A risk chain evaluation method and related apparatus

Abstract

The application belongs to the technical field of artificial intelligence safety evaluation, and aims at the technical problems that the existing risk scene construction lacks real accident mechanism constraints, the evaluation index is difficult to describe the risk chain and the consequence severity, the evaluation process relies on manual work and has low automation degree, and the evaluation result lacks traceable structured records, so it is difficult to form a reproducible and scalable safety evaluation system. A risk chain evaluation method and related device are provided, which constructs a structured risk scene specification based on a risk memory bank, generates and verifies a first frame of image before an accident, and then generates a future prediction video sequence and performs risk chain scoring. The method can improve the authenticity of risk scene construction and the traceability of evaluation results, enhance the description ability of risk causal chain and consequence severity, and improve the evaluation automation degree.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence safety assessment technology, and relates to a risk chain assessment method and related devices. Background Technology

[0002] With the development of embodied intelligence and autonomous robot operation, the training and evaluation of intelligent agents increasingly rely on controllable and large-scale interactive data. Video-generative world models, as neural simulators capable of synthesizing future trajectories based on initial observations and instructions, are gradually being used for tasks such as data augmentation, strategy verification, and security risk simulation. In real-world scenarios, acquiring data for truly high-risk scenarios is costly and risky. Traditional simulations suffer from significant discrepancies between simulation and reality, while the consequences of risks exhibit long-tail characteristics and time-dependent behavior. If the video-generative world model deviates from the triggering conditions and consequence simulations, it may lead to erroneous risk perceptions and affect the security of real-world deployments.

[0003] Existing technologies typically collect data through real-world experiments or construct training and evaluation samples using traditional simulation platforms. They then employ video-generative world models to generate visualized future trajectories, and evaluate these models based on metrics such as generation quality, physical consistency, and temporal coherence. In safety-related research, dangerous scenarios are often artificially constructed, and prediction videos are viewed manually. Initial conditions, triggering actions, and the severity of consequences are labeled, and scores are then tallied using voting or averaging methods to assess the performance of the video-generative world model in risky situations.

[0004] However, artificially constructed risk scenarios lack constraints from real-world accident mechanisms, easily generating unrealistic causal chains and failing to cover long-tail accidents with low probability but high consequences. Furthermore, evaluation metrics tend to focus on the plausibility of visuals and motion, making it difficult to characterize the model's ability to extrapolate risk chains and the severity of consequences. Additionally, the evaluation process heavily relies on manual interpretation and aggregation, resulting in low efficiency and poor consistency, hindering rapid model iteration. The evaluation results also lack structured records of triggering moments, key evidence, and judgment criteria, making it difficult to establish a traceable, reusable, and sustainably scalable safety evaluation system. Summary of the Invention

[0005] This application addresses the technical problems of existing risk scenario construction lacking constraints from real accident mechanisms, evaluation indicators being difficult to characterize risk chains and the severity of consequences, evaluation processes relying on manual labor with low automation, and evaluation results lacking traceable and structured records, making it difficult to form a reproducible and scalable safety evaluation system. It provides a risk chain assessment method and related devices.

[0006] To achieve the above objectives, this application adopts the following technical solution: Part One, this application provides a risk chain assessment method, including: Obtain the evaluation specifications; the evaluation specifications include the target scenario and the capability boundaries of the embodied intelligent agent. The assessment specifications are retrieved based on the risk memory bank to obtain an evidence set, and a structured risk scenario specification is generated under the constraints of the evidence set. Based on the structured risk scenario specifications, a pre-accident first frame image is generated, and the pre-accident first frame image is subjected to consistency verification. The first frame of the pre-accident image that has passed the consistency check is input into the video generative world model to generate a future predicted video sequence. Based on the predicted future video sequence, a risk chain scoring is performed to obtain an initial condition score, a triggering process score, and a consequence severity score, and the evidence fragments corresponding to each score are output. Based on the logical relationship between the initial condition score, the triggering process score, and the consequence severity score, and combined with the evidence fragments corresponding to each score, the final assessment result is determined.

[0007] Furthermore, the method for constructing the risk memory bank includes: The risk incident data is denoised and structurally extracted to form multiple risk memory units; the risk incident data includes real incident reports, safety guidelines, and warning clauses; the risk memory units include an incident mechanism description, physical consequences, prevention or mitigation measures, and source citations; A risk memory bank is constructed by using multiple risk memory units.

[0008] Furthermore, the method for retrieving the assessment specifications based on the risk memory bank includes: For the assessment specifications, a retrieval query is constructed based on the obtained security prompts and the scenario categories corresponding to the target scenario, and the retrieval query is used to retrieve candidate evidence in the risk memory bank; during the retrieval, key actions, key objects and risk domain words are weighted and matched to determine the relevance score of candidate evidence, and then the top-k candidate evidences are obtained according to the relevance scores of candidate evidences to form an evidence set.

[0009] Furthermore, the method for generating structured risk scenario specifications under the constraints of the evidence set includes: Generate candidate structured risk scenario specifications that include object sets, object attributes, spatial topological constraint sets, embodied intelligent agent capability boundaries, executable instruction sequences, and risk interpretations; The hazards, triggering actions, and consequences in the candidate structured risk scenario specifications are converted into observable visual cues and written into the candidate structured risk scenario specifications. The specifications of candidate structured risk scenarios are stored in the form of structured files; Record evidence consistency index, infeasibility ratio index and non-redundancy coverage index as quality measures of candidate structured risk scenario specifications to obtain the final structured risk scenario specifications.

[0010] Furthermore, the method for generating the first frame image of a pre-accident based on the structured risk scenario specifications includes: Image generation prompts are obtained based on the structured risk scenario specifications; wherein, the names of similar objects in the structured risk scenario specifications carry a unique identification number. Based on the image, a prompt word is generated to generate a pre-accident first frame image; the pre-accident first frame image is used to characterize the initial scene state before the accident occurs.

[0011] Furthermore, the method for performing consistency verification on the first frame image of the pre-accident event includes: The first frame image of the pre-accident is subjected to consistency verification of safety-critical physical attributes, spatial topology, and pre-accident state constraints; wherein, the pre-accident state constraint verification is used to determine whether there is a visual representation in the image that the accident has occurred.

[0012] Furthermore, a method for generating future predicted video sequences by inputting the pre-accident first frame image, which has passed consistency verification, into a video generative world model includes: The first frame image of the pre-accident is used as the initial state, and the conditional text derived from the structured risk scenario specification is used as the control signal. These are input into the video generative world model to generate multiple candidate videos corresponding to different random seeds or sampling parameters. Multiple candidate videos are validated, and the candidate videos that pass the validation are combined to form a future predicted video sequence.

[0013] Furthermore, the risk chain scoring method includes: According to the preset default robust strategy, overlay sampling is performed at fixed time intervals along the time axis of the future predicted video sequence to obtain visual segments, and the timestamp index corresponding to each visual segment is retained; wherein, during overlay sampling, encrypted sampling is used at each stage of the future predicted video sequence. All visual segments and their corresponding timestamp indices are input into the multimodal large model evaluator. Combined with the scoring rules constructed based on the structured risk scenario specifications, three independent scoring processes are performed on the initial condition score, the triggering process score, and the consequence severity score. The initial condition score, triggering process score, and consequence severity score are determined based on three independent scores through a voting mechanism, and evidence fragments corresponding to each score are output.

[0014] Part Two, this application discloses a risk chain assessment system, comprising: The data module is used to acquire evaluation specifications; the evaluation specifications include the target scenario and the capability boundaries of the embodied intelligent agent. The retrieval module is used to retrieve the assessment specifications based on the risk memory bank, obtain the evidence set, and generate structured risk scenario specifications under the constraints of the evidence set; The first frame image module is used to generate a pre-accident first frame image based on the structured risk scenario specifications, and to perform consistency verification on the pre-accident first frame image. The sequence generation module is used to input the first frame image of the pre-accident video that has passed the consistency check into the video generative world model to generate future predicted video sequences. The scoring module is used to perform risk chain scoring based on the future predicted video sequence to obtain an initial condition score, a triggering process score, and a consequence severity score, and output evidence fragments corresponding to each score. The evaluation module is used to determine the final evaluation result based on the logical relationship between the initial condition score, the triggering process score, and the consequence severity score, combined with the evidence fragments corresponding to each score.

[0015] Part Three, this application discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the risk chain assessment method described above.

[0016] Compared with the prior art, this application has the following beneficial effects: This application proposes a risk chain assessment method. By acquiring assessment specifications and retrieving these specifications from a risk memory bank, a structured risk scenario specification is generated under the constraint of an evidence set. This ensures that the construction of risk scenarios no longer relies solely on abstract settings or subjective splicing, but is based on risk evidence provided by the risk memory bank. This improves the authenticity, controllability, and reproducibility of risk scenario construction, and enhances the consistency and comparability between different assessment batches. Furthermore, this application generates a pre-accident first-frame image based on the structured risk scenario specification and performs consistency verification on the pre-accident first-frame image. Only the pre-accident first-frame image that passes the consistency verification is input into a video-generating world model to generate a future prediction video sequence. This allows for visual anchoring and consistency confirmation of the initial scenario state of the risk chain before video generation, reducing the impact of missing, replaced, or offset key hazard elements in the initial state on subsequent risk extrapolation. This improves the consistency between the future prediction video sequence and the structured risk scenario specification, and enhances the reliability of subsequent risk chain scoring. Meanwhile, this application performs risk chain scoring based on future predicted video sequences, obtaining initial condition scores, triggering process scores, and consequence severity scores. This moves risk assessment beyond a simple judgment of generation quality, instead conducting a chain-like analysis around initial conditions, triggering processes, and the severity of accident consequences. This allows for a more complete characterization of the risk causal chain and differentiates and portrays the severity of accident consequences, addressing the difficulty of existing evaluation indicators in characterizing risk chains and consequence severity, and improving the applicability of evaluation results to safety-critical scenarios. Furthermore, this application outputs evidence fragments corresponding to each score along with the initial condition score, triggering process score, and consequence severity score. Based on the logical relationship between these scores and the evidence fragments corresponding to each score, the final evaluation result is determined. This ensures that the final evaluation result not only has scoring results but also corresponding evidence support and logical connections, thereby improving the interpretability, traceability, and auditability of the evaluation results. This facilitates subsequent review, analysis, and playback verification, addressing the problem of the lack of traceable structured records in existing technologies. Therefore, this application forms a structured assessment path from risk scenario construction to risk result determination through a complete process, including assessment specification acquisition, risk memory bank retrieval, structured risk scenario specification generation, pre-accident first frame image generation and consistency verification, future prediction video sequence generation, risk chain scoring, and determination of the final assessment result by combining evidence fragments. This helps to reduce the reliance on manual sample-by-sample judgment in the assessment process, improve the degree of automation of the assessment, and support the formation of a reproducible and scalable security evaluation system.

[0017] This application also proposes a risk chain assessment system and an electronic device that possess all the advantages of the aforementioned risk chain assessment methods. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the risk chain assessment method used in this application. Figure 2 This is a schematic diagram of another process for the risk chain assessment method of this application; Figure 3 This is a system block diagram illustrating the risk chain assessment method implemented in this application. Figure 4 This is a schematic diagram illustrating the logic of three self-consistent aggregations of VLM chain scoring results, causal consistency chain break verification, and manual review queue triggering in the embodiments of this application. Figure 5 This is a schematic diagram of the risk chain assessment system of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] With the development of embodied intelligence and autonomous robotic operation technologies, the training, evaluation, and deployment of intelligent agents in complex environments increasingly rely on large-scale, diverse, and controllable interactive data. While collecting data in the real world can yield high-fidelity feedback, it is often constrained by practical factors such as high cost, long cycle time, and high barriers to entry in terms of venue and equipment. Furthermore, in high-risk scenarios such as high temperature, flammable and explosive materials, corrosive chemicals, and machining, directly obtaining failure cases or accident consequences through real experiments is not only extremely costly but may also pose unacceptable risks to personal safety and property. Although traditional simulation platforms possess controllability and repeatability, they are prone to discrepancies between simulation and reality in details such as long-tailed accident chains, material properties and state changes, occlusion, and multi-object interactions, and they struggle to quickly reproduce common triggering conditions and cascading consequences in real accidents. Based on this, in recent years, video-generative world models have gradually been regarded as a kind of neural simulator or data engine, capable of synthesizing future visualized trajectories given initial observations and actions or instructions, to support model-based planning, training data augmentation, and policy robustness evaluation. However, safety-related capabilities exhibit significant long-tail characteristics and temporal dependencies. Many catastrophic consequences do not manifest immediately but rather accumulate or cascade over time due to the combined effects of triggering actions, environmental states, object attributes, and spatial relationships, ultimately resulting in observable destructive outcomes. Therefore, if a video-generative world model exhibits a systematic bias in its deduction of the relationship between triggering actions and dangerous consequences in risky scenarios, even if the generated footage is visually realistic and physically coherent, it may still instill incorrect risk priors in the agent during the planning and training phases. This could lead the agent to learn the erroneous rule that dangerous actions will not lead to serious consequences, thus creating potential safety hazards for real-world deployments. In recent years, mainstream evaluation systems for video-generative world models have focused on generation quality, physical consistency, and temporal coherence as core objectives. Evaluation designs often emphasize whether the generated results closely resemble the real world and whether the motion conforms to basic dynamics, but they do not systematically characterize the model's ability to identify and present the severity of consequences in dangerous situations. In practice, while models may perform adequately in terms of appearance and basic dynamics, they tend to underestimate the severity of consequences, overlook key risk clues, or downplay dangerous triggering processes as harmless outcomes in safety scenarios. This creates visually plausible but safety-distorted assessment blind spots. In other words, even if existing metrics can reflect improvements in generation quality, they struggle to answer the core safety questions of whether the model truly understands the risk chain and accurately presents observable catastrophic consequences. Furthermore, existing safety-related evaluations often focus on safety constraints, rejection strategies, or multimodal harmful content at the agent behavior level. Their evaluation objects and targets are typically not the causal inference reliability of video-generated world models as future simulators.In such settings, even if it's possible to measure whether the system outputs safety warnings or avoids certain dangerous instructions, it's difficult to directly evaluate whether, when the model is used to generate future trajectories, it can correctly identify triggering actions from initial conditions and present matching consequence types and severity. Therefore, such evaluation results are difficult to directly transfer to support conclusions regarding the safety extrapolation capabilities of video-generated world models. Furthermore, the construction methods for risk scenarios and accident samples still lack stable guarantees of realism. Relying on imaginative generation from scratch or piecing together dangerous scenarios based solely on abstract descriptions can easily introduce illusory chains inconsistent with real risk mechanisms. For example, serious consequences may occur even when triggering conditions are not met, or multi-stage cascading accidents may be simplified into a single step. This results in insufficient sample interpretability, insufficient controllability, a narrow coverage of accident types, and difficulty in reproducing experiments. For long-tail accident chains with low probability and high consequences, this weakly mechanism-constrained sample generation significantly weakens the credibility of the assessment, making it difficult to support rigorous comparisons and regression tests between different model versions. More importantly, existing risk simulation and assessment processes generally rely heavily on manual annotation and subjective aggregation, lacking sufficient automation to support large-scale, rapid iterative assessment needs. In a typical process, multiple assessors must watch prediction videos sample by sample and independently judge the initial conditions, triggering actions, and severity of consequences, then obtain the final score through majority voting and mean aggregation. While this typical process is closer to the goal of focusing on key clues and incident chains in safety assessments, its high labor costs and long cycle make it difficult to form sustainable and scalable assessment capabilities when models are frequently updated or large-scale cross-model comparisons are required. Furthermore, assessment results often remain at the level of score output, lacking a closed-loop mechanism for the structured storage and feedback of risk chain structures, evidence fragments, and judgment reasons. Due to the lack of standardized evidence alignment, such as the alignment of trigger times, keyframe intervals, and observable consequence clues, and the lack of traceable records, assessments are difficult to distill into reusable assets, hindering subsequent error analysis, targeted data supplementation, and automatic regression verification, thus limiting the long-term scalability of safety assessment benchmarks and knowledge bases. The aforementioned issues collectively result in the fact that existing technologies cannot yet provide a highly automated, reproducible, and closed-loop evolution assessment framework for the reliability of video-generated world models in safety-critical scenarios.

[0023] In view of the above situation, this application proposes a risk chain assessment method and related apparatus, which will be further described in detail below with reference to embodiments and accompanying drawings: like Figure 1 The diagram shown is a schematic representation of one possible risk chain assessment method for this application, which may include: S101, Obtain the evaluation specifications.

[0024] It should be noted that the evaluation specifications are used to define the basic boundary conditions of the instance to be evaluated, including the target scenario and the capability boundaries of the embodied agent. The target scenario characterizes the type of environment targeted by the risk assessment, such as a home, laboratory, factory, or other scenario with operational objects, environmental constraints, and risk source distribution. The capability boundaries of the embodied agent characterize the range of actions, operational capabilities, mobility capabilities, or perception and execution limitations that the embodied agent participating in the assessment can perform under the current task. By obtaining the evaluation specifications first, the evaluation scenario, the agent being evaluated, and what the agent can and cannot do can be clearly defined before the assessment begins, thus providing a unified input boundary for subsequent risk evidence retrieval and scenario construction. This avoids subsequent assessments being divorced from the specific application environment, reduces scenario distortion caused by unbounded generation, and ensures consistent starting conditions among different test samples, which is conducive to forming a reproducible and comparable evaluation basis.

[0025] In this context, an embodied agent is an intelligent agent capable of interacting with its environment and performing actions in a physical form. The capability boundary of an embodied agent is the operational boundary of that agent under the current evaluation settings, rather than an arbitrary set of actions. By introducing capability boundaries, risky extrapolation results that are inconsistent with the actual capabilities of the agent can be avoided during the evaluation process.

[0026] S102, the assessment specifications are retrieved based on the risk memory bank to obtain an evidence set, and a structured risk scenario specification is generated under the constraints of the evidence set.

[0027] It should be noted that a risk memory bank is a storage medium for a collection of evidence established for knowledge of risk incidents. The evidence collection refers to a set of risk evidence retrieved from the risk memory bank and relevant to the current assessment specifications. A structured risk scenario specification refers to a scenario description document or scenario record that clarifies and structures the risk objects, relationships, actions, and consequences.

[0028] Specifically, the risk memory bank stores historical knowledge, mechanistic descriptions, consequences, or preventative measures related to risk incidents, providing factual constraints for subsequent scenario construction. After the evaluation specifications are input, a set of evidence matching the current target scenario and the capabilities of the embodied intelligent agent is retrieved. Subsequently, under the constraints of this evidence set, a structured risk scenario specification is generated. This specification transforms the abstract risk chain into a structured expression that can subsequently generate images, videos, and executable scores.

[0029] The assessment specification provides constraints on what to retrieve, while the risk memory bank provides the factual basis; together, they form a structured risk scenario specification. This structured risk scenario specification can be directly used for subsequent visual generation and video prediction of structured scenario descriptions. It transforms the process of constructing dangerous scenarios—which were originally difficult to reproduce and easily relied on subjective imagination—into an evidence-constrained structured generation process, thereby improving scenario realism and traceability. It reduces the chain of illusions that deviate from the mechanisms of real accidents, making the construction of risk samples more consistent with actual accident mechanisms and more conducive to inter-model comparison and regression testing.

[0030] S103, Generate the first frame image of the pre-accident based on the structured risk scenario specifications, and perform a consistency check on the first frame image of the pre-accident.

[0031] It should be noted that the pre-accident first frame image represents the visual state before an accident occurs, where hazardous elements are visible but the consequences have not yet manifested. Consistency verification refers to the process of determining whether there is a corresponding relationship between the pre-accident first frame image and the structured risk scenario specifications.

[0032] Specifically, the pre-accident first frame image is used to characterize the initial scene state before the accident occurs. This pre-accident first frame image shows the hazardous elements, key objects, and environmental relationships necessary for the subsequent risk chain to form, but it does not yet show the destructive consequences of an accident already occurring. Subsequently, the generated pre-accident first frame image undergoes a consistency check to determine whether it is consistent with the structured risk scenario specifications.

[0033] The structured risk scenario specification provides the basis for generating the first frame image of the pre-accident scenario. This first frame image then transforms the structured description into a visual starting point. The consistency verification output is the first frame image of the pre-accident scenario that passes the verification. By fixing the starting condition of the risk chain to an observable and verifiable visual anchor point, a clear initial state is established for the generation of subsequent future predictive video sequences. This avoids the risk extrapolation from being broken due to inaccurate initial scenarios, missing hazard elements, or misaligned spatial relationships, thereby improving the credibility and stability of subsequent assessment results.

[0034] S104 takes the first frame image of the pre-accident event that has passed the consistency check and inputs it into the video generative world model to generate a future predicted video sequence.

[0035] It should be noted that a video generative world model is a model capable of predicting future visual states based on a given initial state. The predicted future video sequence is a continuous video output of the video generative world model that progresses over time.

[0036] Specifically, the video-generative world model is used to predict the future evolution of the environment based on initial visual observations and corresponding conditions, outputting a continuous sequence of video frames. By using the first frame image of the pre-accident event, after consistency verification, as the starting point for prediction, the future predicted video sequence can be gradually unfolded along the risk chain on the timeline. This application extends the static risk initiation conditions to a dynamic risk evolution process, providing a temporal basis for subsequent scoring of the triggering process and accident consequences. This allows the assessment to move beyond a single static image or human imagination, and instead directly judge the model's understanding and representation of the risk chain based on the future trajectory generated by the model.

[0037] S105, perform risk chain scoring based on the future prediction video sequence to obtain initial condition score, triggering process score and consequence severity score, and output evidence fragments corresponding to each score.

[0038] It should be noted that evidence fragments are traceable content that can support the corresponding scoring judgment. They can be video clips, keyframe information, or associated time indexes within the corresponding time period.

[0039] Specifically, the initial condition score reflects whether the video sequence maintained the key initial conditions required for the establishment of the risk chain in the initial stage; the triggering process score reflects whether a key triggering process that could lead to an accident occurred in the video sequence; and the consequence severity score reflects whether the accident consequences occurred and to what extent. Evidence fragments corresponding to each score are used to support the scoring conclusions, ensuring that the scoring results are not merely simple scores but also correlated with specific time periods or specific visual evidence.

[0040] The scoring further structures the risk chain in the video into three interrelated dimensions. It transforms the difficult-to-quantify question of whether the model understands risk into a segmented assessment of different stages of the risk chain. This allows for a more granular characterization of whether the model has problems maintaining initial conditions, extrapolating triggering processes, or presenting the severity of consequences, thereby improving the interpretability and auditability of the assessment results.

[0041] S106. Based on the logical relationship between the initial condition score, the triggering process score, and the consequence severity score, and combined with the evidence fragments corresponding to each score, determine the final evaluation result.

[0042] Because the initial conditions, triggering process, and severity of consequences are not isolated but rather interdependent and causally related, the final assessment result is not simply a mechanical summary of the scores. Instead, it requires a comprehensive judgment based on the logical relationships between the three. For example, if the preconditions are insufficient or the triggering process is invalid, the credibility of the severity of consequences score should be limited; conversely, when the three types of scores are logically consistent and the evidence corroborates each other, the credibility of the final assessment result can be improved.

[0043] This application avoids evaluation results remaining at the level of isolated local indicators, forming a complete risk chain judgment conclusion. It can more realistically reflect the overall risk projection reliability of video-generated world models in safety-critical scenarios, thereby solving the problems of existing technologies where evaluation indicators are difficult to characterize risk chains and results lack structured traceability.

[0044] This application transforms the risk assessment process from a subjective and difficult-to-reproduce approach into an assessment process with structured inputs, traceable evidence, and chain-like logical constraints, thereby improving the automation level, interpretability, and consistency of assessments across batches.

[0045] like Figure 2 The diagram shown illustrates another flowchart of the risk chain assessment method used in this application, which may include: S201, Risk Scenario Specification Generation.

[0046] Textual evidence is collected from credible sources such as real accident reports, safety guidelines, and warning clauses. Unstructured text is then denoised and structured extracted to obtain structured risk memory units. :

[0047] in, This describes the accident mechanism. Indicates physical consequences. Indicates preventive or mitigation measures. Indicates source citation.

[0048] A searchable risk memory bank is constructed based on multiple risk memory units to provide traceable factual constraints and causal chain evidence for the construction of subsequent risk scenarios. Risk memories are used to represent traceable safety evidence from real accident reports, safety guidelines, or warning clauses, and are stored in the risk memory bank as structured units, serving as factual anchors for the construction and constraint generation of subsequent risk scenarios.

[0049] S202, Scenario specification generation based on risk memory (structured scenario file).

[0050] Evaluation specifications that include the target scenario and the capability boundaries of the embodied intelligent agent Vector retrieval is performed on the risk memory to obtain the Top-k relevant evidence set. .in, Indicates the target scenario. This represents the boundary of the capabilities of an embodied intelligent agent.

[0051] The target scenario refers to the structured description obtained by mapping the risk mechanism to a specific scenario and agent constraints based on the retrieved risk memory, which is used to reduce freely generated physical illusions and ensure executability.

[0052] In the collection of evidence Under the constraints, generate structured risk scenario specifications. :

[0053] in, Represents a set of objects. Represents a set of topological constraints. This indicates the boundary of the capabilities of an embodied intelligent agent. Indicates a sequence of instructions. This indicates an explanation of the risk.

[0054] object set and topological constraint set Used to explicitly define the attributes and spatial relationships of safety-critical objects; instruction sequences Used to construct an executable chain of operations capable of triggering risks; risk interpretation. Used to characterize verifiable causal chains in the form of initial scenarios, risk triggers, and dangerous consequences.

[0055] The above method can output structured risk scenario specifications for subsequent image and video generation.

[0056] S203, Pre-accident first frame image generation and consistency verification.

[0057] Based on structured risk scenario specifications The visual fields in the model include object sets, object attributes, spatial topological constraint sets, embodied agent boundaries, executable instructions, and candidate structured risk scenario specifications for risk interpretation. Image-generated prompts are constructed, and a text-to-image model is used to generate the first frame image of the pre-accident scenario. .

[0058] To ensure that the first frame image of the pre-accident scenario is consistent with the specifications of the structured risk scenario, a binary pass / fail check is performed on the generated first frame image. The check can include the following three dimensions: consistency of safety-critical physical properties, consistency of spatial topology, and pre-accident state constraints. Among them, consistency of safety-critical physical properties includes material state, temperature-related visual cues, and fluid morphology; consistency of spatial topology is used to determine whether the generated image satisfies the relative relationship constraints defined by the topological constraint set T; and pre-accident state constraints are used to determine whether the image only presents the initial state before the accident occurs, and does not allow any visualization of accident consequences such as flames, smoke, explosions, sparks, damage, or splashes.

[0059] Only retain the first pre-accident image that passes the consistency check. Subsequently, the first frame image of the pre-accident event was displayed. , from instruction sequence Summary of command prompts i The risk assessment unit d consists of risk explanation e: d=( , i e) S204, Future Prediction Video Sequence Generation.

[0060] Risk assessment unit The generative world model, which takes the video as input, generates a sequence of predicted future videos. , This represents the generative world model of the video to be tested. The risk interpretation 'e' is not provided to the generative world model of the video to be tested; it is only used for subsequent scoring reference to avoid information leakage that could lead to evaluation distortion.

[0061] S205, three VLM (Vision-Language Model) subjective scores and evidence fragment outputs.

[0062] Predicting future video sequences A preset, robust strategy is used to sample several visual segments from the timeline and retain their timestamp indices. Then, the Gemini 3 Pro multimodal large model evaluator is invoked to perform three independent scoring operations on the same sample, outputting three sets of (init, trg, out) scoring results, and providing traceable evidence segments (timestamped) for each score. Here, init represents the initial condition score, a binary evaluation; trg represents the triggering process score, also a binary evaluation; and out represents the consequence severity score, a continuous score from 0 to 1, used to characterize the degree of matching between consequence type and severity.

[0063] The default robust strategy is a standardized robust scheme used when extracting evidence fragments or sampling keyframes from future predicted video sequences, such as extracting visual fragments at fixed time intervals, key stage windows, or other robust criteria and retaining timestamps.

[0064] The system performs robust temporal sampling on the predicted video sequences to obtain timestamped key segment sequences. It then uses a multimodal large-scale model, such as Gemini 3 Pro (an advanced AI model), as an evaluator to perform three independent subjective scores on the same input. Each score outputs a set of risk chain indicators and corresponding timestamp indices for the evidence segments. Based on the three evaluation results, a self-consistent aggregation mechanism is employed to perform majority voting on binary labels and mean fusion on consecutive scores. After aggregation, a causal consistency constraint is applied, forcibly setting the consequence score to zero when the initial conditions or triggering actions are not met, thus avoiding unreasonably high scores in cases of chain breakage. For samples with significant disagreements in multiple scores, outputs close to the threshold, or insufficient evidence to support a clear ruling, they are automatically submitted to a manual review queue, with corresponding timestamped evidence segments output simultaneously for rapid manual verification. This forms a human-machine collaborative closed loop, primarily based on automatic evaluation and supplemented by manual verification, reducing the risk of evidence omission due to different video lengths, frame rates, and generation noise.

[0065] S206, Self-consistent aggregation, causal consistency verification and manual review queue.

[0066] The three scoring results are aggregated according to the self-consistency principle: the final binary values ​​of init and trg are determined by majority voting, and the final continuous score is obtained by a mean aggregation of out. After aggregation, a causal consistency constraint is introduced for chain break verification: when init=0 or trg=0, out=0 is forcibly set; for samples with uncertain scores, significant discrepancies among the three scores, or close to the threshold, they are automatically added to the manual review queue, and the timestamp index of the corresponding evidence fragment is output for quick manual verification. The final output includes the risk chain scoring result (init, trg, out), as well as the aggregated comprehensive score and the set of traceable evidence fragments.

[0067] This application fully leverages the advantages of multimodal large-scale models and video-generative world models in understanding complex security scenarios and extrapolating temporal consequences. It proposes a VLM-based automatic risk chain assessment framework to achieve near-fully automated construction, generation, and quantitative assessment of initial hazard conditions, triggering actions, and the chain of accident consequences. The principle is as follows: First, reusable risk knowledge is extracted from accident cases, safety guidelines, and warning clauses based on a risk memory library. A structured risk scenario specification document is then generated under the constraints of the target scenario and the capabilities of the embodied intelligent agent. This document clarifies key objects in the scenario, their safety-critical attributes, spatial topological relationships, and the sequence of instructions that trigger accidents and their expected consequences. Further, a pre-accident first frame image is generated based on the risk scenario specification document. Consistency checks ensure that key hazardous objects and key pre-conditions are visually visible and consistent with the risk scenario specification document, thus solidifying the starting conditions of the risk chain into observable visual anchor points. Subsequently, using the first frame image of the pre-incident scenario as the initial state for generating image-to-video, and constraining the temporal evolution of the video with conditional prompts derived from the risk scenario specification file, a future predicted video sequence is generated. This future predicted video sequence includes the key stages corresponding to the risk chain on the timeline, namely the danger existence stage before the accident, the triggering action execution stage, and the accident consequence presentation stage, thus forming an assessment object that can be automatically scored. This framework is geared towards the future predicted video sequence generation paradigm based on image-to-video (I2V). By introducing risk memory-enhanced scenario specification synthesis, consistency anchoring of the first frame image of the pre-incident scenario, and a multi-round subjective scoring and evidence timestamp backtracking mechanism, it can stably evaluate the reliability of the video-generative world model in safety-critical extrapolation under large-scale, long-tail risk distribution and multi-scenario variant conditions. This application can be widely used in model evaluation and dataset construction in safety-critical environments such as embodied intelligence, industrial robots, laboratories, homes, and factories. By leveraging the closed-loop capabilities of automated generation and automated evaluation, it can reduce the cost of manual annotation while improving the traceability and consistency of risk assessment, thereby providing an objective basis for subsequent safety alignment, training data screening, and model iteration.

[0068] It should be noted that I2V is a generation method that uses a single or a small number of initial observation images as the initial state and combines them with text instructions / action descriptions to generate subsequent video sequences. In this application, I2V is used to generate predictive videos given a pre-accident first frame image and a risk triggering instruction, in order to evaluate the model's ability to present risk consequences.

[0069] The following is a more detailed embodiment of this application: In this embodiment, the overall process consists of risk scenario specification generation, scenario first frame generation and consistency verification, future prediction video sequence generation, VLM chain scoring based on timestamp evidence fragments, self-consistency aggregation and causal consistency chain break verification, and manual review as a fallback, forming a closed loop for scalable automatic assessment and traceable evidence storage. Figure 3 The diagram shown illustrates a system block diagram for implementing the risk chain assessment method of this application. The system may include a risk scenario generation module, a risk scenario image generation module, a video generation module, a multimodal risk assessment module, and a risk aggregation and manual review module. The modules communicate with each other using risk scenario specification documents, the first frame image of the pre-accident event, future prediction video sequences, and evidence fragments (including timestamp indexes) as key data carriers, thereby achieving a structured output of the risk causal chain (initial hazard, triggering event, accident consequences). The inputs, outputs, and key mathematical constructions of each module are detailed below.

[0070] (1) Risk scenario generation module.

[0071] The risk scenario generation module extracts textual evidence from trusted sources into searchable risk memory units and generates a structured risk scenario specification z under risk memory constraints. This module constructs the structured risk scenario specification file, which serves as input for the subsequent risk scenario image generation module to perform first-frame consistency anchoring, and as conditional input for the video generation module to generate future predicted video sequences. It also provides structured risk chain reference information for the subsequent multimodal large model evaluator.

[0072] The risk scenario generation module, based on the principle of risk memory enhancement, uses an accident knowledge base and a safety alert database as evidence sources. It retrieves a set of evidence related to the current safety alert and, when generating structured risk scenario specifications, meets the requirement of visual salience. This ensures that the hazards, triggering actions, and consequences in the risk chain are presented as observable visual signals in the video, such as flames, smoke, sparks, liquid splashes, container bulging, and debris splashing, thereby reducing assessment distortion caused by invisible risks. Specifically: 1) Input / output and data structure construction.

[0073] In some embodiments, the input to the risk scenario generation module may include: the target scenario. Safety tips (t), risk explanations (e), and source citations The target scenario s can be a home, laboratory, or factory, etc.; the risk explanation e can be used to explain the cause of the hazard and emphasize the visible consequences.

[0074] The output of the risk scenario generation module is one or more candidate structured risk scenario specifications. The data is stored in a structured file format. In practical applications, the structured file should include at least the following fields: scene, safety_tip, explanation, type_of_robot, and generated_scenes. The stored structured file is written to a specified directory to form a traceable JSON file.

[0075] 2) Risk memory retrieval and evidence set construction.

[0076] To enhance risk memory, the risk scenario generation module first constructs a retrieval query q=(s,t,e) based on the input safety prompt t, target scenario s, and risk explanation e, and then retrieves the evidence set C from the accident case database based on the retrieval query.

[0077] in, These are the first to the Kth pieces of evidence in evidence set C, in that order.

[0078] The accident case library can include fields such as accident details, consequences, prevention tips, and sources, which are used to provide factual anchors for the generation of structured risk scenario specifications, corresponding to events, consequences, and prevention.

[0079] In some embodiments of this application, during the retrieval, key actions, key objects, and risk domain terms such as electrical and chemical can be weighted and matched to determine the relevance score of candidate evidence. The candidate evidence with the highest relevance score is selected as the evidence set C.

[0080] 3) Generation of structured risk scenario specifications and visual salient constraints.

[0081] After obtaining the evidence set C, the risk scenario generation module calls the scenario specification generator G(·) to generate candidate structured risk scenario specifications. The scene specification generator receives inputs (s, t, e, ...). r ,C), output a structured scene description that meets the requirements for video observability.

[0082] The candidate structured risk scenario specifications should include at least the following: user instructions that induce the robot to perform dangerous actions, a list of operable objects, the relative positions of the objects, the visual attributes of the objects, and a description of the consequences of the accident. The visual attributes of the objects may include material, color, status, and visible temperature indicators; the description of the consequences of the accident is used to emphasize visible consequences such as flames, smoke, explosion debris, and sparks.

[0083] In some specific embodiments, the generated prompts specify that abstract risks must be converted into visual signals, and require the Danger field to describe only the accident type and basic visible consequences, in order to reduce the impact of invisible risks and non-visual descriptions on subsequent assessments.

[0084] 4) Quality metrics and objective function construction.

[0085] To facilitate automated screening and statistical analysis of generated candidate structured risk scenario specifications, the risk scenario generation module can introduce the Information Grounding Rate (IGR), Unfeasible Hypothesis Ratio (UHR), and Diversity Validity Score (DVS) as generation quality indicators. These indicators can be used for quality recording, threshold screening, or subsequent statistical analysis. Specifically: (a) Evidence consistency index.

[0086] For candidate structured risk scenario specifications With the evidence set C, defined by The set of extractable atomic facts is X( As an example, atomic facts may include the existence and attributes of a hazard source object, the location of the triggering action, and the visual signals corresponding to the consequences.

[0087] For each atomic fact x∈X( The degree to which an evidence is supported by the evidence set C is determined according to the scoring rules. The terms in the scoring function `score(x;C)∈{1,0.5,0}` represent support, partial support, and no support, respectively. Therefore, the evidence consistency index can be defined as:

[0088] in, This represents the consistency index of evidence given atomic fact x and evidence set C.

[0089] Evidence consistency index is used to characterize the degree of consistency between the candidate structured risk scenario specification and the evidence set, in order to reduce the illusion of risk setting that is out of sync with evidence.

[0090] (b) Infeasible proportion indicators.

[0091] Candidate structured risk scenarios that do not meet the boundary constraints of the target scenario and the embodied intelligent agent's capabilities are judged as infeasible samples. Infeasible and specific examples include unreachable robot skill sets, conflicting object and location configurations, and invisible risks.

[0092] For a sample set consisting of N candidate structured risk scenario specifications The infeasibility ratio index represents the proportion of infeasible samples, used to characterize the executability and filmability of the generated results. Among them, i Represents the first in the sample set i Serial number.

[0093] (c) Non-redundant coverage metrics.

[0094] To avoid the generated results being concentrated in a few templated scenarios, the generated samples can be clustered, and the proportion of unique clusters can be calculated:

[0095] in, Indicates the candidate structured risk scenario specification Non-redundant coverage metrics This represents the number of unique clusters formed after clustering the candidate structured risk scenario specifications. The non-redundant coverage metric is used to characterize the breadth and non-redundancy of the generated risk cases.

[0096] In some optional embodiments, a comprehensive quality objective function can also be constructed based on the evidence consistency index, the infeasibility ratio index, and the non-redundant coverage index, as an optional reference for screening or iteratively generating candidate structured risk scenario specifications. :

[0097] in, Indicates the candidate structured risk scenario specification The infeasibility ratio index, λ1, λ2, and λ3 are the weight coefficients of the corresponding items, which can be set based on experience or adjusted according to the statistical results of the target field.

[0098] 5) Execution process and file storage.

[0099] In some embodiments, the risk scenario generation module corresponds to the step of generating scenario text descriptions in the processing flow, that is, calling the scenario generation function for each input security prompt to obtain candidate structured risk scenario specifications and writing them into a JSON file.

[0100] In practical applications, the JSON file can also store runtime metadata such as scene category and robot type, so that subsequent modules can generate pre-accident first frame images and future prediction video sequences in batches based on the structured file.

[0101] (2) Risk scenario image generation module.

[0102] The risk scenario image generation module generates image prompts based on the visual fields in the structured risk scenario specification z, and uses a text-to-image model to generate the first frame image of the pre-accident scenario. The consistency checker then performs pass / fail checks on the consistency of safety-critical physical attributes, spatial topology, and pre-accident state constraints, and outputs the first frame image of the pre-accident event. And the verification result marker.

[0103] The risk scenario image generation module further transforms the structured risk scenario specifications output by the risk scenario generation module into a pre-accident first-frame image, and obtains a consistency verification mark. The pre-accident first-frame image serves as the visual initial condition for the generation of subsequent future predicted video sequences and provides a visual anchor point for the initial hazard conditions in subsequent risk chain scoring. The pre-accident first-frame image emphasizes the scenario state before the accident occurs, i.e., the embodied agent has not yet executed the hazard command, but the key objects, key attributes, and key spatial layouts required for the establishment of the risk chain are already present, while avoiding any visual representation of accident consequences, such as damage caused by flames, smoke, explosions, or splashes. The purpose of this setting is to fix the initial state of the risk chain as an observable and verifiable visual starting point, thereby reducing the drift of initial conditions in the subsequent video generation and scoring process. Specifically: 1) Input / output mapping with structured fields.

[0104] In some embodiments, the input to the risk scenario image generation module is a structured risk scenario specification. The structured risk scenario specification should include at least structured fields such as object set, object location description, object attribute description, embodied agent type, and embodied agent location to ensure that the generated image is consistent with the structured risk scenario specification.

[0105] To facilitate the differentiation of multiple objects of the same type, object names can include unique identifiers, such as "knife1" and "knife2". The purpose of setting unique identifiers is to ensure a clear correspondence between the spatial location and attribute descriptions of each object in the image generation prompt, thereby reducing the consistency deviation of the first frame image caused by confusion due to objects with the same name.

[0106] The output of the risk scene image generation module is the first frame image of the pre-accident event. In some embodiments, metadata bound to the first frame image of the pre-accident event may also be output simultaneously. The metadata may include information such as image generation prompts, save path, scene number, and scene name, so as to facilitate traceability in the subsequent future prediction video sequence generation and evaluation stages.

[0107] 2) Prompt word construction and first frame constraints for pre-accident events.

[0108] In some embodiments, the candidate structured risk scenario specifications are first defined. Mapping to generate prompts for images Image generation prompts are used to constrain the image generation model to generate only the initial state image before the accident occurs, and to ensure that the object placement, object attributes, embodied agent type, and embodied agent position are consistent with the structured risk scenario specifications, while ensuring that the image style is realistic, the image is free of text, and no accident consequences are shown.

[0109] The above mapping relationship can be expressed as:

[0110] in, This represents the image-generated prompt word construction function, used to specification candidate structured risk scenarios. The fields such as object, object location, object attributes, embodied agent, embodied agent location, and embodied agent attributes are combined into an executable image generation instruction, and a constraint statement of PRE-INCIDENT INITIAL FRAME (the initial frame before the accident) is explicitly added. This constraint statement is used to emphasize that the embodied agent is already at the target location but has not yet started to perform any actions; the existence of the hazard itself is allowed, but no visualization of the consequences of the accident is permitted.

[0111] For ease of explanation, the above constraint set will be denoted as .in, It can include: (i) Completeness constraint of object set This is used to specify that key objects must appear in the frame; (ii) Spatial layout constraints Used to define the relative positions between objects and the specifications of candidate structured risk scenarios. Consistent; (iii) Attribute salience constraint It is used to define the key attributes of an object, including material, state, and hazard attributes, which can be visually identified in the image; (iv) Consistency constraints of embodied agents This is used to ensure that the type and location of the embodied intelligent agent are correct.

[0112] Therefore, the goal of image-generated cue word construction can be understood as generating cue words that satisfy... The first frame image of the pre-accident event is used as a stable initial condition for generating subsequent future predictive video sequences. It should be noted that a stable initial condition here means that the image can accurately retain the key visual information required at the beginning of the risk chain, thus providing a consistent starting point for assessing the subsequent triggering process and the severity of consequences.

[0113] 3) Generation of the first frame image of the pre-accident event.

[0114] In some embodiments, an image generation model can be invoked. For example, diffusion-based generative models or multimodal generative large models generate prompts based on images. Generate candidate pre-accident first frame image :

[0115] in, Generate model parameters for the image. The first pre-accident image obtained after consistency verification can be denoted as... The first frame image of the pre-accident obtained after consistency verification. As the initial visual input to the video generation module, the generation of subsequent predicted video sequences can be represented as follows:

[0116] in, This represents a video-generated world model. This indicates the specifications of candidate structured risk scenarios. Exported action instructions / condition prompts.

[0117] By introducing a consistent anchoring of the first frame image of the pre-accident scenario before generating future predictive video sequences, chain assessment errors caused by the absence, substitution, or positional shift of key hazardous objects in the initial frame can be reduced, and the initial conditional scoring can have a clear and stable visual basis. This goes beyond simply generating an initial image; it provides a verifiable starting state consistent with the structured risk scenario specifications for subsequent video generation and risk chain scoring.

[0118] (3) Video generation module.

[0119] The video generation module generates a future prediction video sequence for evaluation based on the pre-accident first frame image output by the risk scenario image generation module. This sequence simulates the temporal behavior of the embodied agent under given instructions and the evolution of potential risk consequences. The future prediction video sequence output by the video generation module serves as the direct evaluation object for subsequent risk chain scoring. Therefore, during the control generation process, the video generation module needs to ensure that the initial hazardous conditions, triggering processes, and accident consequences in the risk chain have observable visual evidence on the video timeline, thus meeting the requirements of chain evaluation and timestamped evidence backtracking.

[0120] 1) Input / output and generated object definition.

[0121] In some specific embodiments, the input to the generative world model (WM) of the video under test is limited to the first frame image of the pre-accident event. With the candidate structured risk scenario specification Exported action instructions / condition prompts i 1. The risk interpretation e in the structured risk scenario specification is not used as input to the video generative world model WM to avoid information leakage from biasing the generation results.

[0122] Specifically, the inputs to the video generation module may include: (i) The first frame image of the pre-accident process that has passed the consistency check It satisfies the consistency constraint that key hazardous elements are visible before the accident occurs; (ii) Structured risk scenario specifications This includes at least the task instructions, descriptions of dangerous consequences, and information on key objects, object attributes, and spatial layout.

[0123] The output of the video generation module is a predicted video sequence for the future. , used to characterize and Under the common constraints, the video generative world model generates the future temporal trajectory. To facilitate subsequent sampling and evidence timestamp localization, the video generation module can also record metadata corresponding to the predicted future video sequence, including frame rate, total number of video frames, total duration, and generation parameters. The purpose of recording metadata here is to provide a timeline mapping basis for subsequent visual segment sampling and evidence segment localization.

[0124] 2) Conditional text construction and risk visualization to generate constraints.

[0125] To enable future predicted video sequences to be used for risk chain scoring, the video generation module, based on... Build . This can include action instructions that the embodied agent should execute, prompting the video-generative world model to present observable triggering processes and consequence signals in the video. This construction process can be represented as:

[0126] in, A conditional construction function is used to specify candidate structured risk scenarios. The task instructions, descriptions of dangerous consequences, and key objects and spatial layout points are combined into unified generation conditions.

[0127] In some preferred embodiments, It should include at least the following constraint statements: (a) Action constraints, which are used to define the key triggering actions that the embodied agent needs to perform. These constraints correspond to the scoring of the subsequent triggering process. (b) Stage constraints, which require that the video include three stages: initial conditions, triggering process, and consequence presentation, and that each stage has a identifiable time boundary.

[0128] The above set of constraints can be denoted as The purpose of setting the constraint set is to make the generated video discriminable, sampleable, and traceable for subsequent risk chain scoring. That is, the video should not only show the evolution of actions, but also show the sequential relationship between the stages of the risk chain.

[0129] 3) Video-generated world model invocation and future prediction video sequence generation.

[0130] In one specific embodiment, the video generation module calls the video generative world model WM(·) to obtain the pre-accident first frame image after consistency verification. As the initial state, and with As a control signal, a future predicted video sequence is generated. Its generation process can be represented as:

[0131] in, This represents a sequence of frames arranged chronologically, with the first frame f1 corresponding to the first frame of the pre-accident image. Maintaining or nearly maintaining visual consistency ensures that subsequent initial condition scoring has a clear reference. Emphasis is placed on first-frame consistency because subsequent scoring first requires determining whether the initial hazard conditions are met. If the starting frame of the video deviates significantly from the first frame of the pre-accident image, it will affect the stable judgment of the initial conditions. In some specific embodiments, to reduce the randomness of single video generation and improve the credibility of subsequent risk chain scoring, the video generation module generates multiple candidate videos corresponding to different random seeds or sampling parameters. ,Right now:

[0132] in, Representing different random seeds or sampling parameters, This represents the number of random seeds or sampling parameters. Subsequently, a validity check function Λ( is generated.) , The candidate videos are filtered for validity, and those that do not meet the basic constraints of the structured risk scenario specifications are removed. The multiple candidate videos that pass the validity check are combined to form the sample set Y* of the future predicted video sequence. Y* = { | Λ( , ) ≥ τ, j = 1, …, m} Where τ represents the basic constraints of the structured risk scenario specifications.

[0133] Used for subsequent multi-sample independent scoring and average aggregation:

[0134] In this approach, each candidate video in set Y is used as a sample of the future predicted video sequence and input into the subsequent multimodal risk assessment module. The subsequent multimodal risk assessment module independently performs risk chain scoring on each candidate video and averages and aggregates the scores of all candidate videos. This significantly reduces the randomness of video generation and improves the credibility of the assessment results without relying on any single generation result.

[0135] 4) Interface output with the subsequent multimodal risk assessment module.

[0136] The final output of the video generation module is a set of future predicted video sequence samples Y and its corresponding metadata set M. For each candidate video in Y, the metadata set includes at least records of frame rate, total number of frames, total duration, and conditional text. The subsequent risk chain scoring process will perform default robust strategy sampling on each candidate video in the future predicted video sequence sample set Y and output timestamped evidence segments. By averaging the scoring results of all candidate videos in Y, the randomness of a single generation is reduced, and the credibility of the evaluation results is improved. Therefore, the video generation module needs to ensure that the video timeline is accurately mapped to the timestamp t when outputting, i.e.:

[0137] Here, fps represents the frame rate.

[0138] This allows subsequent evidence fragments to be presented in a way that... Backtracking and locating in the form of [method / method]. and These represent the upper and lower limits of the timestamp interval on the video timeline. This timestamp mapping is used to establish the correspondence between video frame indices and actual time positions, ensuring that subsequent evidence fragments can be accurately retrieved and verified.

[0139] (4) Multimodal risk assessment module.

[0140] The multimodal risk assessment module first uses a default robust strategy to sample future prediction video sequences along the time axis to obtain a set of evidence segments E (including timestamp indices). Then, it performs three independent VLM scoring operations on the same input and outputs three sets of (init, trg, out) scoring results and corresponding evidence timestamps.

[0141] The multimodal risk assessment module performs chain-like risk assessments on the future predicted video sequences output by the video generation module, and outputs structured scoring results and traceable evidence fragments (with timestamps). The module uses a large visual language multimodal model, such as Gemini 3 Pro, as the primary evaluator, combined with a default robust video sampling scheme and Rubric scoring rules, to subjectively determine and quantitatively score the three key links in the risk causal chain: initial hazard conditions, triggering actions, and accident consequences. To reduce the randomness of a single subjective assessment, three independent assessments are performed on the same input video, outputting three sets of scoring results and evidence respectively. These results are then submitted to the subsequent risk aggregation and manual review module for self-consistency fusion and causal consistency verification. Specifically: 1) Definition of input, output and evaluation objects.

[0142] In some embodiments of this application, the inputs to the multimodal risk assessment module include: (i) Future predicted video sequences and their metadata.

[0143] (ii) Structured risk scenario specifications, used to construct Rubric scoring details and prompts; (iii) Default robust sampling parameters, such as sampling step size, maximum number of segments, etc.

[0144] The multimodal risk assessment module outputs a set of three independent assessment results. Each evaluation result All include:

[0145] in, , , , This is a collection of evidence fragments, output in the form of timestamp intervals, used to trace back and locate key video fragments that support the scoring. , , They represent the first The init, trg, and out values ​​obtained from each independent evaluation.

[0146] 2) Default robust sampling scheme and timestamp fragment construction.

[0147] To reduce video length and evaluation costs, and to ensure consistent comparability among different videos during evaluation, the input videos are first subjected to temporal sampling using a default robust scheme. The sampler is denoted as... Regarding the video Generate sampled fragment sequence :

[0148] in, This represents the k-th sample segment, which can be a keyframe or a short segment. Let be the timestamp interval of the k-th sample segment on the original video timeline. The mapping relationship between timestamps and frame indices is as follows:

[0149] In some embodiments of this application, the sampling strategy employs a fixed time interval coverage, such as sampling once every Δt seconds, and can perform encrypted sampling at the beginning, end, and end of the video segment to improve the probability of capturing the triggered actions and their consequences. The timestamp interval of the above sampling output will be directly used as a reference index for subsequent evidence segments to meet traceability requirements.

[0150] 3) Rubric scoring rules construction and prompt format constraints.

[0151] To ensure that VLM outputs meet the requirements of chained assessment, Rubric scoring rules are constructed based on the structured risk scenario specifications, and assessment prompts are generated. Rubric is used to clarify what constitutes a pass, fail, or partial pass, and requires the VLM to provide verifiable evidence (timestamps) along with its scoring conclusion. This construction process is represented as:

[0152] in, Build functions for Rubric to use from Extract key objects, key initial conditions, key triggering actions and expected consequences, and form evaluation criteria for init, trg, and out.

[0153] In some embodiments of this application, Rubric includes at least the following criterion definitions: (a) Initial condition score init(binary).

[0154] The `init` function is used to determine whether critical hazardous objects or conditions exist before an accident occurs. Preferably, it determines whether critical objects have disappeared before the accident, based on pre-defined criteria. Specifically, if critical hazardous objects are continuously visible in the video during the early stages and their state is consistent with the structured risk scenario specifications, `init` is set to 1; if critical hazardous objects are missing, improperly replaced, or disappear in a way that contradicts the specifications, `init` is set to 0.

[0155] (b) Triggering process score trg (binary).

[0156] `trg` is used to determine whether the key action that caused the accident was performed. If there is a clear clip on the video timeline showing that the robot performed the triggering action defined in the structured risk scenario specification, such as throwing a water-filled object into hot oil or putting metal into a microwave oven, then `trg` is set to 1; otherwise, `trg` is set to 0.

[0157] (c) Consequence severity score out (continuous, 0~1).

[0158] `out` is used to assess whether an accident consequence occurred, and the degree to which the consequence's manifestation matches the expected consequences in the structured risk scenario specification in terms of type and severity. This indicator is a continuous score, `out` ∈ [0, 1], where 0 indicates no consequence or no match at all, and 1 indicates a high degree of match between the consequence type and severity.

[0159] 4) Three independent assessments and structured outputs (including evidence timestamps).

[0160] To improve robustness, this module performs sampling on the same sampled segment sequence. Perform three independent VLM assessments. The output of this evaluation is defined as:

[0161] in, This refers to the random seed or cue perturbation parameter used to distinguish the three evaluations, so as to encourage VLM to form independent judgments in subjective evaluation and reduce randomness.

[0162] Collection of evidence Output in timestamp fragment format to support each score, preferably in the following structured form:

[0163] in, , For a brief description of the visible clues in this clip, such as the key object still being on the table on the left side of the screen, the robot performing a dropping action, and the presence of obvious flames / smoke / splashes, etc. This indicates the initial condition scoring evidence. This indicates evidence for the triggering process scoring. This indicates the severity of the consequences as evidence. By outputting the timestamps of the evidence, subsequent modules can directly replay and locate key segments, enabling the assessment conclusions to be auditable, verifiable, and traceable.

[0164] 5) Module output and subsequent aggregation interface.

[0165] The multimodal risk assessment module ultimately outputs three sets of assessment results. and the corresponding set of evidence This data is then used as input to subsequent risk aggregation and manual review modules for performing self-consistent aggregation, causal consistency checks, and triggering of the manual review queue. To ensure that downstream processing can directly perform statistical fusion, the output of the multimodal risk assessment module uses unified fields and a unified value range:

[0166] And ensure that each item has at least one corresponding piece of evidence or is clearly marked as missing evidence, so that downstream modules can make uncertainty judgments and trigger manual review.

[0167] (5) Risk aggregation and manual review module.

[0168] like Figure 4 The diagram shown is a logical schematic of the three self-consistent aggregations of VLM chain scoring results, the causal consistency chain break verification, and the triggering of the manual review queue in this embodiment.

[0169] First, self-consistent aggregation is performed on the three scoring results: majority voting is conducted on init and trg to obtain the final label, and the mean aggregation of out is performed to obtain the continuous score. The evidence timestamps are then fused and stored. Second, causal consistency chain break verification is performed. When init=0 or trg=0 is detected, out is forcibly set to zero. Finally, when there is significant disagreement in the scoring, the score is close to the threshold, or the uncertainty is high, a manual review queue is triggered, carrying evidence timestamps to support traceable review.

[0170] The risk aggregation and manual review module performs self-consistent aggregation of multiple scoring results output by the multimodal risk assessment module. Based on this, it applies causal consistency constraints and uncertainty screening to form a final output that can be directly used for assessment reports and traceable evidence. For scoring discrepancies or critical samples, they automatically enter the manual review queue with timestamped evidence fragments, forming a closed loop combining automatic assessment, interpretable evidence, and manual fallback.

[0171] 1) Input / output and data structures.

[0172] In some embodiments of this application, the input to the risk aggregation and manual review module is three independent evaluation results of the same future prediction video sequence sample. Each evaluation result includes a binary label. Continuous fractions and collection of evidence fragments Evidence fragments are represented using timestamp intervals, preferably... It can be mapped to the time of a video sampling segment or keyframe, and includes an evidence type marker and a brief description for subsequent playback and location.

[0173] The output of the risk aggregation and manual review module is the final result record. It must at least include the final initial condition rating label. Final Triggering Process Rating Tags Final Consequence Severity Rating Labels The aggregated set of evidence And whether it has entered the manual review stage. This record, serving as the smallest unit for subsequent statistical reports and traceable evidence, can be directly written into the results file or database.

[0174] 2) Self-consistent aggregation.

[0175] To reduce the randomness of subjective VLM scoring in a single test, the risk aggregation and manual review modules perform self-consistent aggregation on the three results. For the binary terms init and trg, a majority vote is used to obtain the final judgment.

[0176] in, It indicates voting.

[0177] For continuous terms Preliminary consequence scores were obtained by mean aggregation. :

[0178] The above aggregation method does not rely on complex statistical modeling, which facilitates engineering implementation and review and understanding, while significantly mitigating the impact of fluctuations in a single assessment on the final conclusion.

[0179] 3) Causal consistency constraint.

[0180] To ensure that the output conforms to the causal chain logic of initial hazard, triggering event, and accident consequences, the risk aggregation and manual review module applies causal consistency constraints to the accident consequence item after aggregation. This constraint applies if and only if... and Only when any of the preconditions are met is the consequence score allowed to be retained; if any precondition is not met, the causal chain is considered broken, and the consequence score is forcibly set to zero.

[0181] This constraint avoids illogical outputs that determine an accident result when there is no danger or trigger, thereby improving the interpretability and credibility of chained evaluations.

[0182] 4) Evidence timestamp fusion and traceable evidence preservation.

[0183] The risk aggregation and manual review module performs lightweight fusion of the evidence fragments returned from the three assessments, preferably retaining evidence fragments supported by at least two assessments to reduce noise and enhance verifiability. The fused evidence set... Storing evidence in groups by type. Each piece of evidence includes a timestamp range. The number of times the causal consistency is supported and the corresponding explanatory text. For samples that are ultimately set to zero due to causal consistency, i.e. Furthermore, even if a link break is triggered, the init / trg evidence can still be retained to explain why the link broke, but the out consequence evidence will no longer be output, or it will be marked as invalid consequence evidence. The fused evidence can also retain the corresponding frame index for subsequent manual review and audit playback.

[0184] 5) Uncertainty discrimination and manual review queue.

[0185] In practical implementation, the risk aggregation and manual review module executes as follows: It receives three independent outputs from the multimodal risk assessment module and processes them separately. Perform a majority vote with trg to obtain and Perform mean aggregation on out to obtain Subsequently, the severity score of the consequence was corrected for chain breaks based on the causal consistency constraint. And further calculation (Risk-Chain Causal Consistency Coefficient) and (Failure Scenario Triplet) to generate chained metrics that can be used for batch statistics; simultaneously, the three evidence fragments are aggregated according to rules supported by at least two to obtain corresponding results. And retain timestamps for evidence. Finally, generate based on the divergence signal and the threshold critical signal. If no manual review is required, output directly. If manual review is required, then... The aggregated evidence (including timestamps) is written into the manual review queue, where key evidence fragments are reviewed and adjudicated by humans, and the final result is written back, thereby achieving a combination of automated evaluation and manual backup.

[0186] This application uses enhanced risk memory as the factual anchor for scenario construction, extracting credible sources such as accident cases, safety guidelines, and warning clauses into structured risk memory units, and organizing and retrieving relevant risk evidence through a risk memory bank. Furthermore, it can construct retrieval queries based on safety prompts and the scenario categories corresponding to the target scenario, and obtain relevance scores for candidate evidence by weighted matching of key actions, key objects, and risk domain terms, thus forming an evidence set. Through this more specific risk memory bank construction and retrieval method, it can suppress the illusionary chains and unexplainable sample problems caused by imaginative splicing under weak mechanistic constraints from the source, further improving the authenticity, controllability, reproducibility, comparability across model versions and experimental batches, and evaluation credibility of risk samples.

[0187] Under the constraints of the evidence set, this application can further generate candidate structured risk scenario specifications, including object sets, object attributes, spatial topological constraint sets, embodied intelligent agent capability boundaries, executable instruction sequences, and risk interpretations. Furthermore, the hazard sources, triggering actions, and consequences are converted into observable visual cues and written into the structured risk scenario specifications, ensuring that initial hazards, triggering events, and accident consequences are all presented as observable visual cues. Simultaneously, the candidate structured risk scenario specifications can be stored in the form of structured files, and evidence consistency indicators, infeasibility ratio indicators, and non-redundant coverage indicators can be recorded as quality metrics for the candidate structured risk scenario specifications. Optionally, a quality objective function can be constructed based on the above indicators to screen or iteratively generate candidate samples. Through these more refined generation and screening mechanisms, the executability, photographic feasibility, non-redundant coverage, and visual assessability of risk scenarios can be further enhanced.

[0188] This application introduces a pre-accident first-frame consistency anchoring mechanism. Before video generation, initial observation images prior to the accident are generated, and consistency checks are performed on safety-critical objects, key attributes, and spatial topology. Only the pre-accident first-frame images that pass the check are retained for future prediction video sequence generation. Furthermore, image generation prompts can be constructed based on object location descriptions, object attribute descriptions, embodied agent types, and embodied agent locations in the structured risk scenario specifications. Unique identifiers are introduced into the names of similar objects to reduce object confusion. This more specific first-frame generation and consistency anchoring method further reduces drift during the future prediction video sequence generation process and minimizes subsequent chain breakage misjudgments caused by missing, replaced, or offset key hazard elements in the initial frame, making the starting conditions of the chain assessment more visually verifiable and traceable.

[0189] In terms of generating future predicted video sequences, this application can also adopt an implementation method based primarily on image-generated video. The first frame image of the pre-accident event is used as the initial state, and the conditional text derived from the structured risk scenario specifications is used as a control signal input to the video generative world model to generate future predicted video sequences. Furthermore, multiple candidate videos corresponding to different random seeds or sampling parameters can be generated, and all generated candidate videos can be used as samples for the future predicted video sequence input into subsequent risk chain scoring. By averaging and aggregating the scoring results of multiple samples, the randomness of a single video generation is reduced, and the credibility of the evaluation results is improved.

[0190] Regarding risk chain scoring, this application explicitly structures the risk assessment objective into a chain-like indicator system of initial hazardous conditions, triggering actions, and accident consequences, specifically corresponding to init, trg, and out. Out can use a continuous score of 0-1 to characterize the degree of matching between consequence type and severity, and requires the evaluator to output time-stamped evidence fragments, achieving integrated recording of scoring and evidence. Furthermore, a default robust sampling strategy can be adopted, performing coverage sampling at fixed time intervals along the timeline of the future predicted video sequence, and employing encrypted sampling at each stage of the video to extract key time-stamped fragments. Simultaneously, multimodal large models such as Gemini 3 Pro can be introduced, and chain-like subjective scoring can be performed according to Rubric. Through this more refined scoring implementation, the evaluation results no longer remain at a single score but form an auditable, interpretable, and replayable verifiable chain of evidence, which is beneficial for error analysis, targeted data supplementation, and automatic regression verification, thereby supporting the long-term accumulation and expansion of the evaluated assets.

[0191] This application can also employ three independent scoring methods and perform self-consistent aggregation, perform majority voting on binary labels, and perform mean fusion on continuous scores to reduce the impact of single subjective evaluation or randomness of model output on the conclusion. At the same time, a causal consistency constraint is applied after aggregation, that is, when the current cause is not valid, the consequence score is forced to be set to zero, thereby avoiding the problem of high scores with no cause and no consequence, and improving the robustness and stability of the evaluation output at the logical level.

[0192] Furthermore, this application can automatically submit samples with significant scoring discrepancies, borderline thresholds, or insufficient evidence to a manual review queue through uncertainty discrimination, and include timestamped evidence fragments for rapid verification, thereby achieving a human-machine collaborative mechanism with automatic evaluation as the primary method and manual review as a backup. This mechanism ensures overall automated throughput while concentrating human resources on a small number of high-risk or highly uncertain samples, further improving the reliability and engineering usability of the final output.

[0193] like Figure 5The diagram shown is a schematic representation of the risk chain assessment system of this application, which may include: The data module is used to acquire evaluation specifications; the evaluation specifications include the target scenario and the capability boundaries of the embodied intelligent agent. The retrieval module is used to retrieve the assessment specifications based on the risk memory bank, obtain the evidence set, and generate structured risk scenario specifications under the constraints of the evidence set; The first frame image module is used to generate a pre-accident first frame image based on the structured risk scenario specifications, and to perform consistency verification on the pre-accident first frame image. The sequence generation module is used to input the first frame image of the pre-accident video that has passed the consistency check into the video generative world model to generate future predicted video sequences. The scoring module is used to perform risk chain scoring based on the future predicted video sequence to obtain an initial condition score, a triggering process score, and a consequence severity score, and output evidence fragments corresponding to each score. The evaluation module is used to determine the final evaluation result based on the logical relationship between the initial condition score, the triggering process score, and the consequence severity score, combined with the evidence fragments corresponding to each score.

[0194] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of each block is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple blocks may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0195] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0196] This application also provides an electronic device, which may include one or more processors, memory and communication interfaces.

[0197] The memory, communication interface, and processor are coupled together. For example, the memory, communication interface, and processor can be coupled together via a bus.

[0198] The communication interface is used for data transmission with other devices. The memory stores computer program code. This computer program code includes computer instructions, which, when executed by the processor, cause the electronic device to perform the steps of the aforementioned risk chain assessment method.

[0199] The processor can be a processor or controller, such as a Central Processing Unit (CPU), a general-purpose processor, a Digital Signal Processor (DSP), an Application-Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with this disclosure. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The processor can be used to support an electronic device in performing the method steps provided in the above embodiments.

[0200] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. These buses can be categorized as address buses, data buses, control buses, etc.

[0201] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A risk chain assessment method, characterized in that, include: Obtain the evaluation specifications; the evaluation specifications include the target scenario and the capability boundaries of the embodied intelligent agent. The assessment specifications are retrieved based on the risk memory bank to obtain an evidence set, and a structured risk scenario specification is generated under the constraints of the evidence set. The method for constructing the risk memory bank includes: denoising and structurally extracting risk incident data to form multiple risk memory units; the risk incident data includes real incident reports, safety guidelines, and warning clauses; the risk memory unit includes an incident mechanism description, physical consequences, and source citations; and the risk memory bank is constructed through multiple risk memory units. The method for retrieving the assessment specification based on the risk memory bank includes: for the assessment specification, constructing a search query based on the obtained safety prompts and the scenario category corresponding to the target scenario, and using the search query to retrieve candidate evidence in the risk memory bank; during the search, performing weighted matching on key actions, key objects and risk domain words to determine the relevance score of the candidate evidence, and then obtaining the Top-k candidate evidence based on the relevance score of the candidate evidence to form an evidence set; A method for generating structured risk scenario specifications under the constraints of the evidence set includes: Generate candidate structured risk scenario specifications that include an object set, object attributes, a set of spatial topological constraints, embodied agent capability boundaries, executable instructions, and risk interpretations; convert the hazards, triggering actions, and consequences in the candidate structured risk scenario specifications into observable visual cues and write them into the candidate structured risk scenario specifications; store the candidate structured risk scenario specifications in the form of structured files; record evidence consistency indicators, infeasibility ratio indicators, and non-redundant coverage indicators as quality measures of the candidate structured risk scenario specifications to obtain the final structured risk scenario specifications; Based on the structured risk scenario specifications, a pre-accident first frame image is generated, and the pre-accident first frame image is subjected to consistency verification. The first frame image of the pre-accident, which has passed the consistency check, is input into the video generative world model to generate a future predicted video sequence: the first frame image of the pre-accident is used as the initial state, and the conditional text derived from the structured risk scenario specification is used as the control signal and input into the video generative world model to generate multiple candidate videos corresponding to different random seeds or sampling parameters; the validity of the multiple candidate videos is checked, and the multiple candidate videos that pass the validity check form a future predicted video sequence; Based on the predicted future video sequence, a risk chain scoring method is performed to obtain an initial condition score, a triggering process score, and a consequence severity score, and to output evidence fragments corresponding to each score. The risk chain scoring method includes: performing coverage sampling along the timeline of the predicted future video sequence at fixed time intervals according to a preset default robust strategy to obtain visual fragments, and retaining the timestamp index corresponding to each visual fragment; wherein, during coverage sampling, encrypted sampling is used at each stage of the predicted future video sequence; inputting all visual fragments and their corresponding timestamp indices into a multimodal large model evaluator, and combining the scoring rules constructed based on the structured risk scenario specifications, performing three independent scoring operations on the initial condition score, triggering process score, and consequence severity score; determining the initial condition score, triggering process score, and consequence severity score based on the three independent scores through a voting mechanism, and outputting evidence fragments corresponding to each score. Based on the logical relationship between the initial condition score, the triggering process score, and the consequence severity score, and combined with the evidence fragments corresponding to each score, the final assessment result is determined.

2. The risk chain assessment method according to claim 1, characterized in that, The method for generating the first frame image of a pre-accident scenario based on the structured risk scenario specifications includes: Image generation prompts are obtained based on the structured risk scenario specifications; wherein, the names of similar objects in the structured risk scenario specifications carry a unique identification number. Based on the image, a prompt word is generated to generate a pre-accident first frame image; the pre-accident first frame image is used to characterize the initial scene state before the accident occurs.

3. The risk chain assessment method according to claim 1, characterized in that, The method for performing consistency verification on the first frame image of the pre-accident event includes: The first frame image of the pre-accident is subjected to consistency verification of safety-critical physical attributes, spatial topology, and pre-accident state constraints; wherein, the pre-accident state constraint verification is used to determine whether there is a visual representation in the image that the accident has occurred.

4. A risk chain assessment system, characterized in that, include: The data module is used to acquire evaluation specifications; the evaluation specifications include the target scenario and the capability boundaries of the embodied intelligent agent. The retrieval module is used to retrieve the assessment specifications based on the risk memory bank, obtain the evidence set, and generate structured risk scenario specifications under the constraints of the evidence set; The method for constructing the risk memory bank includes: denoising and structurally extracting risk incident data to form multiple risk memory units; the risk incident data includes real incident reports, safety guidelines, and warning clauses; the risk memory unit includes an incident mechanism description, physical consequences, and source citations; and the risk memory bank is constructed through multiple risk memory units. The method for retrieving the assessment specification based on the risk memory bank includes: for the assessment specification, constructing a search query based on the obtained safety prompts and the scenario category corresponding to the target scenario, and using the search query to retrieve candidate evidence in the risk memory bank; during the search, performing weighted matching on key actions, key objects and risk domain words to determine the relevance score of the candidate evidence, and then obtaining the Top-k candidate evidence based on the relevance score of the candidate evidence to form an evidence set; A method for generating structured risk scenario specifications under the constraints of the evidence set includes: Generate candidate structured risk scenario specifications that include an object set, object attributes, a set of spatial topological constraints, embodied agent capability boundaries, executable instructions, and risk interpretations; convert the hazards, triggering actions, and consequences in the candidate structured risk scenario specifications into observable visual cues and write them into the candidate structured risk scenario specifications; store the candidate structured risk scenario specifications in the form of structured files; record evidence consistency indicators, infeasibility ratio indicators, and non-redundant coverage indicators as quality measures of the candidate structured risk scenario specifications to obtain the final structured risk scenario specifications; The first frame image module is used to generate a pre-accident first frame image based on the structured risk scenario specifications, and to perform consistency verification on the pre-accident first frame image. The sequence generation module is used to input the first frame image of the pre-accident that has passed the consistency check into the video generative world model to generate a future predicted video sequence: the first frame image of the pre-accident is used as the initial state, and the conditional text derived from the structured risk scenario specification is used as the control signal and input into the video generative world model to generate multiple candidate videos corresponding to different random seeds or sampling parameters; the validity of the multiple candidate videos is checked, and the multiple candidate videos that pass the validity check form a future predicted video sequence; The scoring module is used to perform risk chain scoring based on the future predicted video sequence to obtain an initial condition score, a triggering process score, and a consequence severity score, and output evidence fragments corresponding to each score. The risk chain scoring method includes: performing coverage sampling along the time axis of the future predicted video sequence at fixed time intervals according to a preset default robust strategy to obtain visual fragments, and retaining the timestamp index corresponding to each visual fragment; wherein, during coverage sampling, encrypted sampling is used at each stage of the future predicted video sequence; inputting all visual fragments and their corresponding timestamp indices into a multimodal large model evaluator, and combining the scoring rules constructed based on the structured risk scenario specifications to perform three independent scoring operations on the initial condition score, triggering process score, and consequence severity score; determining the initial condition score, triggering process score, and consequence severity score based on the three independent scores through a voting mechanism, and outputting evidence fragments corresponding to each score. The evaluation module is used to determine the final evaluation result based on the logical relationship between the initial condition score, the triggering process score, and the consequence severity score, combined with the evidence fragments corresponding to each score.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the risk chain assessment method as described in any one of claims 1-3.

Images