Methods, devices, and equipment for controlling embodied intelligent agents in autonomous environmental domestication

Abstract

This disclosure provides a method, apparatus, and device for controlling an embodied intelligent agent in an autonomous environment, comprising: controlling the embodied intelligent agent to perform autonomous interactions to obtain multiple interaction trajectories; generating and storing external environment knowledge based on the multiple interaction trajectories; when the embodied intelligent agent performs an embodied task to be processed, determining spatial distribution priors and procedural rules matching the embodied task to be processed from the external environment knowledge according to the task instructions of the embodied task to be processed and the current environmental state observation; constructing contextual prompt information based on the task instructions, the current environmental state observation, and the matching spatial distribution priors and procedural rules, and inputting the contextual prompt information into the basic model of the embodied intelligent agent to obtain a decision action sequence; controlling the embodied intelligent agent to interact with the target environment according to the decision action sequence to complete the embodied task to be processed. This effectively improves the control efficiency of the intelligent agent.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the fields of artificial intelligence and natural language processing technology, and more specifically, to an embodied intelligent agent control method, apparatus, and device suitable for autonomous environment domestication. Background Technology

[0002] Agents based on Large Language Models (LLMs) have demonstrated strong general intelligence in many static reasoning tasks. However, in embodied planning scenarios, agents must actively gather information and make continuous decisions under partially observable conditions.

[0003] In related technologies, knowledge of a specific domain is injected into the intelligent agent through methods such as parameter fine-tuning, imitation learning, or the introduction of self-reflection prompts, in order to control the intelligent agent to perform corresponding tasks.

[0004] However, using existing methods results in inefficient agent control. Summary of the Invention

[0005] The embodiments described herein provide a method, apparatus, and device for controlling an embodied intelligent agent in an autonomous environment, overcoming the aforementioned problems.

[0006] Firstly, based on the content of this disclosure, a method for controlling an embodied intelligent agent in an autonomous environment is provided, comprising: Without updating the basic model parameters of the embodied agent, the embodied agent is controlled to autonomously interact in the training environment instance set to obtain multiple interaction trajectories. Based on the multiple interaction trajectories, external environment knowledge is generated and stored; the external environment knowledge includes spatial distribution priors and procedural rules; the spatial distribution priors are used to characterize the probability of the target object appearing in the candidate space, and the procedural rules are used to constrain the action execution conditions, action sequence, or interaction mode; When the embodied intelligent agent performs the embodied task to be processed, based on the task instructions of the embodied task to be processed and the current environmental state observation, the spatial distribution priors and procedural rules that match the embodied task to be processed are determined from the external environmental knowledge; Based on the task instructions, the current environmental state observations, and the matching spatial distribution priors and procedural rules, contextual prompts are constructed, and the contextual prompts are input into the basic model of the embodied intelligent agent to obtain a decision action sequence; The embodied intelligent agent is controlled to interact with the target environment according to the decision action sequence in order to complete the embodied task to be processed.

[0007] Secondly, according to the present disclosure, an embodied intelligent agent control device for autonomous environmental domestication is provided, comprising: The first control module is used to control the embodied agent to perform autonomous interactions in the training environment instance set without updating the basic model parameters of the embodied agent, so as to obtain multiple interaction trajectories. The generation module is used to generate and store external environment knowledge based on the multiple interaction trajectories; the external environment knowledge includes spatial distribution priors and procedural rules; the spatial distribution priors are used to characterize the probability of the target object appearing in the candidate space, and the procedural rules are used to constrain the action execution conditions, action sequence, or interaction mode; The determination module is used to determine, when the embodied agent is performing the embodied task to be processed, spatial distribution priors and procedural rules that match the embodied task to be processed from the external environment knowledge, based on the task instructions of the embodied task to be processed and the current environmental state observation; The construction module is used to construct contextual prompt information based on the task instructions, the current environmental state observations, and the matching spatial distribution priors and procedural rules, and input the contextual prompt information into the basic model of the embodied intelligent agent to obtain a decision action sequence; The second control module is used to control the embodied intelligent agent to interact with the target environment according to the decision action sequence in order to complete the embodied task to be processed.

[0008] Thirdly, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the embodied intelligent agent control method for autonomous environment domestication as described in any of the above embodiments.

[0009] This application provides a method for controlling an embodied agent based on autonomous environment training. This method controls the embodied agent to autonomously interact within a training environment instance set without updating the agent's base model parameters, thereby obtaining multiple interaction trajectories including environmental state observations, actions, environmental feedback, and task rewards. Based on these multiple interaction trajectories, on the one hand, the frequency of occurrence of the target object in multiple candidate spaces is statistically analyzed, and combined with the semantic probability between the target object and the candidate spaces, a spatial distribution prior is generated to characterize the spatial distribution pattern of the target object. On the other hand, based on the task rewards, comparison trajectory pairs are constructed from the interaction trajectories, and by analyzing the key divergent actions, environmental state differences, and failure reasons in the comparison trajectory pairs, procedural rules are generated to constrain action execution conditions, action order, or interaction methods.

[0010] Furthermore, spatially distributed priors and procedural rules are stored as external environmental knowledge. When the embodied agent executes the embodied task to be processed, the target environment corresponding to the embodied task can be a test environment that has not participated in autonomous interaction, and the test environment and the training environment instance set belong to the same environmental domain. Based on the task instructions and current environmental state observations, spatially distributed priors and procedural rules matching the embodied task to be processed are determined from the external environmental knowledge, and contextual prompts are constructed based on the task instructions, current environmental state observations, spatially distributed priors, and procedural rules. The basic model of the embodied agent outputs a sequence of decision actions based on the contextual prompts and controls the embodied agent to interact with the target environment according to the sequence of decision actions to complete the embodied task to be processed.

[0011] In the above manner, the embodiments of this application provide the spatial distribution priors and procedural rules obtained during the environment acclimatization stage to the embodied agent in the form of external environmental knowledge independent of the model parameters. This enables the embodied agent to transfer the environmental knowledge obtained from the training environment instance set to the target environment in the same environmental domain without updating the basic model parameters. The spatial distribution priors are used to narrow the search range of the target object, and the procedural rules are used to constrain the action planning process, thereby improving the search efficiency, planning stability and control efficiency in the execution of the embodied task.

[0012] The above description is merely an overview of the technical solutions of the embodiments of this application. In order to better understand the technical means of the embodiments of this application and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of this application more obvious and understandable, specific implementation methods of this application are described below. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments will be briefly described below. It should be understood that the drawings described below only relate to some embodiments of this disclosure and are not intended to limit this disclosure, wherein: Figure 1 This is a flowchart illustrating an embodied intelligent agent control method for autonomous environmental domestication provided in this disclosure.

[0014] Figure 2 This is a schematic diagram of the structure of an embodied intelligent agent control device for autonomous environmental domestication provided in this disclosure.

[0015] Figure 3 This is a schematic diagram of the structure of a computer device provided in this disclosure.

[0016] It should be noted that the elements in the attached diagram are schematic and not drawn to scale. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are also within the scope of protection of this disclosure.

[0018] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this subject matter pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having the meaning consistent with their meaning in the context of the specification and in the relevant art, and shall not be interpreted in an idealized or overly formal form unless otherwise explicitly defined herein. As used herein, the statement of “connecting” or “coupling” two or more parts together shall mean that these parts are directly joined together or joined through one or more intermediate components.

[0019] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of the phrase "embodiment" in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists, A and B exist simultaneously, or B exists. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Terms such as "first" and "second" are only used to distinguish one component (or part of a component) from another component (or another part of a component).

[0021] In the description of this application, unless otherwise stated, "multiple" means two or more (including two), and similarly, "multiple groups" means two or more (including two groups).

[0022] This embodiment decouples the agent's task execution process into an "environment interaction phase" and a "test application phase." In the environment interaction phase, the agent actively explores and learns in the target environment or a similar simulated environment without human intervention. During this process, the system collects the agent's interaction trajectories and utilizes the semantic understanding and reasoning capabilities of a large language model to extract the environment-specific "distributed priors" (guiding where to find the target) and "procedural rules" (guiding how to correctly manipulate objects). In the test application phase, the system injects the extracted environmental knowledge as external context prompts into the agent's control commands, guiding the agent to complete the final task planning and execution. The entire solution does not require updating the weight parameters of the agent's underlying model, achieving low-cost, highly adaptable environment alignment.

[0023] To enable those skilled in the art to better understand 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.

[0024] Figure 1 This is a flowchart illustrating an embodied intelligent agent control method for autonomous environment domestication provided in an embodiment of this disclosure. Figure 1 As shown, the specific process of the embodied intelligent agent control method for autonomous environment domestication includes: S110. Without updating the basic model parameters of the embodied agent, control the embodied agent to perform autonomous interactions in the training environment instance set to obtain multiple interaction trajectories.

[0025] Before the formal testing task is executed, the system can control the embodied agent to autonomously explore and experiment within a training environment instance set. The embodied agent can randomly interact with objects in the environment according to a preset heuristic exploration strategy. During the interaction, the system records the complete interaction trajectory, which includes: environmental state observations, actions output by the embodied agent, environmental feedback, and task rewards.

[0026] The interaction trajectory comprises multiple trajectory segments arranged in chronological order. Each trajectory segment includes at least one or more of the following: time step identifier, environmental state observation, task instruction, target object identifier, candidate space identifier, action output by the embodied agent, action execution result, environmental feedback, and task reward. Environmental state observation may include one or more of the following: visual observation, text observation, object detection result, scene graph, current location, visited space identifier, and list of interactive objects. Action execution result may include one or more of the following: action successful, action failed, target object discovered, target object not discovered, environmental state changed, or task terminated.

[0027] In some embodiments, the training environment instance set includes one or more training environment instances, which belong to the same environmental domain or the same task domain; for example, multiple home environments, kitchen environments, office environments, or other indoor environments with the same or similar object categories, spatial functional areas, or task operation constraints. External environmental knowledge is generated based on the interaction trajectories of the embodied agent in the training environment instance set and is used to control the embodied agent to perform embodied tasks in the target environment, wherein the target environment and the training environment instance set belong to the same environmental domain or the same task domain. Specifically, the embodied agent generates external environmental knowledge to characterize the spatial distribution patterns of objects and action execution constraints in the environmental domain by engaging in autonomous trial-and-error interactions within the training environment instance set. This external environmental knowledge can be used by the embodied agent to perform embodied tasks in the target environment; wherein the target environment can be an environment that has participated in autonomous trial-and-error interactions, or an environment that has not participated in autonomous trial-and-error interactions but belongs to the same environmental domain or the same task domain as the training environment instance set.

[0028] It should be noted that external environment knowledge is stored independently of the parameters of the embodied agent base model and is input into the embodied agent base model in the form of contextual prompts.

[0029] S120: Generates and stores external environment knowledge based on multiple interaction trajectories.

[0030] Among them, external environmental knowledge includes spatial distribution priors and procedural rules; spatial distribution priors are used to characterize the probability of the target object appearing in the candidate space, and procedural rules are used to constrain the execution conditions, action sequence, or interaction method of actions.

[0031] Candidate spaces can be determined based on the spatial functional areas, navigable areas, interactive containers, or object-carrying areas of the target environment. Candidate spaces include, but are not limited to, one or more of the following: rooms, cabinets, desktops, drawers, shelves, countertops, refrigerators, microwave ovens, areas near sinks, and areas near desks. The system can generate a set of candidate spaces based on environment maps, scene diagrams, object detection results, or a pre-defined candidate space vocabulary, and assign a candidate space identifier to each candidate space. For different environment instances, candidate spaces with the same or similar semantic functions can be mapped to the same candidate space category.

[0032] In some embodiments, external environment knowledge is generated and stored based on multiple interaction trajectories, including: Obtain the semantic confidence distribution output by the embodied agent or large language model regarding the relationship between the target object and the candidate space, and statistically analyze the frequency of empirical occurrence of the target object relative to each candidate space in multiple interaction trajectories. Normalize the semantic confidence distribution to a semantic probability distribution, and further normalize it according to the sum of the empirical occurrence frequency and the semantic probability distribution weighted by the concentration hyperparameter, to obtain the spatial distribution prior of the target object relative to each candidate space. Guide the embodied agent to perform at least one trial-and-error interaction based on the spatial distribution prior, and construct a comparison trajectory pair from the trajectories obtained from at least one trial-and-error interaction. The comparison trajectory pair includes: a baseline trajectory and a candidate improved trajectory with a higher cumulative reward relative to the baseline trajectory. Analyze the comparison trajectory pair to determine the key divergence actions and corresponding failure reasons of the baseline trajectory relative to the candidate improved trajectory, and generate or update procedural rules based on the key divergence actions and corresponding failure reasons. Store the spatial distribution prior and procedural rules as external environmental knowledge.

[0033] In embodied tasks (such as finding a specific item), the agent often needs to know "where to find it." Specifically, this involves utilizing the common-sense knowledge encoded within the large language model itself to determine the target object. e The semantic probability distribution is obtained by estimating the probability of an apple appearing in different containers or areas (such as refrigerators, cabinets, and tables). P sem ( r | e The statistical agent actually detects target objects at various locations in the corresponding environment. e The frequency of occurrence is denoted as the empirical frequency. N e,r An adaptive Bayesian aggregation of semantic probability distribution and empirical frequency of occurrence is performed to obtain the spatial distribution prior.

[0034] The baseline trajectory is the interaction trajectory obtained by the embodied agent when performing the target embodied task based on the knowledge of the current external environment; the candidate improvement trajectory is the exploration trajectory with the highest cumulative task reward among multiple exploration trajectories.

[0035] Environmental feedback includes at least one of positive, negative, and neutral feedback. Positive feedback includes one or more of the following: target object discovered, subtask completed, action executed successfully, search range narrowed, or task completed. Negative feedback includes one or more of the following: action execution failed, target object not discovered, entering irrelevant space, repeated search, violation of operational constraints, or task timeout. Neutral feedback includes no substantial change in the environmental state or the need for continued observation. The system calculates the cumulative task reward for each interaction trajectory based on the environmental feedback. The cumulative task reward can be determined based on one or more of the following: task completion reward, subtask completion reward, action failure penalty, repeated search penalty, path length penalty, or task time penalty. For example, the system can assign a positive reward when the target object is discovered or the task is completed, and assign a negative reward when the action fails, the same candidate space is repeatedly visited, or a preset number of steps is exceeded. The system then sums the reward values ​​corresponding to each time step in an interaction trajectory to obtain the cumulative task reward for that interaction trajectory.

[0036] Procedural rules are used to define the operational conditions, action order, or interaction constraints that an embodied agent must satisfy when performing a sequence of actions in an environment. By inputting comparative trajectory pairs into the rule analysis process, key divergent actions between the baseline trajectory and candidate improved trajectories, differences in environmental states corresponding to these key divergent actions, and reasons for task failure or low reward can be identified. Rule editing information is generated based on these key divergent actions, environmental state differences, and failure reasons. This information includes at least one of adding, deleting, modifying, or strengthening existing rules. Candidate procedural rules are then generated based on this information to arrive at the final procedural rules. Procedural rules are stored in the form of structured rule items. Each rule item includes one or more of the following: rule identifier, applicable task type, applicable environmental state, trigger condition, recommended action, prohibited action, action order constraint, confidence level, and update time. For example, a procedural rule could be expressed as: when the task type is target object search and the current candidate space is an openable / closable container, first execute the open action, then execute the observe or search action. The system can match corresponding procedural rules based on task instructions and current environmental state observations, and write the matched procedural rules into contextual prompts.

[0037] By storing spatial distribution priors and procedural rules as external environmental knowledge, embodied agents can narrow down the search space through spatial distribution priors and generate planned action sequences for performing embodied tasks through procedural rules.

[0038] In some embodiments, the spatial distribution prior of the target object relative to each candidate space is obtained by normalizing the sum of empirical occurrence frequency and semantic probability distribution weighted by concentration hyperparameters, including: For each candidate space, the product of the empirical occurrence frequency corresponding to the candidate space and the semantic probability and concentration hyperparameter corresponding to the candidate space is added to obtain the fusion value corresponding to the candidate space. The sum of the fusion values ​​corresponding to each candidate space is used as a normalization factor to normalize the fusion value corresponding to each candidate space, so as to obtain the spatial distribution prior of the target object relative to each candidate space.

[0039] Since relying solely on common sense can easily fail in specific environments, this embodiment utilizes Bayes' theorem to smoothly aggregate the semantic priors (i.e., semantic probabilities) of the large language model with empirical statistics of the environment. Spatial distribution priors As shown in formula (1) below.

[0040] (1) In formula (1), This represents the prior spatial distribution of the final generated object after fusion, i.e., the target object after environment alignment. e Located in the candidate space r Aggregate probability in; e This refers to the target object that the intelligent agent needs to find or manipulate during the task; r Represents the specific candidate space in the environment; Representing the candidate space r The corresponding aggregate value; This represents the total probability distribution of the target object across all candidate spaces; P sem ( r | e () represents the semantic probability distribution, i.e., the large model's inference of the target object. e In the candidate space r The probability of it; This represents the frequency of experience (i.e., the empirical probability distribution), which indicates the agent's position in the candidate space within the historical exploration trajectory. r The target object was actually found inside. e The number of times; This represents the concentration hyperparameter, used to adjust the influence weight of the semantic probability distribution in the spatial distribution prior; It represents all possible candidate spaces in the environment; by summing all possible regions, the calculation results are normalized to ensure that the sum of probabilities is valid, effectively converging the agent's search space in the training environment instance set and reducing the probability of blind exploration.

[0041] In some embodiments, constructing a comparison trajectory pair from trajectories obtained from at least one trial-and-error interaction includes: The interaction trajectory obtained by the embodied agent in performing the target embodied task based on the current external environment knowledge is used as the baseline trajectory; the sampling temperature parameter of the embodied agent in the action generation process is adjusted to obtain multiple exploration trajectories; and the exploration trajectory with the highest cumulative task reward is determined as the candidate improvement trajectory; if the cumulative task reward of the candidate improvement trajectory is higher than that of the baseline trajectory, the baseline trajectory and the candidate improvement trajectory are constructed as a comparison trajectory pair.

[0042] The system first controls the embodied agent to execute a target embodied task based on the current external environment knowledge, and uses the resulting interaction trajectory as the baseline trajectory. Then, by adjusting the sampling temperature parameter during the action generation process, the system controls the embodied agent to generate multiple exploration trajectories. For each exploration trajectory, the system calculates its cumulative task reward and selects the trajectory with the highest cumulative task reward from among the multiple exploration trajectories as a candidate for improvement. When the cumulative task reward of the candidate improved trajectory is higher than that of the baseline trajectory, the baseline trajectory and the candidate improved trajectory are constructed as a comparison trajectory pair; otherwise, the procedural rules are not updated based on this round of exploration, or subsequent trial-and-error interactions continue.

[0043] In some embodiments, the comparison trajectory pairs are analyzed to generate or update procedural rules, including: The comparison trajectory pairs are analyzed to obtain the key divergence actions and corresponding failure reasons of the baseline trajectory relative to the candidate improved trajectory; rule editing information is generated based on the key divergence actions, corresponding failure reasons, and / or differences in environmental state, and candidate procedural rules are generated based on the rule editing information; the candidate procedural rules are used to control the embodied agent to perform tasks on the training environment instance set to obtain the candidate overall task performance index; if the candidate overall task performance index is higher than the historical best task performance index, the candidate procedural rule is saved; if the candidate overall task performance index is not higher than the historical best task performance index, the candidate procedural rule is discarded.

[0044] The process involves generating candidate procedural rules based on rule editing information, and using these rules to control the embodied agent to perform tasks on a training environment instance set to obtain candidate overall task performance metrics. If the candidate overall task performance metrics are higher than the historical best task performance metrics, the candidate procedural rules are saved; otherwise, they are discarded to reduce the risk of procedural rule degradation due to accidental trajectories or erroneous reflections.

[0045] For example, in an embodied task of operating a microwave oven, the embodied agent executes the task based on knowledge of the current external environment to obtain a baseline trajectory, and generates multiple exploration trajectories through different sampling temperatures. If one of the exploration trajectories outperforms the baseline trajectory in terms of cumulative task reward, the system uses this exploration trajectory as a candidate for improvement and constructs a comparison trajectory pair with the baseline trajectory. By comparing the two trajectories, the system can identify key divergent actions in the baseline trajectory, such as attempting to start the microwave oven without confirming the power connection status, and generate procedural rules accordingly, such as "before starting the microwave oven, the power connection status and the door closing status should be confirmed first." This procedural rule is only saved if it improves the overall task performance of the embodied agent on the training environment instance set.

[0046] S130. When the embodied agent is performing an embodied task to be processed, it determines the spatial distribution priors and procedural rules that match the embodied task to be processed from the external environment knowledge, based on the task instructions of the embodied task to be processed and the current environmental state observation.

[0047] In some embodiments, based on the task instructions of the embodied task to be processed and observations of the current environmental state, spatial distribution priors and procedural rules matching the embodied task to be processed are determined from external environmental knowledge, including: Semantic encoding is performed on task instructions and current environmental state observations to obtain a task state vector. Knowledge indexes are established for spatial distribution priors and procedural rules in external environmental knowledge. Based on the similarity between the task state vector and the knowledge indexes, spatial distribution priors and procedural rules that match the embodied task to be processed are retrieved from the external environmental knowledge.

[0048] The knowledge index includes at least one of the following: target object identifier, candidate space identifier, task type identifier, environment state identifier, and corresponding semantic vector. By establishing targeted knowledge indexes, the search and matching scope can be effectively narrowed, improving the accuracy and efficiency of external environment knowledge retrieval. This allows the agent to quickly locate the external environment knowledge required for the current task without having to traverse the entire external knowledge base, further shortening the time required for knowledge matching.

[0049] Furthermore, when no external environmental knowledge matching the task instructions and current environmental state observations is found, the system can directly generate a sequence of decision actions based on the task instructions and current environmental state observations using the embodied agent base model, or generate a sequence of decision actions based on the default search strategy. When the retrieved spatial distribution prior is inconsistent with the current environmental state observations, the system can reduce the weight of the spatial distribution prior and prioritize the current environmental state observations to determine executable actions. When the embodied agent continuously outputs illegal actions or the number of failed action executions reaches a preset threshold, the system can re-retrieve external environmental knowledge, reconstruct contextual prompts, or trigger new autonomous trial-and-error interactions to update the external environmental knowledge.

[0050] S140. Based on task instructions, current environmental state observations, and matching spatial distribution priors and procedural rules, construct contextual cue information, and input the contextual cue information into the basic model of the embodied intelligent agent to obtain a decision action sequence.

[0051] The contextual information includes one or more of the following fields: task description, current state, spatial distribution prior, procedural rules, action constraints, and output format. Specifically, the task description field records the task instructions for the embodied task to be processed; the current state field records the current location, currently visible objects, visited candidate spaces, and the current set of executable actions; the spatial distribution prior field records the probability ranking of the target object relative to the candidate spaces or the priority search space; the procedural rules field records the action execution conditions, action order, or interaction constraints matching the current task; the action constraints field restricts the embodied agent from selecting actions from the current set of executable actions; and the output format field constrains the embodied agent to output one or more structured actions. For example, contextual prompts can include the following structured content: the task instruction is "find the target object"; the current state is "located at the kitchen entrance, with the refrigerator, counter, and microwave visible"; the spatial distribution prior is "the probability of the target object being located in the refrigerator, counter, and cabinet decreases in that order"; the procedural rules are "the opening action should be performed before searching for openable containers, and the oven door status should be checked before operating the microwave oven"; and the output requirement is "output the next action and its target object from the current set of executable actions".

[0052] The decision-making action sequence consists of actions from the set of legal actions. The set of legal actions can include one or more of the following: navigation actions, observation actions, search actions, grasping actions, placing actions, opening actions, closing actions, pressing actions, rotating actions, waiting actions, or task completion actions. The system can determine the current set of executable actions based on observations of the current environmental state. When a candidate action output by the embodied agent does not belong to the current set of executable actions, the system can reconstruct contextual hints to instruct the embodied agent to regenerate an action from the current set of executable actions, or map the candidate action to a semantically similar legal action.

[0053] In some embodiments, contextual cue information is constructed based on task instructions, current environmental state observations, and matching spatial distribution priors and procedural rules, including: Based on the spatial distribution prior, the candidate spaces related to the target object are sorted, and the candidate spaces that meet the preset sorting conditions are selected as the priority search spaces. The priority search spaces, procedural rules, task instructions, and current environmental state observations are combined according to the preset prompt template to obtain context prompt information.

[0054] Among them, contextual hints are used to constrain the embodied agent to prioritize searching for target objects in the priority search space and to determine the order or conditions for action execution according to procedural rules.

[0055] S150. Control the embodied intelligent agent to interact with the target environment according to the decision action sequence in order to complete the embodied task to be processed.

[0056] The target environment is the current execution environment corresponding to the embodied task to be processed.

[0057] Taking the task instruction "Help me get the blue water glass from the living room table" as an example, when the embodied agent receives this instruction, it obtains the observation information of the current environment (such as the position and category information of various objects in the living room), transforms the task instruction and environmental observation information into a query vector, and retrieves external environmental knowledge from the external environment knowledge database: "The blue water glass is usually placed on the living room table" and "The action of picking up the object requires first moving to the target position, then grabbing the object, and finally moving back to the starting position". By integrating the information and context prompts into the basic model, the model outputs a complete decision action sequence of "moving 1 meter towards the living room → moving the arm to the table coordinates → grabbing the blue water glass → moving 1.5 meters back to the current position". The embodied agent can then execute this sequence to complete the task.

[0058] Furthermore, for large and complex environments, this embodiment utilizes a Retrieval Augmentation (RAG) architecture to vectorize the extracted distributed priors and procedural rules and store them in a vector database. During test execution, the most relevant top-k rules are dynamically injected into the underlying large model based on the current observation state. The current environmental state observation can also be visual images or point cloud data, and the corresponding underlying large model can be a Vision-Language Model (VLM) to directly extract environmental features from the visual trial-and-error trajectory. This embodiment is not limited to relying on the self-reflection of a single agent; it can also introduce an independent "Critic Model" or establish a multi-agent debate mechanism to specifically compare trial-and-error trajectories and extract procedural rules, thereby improving the reliability of rule generation.

[0059] In this embodiment, without updating the basic model parameters of the embodied agent, the embodied agent is controlled to autonomously interact within the training environment instance set to obtain multiple interaction trajectories. Based on these multiple interaction trajectories, external environment knowledge is generated and stored. This external environment knowledge includes spatial distribution priors and procedural rules. The spatial distribution priors characterize the probability of a target object appearing in the candidate space, while the procedural rules constrain the execution conditions, action sequence, or interaction method. When the embodied agent executes the embodied task to be processed, the spatial distribution priors and procedural rules matching the embodied task are determined from the external environment knowledge based on the task instructions and current environmental state observations. Contextual prompts are constructed based on the task instructions, current environmental state observations, and the matching spatial distribution priors and procedural rules, and these prompts are input into the embodied agent's basic model to obtain a decision action sequence. The embodied agent is then controlled to interact with the target environment according to the decision action sequence to complete the embodied task. In this way, by providing the spatial distribution priors and procedural rules obtained during the environment domestication stage to the embodied agent in the form of external environmental knowledge independent of the model parameters, the embodied agent can transfer the environmental knowledge obtained from the training environment instance set to the target environment in the same environmental domain without updating the basic model parameters. It can use the spatial distribution priors to narrow the search range of the target object and use the procedural rules to constrain the action planning process, thereby improving the search efficiency, planning stability and control efficiency in the execution of the embodied task.

[0060] In summary, the agent receives exploration instructions and executes random or heuristic exploration tasks in the target environment. The system simultaneously records the agent's interaction logs with the environment, including state observations, action sequences, and environmental feedback (such as successful operation, failed operation, collision warnings, etc.), forming historical trajectory data. For the planning requirement of "object location finding," the system combines the inherent semantic intuition of the large language model with empirical statistics from interaction trajectories to construct a distributed prior. The "semantic probability" generated by the large model based on common sense (e.g., the probability of an apple being in the refrigerator) and the "empirical probability" obtained from actual interaction statistics are combined using Bayesian aggregation or weighted fusion. The extracted distributed prior is used to narrow the agent's search space during the execution phase, reducing the probability of blind exploration. For the planning requirement of "complex object interaction," the system uses the large model to compare and analyze successful and failed trajectories, generating procedural rules to regulate action execution. The system extracts key action nodes that lead to task failure (e.g., "directly opening a running microwave oven causes an error") and, combined with successful trajectories, derives environment-specific conditional operation rules (e.g., "the power must be turned off before operating the microwave oven"). With each interaction round, these rules are iteratively optimized and solidified into concise text instructions. When the agent receives a formal test application task, the system retrieves matching distributed priors and procedural rules based on the items and scenarios involved in the current task. The retrieved environmental knowledge is then concatenated with the task description to construct a system prompt, which is input to the agent. The agent generates and executes actions based on the injected external knowledge. This embodiment, by introducing distributed priors and procedural rules as external context, allows the agent to quickly adapt to new environments with unique layouts and specific interaction logic, effectively reducing the rate of repeated trial and error and the risk of target deviation during planning and execution. The entire knowledge extraction and injection process adopts a "plug-and-play" prompt engineering architecture, eliminating the need for expensive parameter fine-tuning of the underlying large language model, effectively avoiding the catastrophic forgetting problem caused by model fine-tuning, and enhancing the system's reusability. By utilizing self-generated comparison trajectories for self-reflection and optimization, the system eliminates the need for manual annotation of large amounts of high-quality operational demonstration data, thereby improving the feasibility of large-scale deployment in diverse scenarios.

[0061] Figure 2 This is a schematic diagram of the structure of an embodied intelligent agent control device for autonomous environment domestication provided in this embodiment. The embodied intelligent agent control device for autonomous environment domestication may include: The first control module 210 is used to control the embodied agent to perform autonomous interactions in the training environment instance set without updating the basic model parameters of the embodied agent, so as to obtain multiple interaction trajectories.

[0062] The generation module 220 is used to generate and store external environment knowledge based on multiple interaction trajectories. The external environment knowledge includes spatial distribution priors and procedural rules. Spatial distribution priors are used to characterize the probability of the target object appearing in the candidate space, and procedural rules are used to constrain the action execution conditions, action sequence, or interaction method.

[0063] The determination module 230 is used to determine, based on the task instructions of the embodied intelligent agent and the current environmental state observation, the spatial distribution priors and procedural rules that match the embodied task to be processed from the external environmental knowledge when the embodied intelligent agent is performing the embodied task to be processed.

[0064] The construction module 240 is used to construct contextual prompts based on task instructions, current environmental state observations, and matching spatial distribution priors and procedural rules, and input the contextual prompts into the basic model of the embodied intelligent agent to obtain a decision action sequence.

[0065] The second control module 250 is used to control the embodied intelligent agent to interact with the target environment according to the decision action sequence in order to complete the embodied task to be processed.

[0066] In some embodiments, the generation module 220 is specifically used for: Obtain the semantic confidence distribution output by the embodied agent or large language model regarding the relationship between the target object and the candidate space, and statistically analyze the frequency of empirical occurrence of the target object relative to each candidate space in multiple interaction trajectories. Normalize the semantic confidence distribution to a semantic probability distribution, and further normalize it according to the sum of the empirical occurrence frequency and the semantic probability distribution weighted by the concentration hyperparameter, to obtain the spatial distribution prior of the target object relative to each candidate space. Guide the embodied agent to perform at least one trial-and-error interaction based on the spatial distribution prior, and construct a comparison trajectory pair from the trajectories obtained from at least one trial-and-error interaction. The comparison trajectory pair includes: a baseline trajectory and a candidate improved trajectory with a higher cumulative reward relative to the baseline trajectory. Analyze the comparison trajectory pair to determine the key divergence actions and corresponding failure reasons of the baseline trajectory relative to the candidate improved trajectory, and generate or update procedural rules based on the key divergence actions and corresponding failure reasons. Store the spatial distribution prior and procedural rules as external environmental knowledge.

[0067] In some embodiments, the generation module 220 is specifically used for: For each candidate space, the product of the empirical occurrence frequency corresponding to the candidate space and the semantic probability and concentration hyperparameter corresponding to the candidate space is added to obtain the fusion value corresponding to the candidate space. The sum of the fusion values ​​corresponding to each candidate space is used as a normalization factor to normalize the fusion value corresponding to each candidate space, so as to obtain the spatial distribution prior of the target object relative to each candidate space.

[0068] In some embodiments, the generation module 220 is specifically used for: The interaction trajectory obtained by the embodied agent performing the target embodied task based on the current external environment knowledge is used as the baseline trajectory. The sampling temperature parameter of the embodied agent during the action generation process is adjusted to obtain multiple exploration trajectories. The exploration trajectory with the highest cumulative task reward is determined as the candidate improved trajectory. If the cumulative task reward of the candidate improved trajectory is higher than that of the baseline trajectory, the baseline trajectory and the candidate improved trajectory are constructed as a comparison trajectory pair. Correspondingly, the generation module 220 is also specifically used for calculating the cumulative task reward.

[0069] In some embodiments, the generation module 220 is specifically used for: The comparison trajectory pairs are analyzed to obtain the key divergence actions and corresponding failure reasons of the baseline trajectory relative to the candidate improved trajectory; rule editing information is generated based on the key divergence actions, corresponding failure reasons, and / or differences in environmental state, and candidate procedural rules are generated based on the rule editing information; the candidate procedural rules are used to control the embodied agent to perform tasks on the training environment instance set to obtain the candidate overall task performance index; if the candidate overall task performance index is higher than the historical best task performance index, the candidate procedural rule is saved; if the candidate overall task performance index is not higher than the historical best task performance index, the candidate procedural rule is discarded.

[0070] In some embodiments, the training environment instance set includes one or more training environment instances, which belong to the same environment domain or the same task domain; the external environment knowledge is generated based on the interaction trajectory of the embodied agent in the training environment instance set and is used to control the embodied agent to perform embodied tasks in the target environment, wherein the target environment and the training environment instance set belong to the same environment domain or the same task domain.

[0071] In some embodiments, the determining module 230 is specifically used for: Semantic encoding is performed on the task instructions and current environmental state observations to obtain a task state vector. Knowledge indexes are established for the spatial distribution priors and procedural rules in the external environmental knowledge, where each knowledge index includes at least one of the following: target object identifier, candidate space identifier, task type identifier, environmental state identifier, and corresponding semantic vector. Based on the similarity between the task state vector and the knowledge indexes, spatial distribution priors and procedural rules matching the embodied task to be processed are retrieved from the external environmental knowledge. Correspondingly, the determination module 230 is also specifically used for anomaly handling when no matching external environmental knowledge is found for the task instructions and current environmental state observations.

[0072] In some embodiments, the construction module 240 is specifically used for: Based on spatial distribution priors, candidate spaces related to the target object are sorted, and candidate spaces that meet preset sorting conditions are selected as priority search spaces. The priority search space, procedural rules, task instructions, and current environmental state observations are combined according to preset prompt templates to obtain context prompt information. The context prompt information is used to constrain the embodied agent to prioritize searching for the target object in the priority search space and to determine the action execution order or action execution conditions according to procedural rules.

[0073] The embodied intelligent agent control device for autonomous environmental domestication provided in this disclosure can execute the above-described method embodiments. Its specific implementation principle and technical effects can be found in the above-described method embodiments, and will not be repeated here.

[0074] This application also provides a computer device. Please refer to the following for details. Figure 3 , Figure 3 This is a basic structural block diagram of the computer device in this embodiment.

[0075] The computer device includes a memory 310 and a processor 320 that are interconnected via a system bus. It should be noted that only a computer device with memory 310 and processor 320 is shown in the figure; however, it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented alternatively. Those skilled in the art will understand that the computer device described herein is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0076] Computer devices can include desktop computers, laptops, handheld computers, and cloud servers. These devices allow for human-computer interaction with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0077] Processor 320 is typically used to perform overall operations of a computer device. In this embodiment, memory 310 is used to store program code or instructions, including computer operation instructions, and processor 320 is used to execute the program code or instructions stored in memory 310 or process data, such as program code that runs the methods described above.

[0078] In this article, the bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. This bus system can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0079] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0080] In the various embodiments of this application, the functional units or modules can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0081] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application.

[0082] In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" as described in this application does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims listing several means, several units of these means may be embodied by the same item of hardware. The use of "first," "second," and "third," etc., does not indicate any order and these words should be interpreted as names. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

[0083] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for controlling an embodied intelligent agent with autonomous environmental domestication, characterized in that, include: Without updating the basic model parameters of the embodied agent, the embodied agent is controlled to autonomously interact within a training environment instance set to obtain multiple interaction trajectories. Each interaction trajectory includes multiple trajectory segments arranged chronologically, and each trajectory segment includes at least one or more of the following: time step identifier, environmental state observation, task instruction, target object identifier, candidate space identifier, action output by the embodied agent, action execution result, environmental feedback, and task reward. The environmental state observation includes one or more of the following: visual observation, text observation, object detection result, scene graph, current position, visited space identifier, and list of interactive objects. The action execution result includes one or more of the following: action successful, action failed, target object discovered, target object not discovered, environmental state changed, or task terminated. Based on the multiple interaction trajectories, external environment knowledge is generated and stored. This external environment knowledge includes spatial distribution priors and procedural rules. The spatial distribution priors characterize the probability of a target object appearing in a candidate space, and the procedural rules constrain action execution conditions, action sequence, or interaction methods. Generating and storing external environment knowledge based on the multiple interaction trajectories includes: obtaining the semantic confidence distribution output by the embodied agent or large language model regarding the relationship between the target object and the candidate space; statistically analyzing the frequency of the target object's empirical occurrence relative to each candidate space in the multiple interaction trajectories; normalizing the semantic confidence distribution into a semantic probability distribution; and weighting it according to the frequency of empirical occurrence and concentration hyperparameters. The sum of the semantic probability distributions is normalized to obtain the spatial distribution prior of the target object relative to each candidate space; based on the spatial distribution prior, the embodied agent is guided to perform at least one trial-and-error interaction, and a comparison trajectory pair is constructed from the trajectories obtained from the at least one trial-and-error interaction; wherein, the comparison trajectory pair includes: a baseline trajectory and a candidate improved trajectory with a higher cumulative reward relative to the baseline trajectory; the comparison trajectory pair is analyzed to determine the key divergence actions and corresponding failure reasons of the baseline trajectory relative to the candidate improved trajectory, and procedural rules are generated or updated based on the key divergence actions and the corresponding failure reasons; the spatial distribution prior and the procedural rules are stored as external environmental knowledge. When the embodied intelligent agent performs the embodied task to be processed, based on the task instructions of the embodied task to be processed and the current environmental state observation, the spatial distribution priors and procedural rules that match the embodied task to be processed are determined from the external environmental knowledge; Based on the task instructions, the current environmental state observations, and the matching spatial distribution priors and procedural rules, contextual prompts are constructed, and the contextual prompts are input into the basic model of the embodied intelligent agent to obtain a decision action sequence; The embodied intelligent agent is controlled to interact with the target environment according to the decision action sequence in order to complete the embodied task to be processed.

2. The method according to claim 1, characterized in that, The spatial distribution prior of the target object relative to each candidate space is obtained by normalizing the sum of the frequency of occurrence of the empirical data and the semantic probability distribution weighted by the concentration hyperparameter, including: For each candidate space, the product of the frequency of occurrence of the experience corresponding to the candidate space and the semantic probability and concentration hyperparameter corresponding to the candidate space is added to obtain the fusion value corresponding to the candidate space. The sum of the fusion values ​​corresponding to each candidate space is used as a normalization factor to normalize the fusion values ​​corresponding to each candidate space, thereby obtaining the spatial distribution prior of the target object relative to each candidate space.

3. The method according to claim 1, characterized in that, Constructing a comparison trajectory pair from the trajectories obtained from the at least one trial-and-error interaction includes: The interaction trajectory obtained by the embodied intelligent agent in performing the target embodied task based on the current external environment knowledge is used as the reference trajectory; The sampling temperature parameters of the embodied intelligent agent during the action generation process are adjusted to obtain multiple exploration trajectories; and the exploration trajectory with the highest cumulative task reward is determined as the candidate improvement trajectory. If the cumulative task reward of the candidate improved trajectory is higher than the cumulative task reward of the baseline trajectory, the baseline trajectory and the candidate improved trajectory are constructed as the comparison trajectory pair.

4. The method according to claim 3, characterized in that, Analyzing the compared trajectory pairs to generate or update procedural rules includes: The comparison trajectory pairs are analyzed to obtain the key divergence actions and corresponding failure reasons of the baseline trajectory relative to the candidate improved trajectory; Based on the key divergent actions, corresponding failure reasons, and / or differences in environmental conditions, rule editing information is generated, and candidate procedural rules are generated based on the rule editing information. The candidate procedural rules are used to control the embodied agent to perform tasks on the training environment instance set in order to obtain candidate overall task performance indicators. If the overall performance index of the candidate task is higher than the historical best task performance index, the candidate procedural rule is saved. If the overall performance metric of the candidate task is not higher than the historical best task performance metric, the candidate procedural rule is discarded.

5. The method according to any one of claims 1 to 4, characterized in that, The training environment instance set includes one or more training environment instances, which belong to the same environment domain or the same task domain; the external environment knowledge is generated based on the interaction trajectory of the embodied agent in the training environment instance set and is used to control the embodied agent to perform embodied tasks in the target environment, wherein the target environment and the training environment instance set belong to the same environment domain or the same task domain.

6. The method according to any one of claims 1 to 4, characterized in that, Based on the task instructions of the embodied task to be processed and the current environmental state observations, spatial distribution priors and procedural rules matching the embodied task to be processed are determined from the external environmental knowledge, including: Semantic encoding is performed on the task instructions and the current environmental state observation to obtain a task state vector; Knowledge indexes are established for the spatial distribution priors and procedural rules in the external environment knowledge, respectively. The knowledge indexes include at least one of the following: target object identifier, candidate space identifier, task type identifier, environment state identifier, and corresponding semantic vector. Based on the similarity between the task state vector and the knowledge index, spatial distribution priors and procedural rules that match the embodied task to be processed are retrieved from the external environment knowledge.

7. The method according to claim 6, characterized in that, Contextual prompts are constructed based on the task instructions, the current environmental state observations, and matching spatial distribution priors and procedural rules, including: Based on the spatial distribution prior, the candidate spaces related to the target object are sorted, and the candidate spaces that meet the preset sorting conditions are selected as the priority search spaces. The priority search space, the procedural rules, the task instructions, and the current environment state observation are combined according to a preset prompt template to obtain the context prompt information; The contextual hints are used to constrain the embodied intelligent agent to prioritize searching for target objects in the priority search space and to determine the order or conditions for action execution according to the procedural rules.

8. A control device for an embodied intelligent agent capable of autonomous environmental domestication, characterized in that, include: The first control module is used to control the embodied agent to autonomously interact within a training environment instance set without updating the basic model parameters of the embodied agent, thereby obtaining multiple interaction trajectories. Each interaction trajectory includes multiple trajectory segments arranged in chronological order, and each trajectory segment includes at least one or more of the following: time step identifier, environmental state observation, task instruction, target object identifier, candidate space identifier, action output by the embodied agent, action execution result, environmental feedback, and task reward. The environmental state observation includes one or more of the following: visual observation, text observation, object detection result, scene graph, current position, visited space identifier, and list of interactive objects. The action execution result includes one or more of the following: action successful, action failed, target object discovered, target object not discovered, environmental state changed, or task terminated. The generation module is used to generate and store external environment knowledge based on the multiple interaction trajectories. The external environment knowledge includes spatial distribution priors and procedural rules. The spatial distribution priors characterize the probability of a target object appearing in the candidate space, and the procedural rules constrain action execution conditions, action order, or interaction methods. Specifically, the generation module is used to: obtain the semantic confidence distribution output by the embodied agent or large language model regarding the relationship between the target object and the candidate space; count the frequency of the target object's empirical occurrence relative to each candidate space in the multiple interaction trajectories; normalize the semantic confidence distribution into a semantic probability distribution; and weight the empirical occurrence frequency with a concentration hyperparameter. The sum of probability distributions is normalized to obtain the spatial distribution prior of the target object relative to each candidate space; based on the spatial distribution prior, the embodied agent is guided to perform at least one trial-and-error interaction, and a comparison trajectory pair is constructed from the trajectories obtained from the at least one trial-and-error interaction; wherein, the comparison trajectory pair includes: a baseline trajectory and a candidate improved trajectory with a higher cumulative reward relative to the baseline trajectory; the comparison trajectory pair is analyzed to determine the key divergence actions and corresponding failure reasons of the baseline trajectory relative to the candidate improved trajectory, and procedural rules are generated or updated based on the key divergence actions and the corresponding failure reasons; the spatial distribution prior and the procedural rules are stored as external environmental knowledge. The determination module is used to determine, when the embodied agent is performing the embodied task to be processed, spatial distribution priors and procedural rules that match the embodied task to be processed from the external environment knowledge, based on the task instructions of the embodied task to be processed and the current environmental state observation; The construction module is used to construct contextual prompt information based on the task instructions, the current environmental state observations, and the matching spatial distribution priors and procedural rules, and input the contextual prompt information into the basic model of the embodied intelligent agent to obtain a decision action sequence; The second control module is used to control the embodied intelligent agent to interact with the target environment according to the decision action sequence in order to complete the embodied task to be processed.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the embodied intelligent agent control method for autonomous environment domestication as described in any one of claims 1 to 7.

Images