A humanoid robot action semantic consistency verification method and system based on cognitive execution descriptor

Abstract

This invention discloses a method and system for verifying the semantic consistency of humanoid robot actions based on cognitive execution descriptors, belonging to the field of robot motion safety technology. This method uses a dynamic cognitive execution descriptor carrying an environmental confidence score as the carrier of action intent. The confidence score is dynamically calculated based on the noise variance of the body's sensors; when the noise increases, the confidence score of the action execution is automatically reduced. The descriptor undergoes three levels of processing: semantic filtering by a motion semantic immune grid, risk scoring, and risk entropy accumulation. Execution is delayed when the risk entropy reaches 80% of the baseline, and mandatory auditing is triggered and an immutable audit chain bound to the body's state is generated when it reaches 150% of the baseline. The scheduler periodically injects three types of active detection signals: predictable disturbances, joint responses, and action loops. Hidden anomaly control is determined based on whether the body is normal, has no response, or has a false response; a false response triggers body freezing. In the frozen state, thawing and recovery are achieved through bridging hash continuity verification, and a three-level budget control system (full / partial / shadowless execution) is introduced. This invention uses risk entropy accumulation to pre-emptively control low-frequency hidden anomalies, effectively detecting hidden anomalies in the action control layer and preventing false recovery.

Description

Technical Field

[0001] This invention relates to the field of robot motion safety technology, specifically to a method and system for verifying the semantic consistency of humanoid robot actions based on cognitive execution descriptors. Background Technology

[0002] The motion control core of a humanoid robot is the foundation for decision-making and execution of all its movements, responsible for translating the upper-level task intent into action commands for each joint. Once the motion control core is abnormally injected or hijacked, attackers can bypass the upper-level security mechanisms, freely control joint movements, and create dangerous motion behaviors.

[0003] In existing technologies, safety verification of humanoid robot movements mainly relies on three types of solutions: Option 1 is a security framework based on permissions and controls. This type of solution assigns action execution permissions to each control process in the form of tokens, and the motion control core verifies the validity of the tokens before executing an action. However, this approach only verifies "who is performing what operation," not "whether the execution semantics are consistent with expectations." For example, a control process with legitimate permissions might be injected with an abnormal action sequence, issued to the leg mechanisms in the form of walking commands, but the issued foot trajectory might be a dangerous, unbalanced trajectory instead of a normal gait. Existing permission frameworks are completely unaware of this semantic-level injection.

[0004] Option 2 is an action verification framework based on static policies. This type of framework determines whether an action is allowed to execute based on predefined action rules. However, the trust weights and permission levels in this framework are fixed and do not adjust with changes in the operating environment. In scenarios with fluctuating environmental noise, such as changes in ground friction or increased observation noise caused by instantaneous sensor disturbances, the fixed policy framework cannot distinguish between "random deviations caused by noise" and "dangerous actions masked by noise in abnormal control," leading to a higher false positive rate.

[0005] Option 3: Log- and audit-based analysis systems. This type of system records action trajectories and then analyzes abnormal patterns offline. However, it suffers from significant latency. Attackers can employ low-frequency operation strategies, dispersing abnormal actions over longer time windows to avoid alarms that are too costly to trigger individually. Furthermore, attackers can hijack the critical path of motion control and forge normal responses, making the "normal" operating state recorded in the audit logs actually a state of abnormal control.

[0006] The common shortcomings of the above three schemes are: lack of continuous verification capability for semantic consistency of action execution, lack of dynamic adjustment of trust assessment with environmental state, inability to identify the disguised behavior of the ontology in high noise environment, and inability to detect hidden abnormal control of the action control layer. Summary of the Invention

[0007] To address the problems in existing technologies, such as trust assessment not dynamically adjusting with environmental conditions, inability to detect hidden abnormal control in the action control layer, and lack of continuous verification of semantic consistency of action execution, this invention provides a method and system for verifying the semantic consistency of humanoid robot actions based on cognitive execution descriptors.

[0008] To achieve the above objectives, the technical solution provided by the present invention is as follows: Firstly, a method for verifying the semantic consistency of humanoid robot actions based on cognitive execution descriptors is provided, comprising: constructing a dynamic cognitive execution descriptor model with environmental confidence scores; inputting the dynamic cognitive execution descriptor model into a motion semantic immune grid, and performing multi-layer verification sequentially through a semantic filtering layer, an execution risk layer, and a risk entropy layer; when the accumulated risk entropy value reaches a mandatory audit threshold, triggering a mandatory memory audit to disk and generating an immutable audit chain, while blocking high-privilege actions; periodically injecting active probe signals through a scheduler, and determining whether it is in a hidden abnormal control state based on the response type; after the body enters a frozen state, generating a bridging hash value, and performing body state verification and matching with the bridging hash value when a defreezing request is received, allowing defreezing only when the verification passes and the match is successful.

[0009] Secondly, a humanoid robot action semantic consistency verification system based on cognitive execution descriptors is provided, including a dynamic cognitive execution descriptor construction module, a motion semantic immune grid module, a risk entropy mandatory audit module, an active detection module, and a bridging hash unfreezing and recovery module.

[0010] Thirdly, a humanoid robot is provided, including a processor and a memory, wherein the processor implements the above-described method when executing a computer program stored in the memory.

[0011] Fourthly, a computer-readable storage medium is provided, wherein a computer program stored thereon implements the above-described method when executed by a processor.

[0012] The beneficial effects of this invention are as follows: 1. Dynamic Trust Assessment and Noise Adaptability. This invention upgrades the traditional static trust model to a dynamic adaptive trust model by introducing an environmental confidence score based on dynamic calculation of the noise variance of the on-body sensor, fundamentally eliminating the security risk of "high-noise pseudo-trust anchors".

[0013] 2. Early interception of low-frequency, covert anomalies. This invention uses a risk entropy accumulation and recording mechanism to accumulate and record each high-risk action as an operational cost, rather than triggering an alarm only when a single action reaches a threshold. If an attacker uses a low-frequency fragmentation strategy to disperse the execution of abnormal actions, the cumulative risk entropy value of multiple operations will exceed the mandatory audit threshold, triggering an audit and blocking subsequent high-privilege actions.

[0014] 3. Effective detection of hidden abnormal control in the action control layer. This invention determines hidden abnormal control by periodically injecting active probe signals through a scheduler and analyzing the ontology's response characteristics, and has the ability to detect hidden hijacking that cannot be detected by frameworks based on static strategies.

[0015] 4. Preventing spurious recovery attacks. This invention generates a bridging hash value when the ontology is frozen. This hash value is bound to the control state snapshot at the time of freezing and the accumulated risk entropy energy value before freezing. If an attacker attempts to perform a spurious recovery by replacing the control state data, the bridging hash match will inevitably fail because it is impossible to simultaneously forge a complete state consistent with the state at the time of freezing. Attached Figure Description

[0016] Figure 1 This is a diagram of the overall system architecture in an embodiment of the present invention.

[0017] Figure 2 This is a detailed diagram of the three-layer structure of the motion semantic immune grid in an embodiment of the present invention.

[0018] Figure 3 This is a flowchart of dynamic trust decay and risk entropy mandatory audit in an embodiment of the present invention.

[0019] Figure 4 This is a flowchart of the active detection and bridging hash unfreezing and recovery process in an embodiment of the present invention.

[0020] The meanings of the markings in the diagram are as follows: Figure 1 : 101-Cognitive execution descriptor capture and construction module, 102-Dynamic cognitive execution descriptor model, 103-Action intention attribute, 104-Ontology capability attribute, 105-Execution context fingerprint, 106-Environment confidence score, 107-Semantic filtering layer, 108-Execution risk layer, 109-Risk entropy layer, 110-Execution path control unit; Figure 2 201-Dynamic cognitive execution descriptor, 202-Semantic filtering layer entry, 203-Legality judgment, 204-Silent discard, 205-Execution risk layer entry, 206-Multi-dimensional risk scoring, 207-Risk judgment, 208-Low-risk path, 209-Risk entropy recording and accumulation, 210-Risk entropy judgment, 211-Release, 212-Delayed execution; Figure 3301 - Normal operating status of the device; 302 - Collect sensor noise variance; 303 - Calculate environmental confidence score; 304 - Decision if confidence score is less than 3 times the standard deviation; 305 - Normal trusted path; 306 - Execute trust downgrade; 307 - Detect high-risk actions; 308 - Record operation cost; 309 - Accumulate and calculate risk entropy value; 310 - Decision if risk entropy reaches the mandatory audit threshold; 311 - Continue monitoring; 312 - Mandatory audit and block high-privilege actions. Figure 4 401 - Scheduler starts detection cycle; 402 - Generate active detection signal; 403 - Inject three types of detection signals; 404 - Collect ontology response data; 405 - Response type determination (diamond judgment box); 406 - Normal response maintenance strategy; 407 - No response marks suspicious; 408 - False response triggers ontology freeze; 409 - Ontology enters frozen state; 410 - Extract state snapshot at the time of freeze; 411 - Calculate cumulative risk entropy energy value; 412 - Combine snapshot with entropy energy value; 413 - Calculate bridging hash value; 414 - Unfreeze hash matching determination (diamond judgment box); 415 - Keep frozen; 416 - Allow unfreezing and resumption of operation. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] like Figure 1 As shown in the embodiment of the present invention, a method for verifying the semantic consistency of humanoid robot actions based on cognitive execution descriptors is provided. The core idea is as follows: a dynamic cognitive execution descriptor is introduced as a carrier of action intent, and the variance of the ontology sensor noise is mapped to an environmental confidence score to achieve adaptive decay of trust; a three-layer filtering structure is used in the motion semantic immune grid to perform legality verification, risk scoring, and risk entropy accumulation on the cognitive execution descriptor; when the risk entropy accumulation reaches a set threshold, an audit is forcibly executed and high-privilege actions are blocked; a detection signal is periodically injected through an active detection mechanism to detect hidden abnormal control; the recovery of the frozen state must be verified by the continuity of the bridging hash.

[0023] Combination Figure 1The cognitive execution descriptor capture and construction module (labeled 101) receives action execution requests from the task planning layer and constructs a dynamic cognitive execution descriptor model (labeled 102). This model contains four core fields: action intent attribute (labeled 103), representing the action intent declared by the planning layer; ontology capability attribute (labeled 104), representing the set of action permissions currently held by the ontology; execution context fingerprint (labeled 105), recording execution environment features such as instruction source, joint state, and sensor snapshots; and environment confidence score (labeled 106), a trust value dynamically calculated based on the ontology's sensor noise variance. The dynamic cognitive execution descriptor model (labeled 102) enters the motion semantic immune grid, sequentially passing through a semantic filtering layer (labeled 107) to determine the legality of the cognitive execution descriptor, through an execution risk layer (labeled 108) to perform risk scoring on the action instructions, and through a risk entropy layer (labeled 109) to accumulate and record high-risk actions. The execution path control unit (labeled 110) outputs control instructions for allowing, silently discarding, or delaying execution. Figure 1 As shown in markings 101 to 110.

[0024] Combination Figure 2 This section details the three-layer filtering structure of the motion semantic immune grid. After the dynamic cognitive execution descriptor (marked 201) enters the motion semantic immune grid, it first reaches the semantic filtering layer entry point (marked 202). The semantic filtering layer performs format integrity checks, field validity checks, and consistency cross-checks on the action intent attribute, ontology capability attribute, and execution context fingerprint. The validity check (marked 203) outputs a silent discard if any check fails (marked 204). If all checks pass, it enters the execution risk layer entry point (marked 205). The execution risk layer performs multi-dimensional risk scoring on the action command (marked 206), with risk scoring dimensions including... This includes the boundary compliance of action parameters, the jump depth of the instruction chain, and the deviation of historical action patterns; risk judgment (marked 207) enters the risk entropy layer entrance along the low-risk path (marked 208) when the risk score is lower than the risk threshold, and enters the risk entropy layer entrance along the high-risk path after weighting when the risk score reaches the risk threshold; the risk entropy layer records and accumulates the operation cost of high-risk actions (marked 209), and risk entropy judgment (marked 210) outputs release (marked 211) when the accumulated risk entropy value is lower than the delay threshold, and outputs delayed execution (marked 212) when the delay threshold is reached. For example... Figure 2 As shown by markings 201 to 212.

[0025] Combination Figure 3This section details the complete process of dynamic trust decay and mandatory risk entropy auditing. The system begins each execution cycle from its normal operating state (marked 301). First, it collects the noise variance of the on-board sensors (marked 302). Based on the current noise variance Σ_current and the baseline noise variance Σ_baseline, it calculates the environment confidence score C_env (marked 303), using the formula C_env = 1 - |Σ_current - Σ_baseline| / (|Σ_baseline| + ε), where ε is taken as 10 to the power of -6 to prevent division by zero. If C_env is lower than 3 times the noise standard deviation (marked 304), execution continues along the normal trust path (marked 305) if C_env is not lower than the threshold. If C_env is lower than the threshold, trust degradation is triggered (marked 306), and the caller's privilege level is reduced. The system continuously detects high-risk actions (marked 307). When a high-risk action is detected, the operation cost is recorded (marked 308), and the risk entropy value D_entropy is accumulated and calculated (marked 309), with the formula D_entropy = Σ Cost(i) - ∫ Noise dt. If the risk entropy reaches the mandatory audit threshold (marked 310), monitoring continues (marked 311) until the threshold is reached. If the threshold is reached, a mandatory memory audit is triggered, an immutable audit chain is generated, and high-privilege actions are blocked (marked 312), and the system enters a secure blocking state. Figure 3 As shown in markings 301 to 312.

[0026] Combination Figure 4This section details the complete process of active probing and bridging hash unfreezing and recovery. The scheduler initiates a probing cycle (marked 401) and generates an active probing signal (marked 402), which includes three types of probing: predictable disturbance signals, joint response test signals, and action loopback test signals (marked 403). The entity receives the probing signal and collects response data (marked 404). Response type determination (marked 405) categorizes the results into three types based on the response data: normal response paths (marked 406) maintain the current execution strategy; no-response paths (marked 407) mark the target as suspicious and add it to the key monitoring list; and pseudo-response paths (marked 408) immediately trigger entity freezing, and the entity enters a frozen state (marked 409). In the frozen state, the bridging hash unfreezing and recovery module performs the following: extracts a snapshot of the control state at the time of freezing (marked 410), calculates the cumulative risk entropy energy value (marked 411), and after splicing the snapshot and the risk entropy energy value (marked 412), calculates the bridging hash value using a one-way hash algorithm (marked 413). Upon receiving an unfreeze request, the current bridging hash value is recalculated and matched against the bridging hash value stored at the time of freezing. If the match is successful, the ontology is allowed to unfreeze and resume operation (marked 416); if the match fails, unfreezing is rejected and the ontology remains frozen (marked 415), requiring manual intervention for security auditing and recovery processing. Figure 4 The unfreezing hash matching check shown in mark 414 is referenced in this hash verification process. For example... Figure 4 As shown in markings 401 to 416.

[0027] Combination Figure 2 The execution path control unit and shadow execution mechanism are further improved by introducing a shadow execution budget control mechanism: the execution budget is allocated into three levels according to the priority level of the cognitive execution descriptor. The high priority is allocated a full shadow execution budget for full execution pre-verification in an independent shadow space. The medium priority is allocated a partial shadow execution budget for shadow execution verification only for high-risk action steps. The low priority is allocated a no-shadow execution budget and is directly released after the semantic filtering layer verification is passed.

[0028] Combination Figure 4 The active detection process also introduces a Gaussian kernel smoothing time topology stretching mechanism: with the token generation clock signal as the reference axis, the micro-hardware performance statistics signal is cascaded and integrated through kernel sampling in each token cycle, and the integral value is smoothed through the Gaussian kernel function, so as to nonlinearly map the micro-time scale of the hardware level to the macro-time axis of the semantic level; only when the smoothed behavior density feature peak and the monotonically increasing slope of the semantic risk entropy collide concurrently in the same token generation clock time window is the transaction circuit breaker of the action control core triggered.

[0029] The effectiveness of the present invention has been confirmed through the verification of this embodiment: the dynamic trust decay mechanism effectively reduces the number of falsely allowed high-privilege actions when the environmental noise variance increases; the risk entropy accumulation mechanism effectively detects hidden abnormal control sequences executed in a decentralized manner; the active detection mechanism maintains a high detection rate in hidden abnormal control scenarios; and the bridging hash unfreezing and recovery mechanism can correctly reject pseudo-recovery attempts.

Claims

1. A method for verifying the semantic consistency of humanoid robot actions based on cognitive execution descriptors, characterized in that, include: A dynamic cognitive execution descriptor model with an environment confidence score is constructed. The cognitive execution descriptor includes action intent attributes, ontology capability attributes, execution context fingerprint, and environment confidence score. The environment confidence score is dynamically calculated based on the ontology sensor noise variance according to C_env = 1 - |Σ_current - Σ_baseline| / (|Σ_baseline| + ε), where ε is a positive constant. When C_env is lower than 3 times the noise standard deviation, the trust factor is dynamically downgraded and the caller's permission level is reduced. The dynamic cognitive execution descriptor model is input into the motion semantic immune grid. It is then sequentially processed through a semantic filtering layer to verify the format integrity, field validity, and cross-consistency of the cognitive execution descriptor; an execution risk layer to score the risk of the action commands based on the compliance of action parameter boundaries, instruction chain jump depth, and deviation from historical action patterns; and a risk entropy layer to accumulate and record high-risk actions according to D_entropy = Σ Cost(i) - ∫ Noise dt, where Σ Cost(i) is the sum of costs of all recorded high-risk actions and ∫ Noise dt is the integral term of the body sensor noise over time. Based on the validity verification result, risk score result, and risk entropy accumulation status, corresponding execution path control commands are output. Specifically, a silent discard command is output when the validity verification result is illegal; a delayed execution command is output when the validity verification result is legal and the risk entropy accumulation value reaches a delay threshold (where the delay threshold is the value when the risk entropy accumulation value exceeds 80% of the baseline risk entropy value); and a release command is output when the validity verification result is legal, the risk score is lower than the risk threshold, and the risk entropy accumulation value is lower than the delay threshold. When the accumulated risk entropy value reaches the mandatory audit threshold, a memory audit is triggered to force disk write and generate an immutable audit chain bound to the current ontology state. At the same time, high-privilege actions are blocked. The mandatory audit threshold is the value when the accumulated risk entropy value exceeds 150% of the baseline risk entropy value. Active probe signals, including predictable disturbance signals, joint response test signals, and motion loop test signals, are periodically injected into the motion control core by the scheduler. The predictable disturbance signals have a deterministic sequence of disturbance parameters. The joint response test signals are used to trigger joint motion response events and detect the corresponding response delays. The motion loop test signals are used to construct self-returning motion commands and detect the integrity of the command chain. The system receives the response of the entity to the active detection signal and determines whether the entity is in a hidden abnormal control state based on the response type. When the response is a normal response, the current execution strategy is maintained. When the response is no response, it is marked as a suspicious state. When the entity generates an output response after receiving the active detection signal, but the deviation between the output response and the expected response characterized by the deterministic perturbation parameter sequence exceeds the tolerable deviation range, it is determined to be a false response and the entity is frozen. After the ontology enters a frozen state, a bridging hash value is generated. The bridging hash value is obtained by performing a one-way hash calculation on the concatenated data of the ontology's control state snapshot at the time of freezing and the accumulated risk entropy energy value before the time of freezing. When a defreezing request is received, ontology state verification is performed and the verification result is matched with the bridging hash value. The ontology is allowed to be defreezed only if the verification is successful and the match is successful; otherwise, defreezing is prohibited. Furthermore, a shadow execution budget control mechanism is introduced into the execution path of the motion semantic immune grid. The execution budget is allocated to three levels—full shadow execution, partial shadow execution, and no shadow execution—based on the priority level of the cognitive execution descriptor. High-priority cognitive execution descriptors are allocated full shadow execution budgets to perform complete execution verification in an independent shadow space. Medium-priority cognitive execution descriptors are allocated partial shadow execution budgets to perform shadow execution verification only for high-risk action steps. Low-priority cognitive execution descriptors are allocated no shadow execution budgets to allow direct passage.

2. The method according to claim 1, characterized in that, The active detection signal judgment process introduces a Gaussian kernel smoothing time topology stretching mechanism. Taking the token generation clock signal as the reference axis, the kernel samples and performs cascaded integration of the micro-hardware performance statistics signal within the token cycle, nonlinearly mapping the micro-time scale of the hardware level to the macro-time axis of the semantic level. Only when the smoothed behavior density feature peak and the monotonically increasing slope of the semantic risk entropy collide concurrently within the same token generation clock time window is the transaction circuit breaker of the action control core triggered.

3. A semantic consistency verification system for humanoid robot actions based on cognitive execution descriptors, characterized in that, include: The dynamic cognitive execution descriptor construction module is used to construct a dynamic cognitive execution descriptor model with an environmental confidence score. The cognitive execution descriptor includes action intention attributes, ontology ability attributes, execution context fingerprint, and environmental confidence score. The environmental confidence score is dynamically calculated based on the ontology sensor noise variance. The motion semantic immune grid module is used to sequentially verify the legality of the cognitive execution descriptor model by passing it through the semantic filtering layer, score the risk of the action instruction by passing it through the execution risk layer, and accumulate and record high-risk actions by passing it through the risk entropy layer. Based on the legality verification result, risk score result and risk entropy accumulation status, the corresponding execution path control instruction is output. The risk entropy mandatory audit module is used to trigger a mandatory memory audit and generate an immutable audit chain bound to the current ontology state when the accumulated risk entropy value is calculated according to D_entropy = Σ Cost(i) - ∫ Noise dt and reaches the mandatory audit threshold. At the same time, it blocks high-privilege actions. Here, Σ Cost(i) is the sum of the costs of all recorded high-risk actions, ∫ Noise dt is the integral term of the ontology sensor noise over time, the delay threshold is the value when the accumulated risk entropy value exceeds 80% of the baseline risk entropy value, and the mandatory audit threshold is the value when the accumulated risk entropy value exceeds 150% of the baseline risk entropy value. The active detection module is used to periodically inject active detection signals into the motion control core through the scheduler, receive the body's response to the active detection signals, and determine whether the body is in a hidden abnormal control state based on the response type. The bridging hash unfreezing and recovery module is used to generate a bridging hash value after the ontology enters a frozen state. When an unfreezing request is received, the ontology state is verified and the verification result is matched with the bridging hash value. The ontology is allowed to be unfrozen only if the verification is successful and the match is successful; otherwise, unfreezing is prohibited.

4. The system according to claim 3, characterized in that, The active detection signal judgment process introduces a Gaussian kernel smoothing time topology stretching mechanism. Taking the token generation clock signal as the reference axis, the kernel samples and performs cascaded integration of the micro-hardware performance statistics signal within the token cycle, nonlinearly mapping the micro-time scale of the hardware level to the macro-time axis of the semantic level. Only when the smoothed behavior density feature peak and the monotonically increasing slope of the semantic risk entropy collide concurrently within the same token generation clock time window is the transaction circuit breaker of the action control core triggered.

5. A humanoid robot, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method of any one of claims 1 to 2.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 2.

Images