Multi-hypothesis timing world model construction method, system, storage medium and device

Abstract

The application discloses a multi-hypothesis timing world model construction method and system, a storage medium and equipment, and relates to the field of embodied intelligent robots. The method comprises the following steps: S1, obtaining a historical observation sequence and an action sequence of a robot, performing timing feature extraction, and generating an implicit space state vector at a current time; S2, inputting the implicit space state vector and a current action vector into a multi-hypothesis prediction module comprising N parameter-independent probability decoders; S3, constructing a total loss function to drive the plurality of parameter-independent probability decoders to maintain prediction differences; S4, determining a best prediction estimate of the system by using an integrated mean of the output of each decoder, and calculating a cognitive uncertainty quantization value by using the dispersion of each decoder relative to the integrated mean; and S5, iteratively performing steps S1 to S4 to identify a high confidence interval and an illusion accumulation interval in a predicted trajectory. The application explicitly quantifies cognitive uncertainty and its long-time domain propagation by means of prediction head divergence.

Description

Technical Field

[0001] This invention relates to the field of embodied intelligent robots, specifically including a method, system, storage medium, and device for constructing a multi-hypothesis temporal world model for embodied intelligence. Background Technology

[0002] In the field of embodied intelligence, world models are considered a key cornerstone for endowing robots with higher-order intelligence. They predict future environmental changes by observing current states and actions, providing support for motion planning and simulation. Related research has proposed an AI robot control system that integrates a world model architecture. This system uses a shared multimodal backbone network to connect the policy head and the world model head respectively, in order to learn the causal relationship between actions and state changes. However, such solutions often use a single deterministic prediction head architecture, whose output is only a point estimate. This makes it impossible to distinguish between the randomness of the environment itself and the cognitive blind spots caused by insufficient training data coverage. As a result, the model still blindly and confidently outputs prediction results in areas outside the distribution, making it difficult to issue risk warnings.

[0003] The shortcomings of deterministic modeling are further amplified in long-term extrapolation. Because a single model lacks the ability to express "ignorance," it cannot quantify the reliability of its predictions when the robot enters a state space not covered by the training data. Simultaneously, in the autoregressive rolling prediction process of the world model, tiny single-step prediction errors are exponentially amplified with each time step iteration, causing the generated future trajectories to gradually deviate from physical laws, producing seemingly reasonable but actually erroneous illusory data. Related research has pointed out that in long-term tasks, due to the inability to predict the long-term consequences of one's own actions, tiny decision-making errors accumulate, ultimately leading to task failure.

[0004] To alleviate the aforementioned problem of insufficient uncertainty quantification, some studies have attempted to employ probabilistic ensemble methods, which involve training multiple models and calculating uncertainty based on their prediction discrepancies. For example, in the field of sequence recommendation, existing approaches integrate multiple independent sub-networks after the encoder and introduce diversity regularization terms to prevent convergence of sub-network outputs, attempting to perceive cognitive uncertainty through prediction differences between sub-networks. However, in such ensemble methods, all sub-models tend to converge to similar parameter spaces, resulting in highly consistent and erroneous predictions of out-of-distribution data, thus losing the ability to detect anomalies through prediction discrepancies. Furthermore, existing methods are mostly limited to single-step prediction scenarios, failing to effectively embed uncertainty measures into long-sequence autoregressive deduction processes, and thus unable to achieve early identification of error accumulation and illusion diffusion.

[0005] In summary, existing data-driven world models are prone to producing hallucinatory trajectories in long-term predictions and are difficult to effectively quantify the cognitive uncertainty caused by insufficient training data coverage. This makes it impossible for robots to proactively issue risk warnings when prediction reliability declines, and it is also difficult to accurately identify the hallucination accumulation interval in long-term extrapolation. This is a bottleneck that urgently needs to be overcome to achieve general embodied intelligence. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a method, system, storage medium, and device for constructing a multi-hypothesis time-series world model. It generates discriminative future hypotheses in parallel through multiple independent decoders, maintains hypothesis diversity through exclusionary regularization constraints, and explicitly quantifies cognitive uncertainty and its long-term propagation with predictor head divergence, thereby establishing a comprehensive risk boundary identification mechanism of predicted value + credibility in long-sequence prediction.

[0007] The technical solution of this invention is as follows:

[0008] Firstly, a method for constructing a multi-hypothesis time-series world model includes the following steps:

[0009] Step S1: Obtain the robot's historical observation sequence and action sequence, and extract temporal features from the historical observation sequence and action sequence using a shared temporal feature encoder to generate the latent space state vector at the current moment;

[0010] Step S2: Input the latent space state vector and the current action vector into the multi-hypothesis prediction module containing N independent probability decoders. Each probability decoder outputs the prediction distribution parameters of the observation at the next time step, where N is an integer greater than 1.

[0011] Step S3: During the model training phase, a total loss function is constructed that includes a prediction accuracy loss term and a diversity regularization loss term. The diversity regularization loss term increases as the proximity between the outputs of each probability decoder increases, so as to drive multiple parameter-independent probability decoders to maintain prediction differences and suppress mode collapse of the multi-hypothesis prediction module.

[0012] Step S4: In the inference phase, the best prediction estimate of the system is determined by using the ensemble mean of the outputs of each decoder, and the cognitive uncertainty quantification value is determined by calculating the dispersion of each decoder relative to the ensemble mean.

[0013] Step S5: In the long-sequence autoregressive extrapolation process, the integrated mean of the current prediction step output is concatenated with the action vector of the next moment as part of the updated historical observation sequence. Steps S1 to S4 are executed iteratively to make the cognitive uncertainty quantification value accumulate and propagate along the time axis. By comparing the cognitive uncertainty quantification value at each moment with the preset threshold, the high confidence interval and hallucination accumulation interval in the prediction trajectory are identified.

[0014] In this method, step S1 corresponds to the data-driven world model, establishing a shared encoder to extract unified dynamic representations, providing a common basis for subsequent multi-hypothesis generation; step S2 introduces a multi-hypothesis parallel decoding structure, providing multiple independent sources of opinion for the quantification of cognitive uncertainty, which is different from the existing single prediction head architecture; step S3 clarifies the training mechanism through an exclusionary diversification regularization term, which directly solves the problem of uncertainty failure caused by ensemble mode collapse; step S4 explicitly defines the calculation formula for cognitive uncertainty, transforming multi-hypothesis divergence into an operable risk metric, enabling the model to have self-examination capabilities; step S5 embeds the uncertainty metric into the long-sequence autoregressive process, allowing the out-of-distribution offset caused by accumulated errors to be captured in real time, thereby accurately identifying the hallucination occurrence interval, and ultimately overcoming both long-sequence hallucinations and the lack of cognitive uncertainty.

[0015] As a preferred embodiment, step S1 specifically includes:

[0016] Collect multi-source heterogeneous data generated by the robot's interaction in real or simulated environments. The observation sequence includes visual observation data (camera images or point cloud data), body perception data (joint angles, joint angular velocities, end effector pose, contact force / torque signals), and inertial measurement unit data.

[0017] The multi-source heterogeneous data is time-aligned and standardized to construct a fixed-length time window sample, which is the observation sequence of H historical time steps. With action sequence As the input sequence;

[0018] Encoder() represents the shared temporal feature encoder. The temporal feature encoder employs a gated recurrent unit (GRU) or Transformer structure to map the historical observation sequence into a compact latent space state vector.

[0019]

[0020] The This is used to characterize the dynamic context features of the environment at the current moment, providing shared conditional input for subsequent multi-hypothesis prediction modules.

[0021] As a preferred embodiment, step S2 specifically includes:

[0022] The N probabilistic decoders have their own independent network weight parameters and do not share the weights of the shared temporal feature encoder. Each probabilistic decoder uses a fully connected neural network with the same structure but independent parameters.

[0023] Each of the probabilistic decoders receives the same latent space state vector. With the current action vector The concatenated joint feature vectors are then used to independently output the next time-step observations. The conditional probability distribution parameters;

[0024]

[0025] in, Indicates the first The conditional probability distribution predicted by each decoder (prediction head). Indicates the observation value at the next moment. Represents a multidimensional Gaussian distribution. Indicates the first case under the current circumstances. The mean vector predicted by each decoder network. Indicates the first case under the current circumstances. The diagonal covariance matrix predicted by each decoder network. This represents the set of learnable parameters of a probabilistic decoder neural network, namely the network weights and biases.

[0026] As a preferred approach, the specific construction of the total loss function in step S3 includes:

[0027] The prediction accuracy loss term The average negative log-likelihood loss of the output distribution of each probability decoder is based on the true observations. Conditional probability distribution of each decoder output calculate;

[0028]

[0029] in Represents the training dataset. This represents the Gaussian probability density function of the output of the k-th probability decoder;

[0030] The exclusionary diversification regularization term Based on the mean vector output by each probability decoder and Euclidean distance between The calculation is performed using the following Gaussian kernel function form:

[0031]

[0032] in These are the preset kernel bandwidth parameters;

[0033] The final total loss function is ,in Set it to 0.08-0.2.

[0034] In this scheme, the functional description of exponential growth is transformed into a reproducible mathematical formula—the Gaussian kernel exponential function. This is used to ensure prediction accuracy and avoid sacrificing accuracy for the sake of discrepancies. This is used to ensure diversity and avoid falling into collapse in the pursuit of precision.

[0035] To define cognitive uncertainty as an interpretable geometric metric or predictive head divergence, the specific calculation of the cognitive uncertainty quantification value in step S4 includes:

[0036] Based on the multi-hypothesis temporal dynamics ensemble model after training convergence, for a given latent space state vector With action vectors Each probability decoder outputs the predicted mean vector of the observations at the next time step. ,in ;

[0037] Calculate the The arithmetic mean of the predicted mean vectors is used as the system's ensemble predicted mean.

[0038] The integrated predicted mean The system outputs its best estimate of the state at the next moment to the downstream planner or controller.

[0039] Calculate the prediction mean vector of each probability decoder output. Relative to the integrated prediction mean square Euclidean distance and for all The probability decoder is taken as an arithmetic mean to obtain the quantification value of cognitive uncertainty. The The cognitive confidence of the model in the current prediction is represented by a higher value, which indicates greater divergence among the prediction heads and less cognitive ability of the model.

[0040]

[0041] Among them, when the When the value is below a preset threshold, the current prediction is determined to be within the coverage area of ​​the training data distribution, and the prediction result is considered reliable; when the value is below a preset threshold, the prediction is considered to be within the coverage area of ​​the training data distribution. When the value exceeds a preset threshold, the model is determined to have entered an out-of-distribution region, and the current prediction carries a risk of hallucination.

[0042] Furthermore, it also includes: fusing the diagonal covariance matrices output from each probability decoder to calculate the comprehensive prediction uncertainty, wherein the comprehensive prediction uncertainty includes the cognitive uncertainty. With chance and uncertainty The random uncertainty is determined by the diagonal covariance matrix output by each decoder. The integrated average is determined;

[0043]

[0044] in, For the first The diagonal covariance matrix output by each probability decoder, where N is the total number of decoders.

[0045] The physical meaning of this formula is: taking the ensemble average of the inherent noise (random uncertainty) of the data predicted by each decoder, characterizing the level of randomness of the environment itself, and related to cognitive uncertainty. (Knowledge blind spots in the representation model) together constitute the overall prediction uncertainty.

[0046] As a preferred option, the current prediction step Output ensemble mean With the action vector at the next moment The concatenated data, as part of the updated historical observation sequence, is re-input into the shared temporal feature encoder, and steps S1 to S4 are iteratively executed to gradually generate future data. Predicted trajectory sequence of steps and the corresponding cognitive uncertainty sequence ;

[0047] The cognitive uncertainty quantification value The model exhibits cumulative propagation characteristics along the time axis. When the model enters a region with insufficient coverage of training data distribution at a certain prediction step, the cognitive uncertainty quantification value of the model and subsequent steps increases significantly, revealing the error accumulation and illusion diffusion process in long sequence inference.

[0048] Furthermore, the high-confidence interval and hallucination accumulation interval in the identified predicted trajectory include:

[0049] Preset uncertainty threshold Quantify the cognitive uncertainty of each prediction step. Compare with the threshold;

[0050] when When the prediction step is in the high confidence interval, the corresponding predicted trajectory segment is considered reliable.

[0051] when If the predicted step enters the illusion accumulation zone, the corresponding predicted trajectory segment is unreliable, and a risk warning is triggered or the further expansion of the trajectory is stopped.

[0052] Furthermore, it also includes: synthetic data generation application: given initial observations and candidate action sequences, the long sequence autoregressive inference is executed to generate future trajectories and their uncertainty sequences; the state transition results within the high confidence interval are retained as high confidence synthetic samples, and the sampling weight of trajectory segments within the illusion accumulation interval is removed or reduced to prevent low-quality illusion data from polluting subsequent training.

[0053] Furthermore, it also includes: model predictive control application, performing the long-sequence autoregressive inference on multiple candidate action sequences respectively, and calculating the cumulative cognitive uncertainty index of each candidate trajectory within the entire prediction window.

[0054]

[0055] The cumulative cognitive uncertainty index and the task benefit index are jointly evaluated, and the action sequence with cumulative uncertainty below the safety threshold and task benefit meeting the requirements is selected as the final control output.

[0056] Secondly, the present invention includes a system for constructing a cognitive uncertainty perception multi-hypothesis temporal world model, comprising:

[0057] The data acquisition module is used to acquire the robot's historical observation sequences and action sequences;

[0058] A shared temporal feature encoder is used to extract temporal features from the historical observation sequence and action sequence to generate the latent space state vector at the current moment.

[0059] The multi-hypothesis prediction module contains N independent probability decoders, which are used to receive the latent space state vector and the current action vector, and each probability decoder outputs the prediction distribution parameters of the observation at the next time step, where N is an integer greater than 1.

[0060] The training module is used to construct a total loss function that includes a prediction accuracy loss term and a diversity regularization loss term. The diversity regularization loss term increases as the proximity between the outputs of each probability decoder increases, so as to drive each probability decoder to maintain prediction differences and suppress mode collapse of the multi-hypothesis prediction module.

[0061] The inference module is used to determine the best prediction estimate of the system based on the output of each probability decoder, and to determine the cognitive uncertainty quantification value based on the degree of dispersion of each probability decoder relative to the best prediction estimate.

[0062] The long sequence inference module is used to take the best prediction estimate output from the current prediction step as part of the input observation for the next step, and iteratively drive the data acquisition module, shared temporal feature encoder, multi-hypothesis prediction module, training module and inference module to perform operations, so that the cognitive uncertainty quantification value accumulates and propagates along the time axis, and by comparing the cognitive uncertainty quantification value at each time with the preset threshold, the high confidence interval and hallucination accumulation interval in the prediction trajectory are identified.

[0063] Thirdly, the present invention is a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is used to implement the aforementioned method for constructing a multi-hypothesis temporal world model, or to implement the aforementioned system for constructing a cognitive uncertainty-perceived multi-hypothesis temporal world model.

[0064] Fourthly, the present invention is a computer device, including a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the aforementioned method for constructing a multi-hypothesis temporal world model, or implements the aforementioned system for constructing a cognitive uncertainty-perceived multi-hypothesis temporal world model.

[0065] According to the above-described solution, the beneficial effects of this invention are as follows:

[0066] This invention, through the collaborative design of a shared temporal feature encoder, a multi-predictor parallel decoding structure, and an exclusionary diversity regularization term, enables the formation of multiple distinct future dynamic hypotheses under the same historical observation-action conditions. Because the training objective explicitly introduces diversity constraints that differentiate the outputs of each predictor, the multiple predictors do not tend to converge homogeneously after training like conventional ensemble methods. Instead, they can maintain necessary predictive dispersion in sparse or out-of-distribution regions of the training data, structurally suppressing ensemble pattern collapse and improving the sensitivity of the world model to unknown regions.

[0067] This invention uses the ensemble mean of the outputs of each predictor as the best estimate of the system, and defines the degree of dispersion of each predictor around this best estimate as a cognitive uncertainty index. Thus, the model can not only output future state predictions but also simultaneously output corresponding confidence evaluation results, achieving an integrated expression of prediction results and confidence information. When the model is in a region with sufficient training sample coverage, the outputs of each predictor tend to be consistent, and cognitive uncertainty remains at a low level. When the model enters regions with sparse samples, state shifts, or complex dynamics, the divergence in the outputs of each predictor increases, and cognitive uncertainty rises accordingly, thereby providing clear risk indications for upper-level planning and control.

[0068] This invention enables cognitive uncertainty to propagate along the time axis during autoregressive long-sequence extrapolation, thereby enabling earlier identification of error accumulation and illusion diffusion in long-term predictions. Compared to traditional deterministic world models that can only provide single-path predictions but cannot determine when the path begins to distort, this invention can distinguish between reliable physical simulation intervals and unreliable prediction intervals, improving the credibility and usability of long-sequence simulations.

[0069] The cognitive uncertainty output by this invention can be directly used as a risk measure in downstream planning, control, trajectory screening, and synthetic data generation. By using it as a penalty, filtering condition, or sample weight, prediction results in low uncertainty regions can be prioritized, and distorted trajectories in high uncertainty regions can be actively avoided. This reduces the dependence of downstream strategies or controllers on erroneous model predictions and improves the safety and robustness of the system during training and deployment.

[0070] This invention can complete world model training entirely based on offline physical interaction data, without relying on real robots for continuous online trial and error, and without the need to pre-build complex analytical physical models. Therefore, it has good data reuse capability and cross-task transfer potential. In application scenarios where real sampling costs are high and dangerous actions should not be performed frequently, this invention can build a high-quality world model with risk perception capability at a lower cost, thereby reducing R&D costs and shortening the development cycle.

[0071] Since the cognitive uncertainty of this invention stems from explicit divergences among multiple predictors around the optimal estimate, rather than uninterpretable black-box scores, it possesses good interpretability. Researchers can combine prediction trajectories with changes in uncertainty to analyze at which stages the model begins to deviate from the training distribution and which state transitions are more likely to induce prediction distortion, thus simultaneously using this invention as a world model building tool, a data coverage analysis tool, and a risk control and early warning tool.

[0072] In summary, this invention improves the long-sequence prediction capability, risk identification capability, and engineering usability of the world model through the joint design of multi-hypothesis prediction structure, exclusionary diversified training, explicit quantification of cognitive uncertainty, and long-term propagation mechanism. It can provide more reliable, safe, and interpretable technical support for scenarios such as robotic arm operation, humanoid robot motion control, offline reinforcement learning, digital twin simulation, and safety assessment. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the data flow in this invention. Detailed Implementation

[0074] The present invention will now be further described with reference to the accompanying drawings and embodiments:

[0075] The method, system, storage medium, and device for constructing a cognitive uncertainty-aware multi-hypothesis temporal world model (MHTD-EM) for embodied intelligence are presented. The overall process is divided into an offline training phase and an online inference / long sequence prediction phase.

[0076] Step S1: Obtain the robot's historical observation sequence and action sequence, and extract temporal features from the historical observation sequence and action sequence using a shared temporal feature encoder to generate the latent space state vector at the current moment;

[0077] Step S2: Input the latent space state vector and the current action vector into the multi-hypothesis prediction module containing N independent probability decoders. Each probability decoder outputs the prediction distribution parameters of the observation at the next time step, where N is an integer greater than 1.

[0078] Step S3: During the model training phase, a total loss function is constructed that includes a prediction accuracy loss term and a diversity regularization loss term. The diversity regularization loss term increases as the proximity between the outputs of each probability decoder increases, so as to drive each probability decoder to maintain prediction differences and suppress mode collapse of the multi-hypothesis prediction module.

[0079] Step S4: In the inference phase, the optimal prediction estimate of the system is determined based on the output of each probability decoder, and the cognitive uncertainty quantification value is determined based on the degree of dispersion of each probability decoder relative to the optimal prediction estimate.

[0080] Step S5: In the long-sequence autoregressive extrapolation process, the integrated mean of the current prediction step output is concatenated with the action vector of the next moment as part of the updated historical observation sequence. Steps S1 to S4 are executed iteratively to make the cognitive uncertainty quantification value accumulate and propagate along the time axis. By comparing the cognitive uncertainty quantification value at each moment with the preset threshold, the high confidence interval and hallucination accumulation interval in the prediction trajectory are identified.

[0081] In this method, step S1 corresponds to the data-driven world model, establishing a shared encoder to extract unified dynamic representations, providing a common basis for subsequent multi-hypothesis generation; step S2 introduces a multi-hypothesis parallel decoding structure, providing multiple independent sources of opinion for the quantification of cognitive uncertainty, which is different from the existing single prediction head architecture; step S3 clarifies the training mechanism through an exclusionary diversification regularization term, which directly solves the problem of uncertainty failure caused by ensemble mode collapse; step S4 explicitly defines the calculation formula for cognitive uncertainty, transforming multi-hypothesis divergence into an operable risk metric, enabling the model to have self-examination capabilities; step S5 embeds the uncertainty metric into the long-sequence autoregressive process, allowing the out-of-distribution offset caused by accumulated errors to be captured in real time, thereby accurately identifying the hallucination occurrence interval, and ultimately overcoming both long-sequence hallucinations and the lack of cognitive uncertainty.

[0082] As a preferred embodiment, step S1 specifically includes:

[0083] Offline physical interaction data acquisition and sample construction: First, offline interaction data of the robot in real or simulated environments is collected; each trajectory data includes at least a continuous observation sequence, a continuous action sequence, and the corresponding real observation value at the next moment; the observations may include joint angles, joint angular velocities, end effector pose, contact state, inertial measurement unit information, and visual features, and the motion quantities may include joint torques, target positions, target velocities, or end control quantities.

[0084] After standardizing all observations and motion quantities, training samples are constructed according to a fixed historical window length H, i.e., historical observation-motion segments. As input, the actual next moment observation This serves as a supervisory label. This step corresponds to the source of the original input in subsequent formulas.

[0085] The second step involves the shared temporal feature encoder calculating the latent space state vector. The historical observation sequence and action sequence are input into the shared temporal feature encoder, and the latent space state vector at the current moment is calculated using the following formula:

[0086]

[0087] in, It is a temporal feature encoder. This is the hidden state representation obtained by compressing historical information. This formula is used at the start of the model's forward propagation, and its function is to map the high-dimensional temporal input into compact dynamic context features, which can be shared and used by all prediction heads.

[0088] The third step involves the multi-predictor probabilistic decoder outputting future state distribution parameters:

[0089] Obtain the hidden state Afterwards, With the current action They are all input into K independent probability decoders.

[0090] The N probability decoders use networks with the same structure but independent parameters. Each probability decoder receives the latent space state vector output by the shared temporal feature encoder as input.

[0091] Output the probability distribution parameters of the observations at the next time step, whose expressions are:

[0092]

[0093] in, } indicates the first The conditional probability distribution predicted by each decoder ) represents a multidimensional Gaussian distribution. Indicates the first The mean vector predicted by each decoder network. Indicates the first The diagonal covariance matrix predicted by each decoder network. This represents the set of learnable parameters of the probabilistic decoder neural network; this formula is used in the second stage of the model's forward propagation, where it generates multiple different future dynamics hypotheses under the same historical conditions.

[0094] The fourth step is to calculate the average negative log-likelihood loss using the actual next-time observations. Substituting the Gaussian distribution corresponding to each predictor, we calculate the prediction error of each predictor and average it over all predictors to obtain:

[0095]

[0096] This formula is used in the accuracy loss calculation stage of the training phase. Its purpose is to constrain each predictor to have basic dynamic prediction capabilities, and to prevent the model from sacrificing prediction accuracy by pursuing dispersion.

[0097] Step 5: Calculate the exclusionary diversification loss using the difference in the mean vectors of the prediction heads. To prevent multiple prediction heads from converging to the same solution after training, calculate the exclusionary diversification loss for the mean vectors of any two prediction heads:

[0098]

[0099] in, is the Gaussian kernel bandwidth hyperparameter. This formula is used in the diversity constraint stage of the training phase. Its function is to increase the loss value when the outputs of multiple predictors are too similar, thereby driving each predictor to maintain differences in the parameter space and output space and suppressing ensemble mode collapse.

[0100] Using the Adam optimizer, End-to-end joint training is performed on the shared temporal feature encoder and all probabilistic decoders until the validation set is reached. Convergence yields the trained multi-hypothesis temporal dynamics ensemble model.

[0101] Step 6: Train the MHTD-EM model using the joint loss function. Combine the prediction accuracy loss and diversity loss to obtain the model's total loss function:

[0102]

[0103] in, For diversity regularization weights, Set to 0.08-0.2, preferably 0.1. This formula is used in the overall objective function construction stage of the training phase. In implementation, it is first calculated based on the forward propagation results. and Then, by using the above formula, we can obtain... and with Backpropagation updates are performed on the shared encoder and all prediction heads until the model converges, resulting in the trained MHTD-EM model.

[0104] Step 7: Utilize the training to calculate the optimal estimate of the system. After model training is complete, for any input sample, first obtain the mean vector of the K predictor outputs using the aforementioned formula, and then calculate the optimal estimate of the system using the following formula:

[0105]

[0106] in, This represents the system's final prediction of the state at the next moment. This formula is used in the output fusion stage of the inference phase, and its function is to integrate multiple future dynamic assumptions into a unified state prediction result, which can be directly called by downstream planners, controllers or simulators.

[0107] Step 8: Calculate cognitive uncertainty based on the dispersion of the predictor head around the best estimate. After obtaining the best estimate, continue to calculate the cognitive uncertainty at that moment using the following formula:

[0108]

[0109] in, This represents the cognitive uncertainty at time t+1; this formula is used in the credibility assessment stage of the inference phase. Its meaning is that the closer each predictor is to the optimal estimate set, the better. The smaller the value, the more reliable the current forecast; the more dispersed the forecast heads, the better. The larger the value, the less the model understands the region, and the higher the risk of the current prediction.

[0110] Step nine involves applying the best estimate and cognitive uncertainty in long-term autoregressive extrapolation. In long-term prediction scenarios, the best estimate obtained in the previous step is used... As part of the next input observation, it is re-inputted into the model along with the action at the next time step, and the process of "time series encoding - multi-head prediction - best estimation - uncertainty calculation" is repeated to gradually obtain... , ... and the corresponding , ... In this step, the aforementioned formula is repeatedly called in each prediction step. Its function is to allow cognitive uncertainty to propagate along the time axis, thereby revealing the error accumulation process in long-sequence extrapolation.

[0111] Step 10: Utilize cognitive uncertainty to screen high-confidence synthetic trajectories. In synthetic data generation applications, given initial observations and candidate action sequences, future state trajectories and their corresponding uncertainty sequences are generated using the aforementioned autoregressive method.

[0112] During implementation, an uncertainty threshold can be preset. When a certain step satisfies When the prediction is considered reliable, the corresponding state transition result can be retained as a high-confidence sample; when If the prediction risk at this step is deemed high, trajectory expansion can be stopped, or this step and subsequent trajectories can be marked as low-confidence samples and removed. In this step, the optimal estimation formula is used to generate the trajectory, and the cognitive uncertainty formula is used to filter the trajectory.

[0113] Step 11: Utilize cognitive uncertainty constraint model predictive control. In model predictive control applications, the controller generates multiple candidate action sequences at the current moment, performs the aforementioned long sequence deduction for each action sequence, and calculates the corresponding cumulative uncertainty index.

[0114]

[0115] in, This represents the cumulative cognitive uncertainty of a candidate trajectory within the entire prediction window. This formula is used in the control decision-making stage, where it serves as a risk constraint in the candidate trajectory evaluation. In implementation, it can be... In conjunction with task benefits, energy consumption, or stability indicators, priority is given to action sequences with higher benefits and lower cumulative uncertainty, thereby reducing the controller's reliance on high-risk distorted trajectories.

[0116] Step 12: Verify the model's capabilities using the alignment relationship of error uncertainty: To verify that the world model described in this invention possesses cognitive boundary awareness capabilities, long-sequence inference can be performed on the validation set, and the true prediction error at each step can be calculated:

[0117]

[0118] in, The error curve is the actual observed value; compare this error curve with the corresponding cognitive uncertainty curve. A comparison is then made. If cognitive uncertainty increases synchronously within the range where the prediction error increases, it indicates that the model can proactively output a higher risk signal when the prediction begins to distort, proving that it has a "self-examination" ability. In this step, the optimal estimation formula is used to construct the error, and the uncertainty formula is used to construct the risk measure. Both are used together to verify the model's performance.

[0119] Furthermore, it also includes: fusing the diagonal covariance matrices output from each probability decoder to calculate the comprehensive prediction uncertainty, wherein the comprehensive prediction uncertainty includes the cognitive uncertainty. With chance and uncertainty The random uncertainty is determined by the diagonal covariance matrix output by each decoder. The integrated average is determined;

[0120]

[0121] in, For the first The diagonal covariance matrix output by each probability decoder, where N is the total number of decoders.

[0122] The physical meaning of this formula is: taking the ensemble average of the inherent noise (random uncertainty) of the data predicted by each decoder, characterizing the level of randomness of the environment itself, and related to cognitive uncertainty. (Knowledge blind spots in the representation model) together constitute the overall prediction uncertainty.

[0123] Example 2

[0124] Based on Embodiment 1, it can also be applied to the following scenarios:

[0125] Synthetic data generation application: Given initial observations and candidate action sequences, the long-sequence autoregressive inference is performed to generate future trajectories and their uncertainty sequences; the state transition results within the high-confidence interval are retained as high-confidence synthetic samples, and the sampling weights of trajectory segments within the illusion accumulation interval are removed or reduced to prevent low-quality illusion data from polluting subsequent training.

[0126] In the model predictive control application, the long-sequence autoregressive inference is performed on multiple candidate action sequences to calculate the cumulative cognitive uncertainty index of each candidate trajectory within the entire prediction window.

[0127]

[0128] The cumulative cognitive uncertainty index and the task benefit index are jointly evaluated, and the action sequence with cumulative uncertainty below the safety threshold and task benefit meeting the requirements is selected as the final control output.

[0129] Furthermore, this embodiment can also be a cognitive uncertainty perception multi-hypothesis temporal world model construction system, which includes:

[0130] The data acquisition module is used to acquire the robot's historical observation sequences and action sequences;

[0131] A shared temporal feature encoder is used to extract temporal features from the historical observation sequence and action sequence to generate the latent space state vector at the current moment.

[0132] The multi-hypothesis prediction module contains N independent probability decoders, which are used to receive the latent space state vector and the current action vector, and each probability decoder outputs the prediction distribution parameters of the observation at the next time step, where N is an integer greater than 1.

[0133] The training module is used to construct a total loss function that includes a prediction accuracy loss term and a diversity regularization loss term. The diversity regularization loss term increases as the proximity between the outputs of each probability decoder increases, so as to drive each probability decoder to maintain prediction differences and suppress mode collapse of the multi-hypothesis prediction module.

[0134] The inference module is used to determine the best prediction estimate of the system based on the output of each probability decoder, and to determine the cognitive uncertainty quantification value based on the degree of dispersion of each probability decoder relative to the best prediction estimate.

[0135] The long sequence inference module is used to take the best prediction estimate output from the current prediction step as part of the input observation for the next step, and iteratively drive the data acquisition module, shared temporal feature encoder, multi-hypothesis prediction module, training module and inference module to perform operations, so that the cognitive uncertainty quantification value accumulates and propagates along the time axis, and by comparing the cognitive uncertainty quantification value at each time with the preset threshold, the high confidence interval and hallucination accumulation interval in the prediction trajectory are identified.

[0136] This embodiment can also be a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it is used to implement the multi-hypothesis temporal world model construction method, or to implement the cognitive uncertainty perception multi-hypothesis temporal world model construction system executed by the system.

[0137] This embodiment can also be a computer device, including a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the multi-hypothesis temporal world model construction method, or implements the cognitive uncertainty perception multi-hypothesis temporal world model construction system executed by the system.

[0138] like Figure 1 As shown, the left side represents the multiple candidate hypotheses generated by multiple predictors for the same future time during the training phase, and the right side represents the propagation process of cognitive uncertainty during autoregressive rolling.

[0139] Figure 1 The left side represents the model training phase: after the historical observation sequence and action sequence are input into a shared temporal encoder, multiple independent prediction heads make parallel predictions for future observations.

[0140] The multiple candidate nodes in the lower left corner, with different colors or shades, are used to illustrate the various future hypotheses given by different prediction heads. At the same prediction time, the more obvious the differences between these candidate nodes in the representation space, the more dispersed the outputs of each prediction head are. This indicates that the exclusionary diversity regularization term effectively suppresses the collapse of the prediction head, thus preserving sufficient hypothesis diversity within the ensemble.

[0141] Figure 1 The right side represents the autoregressive inference stage after model deployment: the model feeds back the current prediction results as the input for the next step, and continuously generates the future trajectory along the time axis.

[0142] The color intensity of the bottom right node corresponds to the cognitive uncertainty at that moment. The values ​​are positively correlated. The darker the color, the greater the divergence among the prediction heads at that moment and the lower the reliability of the prediction at that step. When the rolling step size increases and the color gradually darkens, it indicates that the model is gradually approaching or entering an area with insufficient training data coverage. Risk marking, truncation, or weight reduction should be performed on this segment of the trajectory.

[0143] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

[0144] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the implementation of the present invention is not limited to the above-described manner. Any improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A method for constructing a multi-hypothesis temporal series world model, characterized in that, Includes the following steps: Step S1: Obtain the robot's historical observation sequence and action sequence, and extract temporal features from the historical observation sequence and action sequence using a shared temporal feature encoder to generate the latent space state vector at the current moment; Step S2: Input the latent space state vector and the current action vector into the multi-hypothesis prediction module containing N independent probability decoders. Each probability decoder outputs the prediction distribution parameters of the observation at the next time step, where N is an integer greater than 1. Step S3: During the model training phase, a total loss function is constructed that includes a prediction accuracy loss term and a diversity regularization loss term. The diversity regularization loss term increases as the proximity between the outputs of each probability decoder increases, so as to drive multiple parameter-independent probability decoders to maintain prediction differences and suppress mode collapse of the multi-hypothesis prediction module. Step S4: In the inference phase, the best prediction estimate of the system is determined by using the ensemble mean of the outputs of each decoder, and the cognitive uncertainty quantification value is determined by calculating the dispersion of each decoder relative to the ensemble mean. Step S5: In the long-sequence autoregressive extrapolation process, the integrated mean of the current prediction step output is concatenated with the action vector of the next moment as part of the updated historical observation sequence. Steps S1 to S4 are executed iteratively to make the cognitive uncertainty quantification value accumulate and propagate along the time axis. By comparing the cognitive uncertainty quantification value at each moment with the preset threshold, the high confidence interval and hallucination accumulation interval in the prediction trajectory are identified.

2. The method for constructing a multi-hypothesis temporal world model according to claim 1, characterized in that, Step S1 includes: Collect multi-source heterogeneous data generated by the robot's interaction in a real or simulated environment, wherein the observation sequence includes visual observation data, body perception data, and inertial measurement unit data; Take the observation sequence of H time steps from the historical data With action sequence As the input sequence; The input sequence is mapped into a fixed-dimensional latent space state vector using the shared temporal feature encoder. ； The hidden space state vector Features used to characterize the dynamic context of the environment at the current time t.

3. The method for constructing a multi-hypothesis temporal world model according to claim 2, characterized in that, In step S2: The N probability decoders use networks with the same structure but independent parameters. Each probability decoder receives the latent space state vector output by the shared temporal feature encoder as input. Each of the probabilistic decoders receives the same latent space state vector. With the current action vector The concatenated joint feature vectors are then used to independently output the next time-step observations. The conditional probability distribution parameters; ； in, Indicates the first The conditional probability distribution predicted by each decoder (prediction head). Indicates the observation value at the next moment. Represents a multidimensional Gaussian distribution. Indicates the first case under the current circumstances. The mean vector predicted by each decoder network. Indicates the first case under the current circumstances. The diagonal covariance matrix predicted by each decoder network. This represents the set of learnable parameters for a probabilistic decoder neural network.

4. The method for constructing a multi-hypothesis temporal world model according to claim 3, characterized in that, In step S3: The total loss function is: ； Wherein, the prediction accuracy loss term is ； The diversity regularization loss term is: ； in, These are the regularization weight coefficients. For kernel bandwidth parameters, Represents the training dataset. , Let represent the mean vectors output by the i-th and j-th probability decoders, respectively; When the Euclidean distance When reduced, the diversity regularization loss term It grows exponentially, generating a repulsive force that forces the parameters of each probability decoder to diverge in different parameter spaces.

5. The method for constructing a multi-hypothesis temporal world model according to claim 4, characterized in that, In step S4: The optimal prediction estimate is the ensemble mean of the mean vectors of the outputs of each probability decoder, expressed as: ； The cognitive uncertainty quantification value for ； Wherein, the cognitive uncertainty quantification value Used to characterize the model's cognitive confidence in the current prediction. A larger value indicates greater discrepancy among the predictors and a less comprehensive understanding of the model.

6. The method for constructing a multi-hypothesis temporal world model according to claim 5, characterized in that, Step S5 includes: The integrated mean of the current prediction step output is concatenated with the action vector of the next time step and re-inputted into the shared temporal feature encoder as part of the updated historical observation sequence; Iteratively execute steps S1 to S4 to gradually generate a sequence of predicted trajectories for multiple future steps and a corresponding sequence of cognitive uncertainty. The cognitive uncertainty quantification value The cognitive uncertainty exhibits a cumulative propagation characteristic along the time axis. When the model enters a region where the training data distribution is insufficiently covered, the cognitive uncertainty quantifies for the corresponding prediction step and subsequent steps. Significantly increased; By setting an uncertainty threshold When the cognitive uncertainty quantification value is compared... Less than or equal to the threshold When the prediction step is determined to be in the high confidence interval, it is considered to be above the threshold. The system will determine when a user enters the hallucination accumulation zone and trigger a risk warning or stop the further expansion of the trajectory.

7. The method for constructing a multi-hypothesis temporal world model according to claim 6, characterized in that, Step S5 also includes: Based on the identification results of the high confidence interval and the hallucination accumulation interval, risk constraints are imposed on downstream tasks; The downstream tasks include at least one of synthetic data generation and model predictive control; In the generation of the synthetic data, the state transition results within the high confidence interval are retained as high confidence synthetic samples, and the sampling weight of the trajectory segments within the hallucination accumulation interval is removed or reduced. In the model predictive control, the long-sequence autoregressive inference is performed on multiple candidate action sequences to calculate the cumulative cognitive uncertainty index of each candidate trajectory. ； The cumulative cognitive uncertainty index and the task benefit index are jointly evaluated, and the action sequence with cumulative uncertainty below the safety threshold and task benefit meeting the requirements is preferentially selected as the control output.

8. A system for constructing a multi-hypothesis temporal world model for perceiving cognitive uncertainty, characterized in that, include: The data acquisition module is used to acquire the robot's historical observation sequences and action sequences; A shared temporal feature encoder is used to extract temporal features from the historical observation sequence and action sequence to generate the latent space state vector at the current moment. The multi-hypothesis prediction module contains N independent probability decoders, which are used to receive the latent space state vector and the current action vector, and each probability decoder outputs the prediction distribution parameters of the observation at the next time step, where N is an integer greater than 1. The training module is used to construct a total loss function that includes a prediction accuracy loss term and a diversity regularization loss term. The diversity regularization loss term increases as the proximity between the outputs of each probability decoder increases, so as to drive each probability decoder to maintain prediction differences and suppress mode collapse of the multi-hypothesis prediction module. The inference module is used to determine the best prediction estimate of the system based on the output of each probability decoder, and to determine the cognitive uncertainty quantification value based on the degree of dispersion of each probability decoder relative to the best prediction estimate. The long sequence inference module is used to take the best prediction estimate output from the current prediction step as part of the input observation for the next step, and iteratively drive the data acquisition module, shared temporal feature encoder, multi-hypothesis prediction module, training module and inference module to perform operations, so that the cognitive uncertainty quantification value accumulates and propagates along the time axis, and by comparing the cognitive uncertainty quantification value at each time with the preset threshold, the high confidence interval and hallucination accumulation interval in the prediction trajectory are identified.

9. A computer-readable storage medium having a computer program stored thereon, wherein when executed by a processor, the computer program implements a method for constructing a multi-hypothesis temporal world model as described in any one of claims 1 to 7, or implements a system for constructing a cognitive uncertainty-perceived multi-hypothesis temporal world model as described in claim 8.

10. A computer device comprising a processor and a memory, the memory storing a computer program, wherein when the processor executes the computer program, it implements a method for constructing a multi-hypothesis temporal world model as described in any one of claims 1 to 7, or implements a system for constructing a cognitive uncertainty-perceived multi-hypothesis temporal world model as described in claim 8.

Images