Digital human dynamic rule adjusting method and device based on causal relationship

Abstract

The invention provides a digital human dynamic rule adjustment method and device based on a causal relationship, and the method comprises the steps: obtaining the historical data, related to the interaction of a digital human and a user, in a mixed reality scene, and the historical data comprise historical behavior data and historical environment data; determining a control step length and a sensing sliding window length of the behavior decision of the digital human based on the historical data, and constructing a behavior decision data matrix based on the control step length and the sensing sliding window length; performing causal structure extraction processing on the behavior decision data matrix to obtain a block matrix under causal constraint; generating a behavior adjustment rule set satisfying a causal relationship based on the block matrix, and constructing a causal-driven multi-walk predictor based on the behavior adjustment rule set; based on a multi-walk predictor, the behavior rule of the digital human is dynamically adjusted. According to the method and the device, the technical problem of poor adaptive behavior adjustment capability caused by lack of causal modeling in a complex interaction scene of a digital human is solved.

Description

Technical Field

[0001] The present invention relates to the field of digital humans, and in particular, to a method and device for dynamically adjusting digital human rules based on causal relationships. Background Art

[0002] With the wide application of mixed reality technology, digital humans, as the core medium for users to interact with virtual environments, their behavior adaptability has become the key to enhancing the immersive experience. Currently, most mainstream digital human systems adopt predefined behavior rule libraries or dynamic decision-making models based on statistical learning (such as deep reinforcement learning). However, in complex and ever-changing interaction scenarios, such methods have obvious limitations. On the one hand, static rule libraries are difficult to cover the diversity of user behaviors in dynamic environments. When encountering unpreset interaction situations, the behaviors of digital humans are prone to logical rigidity or inaccurate responses. On the other hand, although data-driven models can learn historical interaction patterns, due to their reliance on statistical correlations rather than causal relationships, they are easily interfered by environmental noise and confounding factors, resulting in pseudo-correlation biases in behavior decisions.

[0003] For the above problems, no effective solutions have been proposed yet. Summary of the Invention

[0004] Embodiments of the present invention provide a method and device for dynamically adjusting digital human rules based on causal relationships, so as to at least solve the technical problem that digital humans have poor adaptive behavior adjustment ability due to the lack of causal modeling in complex interaction scenarios.

[0005] According to one aspect of the embodiments of the present invention, a method for dynamically adjusting digital human rules based on causal relationships is provided, including: obtaining historical data related to the interaction between a digital human and a user in a mixed reality scenario, where the historical data includes historical behavior data and historical environmental data; determining a control step length and a perception sliding window length for the behavior decision of the digital human based on the historical data, and constructing a behavior decision data matrix based on the control step length and the perception sliding window length; performing causal structure extraction processing on the behavior decision data matrix to obtain a block matrix under causal constraints; generating a behavior adjustment rule set that satisfies causal relationships based on the block matrix, and constructing a multi-step behavior predictor driven by causality based on the behavior adjustment rule set; and dynamically adjusting the behavior rules of the digital human based on the multi-step behavior predictor.

[0006] According to another aspect of the embodiments of the present invention, there is provided a digital human dynamic rule adjustment device based on causality, including: an acquisition module configured to acquire historical data related to the interaction between the digital human and the user in a mixed reality scenario, where the historical data includes historical behavior data and historical environment data; a construction module configured to determine a control step length and a perception sliding window length for the behavior decision of the digital human based on the historical data, and construct a behavior decision data matrix based on the control step length and the perception sliding window length; an extraction module configured to perform causal structure extraction processing on the behavior decision data matrix to obtain a block matrix under causal constraints; a generation module configured to generate a behavior adjustment rule set that satisfies the causal relationship based on the block matrix, and construct a multi-step behavior predictor driven by causality based on the behavior adjustment rule set; and an adjustment module configured to dynamically adjust the behavior rules of the digital human based on the multi-step behavior predictor.

[0007] In the embodiments of the present invention, historical data related to the interaction between the digital human and the user in a mixed reality scenario is acquired, where the historical data includes historical behavior data and historical environment data; based on the historical data, a control step length and a perception sliding window length for the behavior decision of the digital human are determined, and a behavior decision data matrix is constructed based on the control step length and the perception sliding window length; causal structure extraction processing is performed on the behavior decision data matrix to obtain a block matrix under causal constraints; a behavior adjustment rule set that satisfies the causal relationship is generated based on the block matrix, and a multi-step behavior predictor driven by causality is constructed based on the behavior adjustment rule set; and the behavior rules of the digital human are dynamically adjusted based on the multi-step behavior predictor. Through the above solution, the technical problem that the digital human has poor adaptive behavior adjustment ability due to the lack of causal modeling in complex interaction scenarios is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0009] Figure 1 is a flowchart of a method for dynamically adjusting the rules of a digital human based on causality according to an embodiment of the present invention;

[0010] Figure 2 is a flowchart of another method for dynamically adjusting the rules of a digital human based on causality according to an embodiment of the present invention;

[0011] Figure 3 is a flowchart of a method for constructing a behavior decision data matrix according to an embodiment of the present invention;

[0012] Figure 4It is a flowchart of a method for generating a behavior adjustment rule set that satisfies causal relationships according to an embodiment of the present invention;

[0013] Figure 5 It is a schematic structural diagram of a digital human dynamic rule adjustment device based on causal relationships according to an embodiment of the present invention;

[0014] Figure 6 It shows a schematic structural diagram of an electronic device suitable for implementing the embodiments of the present disclosure. Detailed implementation manners

[0015] In order to enable those skilled in the art of the present technology to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0016] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order different from those illustrated or described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0017] According to an embodiment of the present invention, a method embodiment of a digital human dynamic rule adjustment method based on causal relationships is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0018] Figure 1 It is a digital human dynamic rule adjustment method based on causal relationships according to an embodiment of the present invention. As Figure 1 shown, the method includes the following steps:

[0019] Step S102, obtain historical data related to the interaction between the digital human and the user in the mixed reality scenario, where the historical data includes historical behavior data and historical environmental data.

[0020] Step S104: Based on the historical data, determine the control step length and the perception sliding window length for the behavior decision of the digital human, and construct a behavior decision data matrix based on the control step length and the perception sliding window length.

[0021] First, determine the control step length and the perception sliding window length. For example, analyze the temporal relationship between the change rate of interaction events and the behavior response in the historical data, and extract dynamic time features reflecting the behavior response trend of the digital human; based on the dynamic time features, evaluate the behavior stability index and the environmental perturbation amplitude of the digital human, and determine the control step length and the perception sliding window length based on the behavior stability index and the environmental perturbation amplitude.

[0022] Next, construct a behavior decision data matrix. For example, based on the control step length and the perception sliding window length, perform sliding window segmentation on the historical data, and respectively extract the input trajectory, state trajectory, and interference features within each sliding window to construct multiple joint data blocks integrating behavior, environment, and response; arrange the multiple joint data blocks in chronological order to form the behavior decision data matrix with a multi-time period causal structure.

[0023] Step S106: Perform causal structure extraction processing on the behavior decision data matrix to obtain a block matrix under causal constraints.

[0024] For example, perform lower triangular orthogonal decomposition on the behavior decision data matrix to obtain a lower triangular causal transfer matrix representing the causal path of behavior response; perform structural decoupling of the perturbation feedback path on the lower triangular causal transfer matrix, retain the lower triangular block in the lower triangular causal transfer matrix and mask the upper triangular block to obtain the block matrix under causal constraints.

[0025] Step S108: Generate a set of behavior adjustment rules that satisfy causal relationships based on the block matrix, and construct a multi-step behavior predictor driven by causality based on the set of behavior adjustment rules.

[0026] First, generate a set of behavior adjustment rules. For example, perform causal rule induction processing on the block matrix, and extract the perturbation feedback path and causal influence relationship of the digital human under different perception inputs and historical behavior conditions; based on the perturbation feedback path and the causal influence relationship, construct the set of behavior adjustment rules including behavior trigger conditions and corresponding behavior response strategies.

[0027] Next, construct a multi-step behavior predictor. For example, based on the set of behavior adjustment rules, adopt a lower triangular prediction structure based on the perturbation feedback path to construct the multi-step behavior predictor.

[0028] Step S110: Dynamically adjust the behavior rules of the digital human based on the multi-step behavior predictor.

[0029] The multi-step behavior predictor outputs the behavior prediction results for multiple future time steps based on the perceptual inputs, historical behavior states, and environmental interference characteristics within the current sliding window. Based on the behavior probability distribution, response delay interval, and perturbation sensitivity score included in the prediction results, the behavior rules of the digital human are dynamically adjusted online.

[0030] Figure 2 It is another method for dynamically adjusting the digital human's rules based on causal relationships according to an embodiment of the present invention. As Figure 2 shown, this method includes the following steps:

[0031] Step S202: Obtain the historical data related to the interaction between the digital human and the user in the mixed reality scenario.

[0032] In this embodiment, a multi-modal sensor system (including a visual sensor, a depth camera, a voice collector, and an environmental sensor) deployed in a mixed reality (MR) scenario is used to continuously collect user behaviors (such as gaze trajectories, voice commands, gesture trajectories, etc.) and environmental information (such as scene lighting, object distribution, spatial sound effects, etc.). After the sensor data is fused by the synchronization fusion module, historical behavior data and historical environmental data in a standard format are generated. The acquisition period can be set from 10 frames per second to 30 frames per second, and is dynamically adjusted according to the processing capacity of the MR device.

[0033] In addition, to ensure the accuracy and integrity of the data, the sensors are calibrated before data acquisition, and time stamps are used to align the multi-source data. The visual sensor uses YOLOv8+DeepSORT for person and object recognition and tracking, the depth information uses a point cloud library (such as PCL) to build a spatial reconstruction model, and the voice information is transcribed into text instructions through a pre-trained speech recognition model (such as Wav2Vec2.0). Finally, a cross-modal unified behavior perception vector sequence is extracted through a temporal fusion network (such as a Transformer Encoder).

[0034] Step S204: Based on the historical data, determine the control step length and the perceptual sliding window length of the digital human's behavior decision, and construct a behavior decision data matrix based on the control step length and the perceptual sliding window length.

[0035] As Figure 3 shown, the method for constructing the behavior decision data matrix includes the following steps:

[0036] Step S2042: Extract dynamic time features.

[0037] In this embodiment, first, a time series feature vector group of interaction events is constructed based on historical data, including a fixation point movement speed sequence v_gaze(t), a gesture direction change sequence v_gesture(t), a voice phoneme start and end sequence v_voice(t), etc.; at the same time, the response behavior delay sequence r_delay(t), action start time r_start(t), reaction peak time r_peak(t), etc. of the digital human in the corresponding time period are extracted.

[0038] The dynamic time warping (DTW) method is used to align the above interaction-response pairs, construct a similarity matrix, and calculate its time series offset index Δτ, which represents the typical delay structure of the digital human's reaction triggered by interaction changes.

[0039] Subsequently, the statistical features (mean, standard deviation, local kurtosis, etc.) of these time series offset indexes are calculated for multiple rounds of interaction samples to form the dynamic time feature T_dyn, which reflects the trend stability and variability of the digital human's response time series.

[0040] Step S2044, determine the control step size and the perception sliding window length.

[0041] In this embodiment, a behavior stability index S_behave is constructed, which includes the following components: mean response delay μ_delay, standard deviation σ_delay, response amplitude consistency ρ_act, response mode confidence score_pattern; at the same time, the environmental perturbation amplitude E_var is evaluated, and the light intensity fluctuation value ΔI(t), environmental sound background change rate ΔS(t), object displacement speed v_obj(t) in the scene, etc. are collected.

[0042] T_dyn, S_behave, and E_var are combined to form a feature set Z_feat, which is input into a lightweight causal structure perception network (such as Causal-GRNN). This network is trained with a labeled interaction data set (such as Multimodal-MR-Sim) and can predict the appropriate control step size h (behavior update granularity) and perception sliding window length l (perception fusion time window) in the current scene.

[0043] To enhance the generalization ability of the estimation, adversarial sample training based on perturbation distribution simulation (such as adding high-frequency light changes or sudden sound effects interference) is introduced in the model training stage to improve the model's adaptability to extreme scenes.

[0044] Step S2046, construct a behavior decision data matrix.

[0045] Set the perception sliding window length to l and the control step size to h, and gradually divide several time periods from the historical data sequence in a sliding window manner. Each time period is represented as: [t, t + l], and it slides at an interval of h steps.

[0046] For each sliding window segment, the following are extracted separately: 1) the input trajectory u[t:t+l], which includes user instructions, environmental changes, etc.; 2) the state trajectory x[t:t+l], which includes the state of the digital human body, such as position, posture, expression parameters, etc.; 3) the interference features d[t:t+l), such as noise in the scene, light changes, abnormal user behaviors, etc.

[0047] The above three types of data are normalized and time-aligned to form a joint data block F_i that fuses behavior-environment-response. The set of joint data blocks corresponding to all sliding window segments is {F_1, F_2, ..., F_N}, where N depends on the number of sliding windows.

[0048] To improve the structural consistency and noise resistance, in some embodiments, the following processing can be further added when constructing each data block: use an occlusion-aware mechanism to perform multimodal imputation on missing / abnormal values; introduce a sine-cosine embedding of the timestamps within the sliding window (temporal positional embedding) to maintain time continuity; adopt a cross-channel attention module to perform weighted fusion on u, x, and d to generate a context-aware multi-channel joint vector as multiple joint data blocks.

[0049] Subsequently, the multiple joint data blocks are arranged in chronological order to form the behavior decision data matrix with a multi-time period causal structure. Specifically, the joint data blocks {F_1, ..., F_N} generated by the above N sliding window segments are stacked in the original chronological order to form a behavior decision data matrix D with dimensions N×M, where M is the multi-dimensional fusion feature dimension of each data block.

[0050] To enhance the causal structure expressiveness of the matrix, in this embodiment, a strict chronological order is maintained during the construction process without being disrupted, and a lower-triangular masking method is used to explicitly encode the causal direction of history-current.

[0051] Step S206, perform causal structure extraction processing on the behavior decision data matrix D to obtain a block matrix under causal constraints.

[0052] This embodiment introduces a causal structure orthogonal decomposition algorithm. First, perform lower-triangular orthogonal decomposition (LT-SVD) on the behavior decision data matrix D to obtain a lower-triangular causal transfer matrix L_causal and a residual term E_res. Use a perturbation feedback path identifier (constituted by a multi-head self-attention mechanism) to decouple the structure of L_causal, extract the lower-triangular block L_lower and mask the upper-triangular block to construct a block matrix B_cause under causal constraints.

[0053] To improve the modeling robustness, a perturbation path compression regularization term can be added during the decomposition process, and the objective function is as follows:

[0054] Min ||D - L_lower||² + λ1 * ||L_lower|| + λ2 * ||E_res||²,

[0055] Among them, the first term is the causal fitting error, the second term is the perturbation path sparsity regularization term, and the third term is the residual control term. In the formula, λ1 is the sparsity regularization coefficient, which is used to balance the weights of the main fitting term and the sparse term. The larger λ1 is, the fewer non-zero causal paths in L_lower will be. λ2 is the residual penalty term coefficient, which is used to control the penalty intensity of the fitting error. The larger λ2 is, the more strict the model's reconstruction requirement for D will be. The objective function in this embodiment can not only significantly weaken the pseudo-correlated paths, but also has scalability to adapt to different task scales.

[0056] In some other embodiments, to address the numerical stability problem in the high-dimensional matrix decomposition process, a block matrix recursive decoupling strategy can also be adopted. Only a part of the causal paths in the time window are processed each time for segmented compression, and then the overall causal graph is reconstructed through weighted averaging.

[0057] Step S208, generate a set of behavior adjustment rules that satisfy the causal relationship based on the block matrix.

[0058] As Figure 4 shown, the method for generating a set of behavior adjustment rules that satisfy the causal relationship includes the following steps:

[0059] Step S2082, extract the perturbation feedback path and the causal influence relationship.

[0060] In this embodiment, the perturbation feedback path is a causal path where the perceived input variable does not directly affect the behavior response, but rather has an indirect regulatory effect through the interference feature variable as an intermediate channel. First, a path tracking graph is constructed for each pair of variables in the block matrix B_cause, and the intermediate links involving the interference feature variable in the multi-hop paths are identified using the structure decomposition algorithm. If the combined weight of the path "perceived input → interference feature → behavior variable" exceeds the threshold, then this path is marked as a perturbation feedback path.

[0061] Next, the causal weight matrix W_c of each behavior decision dependent variable and the input variable in the block matrix B_cause is statistically calculated. This matrix represents the causal influence intensity between each pair of variables, and a numerically stable expression form is obtained through processing such as weighted summation and normalization of the causal graph structure.

[0062] To identify potential types of causal behaviors, an unsupervised clustering method is used to cluster the column vectors or subspaces of W_c. The clustering algorithm can be selected from spectral clustering, the EM algorithm based on Gaussian mixture models, or the sparse subspace clustering method based on mutual information distance metrics. Each cluster represents a behavioral response pattern with a similar causal structure. The final output is a set of causal path sets (i.e., causal influence relationships), and each type can correspond to a set of candidate behavior rules.

[0063] Step S2084, based on the perturbation feedback path and the causal influence relationship, construct the behavior adjustment rule set including the behavior trigger condition and the corresponding behavior response strategy.

[0064] According to the extracted perturbation feedback path and causal influence relationship, analyze the dynamic relationship between the input variables, interference intermediate variables, and behavior output involved in each path in the causal path set one by one, and convert them into rule entries in an interpretable structural form. For each perturbation feedback path, extract the joint conditions of its causal chain. For example, "change in speech intonation" combined with "sudden increase in background noise" can form a predictive causal condition for "speech response delay". If the path appears frequently in the training samples and has a significant causal weight, it is refined into a formal rule.

[0065] When constructing rules, use multi-dimensional constraints to represent the trigger boundary and behavior strategy output of each rule, including: 1) Perceived input conditions: generated by the fuzzy logic partition of the perception vector; 2) Historical behavior states: extract subclasses from the historical behavior label space; 3) Interference feedback states: determined by the activity distribution of the perturbation path nodes; 4) Behavior response strategies: including behavior categories, action parameters, response probability ranges, etc.; 5) Additional execution conditions: such as confidence thresholds, priority levels, time delay tolerances, etc.

[0066] All generated rules are organized into a hierarchical rule set, divided into a regular rule layer (processing basic stable paths), a perturbation adjustment layer (processing rules with intermediate feedback), and an anomaly avoidance layer (processing extreme / sudden inputs). In addition, a perturbation feedback scoring function can be introduced to weight each rule according to environmental fluctuations during the operation phase, enabling it to have the ability of online dynamic adjustment.

[0067] The finally generated behavior adjustment rule set can not only describe the response logic in complex MR scenarios, but also directly support the subsequent prediction scheduling and multi-strategy behavior control module calls.

[0068] Step S210, construct a causality-driven multi-step behavior predictor based on the behavior adjustment rule set.

[0069] In this embodiment, a recursive multi-step prediction network is constructed using a lower triangular prediction structure. The network input is the behavior-perception joint feature within the current sliding window, and the output is the behavior probability distribution for the next T steps. Each prediction step uses a gated recurrent unit (such as GRU) to encode the previous step state, and automatically selects the prediction path in combination with the causal trigger rules in Rule_base.

[0070] To enhance the multi-modal alignment ability of the model, an attention fusion module (Cross-Modal Attention) is introduced into the network, which assigns dynamic weights to the visual, speech, and environmental channels respectively. During the training phase, a multi-task loss function is used:

[0071] L_total = αL_pred + βL_rule + γL_cons

[0072] where L_pred is the behavior prediction error, L_rule is the rule deviation penalty, L_cons is the perturbation path consistency loss, and the hyperparameters α, β, and γ are automatically adjusted through Bayesian optimization.

[0073] During the training process, the Monte Carlo perturbation resampling mechanism is used to improve the robustness to uncertain scenarios, and at the same time, multi-source data augmentation (such as background change, noise superposition) is introduced to improve the generalization of the model.

[0074] Step S212: Dynamically adjust the behavior rules of the digital human based on the multi-step behavior predictor.

[0075] During the running phase, the MR device collects the current behavior-perception features per second and sends them to the predictor to obtain the candidate behavior sequences for the next T steps. The benefit of each candidate sequence is evaluated (such as the task completion rate, the estimated user response satisfaction), and the behavior plan with the maximum expected benefit is selected to dynamically update the current behavior rule table of the digital human and push it to the behavior execution module.

[0076] The dynamic update of the behavior rules is realized through a rule cache table. The system classifies the rules into three categories: high-frequency, low-frequency, and instant rules according to the change frequency, and configures different refresh frequencies and expiration policies for different categories. Instant rules support fast insertion and replacement within 15ms to ensure the responsiveness in high-interaction scenarios.

[0077] At the same time, when the environmental perturbation exceeds the stable threshold range, the behavior rule re-learning mechanism is triggered: the system sends the current behavior-environment data back to the behavior decision module to re-train the lightweight prediction model to achieve online fine-tuning of the behavior rules. This mechanism relies on the TinyML model deployed at the edge to quickly complete micro-parameter adjustment, avoiding frequent migration and update of the overall model, and improving the long-term operation stability of the system.

[0078] Through the above embodiments, the present invention not only improves the causal modeling ability and behavior prediction accuracy of the digital human in the complex MR environment, but also significantly enhances the adaptability of the digital human to sudden disturbances and complex user behaviors, with higher interaction naturalness and system robustness.

[0079] The present application also provides a digital human dynamic rule adjustment device based on causal relationship, as Figure 5 shown, including: an acquisition module 52, configured to acquire historical data related to the interaction between the digital human and the user in the mixed reality scenario, where the historical data includes historical behavior data and historical environment data; a construction module 54, configured to determine the control step length and the perception sliding window length of the behavior decision of the digital human based on the historical data, and construct a behavior decision data matrix based on the control step length and the perception sliding window length; an extraction module 56, configured to perform causal structure extraction processing on the behavior decision data matrix to obtain a block matrix under causal constraints; a generation module 58, configured to generate a behavior adjustment rule set that satisfies the causal relationship based on the block matrix, and construct a multi-step behavior predictor driven by causality based on the behavior adjustment rule set; an adjustment module 59, configured to dynamically adjust the behavior rules of the digital human based on the multi-step behavior predictor.

[0080] It should be noted that: for the digital human dynamic rule adjustment device based on causal relationship provided in the above embodiments, only the above division of each functional module is used as an example. In actual application, the above functions can be allocated to different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the digital human dynamic rule adjustment device based on causal relationship provided in the above embodiments and the embodiments of the digital human dynamic rule adjustment method based on causal relationship belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0081] Figure 6 The structure diagram of the electronic device suitable for implementing the embodiments of the present disclosure is shown. It should be noted that Figure 6 The electronic device shown is only an example and should not bring any limitation to the functions and usage scope of the embodiments of the present disclosure.

[0082] As Figure 6As shown, the electronic device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to the programs stored in the read-only memory (ROM) 1002 or the programs loaded from the storage section 1008 into the random access memory (RAM) 1003. In the RAM 1003, various programs and data required for system operation are also stored. The CPU 1001, ROM 1002, and RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0083] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc. and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as required. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as required so that a computer program read from it can be installed into the storage section 1008 as required.

[0084] The above description is only the preferred embodiment of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A method for dynamically adjusting the rules of a digital human based on causal relationships, characterized in that Including: Obtain historical data related to the interaction between the digital human and the user in the mixed reality scenario, where the historical data includes historical behavior data and historical environment data; Based on the historical data, determine the control step length and the perception sliding window length of the behavior decision of the digital human, and construct a behavior decision data matrix based on the control step length and the perception sliding window length; Perform causal structure extraction processing on the behavior decision data matrix to obtain a block matrix under causal constraints; Generate a behavior adjustment rule set that satisfies the causal relationship based on the block matrix, and construct a multi-step behavior predictor driven by causality based on the behavior adjustment rule set; Based on the multi-step behavior predictor, dynamically adjust the behavior rules of the digital human.

2. The method according to claim 1, characterized in that, Based on the historical data, determining the control step length and the perception sliding window length of the behavior decision of the digital human includes: Analyze the temporal relationship between the change rate of interaction events and the behavior response in the historical data, and extract dynamic time features reflecting the behavior response trend of the digital human; Based on the dynamic time features, evaluate the behavior stability index and the environmental disturbance amplitude of the digital human, and determine the control step length and the perception sliding window length based on the behavior stability index and the environmental disturbance amplitude.

3. The method according to claim 1, wherein Constructing a behavior decision data matrix based on the control step length and the perception sliding window length includes: Based on the control step length and the perception sliding window length, perform sliding window segmentation on the historical data, and respectively extract the input trajectory, state trajectory and interference features within each sliding window to construct multiple joint data blocks that integrate behavior, environment and response; Arrange the multiple joint data blocks in chronological order to form the behavior decision data matrix with a multi-time period causal structure.

4. The method according to claim 3, characterized in that, Performing causal structure extraction processing on the behavior decision data matrix to obtain a block matrix under causal constraints includes: Perform lower triangular orthogonal decomposition on the behavior decision data matrix to obtain a lower triangular causal transfer matrix representing the causal path of the behavior response; Perform structural decoupling of the perturbation feedback path on the lower triangular causal transfer matrix, retain the lower triangular block in the lower triangular causal transfer matrix and mask the upper triangular block to obtain the block matrix under causal constraints.

5. The method according to claim 4, wherein Generating a behavior adjustment rule set that satisfies the causal relationship based on the block matrix includes: Perform causal rule induction processing on the block matrix, and extract the perturbation feedback path and the causal influence relationship of the digital human under different perception inputs and historical behavior conditions; Based on the perturbation feedback path and the causal influence relationship, construct the behavior adjustment rule set including behavior trigger conditions and corresponding behavior response strategies.

6. The method according to claim 1, wherein Constructing a multi-step behavior predictor driven by causality based on the behavior adjustment rule set includes: Based on the behavior adjustment rule set, adopt a lower triangular prediction structure based on the perturbation feedback path to construct the multi-step behavior predictor.

7. A digital human dynamic rule adjustment device based on causal relationships, characterized in that, Including: An acquisition module configured to acquire historical data related to the interaction between the digital human and the user in the mixed reality scenario, where the historical data includes historical behavior data and historical environment data; A construction module, configured to determine a control step length and a perception sliding window length for the behavior decision of the digital human based on the historical data, and construct a behavior decision data matrix based on the control step length and the perception sliding window length; An extraction module, configured to perform causal structure extraction processing on the behavior decision data matrix to obtain a block matrix under causal constraints; A generation module, configured to generate a set of behavior adjustment rules that satisfy causal relationships based on the block matrix, and construct a multi-step behavior predictor driven by causality based on the set of behavior adjustment rules; An adjustment module, configured to dynamically adjust the behavior rules of the digital human based on the multi-step behavior predictor.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein when the program runs, it controls the device where the computer-readable storage medium is located to execute the method according to any one of claims 1 to 6.

9. A computer device, characterized in that, Comprising: A memory and a processor, the memory stores a computer program; the processor is used to execute the computer program stored in the memory, and when the computer program runs, it causes the processor to execute the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 6.

Images