the propagation path, while the serial guard layer is used to additionally block dangerous actions reaching the execution end.

[0123] Step S7: Perform serial execution guard checks on execution requests outside the serial execution guard set, and block dangerous execution requests or allow the execution command to be issued based on the check results.

[0124] Suppose an execution request Source chain inspection Task authorization check is Parameter whitelist check Performing canary checks Then the actual action issued satisfies:

[0125]

[0126] in, For inaction, speed limit, maintaining position, or predefined safety actions. Because the guard only... Subtract non-negative execution exposure It will not increase the exposure of unguarded execution; as long as the selected host exists. Unguarded execution exposure is significantly reduced. The network protection layer retains the core of the equal-cost path optimal for propagation load, while the execution guard layer additionally blocks serial events to prevent both from competing for the same node budget.

[0127] Step S8: Record the pollution status, propagation operator status, operator version, path score, protection set, guard set, and action result of the current cycle, and enter the next control cycle based on the updated status of the current cycle.

[0128] Figure 2 This is a flowchart illustrating the key security determination and anomaly rollback in an embodiment of the present invention.

[0129] like Figure 2 As shown, in one cycle, after obtaining the controlled prediction result, it is determined whether the execution is feasible if the load is below the threshold and the block decay condition is met. If so, the source chain, task authorization, parameter whitelist, and canary check are performed. If not, any one of the following is executed: expanding the network protection budget and serial guard budget, isolating the subnet, or safe shutdown. Simultaneously, no action, rate limiting, or maintaining position is output. When the source chain, task authorization, parameter whitelist, and canary check all pass, the verified action is published; otherwise, no action, rate limiting, or maintaining position is output. The cycle ends after the version, set, action, and rollback reason are written.

[0130] In this embodiment, the state input includes at least the object identifier, acquisition time, period number, robot type, semantic channel, contamination intensity, parameter version, and validity flag. The topology input includes at least the sending node, receiving node, contact weight, effective time, and task dependency type. The actuator mapping includes at least the host, tool interface, serial port, allowed command set, and safety actions. All matrices are checked for dimension, non-negativity, and row normalization conditions before entering the calculation.

[0131] When messages are out of order, they are rearranged by period number and timestamp. When a message is missing, the most recently valid conservative upper bound is used within a limited holding window; if the message exceeds the window, the vulnerability and reachability of the relevant nodes are increased to a preset conservative value. When the row sum of the semantic matrix is ​​not equal to one, non-negative truncation is performed before normalization. When the executor mapping is ambiguous, it is treated as reachable. When optimization or iteration times out, only the candidate set that has passed budget and safety checks is used. No action should be issued directly without bypassing the serialization guard in case of any exception.

[0132] Concurrent updates employ a periodic snapshot-computation-atomic release approach. A consistent snapshot of the topology, state, parameters, and executor mappings is created at the beginning of each period. Mid-cycle version replacements are not accepted during control computation. New parameters only take effect at the boundary of the next period. The log saves a summary of the input snapshots, parameters used in the formulas, candidate sorting, the final set, and the reason for rollback to support replay.

[0133] In this embodiment, the semantic channels can be further subdivided from the five channels into visual annotations, code, planning trajectories, database records, or different executor categories. Channels can also be merged while maintaining the "language state—tool state—execution state" transition relationship. The semantic variation matrix can be generated by fixed rules, statistical frequencies, language model calibration, or constrained online estimation. Regardless of the method used, online values ​​must be non-negative, row-normalized, and include a version number.

[0134] Reverse path pressure can be implemented using fixed-order exponential iteration, convergence threshold iteration, sparse matrix propagation, or distributed message passing. Equal-cost protection packages can also be replaced with cost-based knapsack or integer programming, as long as the optimization objective remains achievable load reduction and the guard budget remains independent. Serial guards can be deployed on robot hosts, tool gateways, controller agents, or external security modules, and can be replaced by blocking, rate limiting, whitelisting, dual-authorization, or security actions.

[0135] In this embodiment, five categories of safe canary text—cue, memory, map, tool, and execution—are used to label the semantic variation matrix. Two text-to-text models are employed, with the smaller model having 60.5 million parameters and the larger model having 222.9 million parameters. The four canary cues for each channel are repeated 64 times on each model, generating 320 rewritten outputs per model, for a total of 640 outputs. The generation temperature is set to 0.85, the kernel sampling probability to 0.92, and the maximum number of new labels to 48. A fixed-channel keyword counting classifier is used to determine the target channel, and a pseudo-count is added to each target channel for Laplace smoothing to obtain row randomization. This matrix is ​​only used to estimate semantic channel movement and is not used as a single text malice detector.

[0136] Simulation setting of robot quantity The scenarios included four types of temporal contact: warehousing, coverage, formation, and task assignment. The robot type probabilities were 0.22 for traditional nodes, 0.45 for general nodes, 0.23 for edge agents, and 0.10 for execution hosts. Each scenario and scale involved 64 matching trials, with 180 control steps per trial. Initial contamination included 2.5% of cue channel nodes, half of which also had memory contamination; an additional tool channel canary was added when an execution host was present.

[0137] The network budget is:

[0138]

[0139] Therefore when hour, The values ​​are approximately 8, 17, 33, and 67, respectively. The channel propagation coefficients are set according to the order of prompts, memories, maps, tools, and execution:

[0140]

[0141] use , And the aforementioned protection package parameters. After aggregating all scenarios, scales, and matched tests, the uncontrolled peak execution load is 1.064, while the CAGMC (Constrained ActuationGuarded Mixed Control) of this invention is 0.943, with a relative decrease of:

[0142]

[0143] The final load of CAGMC was 0.939, corresponding to a task score of 0.493. This result is the same as the simulated load of the path core because serial guards do not change the simulated network operators. The additional effect of serial guards is reflected in the hardware execution event metrics.

[0144] Example 2

[0145] To verify the scenario where contamination lingers in the prompting, memory, and map channels before being converted into an execution request, the elements of the mutation matrix from the three source channels to the tool channel are multiplied by 1.55, and the elements to the execution channel are multiplied by 1.95, and then row normalization is performed again. Robot contact, budget, scale, and matching experiments are kept constant.

[0146] Under this stress scenario, CAGMC's peak load is 0.947, single-step load control is 0.964, current channel rule filtering is 1.008, and control without using actuator reachable items is 1.007. Example 2 shows that before the execution channel becomes the current dominant state, the reverse execution path pressure can select prompt, memory, or map nodes leading to the execution host.

[0147] Table 1: Simulation results of different control methods under delayed execution pressure scenario

[0148]

[0149] Example 3

[0150] In a wired heterogeneous test network, the terminals include Raspberry Pi 2, Raspberry Pi 3B+, Raspberry Pi 4, Raspberry Pi 5, RDK X5, and Jetson Nano, forming a test topology via a managed switch. Selected ports connect to CH340 or Arduino-like serial microcontrollers, and the firmware only accepts secure commands such as identity queries, heartbeats, and analog reads. Dangerous execution contamination is indicated by canary commands that should be blocked; unsafe real physical commands are not sent.

[0151] The terminal runs a micro or medium-sized decoding model locally, with 0.85 million and 2.75 million parameters, respectively. A total of 80 endpoint outputs are generated across the six device types, with a source-to-generation channel matching rate of 1.00 under secure canary conditions. Role cross-validation uses six role configurations, each performing 120 trials and 36 control steps, allowing the same hardware to act as a relay node, edge language node, or actuator reachable from the host in different trials.

[0152] The peak reachable contamination load for the undefended method is 0.490, while the peak reachable contamination load for both the path core method and the CAGMC method is 0.233. The path core method, without serial execution guards, has a canary event blocking ratio of 0, while the CAGMC method has a blocking ratio of 1.00. This verifies that the network layer is responsible for reducing propagation and execution reachability load, and the independent guard layer blocks canary events reaching the serial execution layer without changing the network operators and network protection set.

[0153] Table 2: Hardware Role Cross-Validation Results

[0154]

[0155] Example 4

[0156] The transport robot reads the contaminated task description from an external tag and writes it to the shared map. The system first records this content as a prompt and map channel status, and then predicts, based on the receiver's semantic variation matrix, whether it might be converted into scheduling tool parameters. A certain scheduling robot does not have the highest communication degree, but its tool and execution channel are directly connected to the forklift controller, therefore... , and Higher, included The system reduces the weights of high-risk inbound and outbound edges, enhances detection and recovery, and cleans the map cache, while retaining video and low-risk location data channels.

[0157] After network protection, the system calculates the exposure of the two-step execution. If the scheduling robot still belongs to... If the tool gateway requires the source chain, task authorization, and forklift parameter whitelist to pass simultaneously, the "hold position" action will be output if any condition fails. The handling log records the channel changes from the initial warning to the map and then to the tool, the path pressure iteration results, protection packet parameters, and the reason for the guard.

[0158] When the robot has no entered edges, the normalized denominator is used. The path propagation term for that row is set to zero. When all actuator reachability coefficients are zero, meaningless path sorting is not continued to be calculated; instead, monitoring is retained and the network protection set is set to empty. When the semantic variation matrix has negative values ​​or row sum deviations, the version is rejected first, or a new version is formed after non-negative truncation and row normalization if auditing permits.

[0159] When communication latency causes inconsistencies between the topology and state versions, data from two periods are not mixed. The system recalculates using the most recent consistent snapshot. If a control computation timeout occurs, only the set of protections that have passed budget and security checks are published. If execution feedback is lost, unknown states are treated as incomplete, and the execution end status is queried before deciding whether to resend or rescind. If the safety threshold is exceeded or the block decay condition is not met consecutively, the risk level will be increased sequentially. or Isolate related subnets, stop tool calls, and perform a safe shutdown.

[0160] Figure 3 This is a schematic diagram of the semantic execution contamination propagation control system of the temporal robot intelligent agent network in an embodiment of the present invention.

[0161] like Figure 3 As shown, this embodiment also provides a semantic execution contamination propagation control system 100 for a temporal robot intelligent agent network, including: a data acquisition module 1, a boosting propagation operator generation module 2, a calculation module 3, a path score generation module 4, a network protection packet configuration module 5, a serial execution guard construction module 6, and an inspection module 7.

[0162] The acquisition module 1 uses the above step S1 to acquire the type, online status, temporal directed contact relationship, and actuator interface mapping relationship of each robot node in the network; and to acquire the pollution intensity of the robot node in the semantic channel, as well as the corresponding semantic variation matrix, channel susceptibility, detection intensity, recovery rate, and actuator reachability coefficient.

[0163] The boosting propagation operator generation module 2 adopts the above step S2 to construct a non-negative boosting propagation operator and update the contamination upper bound based on the temporal contact matrix, the basic propagation coefficient of the semantic channel, the semantic variation matrix, the channel susceptibility, the detection intensity and the recovery rate.

[0164] The calculation module 3 uses the above step S3 to calculate the actuator reachability contamination load, task utility, ordered product of the time-series lifting operator, and finite-time domain growth in the finite prediction time domain based on the non-negative lifting propagation operator.

[0165] The path score generation module 4 uses the above step S4 to perform reverse iterative propagation on the normalized temporal contact matrix based on the actuator reachability coefficient to obtain the reverse execution path pressure of each node, and obtain the path score of the robot node based on the reverse execution path pressure, the contamination intensity of the semantic channel and the actuator reachability coefficient.

[0166] The network protection package configuration module 5 uses the above step S5 to select a corresponding number of nodes as network protection core nodes according to the preset network protection budget and the path score from high to low; configure equal-cost network protection packages for the selected network protection core nodes, and use equal-cost network protection packages to update the non-negative boosting propagation operator to obtain the controlled boosting propagation operator.

[0167] The serial execution guard construction module 6 adopts the above step S6 to calculate the two-step prediction state based on the controlled lifting propagation operator, calculate the execution exposure of each serially reachable execution host, construct the serial guard budget, and select several reachable execution hosts with the highest execution exposure to form a serial execution guard set.

[0168] The inspection module 7 uses the above step S7 to perform serial execution guard checks on execution requests outside the serial execution guard set, and blocks dangerous execution requests or allows the execution command to be issued based on the inspection results.

[0169] The role and effect of the embodiments

[0170] The semantic execution contamination propagation control method and system for temporal robot intelligent agent networks according to the present invention includes: Step S1, acquiring the type, online state, temporal directed contact relationship, and actuator interface mapping relationship of each robot node in the network; collecting the contamination intensity of the robot node in the semantic channel and the corresponding semantic variation matrix, channel susceptibility, detection intensity, recovery rate, and actuator reachability coefficient; Step S2, constructing a non-negative boosting propagation operator and updating the contamination upper bound based on the temporal contact matrix, the basic propagation coefficient of the semantic channel, the semantic variation matrix, the channel susceptibility, detection intensity, and recovery rate; Step S3, calculating the actuator reachability contamination load, task utility, ordered product of the temporal boosting operator, and finite time domain growth within the finite prediction time domain based on the non-negative boosting propagation operator; Step S4, iteratively propagating backward on the normalized temporal contact matrix based on the actuator reachability coefficient to obtain the backward execution path pressure of each node, and obtaining the path score of the robot node based on the backward execution path pressure, the contamination intensity of the semantic channel, and the actuator reachability coefficient; Step S5, selecting the robot node according to the path score from high to low based on the preset network protection budget. Select a corresponding number of nodes as core nodes for network protection; configure equal-cost network protection packages for the selected core nodes, and update the non-negative lift propagation operator using the equal-cost network protection packages to obtain the controlled lift propagation operator; Step S6, calculate the two-step prediction state based on the controlled lift propagation operator, and calculate the execution exposure of each serially reachable execution host; construct a serial guard budget, and select several reachable execution hosts with the highest execution exposure to form a serial execution guard set; Step S7, perform serial execution guard checks on the execution requests to be issued outside the serial execution guard set, and block dangerous execution requests or allow the issuance of execution commands based on the check results. Therefore, the semantic execution pollution propagation control method and system of the temporal robot intelligent agent network of the present invention retains the semantic position of the pollution in the multi-channel semantic pollution state modeling, uses the non-negative lift propagation operator to unify and integrate topology, semantic variation and actuator reachability, accurately identifies nodes on dangerous execution paths by means of reverse path pressure, and avoids resource competition by separating the network protection budget and the serial guard budget, thereby achieving verifiable execution pollution reduction, dangerous action blocking and task utility balance maintenance.

[0171] This embodiment improves the semantic position of the state retention pollution through the robot-semantic channel, which can handle the rewriting and delayed propagation of content between prompts, memories, maps, tools and execution.

[0172] This embodiment integrates topology, semantic variation, susceptibility, detection, recovery, and actuator reachability through temporal nonnegative operators and finite time-domain products, making the control basis computable and verifiable.

[0173] This embodiment selects the protection core by using path pressure propagation from the execution host in reverse, which can prioritize protecting nodes that are actually on dangerous execution paths, rather than just protecting nodes with higher degrees.

[0174] This embodiment separates the network budget and the serial guard budget. The network layer maintains the optimal cost-performance selection for the load loader, while the guard layer further reduces the exposure of unguarded executions, without them crowding out each other's slots.

[0175] This embodiment provides a basis for verifying that the contamination state does not exceed the limit and that the block time decreases geometrically by using nonlinear state invariance, comparison upper bound, and weighted block decay conditions.

[0176] This embodiment preserves low-risk collaboration by selectively gating high-risk channels and nodes, which is more conducive to maintaining task effectiveness than network-wide isolation.

[0177] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for controlling the propagation of semantic execution contamination in a temporal robot agent network, characterized in that, include: Step S1: Obtain the type, online status, temporal directed contact relationship, and actuator interface mapping relationship of each robot node in the network; The contamination intensity of the robot node in the semantic channel, as well as the corresponding semantic variation matrix, channel susceptibility, detection intensity, recovery rate, and actuator reachability coefficient, are collected. Step S2: Construct a non-negative boosting propagation operator and update the contamination upper bound based on the temporal contact matrix, the basic propagation coefficient of the semantic channel, the semantic variation matrix, channel susceptibility, detection intensity, and recovery rate; Step S3: Calculate the actuator reachability contamination load, task utility, ordered product of the time-series lifting operator, and finite-time domain growth within the finite prediction time domain based on the non-negative lifting propagation operator; Step S4: Based on the actuator reachability coefficient, perform reverse iterative propagation on the normalized temporal contact matrix to obtain the reverse execution path pressure of each node, and obtain the path score of the robot node according to the reverse execution path pressure, the contamination intensity of the semantic channel and the actuator reachability coefficient. Step S5: Based on the preset network protection budget, select a corresponding number of nodes as network protection core nodes according to the path scores from high to low; configure equal-cost network protection packets for the selected network protection core nodes, and use the equal-cost network protection packets to update the non-negative boosting propagation operator to obtain the controlled boosting propagation operator; Step S6: Calculate the two-step prediction state based on the controlled boosting propagation operator, and calculate the execution exposure of each serially reachable execution host; Construct a serial guard budget, and select several reachable execution hosts with the highest execution exposure to form a serial execution guard set; Step S7: Perform serial execution guard checks on execution requests outside the serial execution guard set, and block dangerous execution requests or allow the execution command to be issued based on the check results.

2. The semantic execution contamination propagation control method for temporal robot agent networks according to claim 1, characterized in that, Also includes: Step S8: Record the pollution status, propagation operator status, operator version, path score, protection set, guard set, and action result of the current cycle, and enter the next control cycle based on the updated status of the current cycle.

3. The semantic execution contamination propagation control method for temporal robot agent networks according to claim 1, characterized in that: in, The semantic channels include: prompt semantic channel, memory semantic channel, map semantic channel, tool semantic channel, and execution semantic channel.

4. The semantic execution contamination propagation control method for temporal robot agent networks according to claim 1, characterized in that: in, The semantic variation matrix is ​​a row-normalized non-negative matrix, with rows corresponding to source channels and columns corresponding to destination channels.

5. The semantic execution contamination propagation control method for temporal robot agent networks according to claim 1, characterized in that: in, The equal-cost network protection package includes: reducing the contact weights of incoming and outgoing edges, increasing detection intensity, improving recovery rate, and cleaning up some local contamination.

6. The semantic execution contamination propagation control method for temporal robot agent networks according to claim 1, characterized in that: in, In step S2, the non-negative boosting propagation operator consists of two parts: the first part is the cross-node propagation term, which corresponds to the amount of propagation after the contaminant reaches the receiving node via directed contact and undergoes channel variation, and is adjusted by susceptibility and detection intensity; the second part is the intra-node residual term, which corresponds to the amount of local contaminant remaining after recovery.

7. The semantic execution contamination propagation control method for temporal robot agent networks according to claim 1, characterized in that: in, In step S3, the actuator reachability contamination load is the weighted average of the contamination intensity of the semantic channel and the corresponding actuator reachability coefficient, used to measure the overall impact of contamination on the physical execution end.

8. The semantic execution contamination propagation control method for temporal robot agent networks according to claim 1, characterized in that: in, In step S6, the execution exposure of each serially reachable execution host is obtained by weighted summation of the contamination intensity of the tool semantic channel and the execution semantic channel at the current time and the predicted time, and the corresponding executor reachability coefficient.

9. The semantic execution contamination propagation control method for temporal robot agent networks according to claim 1, characterized in that: in, The network protection budget and the serial guard budget are independent of each other. The network protection layer is used to optimize load reduction on the propagation path, while the serial guard layer is used to additionally block dangerous actions reaching the execution end.

10. A semantic execution contamination propagation control system for a temporal robot agent network, characterized in that, include: The data acquisition module is used to obtain the type, online status, time-series directed contact relationship, and actuator interface mapping relationship of each robot node in the network; The contamination intensity of the robot node in the semantic channel, as well as the corresponding semantic variation matrix, channel susceptibility, detection intensity, recovery rate, and actuator reachability coefficient, are collected. The propagation operator generation module is used to construct a non-negative propagation operator and update the contamination upper bound based on the temporal contact matrix, the basic propagation coefficient of the semantic channel, the semantic variation matrix, channel susceptibility, detection intensity and recovery rate. The calculation module is used to calculate the actuator reachable contamination load, task utility, ordered product of the time-series lifting operator, and finite-time domain growth in the finite prediction time domain based on the non-negative lifting propagation operator. The path score generation module is used to iteratively propagate backward on the normalized temporal contact matrix based on the actuator reachability coefficient to obtain the backward execution path pressure of each node, and to obtain the path score of the robot node according to the backward execution path pressure, the contamination intensity of the semantic channel and the actuator reachability coefficient. The network protection package configuration module is used to select a corresponding number of nodes as core network protection nodes according to the path score from high to low based on the preset network protection budget. Configure an equal-cost network protection package for the selected network protection core node, and use the equal-cost network protection package to update the non-negative boosting propagation operator to obtain the controlled boosting propagation operator; The serial execution guard construction module is used to calculate the two-step prediction state based on the controlled lift propagation operator and to calculate the execution exposure of each serially reachable execution host. Construct a serial guard budget, and select several reachable execution hosts with the highest execution exposure to form a serial execution guard set; The inspection module is used to perform serial execution guard checks on execution requests outside the serial execution guard set, and block dangerous execution requests or allow the execution command to be issued based on the inspection results.

Images