Personality-driven virtual intelligent body long-time behavior generation system

Abstract

The invention provides a virtual agent long-time behavior generation system based on personality driving, which is used for guiding a virtual agent to continuously and dynamically carry out a series of activities in a three-dimensional environment by fully utilizing the reasoning capability of a large language model and combining a customized multi-layer behavior structure. According to the method, self-adaptive behavior planning of a complex environment is achieved through modeling of personality elements and real-time monitoring of scenes (including space, time, objects and historical activities), and finally the virtual intelligent agent shows vivid and lifelike personalized long-time behaviors; that is to say, multi-round iterative analysis is performed on the environment state by using the decision-making ability of the large language model in conditional reasoning and behavior sampling, and a corresponding behavior sequence is generated at each time step, so that the intelligent agent is ensured to maintain high sensitivity and elastic adaptation to scene change in a long time period.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of virtual reality, artificial intelligence and human - computer interaction, and particularly relates to a long - term behavior generation system of virtual intelligent agents driven by personality. Background Art

[0002] With the increasing maturity of VR and AR technologies, virtual intelligent agents with realistic appearance, credible behaviors and the ability to interact complexly with users or the environment in virtual scenarios have become one of the core elements to enhance user immersion and interaction experience. In recent years, large language models (LLMs) have demonstrated powerful reasoning and expression capabilities in natural language understanding and generation, attracting extensive attention in the academic and industrial fields. Some studies have begun to integrate LLMs into virtual intelligent agents to simulate language conversations, emotional expressions or simple situation reasoning of human roles. However, there are still obvious deficiencies in the issue of "how to generate long - term virtual intelligent agent behaviors with both situation awareness and personality drive for real - world three - dimensional scenes". Most existing methods focus on the following aspects:

[0003] 1. Single - action or short - term behavior generation

[0004] Many technical paths still remain at the behavior synthesis of instantaneous actions or limited - time - sequence interactions, such as action recognition and generation based on text, audio or scene information. Such methods usually lack the modeling of the "personality traits" (such as extroversion, introversion, prudence, friendliness, etc.) of virtual intelligent agents and are difficult to reflect the behavior evolution of intelligent agents in a relatively long time range.

[0005] 2. Pre - scripted or rule - driven

[0006] Some existing methods rely on pre - prepared scripts, rules or limited scene configurations during behavior generation. Although they can present a certain degree of intelligent agent behavior in specific and small - scale scenarios, when the environment changes dynamically and the task process is relatively complex, the writing and maintenance costs of scripts and rules are high, and both flexibility and scalability are limited.

[0007] 3. Lack of effective perception of dynamic elements in 3D scenes

[0008] Currently, some research applies large language models (LLMs) to a two-dimensional sandbox environment (hereinafter referred to as the 2D sandbox environment) to provide certain interaction and behavior generation capabilities for digital characters or agents. The so-called 2D sandbox environment refers to a virtual space mainly composed of planar coordinates, where the activities of multiple characters (or agents) are simulated through simple icons, pixel grids, or symbol systems. This environment usually focuses on the interaction logic between characters or between characters and objects, but does not have a high-fidelity modeling of factors such as precise position, orientation, and physical properties in three-dimensional space. Due to the limitations in scene complexity and physical simulation in the 2D sandbox environment, it pays more attention to aspects such as group behavior, social networks, or simplified resource management, and cannot comprehensively reflect the key elements such as "three-dimensional space layout, object interaction, and user operation behavior" required in real VR / AR scenarios.

[0009] In immersive three-dimensional scenarios such as VR / AR, virtual agents need to understand and respond to more complex environmental elements, such as the structure of different rooms, the spatial position of objects, passable paths and usage methods, the position and direction of users, etc. If only the methods in the 2D sandbox environment are used, it is often impossible to accurately plan paths and conduct dynamic interactions for agents in three-dimensional space, and there is also a lack of fine-grained perception of the object attributes and multi-dimensional context changes (such as time, historical activities, etc.) in the environment, resulting in the generated behaviors lacking the necessary realism and adaptability.

[0010] 4. The research on personalized behaviors of virtual agents is not sufficient

[0011] A few studies have indeed considered personalized elements, such as using the Big Five personality model in psychology (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) to add personality labels to agents. However, most of them only focus on short-term or single situations, and have not yet formed an automated and scalable method to simulate how agents make continuous behavior decisions according to their own personalities in a dynamic environment over a long period.

[0012] It can be seen that currently, most research methods focus on generating short-term or single-step actions, and are unable to effectively combine the personality traits of virtual agents with long-term dynamic behavior decision-making. Even if some solutions attempt to introduce scripts and preset rules to simulate character behaviors in two-dimensional or simple environments, it is difficult to update context information and dynamically evaluate the priority of requirements in more complex three-dimensional environments, nor can they adaptively maintain the personality consistency and context rationality of behaviors in long-term sequences. At the same time, existing technologies are still not flexible enough in environmental perception and interaction. Even using predefined scripts or rules can achieve certain results in small or fixed scenarios, but when the scene structure changes frequently, user interactions are diverse, or the task process is complex, the maintenance cost of scripts or rules increases exponentially, and the behavioral flexibility of agents is greatly reduced. There is also a lack of a hierarchical system expression for behavioral activities and an efficient model to evaluate the behavioral strategies that agents should adopt in different situations and time points. Summary of the Invention

[0013] To solve the problem that existing technologies often have difficulty maintaining the consistency and rationality of behavioral decisions when sudden changes occur in the scene, the present invention provides a long-term behavior generation system for virtual agents driven by personality, which uses a large language model to assist conditional reasoning and behavior sampling, can quickly reason new conditions under the updated world state, and generate new behavior choices, significantly improving the adaptability to complex situations and the flexible expression ability of personality preferences.

[0014] A long-term behavior generation system for virtual agents driven by personality, comprising a world state module and a behavior planner module based on a large language model; wherein, the world state module includes an internal factor unit and an external factor acquisition unit;

[0015] The internal factor unit is used to store the basic attributes of the agent and numerically represent the personality of the agent according to the Big Five personality model in psychology. Among them, the basic attributes include physiological characteristics, social attributes, interests and hobbies, and preferences; the five indicators of the Big Five personality include openness, conscientiousness, extraversion, agreeableness, and neuroticism;

[0016] The external factor acquisition unit is used to obtain the three-dimensional scene information of the environment where the agent is located, the current time, and the behaviors that the agent has completed;

[0017] The behavior planner module uses the large language model to perform conditional reasoning on the basic attributes of the agent, the numerically represented personality of the agent, and the external factors respectively, and obtains the reasoning results. Among them, the reasoning results corresponding to the basic attributes include the behavior tendencies, preferences, and personal habits of the agent, the reasoning results corresponding to the agent's personality are the behavior characteristics generated by the corresponding personality, and the reasoning results corresponding to the external factors include the information contained in the space where the agent is located, the current position of the agent, the spatial objects, and the navigable distance between the agent and the spatial objects;

[0018] The behavior planner module uses a large language model and combines Maslow's hierarchy of needs theory to conduct a needs analysis on the inference results, determining the current most urgent needs of the agent; then, among the predefined task types under the most urgent needs, it determines the most likely task with the highest execution probability at the current time; finally, among the various activities predefined under the most likely task, it determines the most likely activity with the highest execution probability as the final behavior, and predicts the interactive object selected by the final behavior and the activity duration.

[0019] Furthermore, when the behavior planner module predicts the final behavior at the current time, the agent executes the final behavior in the three-dimensional scene, and after completing the final behavior, records the execution log into the variable of the agent's completed behaviors in the external factors of the world state. The execution log includes the start time, end time, object used, and location information.

[0020] Furthermore, after the agent executes the final behavior in the three-dimensional scene, the external factor acquisition unit updates the scene information of the agent's environment, the current time, and the agent's completed behaviors according to the behavior impact. The behavior planner module re-conducts conditional reasoning and generates the next final behavior, thereby continuously generating a long-time sequence of behaviors until the user-specified duration is reached or the preset task goal is completed.

[0021] Furthermore, the needs of the agent include physiological needs, safety needs, and social needs.

[0022] Furthermore, the behavior planner module also sets different weights for the predefined task types under each need according to the numericalized personality of the agent. Among them, the higher the personality value of the agent's personality, the higher the weight of the corresponding task type.

[0023] Beneficial effects:

[0024] 1. The present invention provides a long-time behavior generation system for virtual agents driven by personality, which makes full use of the inference ability of the large language model and combines a custom multi-layer behavior structure to guide the virtual agent to continuously and dynamically carry out a series of activities in a three-dimensional environment; the present invention hopes to realize adaptive behavior planning for complex environments through the modeling of personality elements and the real-time monitoring of the scene (including space, time, object, historical activities), and finally enable the virtual agent to show vivid and realistic personalized long-time behaviors; that is to say, the present invention uses the decision-making ability of the large language model in "conditional reasoning" and "behavior sampling" to conduct multi-round iterative analysis of the environmental state, and generates a corresponding behavior sequence at each time step, ensuring that the agent remains highly sensitive and elastically adaptable to scene changes during a long period.

[0025] 2. The present invention provides a long-term behavior generation system for virtual intelligent agents driven by personality. Through a three-level structure of needs-tasks-activities (N-T-A), the personality characteristics of the intelligent agent are coupled with the dynamic elements of the three-dimensional scene, so as to accurately reflect its preferences, habits, and behavior motives in different situations. That is to say, the present invention sorts out the long-term behavior generation path through the N-T-A three-level structure system, enabling the intelligent agent to make timely decisions and adjustments as time and situation change while meeting multiple needs. Existing technologies often lack hierarchical modeling of the needs layer, task layer, and activity layer, and are also unable to dynamically reflect the continuous impact of personality characteristics on behavior within a longer time series range.

[0026] 3. The present invention provides a long-term behavior generation system for virtual intelligent agents driven by personality, emphasizing continuous perception and iterative update of three-dimensional environmental information. It can be adjusted at any time according to the object positions in the scene, user interactions, and the intelligent agent's own historical activities, thus greatly improving the realism, feasibility, and interactivity of the virtual intelligent agent's behavior. Existing technologies with 2D sandbox environments or simple script configurations usually lack comprehensive support for spatial positions, object connectivity relationships, and VR / AR requirements.

[0027] 4. The present invention provides a long-term behavior generation system for virtual intelligent agents driven by personality, which establishes an extensible and easily migratable general framework. Through an abstract natural language prompt structure and a unified description of the N-T-A hierarchical model, it can be applied to different VR / AR scenarios (such as virtual home, virtual hospital, educational simulation environment, etc.) without large-scale modification of the underlying logic. Compared with existing solutions that require separate scripting or custom rules for each scenario, it greatly saves maintenance and R & D costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a schematic diagram of the overall architecture provided by the present invention;

[0029] Figure 2 is a schematic diagram of the hierarchical representation of human behavior (N-T-A structure) provided by the present invention;

[0030] Figure 3 is a flowchart of "conditional reasoning" provided by the present invention;

[0031] Figure 4 is a flowchart of "behavior sampling" provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.

[0033] The present invention relates to a technical method for long-term behavior generation of virtual agents (Virtual Agents refer to digital characters or entities that can perform behaviors and interactions in virtual scenarios, hereinafter referred to as agents) using large language models (Large Language Models, hereinafter referred to as LLM) in virtual reality (VR) and augmented reality (AR) environments, and is a complete method integrating hierarchical behavior modeling and large language model (LLM) reasoning.

[0034] Specifically, a long-term behavior generation system for virtual agents driven by personality provided by the present invention includes a world state module and a behavior planner module based on a large language model; wherein, the world state module includes an internal factor unit and an external factor acquisition unit;

[0035] The internal factor unit is used to store the basic attributes of the agent and numerically represent the personality of the agent according to the Big Five personality model in psychology. Among them, the basic attributes include physiological characteristics, social attributes, hobbies and preferences; the five indicators of the Big Five personality include openness, conscientiousness, extraversion, agreeableness, and neuroticism;

[0036] The external factor acquisition unit is used to obtain the three-dimensional scene information of the environment where the agent is located, the current time, and the behaviors completed by the agent;

[0037] The behavior planner module uses the large language model to perform conditional reasoning on the basic attributes of the agent, the numerically represented personality of the agent, and the external factors respectively, and obtains the reasoning results. Among them, the reasoning results corresponding to the basic attributes include the behavior tendency, preferences, and personal habits of the agent, the reasoning results corresponding to the agent's personality are the behavior characteristics generated by the corresponding personality, and the reasoning results corresponding to the external factors include the information contained in the space where the agent is located, the current position of the agent, the spatial objects, and the navigable distance between the agent and the spatial objects;

[0038] The behavior planner module uses the large language model and combines Maslow's hierarchy of needs theory to perform a needs analysis on the reasoning results to determine the most urgent need of the agent at present; then determines the most likely task with the highest execution probability at the current time among the predefined task types under the most urgent need; finally determines the most likely activity with the highest execution probability among the predefined activities under the most likely task as the final behavior, and predicts the interactive object selected by the final behavior and the activity duration.

[0039] Further, as Figure 1As shown, the interaction relationship between the "World State" module and the "Behavior Planner" module is marked respectively, as well as the loop structure of the initial input and iterative output. The user-defined personality and scenario information will be directly connected to the behavior planning in this flowchart and generate a long-term activity sequence after multiple rounds of loops. Specifically, the method of the present invention mainly includes the following core links: Figure 1 In (a), the initialization input is defined, including the user-defined personality traits and basic attributes of the intelligent agent, and the structural information of the three-dimensional scene. The personality of the virtual intelligent agent is represented by the Big Five Personality, and the basic attributes are described in natural language, covering physiological characteristics, social attributes, interests and preferences. The 3D scene information includes the scene layout (represented by orange and blue dots), the relationship between scene objects, and the spatial position of the virtual intelligent agent and the objects. Figure 1 In the personality-driven behavior generator of (b), the "World State" continuously monitors the environmental state, while the "Behavior Planner" generates an activity sequence in an autoregressive manner. The world state module on the left continuously maintains and updates the global information (time, historical activities, current scene distribution, object positions, etc.); the behavior planner module selects the specific activity of the intelligent agent in the next step through conditional reasoning and behavior sampling, combined with a hierarchical behavior structure. Figure 1 In (c), the continuously generated long-term behavior sequence is shown, which is executed according to the selected activity and the state is continuously updated. The output is a series of activity sequences, where the activities marked in blue represent the current activity, and the rest represent the completed activities.

[0040] In Figure 1 , the arrow direction illustrates the flow of data or state: the updated result of the world state will be input into the behavior planner in the next iteration, and the behavior selected by the behavior planner will further affect the world state, realizing the automatic and cyclic generation of long-term behavior. The key improvements of the present invention are:

[0041] · Using the three-level Needs-Task-Activity (collectively referred to as N-T-A below) to express the long-term behavior of the virtual intelligent agent;

[0042] · Incorporating the Chain of Thought of the LLM into conditional reasoning and behavior sampling, enabling more delicate semantic understanding and decision-making in dynamic scenarios;

[0043] · Supporting personalized intelligent agents to perform adaptive behavior planning in a changing three-dimensional environment.

[0044] After the behavior planner module predicts the final behavior at the current time, the agent executes the final behavior in the three-dimensional scene, and after completing the final behavior, records the execution log in the variable of the agent's completed behavior in the external factors of the world state, where the execution log includes the start time, end time, used objects, and location information.

[0045] After the agent executes the final behavior in the three-dimensional scene, the external factor acquisition unit updates the scene information of the environment where the agent is located, the current time, and the agent's completed behavior according to the behavior impact. The behavior planner module re-performs conditional reasoning and generates the next final behavior, so as to continuously generate a long-time sequence of behaviors until the user-specified duration is reached or the preset task goal is completed.

[0046] Taking specific requirements and specific activities as examples, a long-time behavior generation system for a virtual intelligent agent based on personality drive of the present invention will be described in detail by way of example.

[0047] The first stage: World state initialization and data preparation

[0048] In this stage, the initialization of the agent's personality and environment is mainly completed. The world state is divided into two modules: internal factors (the personality and basic attributes of the virtual intelligent agent) and external factors (scene information, current time, and completed events). The agent's personality is numerically represented using the "Big Five Personality" model in psychology, including Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. The basic attributes include physiological characteristics, social attributes, hobbies, and preferences, etc. At the same time, scene information (such as rooms, areas, objects, passable paths, etc.) is also input and saved in the world state together with the agent's initial position, initial time, etc.

[0049] The second stage: Hierarchical behavior modeling

[0050] As Figure 2 shown, the present invention hierarchically subdivides long-time behaviors into three layers: the requirement layer N - the task layer T - the activity layer A from top to bottom in sequence. Each layer contains specific definitions of multiple optional elements. This hierarchical behavior representation method helps the implementation of the generation process, gradually advancing from broad goals to specific activities, as Figure 2 shown by the colored arrow lines in

[0051] · Requirement layer: Referring to Maslow's hierarchy of needs theory, the needs strongly associated with time periods in family / daily scenarios are abstracted as "physiological needs", "safety needs", and "social needs";

[0052] · Task layer: Define several specific task types under each type of requirement, such as task elements like "eating", "resting", "working", "socializing", etc.

[0053] · Activity layer: Each task is further refined into a list of executable activities. For example, the "eating" task can be split into activity elements such as "cooking", "ordering takeout", "eating refrigerated food", etc.

[0054] To ensure the richness of behaviors under various personality traits, the present invention pre-configures possible activity options in the task layer and the activity layer respectively. During the subsequent reasoning process, different weight settings will be made in the activity probability distribution or preference according to the different personalities of the agents. For example, an extroverted agent (with a high score in Extraversion) is more likely to choose the activity of "going out to meet friends" when the "social need" appears, while an introverted agent is more likely to choose activities that do not require too much social interaction, such as "chatting online" or "reading books".

[0055] The third stage: Conditional reasoning

[0056] Figure 3 shows how the present invention conducts multi-round reasoning based on the world state in the behavior planner; Figure 3 schematically shows that after receiving several input parameters in the LLM prompt of the LLM, the inferred results of each factor variable in the world state are output to guide the next sampling; The following gives Figure 3 an example of a round of conditional reasoning process included in: Q represents the input provided to the large language model (LLM) in the current step, and R represents the response generated by the LLM; The dashed box labeled "Symbol Definition" explains the specific meaning of each symbol; In this stage, the behavior planner uses the LLM to reason about the world state information and generates the inferred results of several "conditions", and these conditions cover the following specific external and internal influencing factors in the world state:

[0057] (1) The stage at the current time (such as early morning, afternoon, late at night, etc.), which has an obvious impact on the demand intensity;

[0058] (2) The personality traits of the agent, such as whether it is inclined to extroverted social interaction, etc.;

[0059] (3) Historical activity records and whether there are unfinished or ongoing tasks;

[0060] (4) The positions of objects and feasible paths in the scene (such as whether there is an empty room for exercise, or whether the kitchen can be used, etc.).

[0061] In Figure 3In this process, through natural language prompts and chain of thought, the LLM will integrate the above information, generate multiple detailed reasoning processes, and output the reasoning results, providing a basis for screening the specific activities in the next stage.

[0062] Phase 4: Behavior Sampling

[0063] Figure 4 It shows how to perform probabilistic sampling on activity options in combination with the LLM under a specific requirement or task type, and finally determine the next activity of the virtual agent step by step. The figure gives the key semantic information of activity selection and LLM generation examples, providing intuitive guidance for the subsequent execution. Taking the first activity of the virtual agent in the morning as an example, an iterative process is illustrated below; Q represents the input provided to the large language model (LLM) in the current step, and R represents the response generated by the LLM. The results after each layer of sampling are indicated by blue arrows; the meanings of specific symbols are explained in the dashed box labeled "Symbol Definition" in Figure 3 It is described in the dashed box labeled "Symbol Definition".

[0064] After completing the conditional reasoning, the behavior planner enters the behavior sampling stage. At this time, the system will generate probabilistic sampling results for the candidate options of the N-T-A three layers and make selections in sequence: First, select the most urgent or the most representative demand among the three demand types (physiological demand, safety demand, social demand) in the demand layer;

[0065] Secondly, conduct a probability assessment among the task layer elements available for selection under the corresponding demand level to determine the specific task most suitable for this period; after finally selecting the task, select a specific activity layer element that is most likely to be executed at the current time under this task level to execute, and predict the object that will be selected to complete the activity and its activity duration.

[0066] Phase 5: Activity Execution and World State Update

[0067] The agent executes the sampled activity in the three-dimensional scene and, after completion, records the execution log (start time, end time, objects used, location changes, etc.) into the "completed events" variable in the world state. The time variable will also be incremented accordingly, and new state changes (such as "position change", etc.) may occur to the scene objects. After that, it returns to the second phase to carry out the next round of conditional reasoning and activity sampling, thereby continuously generating a long-time sequence of behaviors. The entire process ends only when the specified duration by the user (such as 24 hours a day) or the preset task goal is completed.

[0068] It is worth noting that a core improvement of the present invention lies in the elaborate design of the LLM prompt, especially for the multi-round dialogue and the access to intermediate reasoning results during conditional reasoning and activity sampling, avoiding the limitations of traditional scripts or finite rules that cannot handle complex scenarios. At the same time, the present invention clearly splits each layer of N-T-A into levels, facilitating the expansion of new behavior types or customized personality configurations.

[0069] Thus, compared with ordinary rule systems, the main improvement points of the present invention in terms of method / modeling are as follows:

[0070] (1) Integrate the two elements of meeting personality traits and dynamic scenario changes into the hierarchical representation. Starting from the requirement layer, probabilistically select the task layer and the activity layer, fully reflecting the personalized behavior tendencies of the agent;

[0071] (2) In the reasoning process, utilize the natural language reasoning ability of the LLM, allowing for the introduction of richer context judgments and complex condition analyses, breaking the limitations of traditional reliance on if-then-else-style finite rules;

[0072] (3) Through the continuous iterative world state update mechanism, the generated behaviors can reflect the evolution of scenarios and states over a long time span, adapting to the environmental changes or user interaction requirements that may occur at any time in the VR / AR environment.

[0073] In summary, compared with the existing technical methods that mainly rely on predefined scripts or finite rules and often focus on short-term behavior generation, the present invention has the following outstanding advantages:

[0074] 1. Dynamic generation mechanism combining the three-level behavior structure of N-T-A with personality traits

[0075] By hierarchically representing the agent's requirements, tasks, and activities and incorporating the influence of personality characteristics in each layer of decision-making, the problems of "personality consistency" and "flexible adaptation to scenarios" in long-term behavior generation are solved.

[0076] 2. Continuous update mechanism of the world state module

[0077] By recording and iteratively updating external factors in the world state, including time, historical activities, and scene object information, the agent can always obtain the latest environmental changes and generate context-related behavior decisions.

[0078] 3. Using the LLM for conditional reasoning and behavior sampling

[0079] Adopt the reasoning method of natural language prompts and chain of thought, enhancing the adaptability to complex environments and personality requirements, and being more robust and scalable than traditional script or rule systems.

[0080] 4. Behavior Execution and Scene Interaction Adapted to the VR / AR Environment

[0081] In view of the characteristics of the three-dimensional virtual scene, this solution realizes multi-angle scheduling and adaptive decision-making of the behavior of virtual agents, expanding the application value in the fields of virtual training, educational simulation, social interaction, game entertainment, etc.

[0082] Certainly, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can certainly make various corresponding changes and deformations according to the present invention. However, these corresponding changes and deformations should all fall within the protection scope of the appended claims of the present invention.

Claims

1. A personality-driven virtual agent long-term behavior generation system, characterized in that: It includes a world state module and a behavior planner module based on a large language model; wherein the world state module includes an internal factor unit and an external factor acquisition unit; The internal factor unit is used to store the basic attributes of the intelligent agent and to digitize the personality of the intelligent agent according to the Big Five personality model of psychology, wherein the basic attributes include physiological characteristics, social attributes, interests and preferences; the five indicators of the Big Five personality include openness, conscientiousness, extroversion, agreeableness, and neuroticism; The external factor acquisition unit is used to acquire the three-dimensional scene information of the environment in which the agent is located, the current time, and the completed behavior of the agent; The behavior planner module uses a large language model to perform conditional reasoning on the basic attributes of the agent, the digitized agent personality, and external factors to obtain reasoning results, wherein the reasoning results corresponding to the basic attributes include the agent's behavioral tendencies, preferences, and personal habits; the reasoning results corresponding to the agent's personality are the behavioral characteristics generated by the corresponding personality; and the reasoning results corresponding to the external factors include the inclusion information of the space where the agent is located, the agent's current position, and the navigable distance between the space object and the agent; The behavior planner module uses a large language model and combines Maslow's need theory to perform demand analysis on the reasoning results to determine the most urgent needs of the intelligent agent at present; then determines the most likely task with the highest probability of execution at the current time among the task types predefined under the most urgent needs; finally, determines the most likely activity with the highest probability of execution as the final behavior among the various activities predefined under the most likely task, and predicts the interactive object selected by the final behavior and the duration of the activity.

2. A personality-driven virtual agent long-term behavior generation system as claimed in claim 1, characterized in that: When the behavior planner module predicts the final behavior at the current time, the agent executes the final behavior in the three-dimensional scene, and after completing the final behavior, records the execution log to the agent's completed behavior variable in the world state external factors, where the execution log includes the start time, end time, objects used, and location information.

3. A personality-driven virtual agent long-term behavior generation system as claimed in claim 2, characterized in that: After the agent performs the final behavior in the three-dimensional scene, the external factor acquisition unit updates the scene information of the agent's environment, the current time, and the completed behavior of the agent according to the influence of the behavior. The behavior planner module re-performs conditional reasoning and generates the next final behavior, thereby continuously generating long-term series of behaviors until the user-specified duration is reached or the preset task goal is completed.

4. A personality-driven virtual agent long-term behavior generation system as claimed in claim 1, characterized in that: The needs of intelligent agents include physiological needs, safety needs, and social needs.

5. A personality-driven virtual agent long-term behavior generation system as claimed in claim 1, characterized in that: The behavior planner module also sets different weights for each predefined task type under each requirement according to the numerical agent personality, wherein the agent personality with a higher personality value has a higher weight for the task type.

Images