Real-machine data processing methods and related equipment for intelligent agents

Abstract

This invention provides a method for processing real-device data of an intelligent agent and related equipment. The method includes: acquiring a target action acquisition task and a target action type; in response to the target action type being a long-term task, acquiring key continuous operation segment data in the long-term task; acquiring failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task; identifying and deleting invalid static frames from the key continuous operation segment data and recovery operation data using a pre-trained invalid frame recognition model to obtain cleaned data; evaluating the quality of the cleaned data based on a preset data quality assessment model to obtain a quality assessment score; and training the target intelligent agent using the cleaned data in response to the quality assessment score meeting a preset quality score threshold. This application can effectively reduce the redundancy of real-device data for intelligent agents and enrich data diversity, thereby improving the generalization ability of visual language action models.

Description

Technical Field

[0001] This application relates to, but is not limited to, the field of intelligent robot technology, and in particular to a method for processing real machine data of an intelligent agent and related equipment. Background Technology

[0002] With the development of artificial intelligence technology, the action capabilities of humanoid intelligent agents rely on the training of visual language action models, and the training effect of visual language action models is directly affected by the quality of the real machine data used.

[0003] In related technologies, existing real-device data acquisition and processing solutions have shortcomings in terms of efficiency and quality. On the one hand, the acquired data often contains a large amount of redundant information, resulting in low data utilization. On the other hand, datasets often lack diversity, especially in efficiently covering the continuous operation logic of long-term tasks, and lack data on recovery operations after failures, which limits the generalization ability of models in practical applications. Summary of the Invention

[0004] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application provides a method for processing real-machine data of intelligent agents and related equipment, which can effectively reduce the redundancy of real-machine data of intelligent agents and enrich data diversity, thereby improving the generalization ability of visual language action models.

[0005] In a first aspect, embodiments of this application provide a method for real-machine data processing of an intelligent agent, the method comprising:

[0006] Obtain the target action acquisition task and the target action type;

[0007] In response to the target action type being a long-term task, key continuous operation segment data in the long-term task are collected;

[0008] Collect failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task;

[0009] The cleaned data is obtained by identifying and deleting invalid static frames in the key continuous operation segment data and the recovery operation data through a pre-trained invalid frame identification model.

[0010] The purified data is evaluated based on a preset data quality assessment model to obtain a quality assessment score.

[0011] In response to the quality assessment score meeting a preset quality score threshold, the target agent is trained using the purified data.

[0012] In some embodiments, the step of collecting data on multiple key continuous operation segments in the long-term task in response to the target action type being a long-term task includes:

[0013] Identify repetitive operation units in the long-term task, wherein the repetitive operation units need to be executed repeatedly a first number of times in the long-term task;

[0014] The number of times the key continuous operation segment data is collected is determined to be a second number; wherein, the second number is less than the first number;

[0015] Data from the agent continuously executing the second number of repetitive operation units is collected and used as the key continuous operation segment data.

[0016] In some embodiments, the step of identifying and deleting invalid static frames from the key continuous operation segment data and the restored operation data using a pre-trained invalid frame identification model to obtain cleaned data includes:

[0017] When the motion change of the continuous frame sequence in the key continuous operation segment data or the recovery operation data is detected to be lower than the preset motion change threshold, a representative frame in the continuous frame sequence is extracted.

[0018] The invalid frame identification model is used to perform semantic analysis on the representative frames to determine whether the continuous frame sequence is a task-necessary still frame.

[0019] In response to the determination that the continuous frame sequence is a non-task-necessary static frame, it is identified as an invalid static frame and deleted.

[0020] In some embodiments, the pre-training process of the invalid frame identification model includes:

[0021] Acquire multimodal training data containing agent operations, wherein the multimodal training data includes scene images, action sequences, and corresponding natural language description labels;

[0022] The invalid frame recognition model is pre-trained by performing cross-modal feature alignment between the scene image and the natural language description label, and by performing a temporal context prediction task based on the action sequence data.

[0023] Using a set of labeled agent operation data that includes task-necessary stillness and non-task-necessary stillness, the invalid frame recognition model, which has been pre-trained, is fine-tuned for classification, so that the invalid frame recognition model can distinguish whether the continuous frame sequence belongs to task-necessary stillness or non-task-necessary stillness based on semantic information.

[0024] In some embodiments, before performing a quality assessment on the cleaned data based on a preset data quality assessment model to obtain a quality assessment score, the method further includes:

[0025] Calculate the inter-frame difference in the key continuous operation segment data and the recovery operation data;

[0026] If the inter-frame difference is greater than a first preset threshold, the data is processed by frame extraction at a first frame extraction frequency.

[0027] In response to the inter-frame difference being less than a second preset threshold, the data is processed by frame extraction at a second frame extraction frequency; wherein the first frame extraction frequency is less than the second frame extraction frequency.

[0028] In some embodiments, the step of assessing the quality of the cleaned data based on a preset data quality assessment model to obtain a quality assessment score includes:

[0029] The temporal consistency index is calculated based on the data video stream, joint angle sequence data, and torque data in the purification data.

[0030] Based on the overall characteristics of the purified data, the information entropy is calculated;

[0031] After training the visual language action model for controlling the target agent based on the purified data, the success rate of the visual language action model controlling the target agent to perform the target action is obtained.

[0032] The quality assessment score is calculated based on the time sequence consistency index, the information entropy, and the action success rate.

[0033] In some embodiments, after training the target agent with the cleaned data in response to the quality assessment score meeting a preset quality score threshold, the method further includes:

[0034] Based on the action success rate obtained by the target intelligent agent after training with the purified data, the adjustment weight of the temporal consistency index and the information entropy in the next round of quality assessment is calculated.

[0035] Based on the success rate of the action, determine the motion change threshold and frame sampling frequency to be used by the invalid frame identification model in the next round of data processing.

[0036] Secondly, embodiments of this application provide a real-machine data processing device for an intelligent agent, comprising:

[0037] The acquisition module is used to acquire the target action acquisition task and the target action type;

[0038] The first acquisition module is used to acquire key continuous operation segment data in the long task in response to the target action type being a long-term task;

[0039] The second acquisition module is used to acquire failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task.

[0040] The identification module is used to identify and delete invalid static frames in the key continuous operation segment data and the recovery operation data through a pre-trained invalid frame identification model to obtain cleaned data;

[0041] The quality assessment module is used to assess the quality of the purified data based on a preset data quality assessment model and obtain a quality assessment score.

[0042] The training module is used to train the target agent using the cleaned data in response to the quality assessment score meeting a preset quality score threshold.

[0043] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the real-machine data processing method of the intelligent agent as described in any one of the embodiments of the first aspect of this application.

[0044] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that is executed by a processor to implement the real-machine data processing method for an intelligent agent as described in any one of the embodiments of the first aspect of this application.

[0045] The real-machine data processing method for intelligent agents according to the embodiments of this application has at least the following beneficial effects:

[0046] The real-device data processing method for an intelligent agent according to embodiments of this application includes: acquiring a target action acquisition task and a target action type; in response to the target action type being a long-term task, acquiring key continuous operation segment data in the long-term task; acquiring failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task; identifying and deleting invalid static frames in the key continuous operation segment data and recovery operation data using a pre-trained invalid frame recognition model to obtain cleaned data; performing quality assessment on the cleaned data based on a preset data quality assessment model to obtain a quality assessment score; and training the target intelligent agent using the cleaned data in response to the quality assessment score meeting a preset quality score threshold.

[0047] This application clarifies the specific scenario and task attributes for data collection by acquiring the target action acquisition task and the target action type. Then, by recognizing the target action type as a long-term task, it collects key continuous operation segments within the long-term task, avoiding the large amount of repetitive and redundant information inherent in full-process acquisition, and significantly improving acquisition efficiency while preserving the core temporal logic of the task. Next, by collecting failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task, it incorporates possible anomalies and error correction processes from real-world operations into the training scope, compensating for the lack of negative samples and recovery logic in existing datasets. Subsequently, a pre-trained invalid frame recognition model identifies and removes invalid static frames from key continuous operation segments and recovery operation data, resulting in purified data. This purified data can be precisely removed based on semantic analysis, further improving the effective density of the data. Finally, the purified data is evaluated using a pre-set data quality assessment model, and the target agent is trained when a threshold is met, ensuring that the data used for model learning has both high signal-to-noise ratio and high reliability. Compared with existing technologies, this application can effectively reduce the redundancy of real machine data for intelligent agents and enrich data diversity, thereby improving the generalization ability of visual language action models.

[0048] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0049] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0050] Figure 1 A flowchart illustrating an optional real-machine data processing method for an intelligent agent, provided as an embodiment of this application;

[0051] Figure 2 A flowchart of another optional real-machine data processing method for an intelligent agent provided in an embodiment of this application;

[0052] Figure 3 A flowchart of another optional real-machine data processing method for an intelligent agent provided in an embodiment of this application;

[0053] Figure 4 A flowchart of another optional real-machine data processing method for an intelligent agent provided in an embodiment of this application;

[0054] Figure 5 A flowchart of another optional real-machine data processing method for an intelligent agent provided in an embodiment of this application;

[0055] Figure 6A flowchart of another optional real-machine data processing method for an intelligent agent provided in an embodiment of this application;

[0056] Figure 7 A flowchart of another optional real-machine data processing method for an intelligent agent provided in an embodiment of this application;

[0057] Figure 8 A schematic diagram of a real-machine data processing device for an intelligent agent provided in an embodiment of this application;

[0058] Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0059] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0060] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," "exceeding," etc. are understood to exclude the stated number, while "above," "below," "within," etc. are understood to include the stated number. Where "first" or "second" is mentioned, it is only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the sequential relationship of the indicated technical features.

[0061] In the description of this application, it should be understood that the orientation descriptions, such as up, down, left, right, front, and back, are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0062] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0063] In the description of this application, it should be noted that, unless otherwise explicitly defined, terms such as "setting," "installation," and "connection" should be interpreted broadly. Those skilled in the art can reasonably determine the specific meaning of the above terms in this application based on the specific content of the technical solution. Furthermore, the identification of specific steps in the following text does not imply a limitation on the order of steps or execution logic. The execution order and logic between each step should be understood and inferred from the content described in the embodiments.

[0064] With the development of artificial intelligence robot technology, humanoid robots, as carriers of embodied intelligence, rely on the training of visual language action models for their motor capabilities. The training effect of visual language action models is directly affected by the quality of the real machine data used.

[0065] In related technologies, existing real-device data acquisition and processing solutions have shortcomings in terms of efficiency and quality. On the one hand, the acquired data often contains a large amount of redundant information, resulting in low data utilization. On the other hand, datasets often lack diversity, especially in efficiently covering the continuous operation logic of long-term tasks, and lack data on recovery operations after failures, which limits the generalization ability of models in practical applications.

[0066] Based on this, this application, by acquiring the target action acquisition task and the target action type, can clearly define the specific scenario and task attributes for data acquisition. Then, by responding to the target action type being a long-term task, key continuous operation segments within the long-term task are acquired, avoiding a large amount of repetitive and redundant information from full-process acquisition, and significantly improving acquisition efficiency while preserving the core temporal logic of the task. Next, by acquiring failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task, possible anomalies and error correction processes in real-world operations can be incorporated into the training scope, compensating for the lack of negative samples and recovery logic in existing datasets. Subsequently, a pre-trained invalid frame recognition model identifies and deletes invalid static frames from key continuous operation segments and recovery operation data, resulting in purified data. This purified data can be precisely removed based on semantic analysis, further improving the effective density of the data. Finally, the purified data is evaluated for quality using a pre-set data quality assessment model, and the target agent is trained when a threshold is met, ensuring that the data used for model learning has both high signal-to-noise ratio and high reliability. Compared with existing technologies, this application can effectively reduce the redundancy of real machine data for intelligent agents and enrich data diversity, thereby improving the generalization ability of visual language action models.

[0067] Please see Figure 1 The present invention provides a real-device data processing method for an intelligent agent, which may include, but is not limited to, the following steps 101 to 106:

[0068] Step 101: Obtain the target action acquisition task and the target action type.

[0069] Step 102: In response to the target action type being a long-range task, collect key continuous operation segment data from the long-range task.

[0070] Step 103: Collect failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task.

[0071] Step 104: Identify and delete key continuous operation segment data and invalid static frames in the restored operation data using a pre-trained invalid frame identification model to obtain cleaned data.

[0072] Step 105: Based on the preset data quality assessment model, the purified data is assessed to obtain a quality assessment score.

[0073] Step 106: In response to the quality assessment score meeting the preset quality score threshold, the target agent is trained by cleaning the data.

[0074] The target intelligent agent includes, but is not limited to, robots (such as wheeled robots, legged robots, and flexible robots), robotic arms (such as industrial robotic arms and collaborative robotic arms), and end effectors (such as dexterous hands and grippers), etc., which are embodied intelligent devices. To facilitate understanding and explanation of the technical solutions of this application, the specification and drawings specifically use robots (such as flexible robots or humanoid robots) as typical application examples of intelligent agents for detailed description; however, this exemplary description should not be regarded as a limitation on the scope of application of the intelligent agents of this application, and those skilled in the art can extend and apply this solution to other intelligent agent forms such as robotic arm collaboration or end effector control according to actual needs.

[0075] In step 101 of some embodiments, the target action acquisition task may specifically include complex operation processes such as package sorting, grasping, left-right hand interaction, placement, and button pressing by a humanoid robot in a smart warehousing scenario; or tasks such as clothing sorting in a home service scenario. The target action type can be specifically divided into types for isolated actions or short-range tasks, and types for long-range tasks requiring continuous operation. Long-range tasks refer to a sequence of tasks containing multiple repetitive operation units (such as continuously grasping all clothes in a box) and having a strong temporal dependency, which is different from single, discrete robotic arm movements or static grasping actions.

[0076] In step 102 of some embodiments, for complex tasks involving a large number of repetitive action units (e.g., continuously grabbing dozens of items from a box), a sparse sampling strategy can be used instead of full recording. That is, instead of recording the entire complete task process, representative sequences of continuous actions are extracted. These sequences fully preserve the start and end logic and causal relationships of the actions (e.g., the complete chain of "identification-grabbing-moving-placement"), but eliminate redundant repetitive loops, thereby significantly reducing data storage costs while preserving the core temporal logic of long-term tasks.

[0077] Please see Figure 2 In some embodiments, step 102 may include, but is not limited to, steps 201 to 203.

[0078] Step 201: Determine the repetitive operation unit in the long-running task. The repetitive operation unit needs to be executed repeatedly a first number of times in the long-running task.

[0079] Step 202: Determine the number of times the key continuous operation segment data is collected as the second number.

[0080] Step 203: Collect data from the agent's continuous execution of a second number of repetitive operation units as key continuous operation segment data.

[0081] In step 201 of some embodiments, the repetitive operation unit specifically refers to the basic action loop that constitutes a cyclical operation in a long-term task. For example, in the task of a robot sorting clothes, the complete action flow of "identifying clothes - grasping clothes - folding - placing clothes" constitutes a repetitive operation unit. The first number refers to the total number of repetitions required to complete the current long-term task. For example, if there are 50 pieces of clothing to be sorted in the box, the first number is 50. By decomposing the long-term task through a task analysis algorithm that parses the task instructions or demonstrations, the homogeneous sub-task structures and the total amount of data that need to be executed in the complete business logic are accurately identified, thereby establishing the time span and repetition frequency benchmark of the full data.

[0082] In step 202 of some embodiments, the number of times the key continuous operation segment data is collected (i.e., the second number) is set to a value much smaller than the first number. For example, in the scenario of sorting 50 pieces of clothing, the second number is set to 2 or 3 times. The determination of the second number is based on a balance between model learning efficiency and data redundancy, aiming to retain only the minimum sample set sufficient to cover the core logic and action connection features of the task through a sparse sampling strategy. This value ensures that the amount of data collected is significantly lower than the amount of data collected during full execution, thereby filtering out a large number of repetitive samples that contribute very little to the model's generalization ability.

[0083] In step 203 of some embodiments, collecting data on the robot continuously executing a second number of repetitive operation units specifically refers to controlling the robot to sequentially complete a set number of repetitive operation units (e.g., twice consecutively) during a continuous, uninterrupted operation, and simultaneously recording multimodal sensor data (e.g., vision, joint angles, torque, etc.) during this process. This continuous acquisition differs from discrete acquisition of single actions; it fully preserves the transitional connection between two adjacent repetitive units (e.g., the reset path of the robotic arm moving from the placement area back to the grasping area after the first placement), thereby ensuring that the generated key continuous operation segment data contains the complete temporal dependencies and causal logic chain of task execution.

[0084] Through steps 201 to 203 described above, this embodiment first establishes the basis for data compression by identifying repetitive structures in the task; secondly, by setting a collection number much smaller than the total task volume, efficient compression of long-term task data is achieved; finally, by continuously collecting a specified number of repetitive units, both the loss of temporal information caused by fragmented collection and the data redundancy caused by full collection are avoided. This method enables the robot model to learn a long-term task execution strategy containing the complete logic of "action-reset-reaction" at extremely low data cost, improving data collection efficiency and model training effect.

[0085] In step 103 of some embodiments, operational errors that occur during actual operation due to environmental interference or control deviations, as well as the subsequent error correction process, are actively recorded. Unlike idealized datasets that only contain successful cases, this method specifically collects negative samples (such as grasping slippage, movement collisions, etc.) and posture adjustments and path replanning operations performed by the operator or control algorithm to remedy errors. This enables the model to autonomously escape invalid operation loops when faced with similar anomalies, enhancing the robot's adaptability in special environments.

[0086] In step 104 of some embodiments, an intelligent model with semantic understanding capabilities is used to process the raw data stream to filter out meaningless pauses in the screen caused by operator thinking, waiting for instructions, or equipment debugging during remote operation. This process is not only based on pixel-level motion detection, but also focuses on analyzing the intention behind the stillness of the screen, thereby accurately distinguishing between "non-task-necessary invalid pauses" that should be deleted and "task-necessary waiting or observation" that should be retained. Through this deep semantic analysis, redundant noise can be accurately removed, while ensuring that key timing nodes are not mistakenly deleted, resulting in high-purity data with higher information density.

[0087] Please see Figure 3 In some embodiments, step 104 may include, but is not limited to, steps 301 to 303.

[0088] Step 301: When the motion change of the continuous frame sequence in the key continuous operation segment data or recovery operation data is detected to be lower than the preset motion change threshold, extract representative frames from the continuous frame sequence.

[0089] Step 302: Use the invalid frame identification model to perform semantic analysis on representative frames to determine whether the continuous frame sequence is necessary for the task to remain still.

[0090] Step 303: In response to the determination that the continuous frame sequence is not necessary for the task to remain still, it is determined as an invalid static frame and deleted.

[0091] In step 301 of some embodiments, for the raw video stream data generated by the robot during long-term tasks or failure recovery, a motion detection algorithm based on pixel differences can be used to quantify the dynamic changes between frames. Specifically, the average Euclidean distance between feature points of adjacent image frames in a continuous frame sequence or the mean square error of all pixels in the image is calculated as the motion change value. When this value is lower than a preset motion change threshold (e.g., set to 2% of the total pixel change or a specific displacement of 5 pixels), it is determined that the scene in the current time window is physically still. At this time, the entire segment of still data is not directly processed. Instead, a downsampling strategy is used to extract one or more frames (such as the median frame of the sequence or the frame with the clearest key features) from the continuous frame sequence as representative frames. These representative frames carry the static visual scene information within that time period.

[0092] In step 302 of some embodiments, the extracted representative frames are input into a pre-trained invalid frame recognition model. This model is built on a vision-language multimodal architecture and has the reasoning ability to map visual signals to a semantic space. The model performs semantic analysis by calculating the cosine similarity or classification probability between the image feature vector of the representative frame and the text feature vector corresponding to the preset state labels (such as "waiting for collaboration", "observing the environment", "operator pause", "equipment debugging"). This determines whether the current continuous frame sequence belongs to "task-necessary stillness" (i.e., reasonable waiting that is part of the task logic, such as waiting for the conveyor belt to arrive) or "non-task-necessary stillness" (i.e., redundant stagnation that is unrelated to the task logic, such as the operator's thinking gap or communication delay). For example, if the matching confidence of the representative frame with the "waiting for collaboration" label is 0.92, while the matching confidence with "operator thinking gap" is only 0.05, it is determined to be task-necessary stillness.

[0093] In step 303 of some embodiments, selective pruning is performed on the original data stream based on the semantic decision result of the invalid frame identification model. Once a continuous frame sequence is identified as "non-task-necessary static" (i.e., invalid static frames), a deletion operation is performed, completely removing all frames within that time period from the temporal data. The valid action segments before and after the truncation are then temporally spliced ​​or smoothly interpolated to obtain cleaned data. Conversely, sequences determined to be "task-necessary static" are retained intact. This processing ensures that the dataset contains only static states (such as observations) that are meaningful to the robot's learning task logic, while eliminating the invalid time cost introduced by human teleoperation.

[0094] In some embodiments, a multimodal joint discrimination mechanism combining motion analysis and semantic understanding can be used to lock invalid static frames. Specifically, the fine-grained motion features extracted in step 301 are cross-validated with the high-level semantic labels in step 302: when the semantic model initially determines that the current frame has the attribute of "static", it will further trace back its motion features; if a slight tremor that matches the characteristics of contact operation or a specific frequency jitter during servo locking is detected in the image, even if the semantic level shows static, it is still corrected to "task-necessary static" and retained; only when the semantic recognition is "operator pause" and the motion features show irregular environmental noise or absolute zero displacement is it finally determined to be "non-task-necessary static". This dual-verification logic effectively avoids misjudgment caused by visual occlusion or similar textures under a single modality, ensuring that micro-motion segments containing key physical interaction information are not mistakenly deleted.

[0095] Through steps 301 to 303 above, this embodiment of the application first utilizes a low-computational-cost motion threshold to quickly locate potential static segments, achieving a preliminary coarse screening; secondly, it introduces semantic computation to accurately distinguish between logical waiting and invalid stagnation behind physical stillness, overcoming the technical defects of traditional fixed-threshold-based filtering methods that are prone to accidentally deleting key waiting frames or retaining invalid frames; finally, by deleting non-task-necessary static frames, it improves the purity and information density of the data, ensuring that the trained model's actions are more fluid and coherent, and that it can correctly understand the temporal waiting logic in complex tasks.

[0096] Please see Figure 4 In some embodiments, the pre-training process of the invalid frame identification model may include, but is not limited to, steps 401 to 403.

[0097] Step 401: Obtain multimodal training data containing agent operations. The multimodal training data includes scene images, action sequences, and corresponding natural language description labels.

[0098] Step 402: Perform basic pre-training on the invalid frame recognition model by performing cross-modal feature alignment between scene images and natural language description labels and performing temporal context prediction tasks based on action sequence data.

[0099] Step 403: Using a set of labeled agent operation data containing task-necessary stillness and non-task-necessary stillness, the invalid frame recognition model that has been pre-trained is fine-tuned for classification, so that the invalid frame recognition model can distinguish whether a continuous frame sequence belongs to task-necessary stillness or non-task-necessary stillness based on semantic information.

[0100] In step 401 of some embodiments, acquiring multimodal training data containing robot operations specifically refers to collecting and constructing a large-scale heterogeneous dataset that covers a vast amount of robot operation records in different scenarios (such as factories and homes). The scene images include first-person view video streams from the robot's head-mounted camera and third-person view images from fixed environmental camera positions, encompassing depth maps and RGB images; the action sequences consist of time-series data from the robot's joint angles, angular velocities, end effector poses, and torque sensor feedback; and the natural language description tags are textual annotations of the aforementioned images and action content, including both high-level task instructions (such as "put the red cube into the blue box") and low-level action descriptions (such as "the robotic arm translates to the left and decelerates before contacting the object").

[0101] In step 402 of some embodiments, the basic pre-training of the invalid frame recognition model employs self-supervised learning, mainly including two parallel tasks: cross-modal feature alignment and temporal context prediction. Cross-modal feature alignment aims to teach the model to associate visual scenes with linguistic descriptions. Through a contrastive learning mechanism, the visual embedding vectors of scene images and the text embedding vectors of natural language labels are mapped to the same high-dimensional feature space, narrowing the distance between matching image-text pairs and widening the distance between mismatched pairs, thereby endowing the model with the ability to understand image semantics. The temporal context prediction task aims to teach the model to understand the temporal patterns of actions, allowing the model to predict future action states or masked intermediate frames based on past action sequences and image frames. This forces the model to learn the physical laws, causal relationships, and temporal dependencies of robot movements, constructing a general embodied intelligence foundation representation.

[0102] In step 403 of some embodiments, the model is supervised and fine-tuned using a set of labeled, dedicated datasets. These datasets explicitly label the attributes of various types of still frames; for example, "waiting for the conveyor belt to arrive" and "observation in alignment" are labeled as task-necessary still frames, while "operator pauses" and "network latency issues" are labeled as non-task-necessary still frames. By adjusting the model's weights on specific classification layers or overall parameters, the model can transfer its general visual language understanding capabilities to a specific "still intent recognition" task, thereby enabling the determination of the nature of still frames.

[0103] Through steps 401 to 403 above, this embodiment first constructs a basic model with general visual semantic understanding and physical temporal cognition using massive multimodal data; secondly, through pre-training with cross-modal alignment and temporal prediction, it ensures that the model can understand both the content of the image and the logic of actions; finally, through targeted classification fine-tuning, the general model is transformed into an invalid frame filtering model. This enables the model to accurately distinguish between logical waiting and invalid stagnation in an image based on semantic reasoning.

[0104] Please see Figure 5 In some embodiments, steps 501 to 503 may be included before step 105.

[0105] Step 501: Calculate the inter-frame difference in the key continuous operation segment data and the recovery operation data.

[0106] Step 502: In response to the inter-frame difference being greater than the first preset threshold, the data is processed by frame extraction at the first frame extraction frequency.

[0107] Step 503: In response to the inter-frame difference being less than the second preset threshold, the data is processed by frame extraction at the second frame extraction frequency.

[0108] In step 501 of some embodiments, the inter-frame difference specifically refers to a numerical metric that quantifies the magnitude of content change between adjacent time steps in a video stream or action sequence. This metric can be calculated using pixel-level mean squared error changes or by tracking the displacement distance of key feature points (such as the end effector of a robotic arm) on the image plane. This value reflects in real time the intensity of the robot's movement or the rate of environmental change (e.g., whether it's a high-speed arm swing or a slow approach to an object).

[0109] In step 502 of some embodiments, a first preset threshold sets the boundary for determining high-dynamic scenes. When the calculated inter-frame difference exceeds this threshold, it indicates that the robot is in a state of rapid movement, intense interaction, or a sudden change in the environment (such as a closing action during grasping). At this time, in order to prevent motion blur or loss of key trajectory points, a smaller first frame extraction frequency (i.e., a lower deletion frequency, such as deleting only 1 frame every 10 frames, or even no deletion) can be used to process the data. This "low-frequency deletion" strategy preserves the original frame sequence to the greatest extent, ensuring that every subtle transient action detail can be completely recorded during high-dynamic changes.

[0110] In step 503 of some embodiments, a second preset threshold sets the boundary for determining low-dynamic or stable motion scenarios. When the inter-frame difference is less than this threshold, it indicates that the robot is in a slow-moving, fine-tuning posture phase, with extremely high redundancy between adjacent frames. At this time, a larger second frame-skipping frequency (i.e., a higher deletion frequency, such as deleting 5 frames every 10 frames) can be switched for processing. Given that the first frame-skipping frequency is lower than the second frame-skipping frequency, this deletion strategy in a stable state significantly eliminates redundant data without affecting the continuity of the motion trajectory, achieving efficient utilization of storage space.

[0111] Through steps 501 to 503 above, this embodiment of the application first achieves real-time perception of motion dynamic characteristics by calculating the inter-frame difference degree; secondly, it uses a low frame-dropping frequency in the high dynamic range to ensure high fidelity of motion details, and uses a high frame-dropping frequency in the low dynamic range to eliminate redundant information; finally, this dynamic balancing strategy solves the contradiction between high-speed detail loss and low-speed data redundancy faced by traditional uniform frame-dropping when processing non-uniform motion speed, and improves the comprehensive utilization efficiency and signal-to-noise ratio of real machine data.

[0112] In step 105 of some embodiments, the cleaned data undergoes multi-dimensional quantitative verification to calculate a specific score reflecting its value for downstream training. This process comprehensively considers indicators such as sample diversity, temporal continuity, and information richness, changing the inefficient model that previously relied on manual subjective sampling. By establishing this objective and automated quality control mechanism, decision support can be provided for determining whether the current batch of data is qualified to enter the training library, ensuring that only high-quality data resources are used for subsequent model building.

[0113] Please see Figure 6 In some embodiments, step 105 may include, but is not limited to, steps 601 to 604.

[0114] Step 601: Calculate the time-series consistency index based on the data video stream, joint angle sequence data, and torque data in the purification data.

[0115] Step 602: Calculate the information entropy based on the overall characteristics of the purified data.

[0116] Step 603: Obtain the success rate of the visual language action model controlling the target agent to perform the target action after training the visual language action model used to control the target agent based on the purification data.

[0117] Step 604: Calculate the quality assessment score based on the temporal consistency index, information entropy, and action success rate.

[0118] In step 601 of some embodiments, the purified multimodal data stream is first subjected to deep analysis. The video stream records visual information about the external environment and the operational process; the joint angle sequence data reflects the motion trajectory of each joint of the robot over time; and the torque data characterizes the force experienced by the end effector when it contacts an object. Calculating the temporal consistency index specifically refers to detecting the synchronicity of different modal data in the time dimension. For example, this is done by calculating the cross-correlation coefficient between visual features and joint angular velocities, or by detecting the time stamp deviation between torque abrupt changes and contact events in the video. If the two match and the trajectory is smooth without abrupt changes, the temporal consistency index score is high. Conversely, if there is multimodal asynchrony (such as video stuttering causing the image to lag behind the action) or trajectory jitter, the temporal consistency index score is low, thereby quantifying the physical authenticity and coherence of the data.

[0119] In step 602 of some embodiments, a statistical analysis is performed on the overall characteristic distribution of the cleanup data. For example, the robot's end-effector position space is divided into several grids, the probability distribution of trajectory points falling within each grid is statistically analyzed, and the information entropy value is calculated using the Shannon entropy formula. This value intuitively reflects the amount of information contained in the data. For example, if the robot always repeats a single action, the data distribution is highly concentrated, and the calculated entropy value is low; if the robot performs complex actions covering a large workspace and involving interactions with multiple objects, the data distribution is uniform and widespread, and the calculated information entropy value is high, indicating that the data has greater value for improving the generalization boundary of the model.

[0120] In step 603 of some embodiments, obtaining the action success rate is a post-hoc evaluation process based on practical verification. The current cleaned data is used as a fine-tuning sample set and input into the Visual Language Action Model (VLA) for rapid training. Subsequently, the target agent is controlled to perform the same target action in a simulation environment or on a real machine, and the percentage of times the robot successfully completes the task is counted. For example, if it succeeds 85 times out of 100 attempts, the action success rate is 85%. This metric directly relates to the causal relationship between data quality and the final model performance, reflecting the effectiveness of the data in practical applications.

[0121] In step 604 of some embodiments, a weighted fusion algorithm is used to combine the physical layer coherence, the information layer richness, and the result layer effectiveness. Specifically, based on preset weight coefficients (such as 30% for temporal consistency, 20% for information entropy, and 50% for action success rate), the temporal consistency index, information entropy, and action success rate are weighted and summed to calculate the final quality assessment score.

[0122] Through steps 601 to 604 above, this embodiment first ensures the physical synchronization and coherence of multimodal data through a temporal consistency index, eliminating low-quality sensor noise; secondly, it quantifies sample diversity through an information entropy index, preventing the inclusion of inefficient and repetitive data; thirdly, it introduces action success rate as posterior feedback; and finally, by comprehensively calculating a quality assessment score, it achieves accurate measurement of the value of real machine data from multiple dimensions. This method can effectively reduce the redundancy of robot real machine data and enrich data diversity, thereby improving the generalization ability of the visual language action model.

[0123] In step 106 of some embodiments, in response to the quality assessment score meeting a preset quality score threshold, training the target agent by purifying the data refers to packaging the filtered high signal-to-noise ratio data into a standardized dataset and inputting it into the robot's visual language action model for training or fine-tuning. The preset threshold serves as a data admission condition to prevent low-quality data from contaminating the model parameters. Utilizing this high-quality data rich in key temporal logic and failure recovery strategies, the model can more efficiently master the operational skills of complex tasks and exhibit smoother action execution capabilities and stronger anti-interference capabilities in actual deployment.

[0124] Please see Figure 7 In some embodiments, after step 106, steps 701 to 702 may also be included, but are not limited to.

[0125] Step 701: Based on the action success rate obtained by the target agent after training with cleaned data, calculate the adjustment weights of the temporal consistency index and information entropy in the next round of quality assessment.

[0126] Step 702: Based on the action success rate, determine the motion change threshold and frame sampling frequency to be used by the invalid frame identification model in the next round of data processing.

[0127] In step 701 of some embodiments, the success rate of the target agent's actions in performing tasks in a real or simulated environment is used as a feedback signal to correct the parameter configuration in the data quality assessment model through an adaptive weighted algorithm. Specifically, if the model trained with high-scoring data in the current batch has a low success rate in actual testing, and analysis reveals that this is due to excessive data clutter, the weight of information entropy (representing diversity) will be automatically reduced in the next round of quality assessment, while the adjustment weight of the "temporal consistency index" (representing coherence) will be increased. This dynamic weight adjustment mechanism ensures that the quality assessment criteria are no longer immutable rules, but rather dynamic weights that evolve in real time with fluctuations in model training performance, thereby consistently selecting the data features that best meet the needs of improving model capabilities at the current stage.

[0128] In step 702 of some embodiments, the hyperparameters of the front-end data cleaning and acquisition stage are directly optimized in a closed loop based on the success rate of the action. For example, if feedback shows that the model fails due to the loss of high-frequency action details, the frame extraction frequency will be reduced in the next round of processing (i.e., the proportion of deleted frames will be reduced to retain more original details), or the motion change threshold in the invalid frame recognition model will be increased (i.e., the standard for judging stillness will be more stringent, thereby retaining more frames with minute movements). Conversely, if the success rate is extremely high but the training time is too long, the frame extraction frequency will be appropriately increased to remove more redundancy. This adaptive parameter adjustment can optimize the acquisition strategy according to the final usage effect.

[0129] Please see Figure 8 This application also provides a real-device data processing apparatus for an intelligent agent, which can implement the above-described real-device data processing method for an intelligent agent, including:

[0130] The acquisition module is used to acquire the target action acquisition task and the target action type;

[0131] The first acquisition module is used to acquire key continuous operation segments in a long-term task in response to the target action type being a long-term task.

[0132] The second acquisition module is used to acquire data on failed scenarios and corresponding recovery operations during the execution of the target action acquisition task.

[0133] The identification module is used to identify and delete key continuous operation segments and restore invalid static frames in the operation data through a pre-trained invalid frame identification model to obtain clean data.

[0134] The quality assessment module is used to assess the quality of the cleaned data based on a preset data quality assessment model and obtain a quality assessment score.

[0135] The training module is used to train the target agent by cleaning up data in response to the quality assessment score meeting a preset quality score threshold.

[0136] The real-device data processing method for an intelligent agent according to embodiments of this application includes: acquiring a target action acquisition task and a target action type; in response to the target action type being a long-term task, acquiring key continuous operation segment data in the long-term task; acquiring failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task; identifying and deleting invalid static frames in the key continuous operation segment data and recovery operation data using a pre-trained invalid frame recognition model to obtain cleaned data; performing quality assessment on the cleaned data based on a preset data quality assessment model to obtain a quality assessment score; and training the target intelligent agent using the cleaned data in response to the quality assessment score meeting a preset quality score threshold.

[0137] This application clarifies the specific scenario and task attributes for data collection by acquiring the target action acquisition task and the target action type. Then, by recognizing the target action type as a long-term task, it collects key continuous operation segments within the long-term task, avoiding the large amount of repetitive and redundant information inherent in full-process acquisition, and significantly improving acquisition efficiency while preserving the core temporal logic of the task. Next, by collecting failure scenario data and corresponding recovery operation data during the execution of the target action acquisition task, it incorporates possible anomalies and error correction processes from real-world operations into the training scope, compensating for the lack of negative samples and recovery logic in existing datasets. Subsequently, a pre-trained invalid frame recognition model identifies and removes invalid static frames from key continuous operation segments and recovery operation data, resulting in purified data. This purified data can be precisely removed based on semantic analysis, further improving the effective density of the data. Finally, the purified data is evaluated using a pre-set data quality assessment model, and the target agent is trained when a threshold is met, ensuring that the data used for model learning has both high signal-to-noise ratio and high reliability. Compared with existing technologies, this application can effectively reduce the redundancy of real machine data for intelligent agents and enrich data diversity, thereby improving the generalization ability of visual language action models.

[0138] Reference Figure 9 , Figure 9 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0139] The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0140] The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 using the real-machine data processing method of the intelligent agent in the embodiments of this application.

[0141] The input / output interface 903 is used to implement information input and output;

[0142] The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, Wi-Fi, Bluetooth, etc.).

[0143] Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904);

[0144] The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0145] This application also provides a computer program product, which includes a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the real-machine data processing method for implementing the above-described intelligent agent.

[0146] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in this disclosure and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “including,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatuses.

[0147] It should be understood that in this disclosure, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0148] It should be understood that in the description of the embodiments of this application, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0149] In the several embodiments provided in this disclosure, 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 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, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0150] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0151] Furthermore, the functional units in the various embodiments of this disclosure 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.

[0152] 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 disclosure, 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.) to execute all or part of the steps of the methods of the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0153] It should also be understood that the various implementation methods provided in this application can be combined arbitrarily to achieve different technical effects.

[0154] The above is a detailed description of the embodiments of this disclosure. However, this disclosure is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this disclosure. All such equivalent modifications or substitutions are included within the scope defined by the claims of this disclosure.

Claims

1. A method for real-machine data processing of an intelligent agent, characterized in that, The method comprises the following steps: acquiring a target action collection task and a target action type; in response to the target action type being a long-range task, collecting key continuous operation fragment data in the long-range task; collecting failure scenario data and corresponding recovery operation data in the execution process of the target action collection task; identifying and deleting invalid static frames in the key continuous operation fragment data and the recovery operation data by using a pre-trained invalid frame identification model to obtain purified data; wherein the pre-training process of the invalid frame identification model comprises: acquiring multi-modal training data containing agent operations, the multi-modal training data comprising scene images, action sequences and corresponding natural language description labels; performing cross-modal feature alignment between the scene images and the natural language description labels, and performing a timing context prediction task based on the action sequences to perform basic pre-training on the invalid frame identification model; using a set of labeled agent operation data containing task necessary static and non-task necessary static to perform classification fine-tuning on the invalid frame identification model that has been subjected to basic pre-training, so that the invalid frame identification model can distinguish whether the continuous frame sequence belongs to task necessary static or non-task necessary static according to semantic information; wherein the task necessary static is waiting that belongs to part of the task logic, and the non-task necessary static is redundant stagnation irrelevant to the task logic; performing quality evaluation on the purified data based on a preset data quality evaluation model to obtain a quality evaluation score; in response to the quality evaluation score meeting a preset quality score threshold, training a target agent by using the purified data.

2. The method of claim 1, wherein, The response to the target action type being a long-range task and the collection of multiple key continuous operation fragment data in the long-range task comprise: determining a repeated operation unit in the long-range task, which needs to be executed a first number of times in the long-range task; determining that the collection number of the key continuous operation fragment data is a second number; wherein the second number is less than the first number; collecting data of continuously executing the repeated operation unit for the second number of times as the key continuous operation fragment data.

3. The method of claim 1, wherein, The identification and deletion of invalid static frames in the key continuous operation fragment data and the recovery operation data by using the pre-trained invalid frame identification model to obtain purified data comprise: when detecting that the motion change of a continuous frame sequence in the key continuous operation fragment data or the recovery operation data is lower than a preset motion change threshold, extracting a representative frame in the continuous frame sequence; performing semantic analysis on the representative frame by using the invalid frame identification model to determine whether the continuous frame sequence is task necessary static; in response to determining that the continuous frame sequence is non-task necessary static, determining it as the invalid static frame and performing a deletion operation.

4. The method of claim 1, wherein, Before performing quality evaluation on the purified data based on the preset data quality evaluation model to obtain a quality evaluation score, the method further comprises: calculating the inter-frame difference degree in the key continuous operation fragment data and the recovery operation data; In response to the frame difference being greater than a first preset threshold, frame data is processed at a first frame extraction frequency; In response to the frame difference being less than a second preset threshold, frame data is processed at a second frame extraction frequency; wherein the first frame extraction frequency is less than the second frame extraction frequency.

5. The method of claim 1, wherein, The quality evaluation model is used to evaluate the quality of the purified data, and a quality evaluation score is obtained, including: According to the data video stream, joint angle sequence data, and torque data in the purified data, a time sequence consistency index is calculated; According to the overall characteristics of the purified data, an information entropy is calculated; After the visual language action model for controlling the target agent is trained based on the purified data, an action success rate of the target agent executing a target action is obtained; According to the time sequence consistency index, the information entropy, and the action success rate, the quality evaluation score is calculated.

6. The method of claim 5, wherein, After the target agent is trained by the purified data in response to the quality evaluation score meeting a preset quality score threshold, the following steps are further included: According to the action success rate of the target agent after being trained by the purified data, adjustment weights of the time sequence consistency index and the information entropy in the next round of quality evaluation are calculated; According to the action success rate, a motion change threshold and a frame extraction frequency used by the invalid frame identification model in the next round of data processing are determined.

7. A real-machine data processing device for an intelligent agent, characterized in that, The device includes: An acquisition module is configured to acquire a target action collection task and a target action type; A first collection module is configured to collect key continuous operation segment data in a long-range task in response to the target action type being a long-range task; A second collection module is configured to collect failure scenario data and corresponding recovery operation data in a target action collection task execution process; An identification module is configured to identify and delete invalid static frames in the key continuous operation segment data and the recovery operation data by using a pre-trained invalid frame identification model, and obtain purified data; wherein the pre-training process of the invalid frame identification model includes: acquiring multi-modal training data containing agent operations, the multi-modal training data including scene images, action sequences, and corresponding natural language description labels; performing cross-modal feature alignment between the scene images and the natural language description labels, and performing a time sequence context prediction task based on the action sequences, to perform basic pre-training on the invalid frame identification model; using a set of labeled agent operation data containing task necessary static and non-task necessary static, to perform classification fine-tuning on the invalid frame identification model that has been subjected to basic pre-training, so that the invalid frame identification model can distinguish whether the continuous frame sequence belongs to task necessary static or non-task necessary static according to semantic information; wherein the task necessary static refers to waiting that belongs to part of the task logic, and the non-task necessary static refers to redundant stagnation irrelevant to the task logic; A quality evaluation module is configured to evaluate the quality of the purified data based on a preset data quality evaluation model, and obtain a quality evaluation score. The training module is configured to train the target agent by using the purified data in response to the quality evaluation score satisfying a preset quality score threshold.

8. An electronic device, comprising: The method comprises the following steps: The memory stores a computer program, and the processor executes the computer program to realize the method for processing real machine data of the agent according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a program, and the program is executed by the processor to realize the method for processing real machine data of the agent according to any one of claims 1 to 6.

Images