Multi-dimensional body data quality verification and processing method, system, equipment and medium

Abstract

The invention provides a multi-dimensional body data quality verification and processing method, system and device and a medium. The method comprises the following steps: receiving an original multi-modal data stream collected by a body intelligent robot; performing time sequence alignment on the original multi-modal data flow in a pre-acquisition buffer pool, and aggregating the original multi-modal data flow into a joint data sample; extracting a multi-dimensional feature vector in real time from the joint data sample; based on the time sequence features, the spatial geometric features and the ontology state features, multi-dimensional quality indexes are calculated, weighted aggregation is carried out on the multi-dimensional quality indexes, and a comprehensive quality score is generated; the comprehensive quality score is compared with a preset high-quality threshold value and a preset low-quality threshold value, a storage routing decision is made according to a comparison result, the data availability rate is improved through source real-time quality control, and the collection and labeling cost of mass invalid data is fundamentally reduced; and the comprehensive quality score is calculated through the multi-dimensional feature vector, so that the credibility and robustness of the data are guaranteed, and the conversion from passive cleaning to active intelligent treatment is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system, device and medium for multi-dimensional embodied data quality verification and processing. Background Technology

[0002] With the rapid development of artificial intelligence technology, especially large-scale model technology, embodied AI has become a key direction for the evolution of next-generation artificial intelligence. Unlike large language models that rely on massive amounts of internet text data for training, embodied AI requires intelligent agents (such as robots) to interact with their environment in real time through their physical bodies, actively collecting multimodal perception data (such as visual images, force signals, joint motion trajectories, etc.) in the process. This type of embodied data has significant dynamic interactivity, multimodal synchronization, and physical feedback, making it an indispensable foundation for training and fine-tuning high-performance robot models. Therefore, building high-quality, high-reliability embodied multimodal datasets is a core prerequisite for promoting the development of embodied AI. However, existing technologies have a series of inherent and systemic shortcomings in the acquisition and quality assurance of embodied data, which restrict the usability of datasets and the improvement of final model performance. First, embodied data is directly stored after acquisition, lacking the ability to perform real-time quality judgment and intervention before being stored on disk. This results in a large amount of invalid or low-quality data entering the downstream process, causing a huge waste of storage and computing resources and transferring high costs to the subsequent data cleaning and labeling stages. Secondly, embodied intelligence tasks demand stringent synchronization accuracy from multimodal data, often requiring millisecond or even nanosecond levels. However, existing general-purpose data processing systems struggle to perform real-time online verification of this high-precision temporal consistency. Minor misalignments between multiple data streams can accumulate to generate significant temporal errors, thereby corrupting data processing results. Furthermore, existing technologies often employ fragmented solutions addressing single problems. For example, to address frame loss, high-speed caching and multi-threaded processing are used to reduce data loss; for duplicate frame detection, content quality inspection systems employ inter-frame similarity thresholds. However, these solutions lack a systematic closed-loop verification framework, failing to uniformly cover multiple key quality dimensions such as time synchronization, and thus failing to meet the stringent data quality requirements of large-scale embodied intelligence models. Summary of the Invention

[0003] This invention provides a method, system, device, and medium for multi-dimensional embodied data quality verification and processing, which addresses the lack of a real-time and systematic quality verification framework in existing technologies for embodied data, resulting in high costs, poor consistency, and fragmented management of embodied data at the source.

[0004] This invention provides a method for multi-dimensional embodied data quality verification and processing, including: Receive raw multimodal data streams collected by embodied intelligent robots; The original multimodal data stream is time-aligned in the pre-acquisition buffer pool and aggregated into a joint data sample; Multidimensional feature vectors are extracted in real time from the joint data samples. The multidimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features. Based on the temporal features, spatial geometric features, and ontological state features, multidimensional quality indicators are calculated, and the multidimensional quality indicators are weighted and aggregated to generate a comprehensive quality score for the joint data sample. The overall quality score is compared with preset high-quality thresholds and low-quality thresholds, and a storage routing decision is made based on the comparison results.

[0005] According to the multi-dimensional embodied data quality verification and processing method provided by the present invention, the step of comparing the comprehensive quality score with preset high-quality thresholds and low-quality thresholds, and making storage routing decisions based on the comparison results, includes: When the overall quality score is greater than or equal to the high quality threshold, it is determined to pass, and the joint data sample is stored in the main training database; When the overall quality score is between the high quality threshold and the low quality threshold, it is determined to be degraded for storage, and the joint data sample is stored in the archive library; When the overall quality score is less than the low quality threshold, it is determined to be rejected.

[0006] According to the multi-dimensional embodied data quality verification and processing method provided by the present invention, the step of performing time-series alignment of the original multimodal data stream in a pre-acquisition buffer pool and aggregating it into a joint data sample includes: Based on a hardware clock or precision clock synchronization protocol, timestamps with a unified time base are applied to the raw data frames of all modes. In the pre-acquisition buffer pool, triggered by a time window or key event, each modal data frame falling within the same time window is resampled onto a unified time series using a timestamp-based linear or spline interpolation algorithm to generate joint data samples.

[0007] According to the multi-dimensional embodied data quality verification and processing method provided by the present invention, the step of extracting multi-dimensional feature vectors from the joint data samples in real time, wherein the multi-dimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features, including: Calculate the maximum timestamp deviation between each modal data frame within the joint data sample, and calculate the instantaneous frame rate of at least one key sensor and its stability deviation from the nominal frame rate to extract temporal features; Extract the image coordinates or 3D coordinates of a pre-defined calibration reference from visual or depth data, and calculate the residual between the reference and the calibration position; at the same time, calculate the sparsity ratio of the depth map or the optical distortion measure of the image to extract spatial geometric features. Based on the robot's kinematic model, the expected pose of the end effector is calculated according to the joint encoder readings, and the actual pose of the end effector is identified from the visual data. The instantaneous spatial error between the two is calculated. The smoothness feature score is quantized according to the instantaneous spatial error to extract the body state features.

[0008] According to the multi-dimensional embodied data quality verification and processing method provided by the present invention, the comprehensive quality score of the joint data sample includes a calibration health score, and the calibration health score generation method includes: For at least one target sensor in the joint data sample, its current reading is obtained, and the corresponding predicted reading is calculated based on the calibration model of the target sensor; Calculate the residual between the current reading and the predicted reading, and store the residual in a residual sequence; The distribution of the residual sequence is monitored in real time using a statistical estimation algorithm to obtain the mean or variance of the residuals; The calibrated health score is generated based on the changes in the mean or variance of the monitored residual sequence.

[0009] According to the multi-dimensional embodied data quality verification and processing method provided by the present invention, the comprehensive quality score of the joint data sample includes a cross-modal physical consistency score, and the method for generating the cross-modal physical consistency score includes: Obtain predefined physical association rules, and extract first modal features and second modal features corresponding to the physical association rules from the joint data samples; wherein, the first modal features include at least visual features, and the second modal features are robot body state features; The first modal feature and the second modal feature are input into a lightweight feature association model to obtain the predicted value of the target physical quantity; From the joint data sample, extract the actual observed values ​​corresponding to the target physical quantity; calculate the residual between the actual observed values ​​and the predicted values; The residual is compared with a dynamic threshold, and the cross-modal physical consistency score is generated based on the comparison result.

[0010] The multi-dimensional embodied data quality verification and processing method provided by the present invention further includes: Receive feedback information from downstream model training tasks; Based on the feedback information, the preset weights of the multidimensional quality indicators and / or the high-quality threshold and low-quality threshold are dynamically adjusted.

[0011] The present invention also provides a quality decision gate system for embodied data before disk write-to-disk, comprising: The receiving module is used to receive the raw multimodal data stream collected by the embodied intelligent robot; The alignment module is used to perform time-series alignment of the original multimodal data stream in the pre-acquisition buffer pool and aggregate it into a joint data sample; The extraction module is used to extract multi-dimensional feature vectors from the joint data samples in real time. The multi-dimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features. The generation module is used to calculate multi-dimensional quality indicators based on the temporal features, spatial geometric features and ontological state features, and to perform weighted aggregation of the multi-dimensional quality indicators to generate a comprehensive quality score for the joint data sample. The decision module is used to compare the comprehensive quality score with preset high quality thresholds and low quality thresholds, and make storage routing decisions based on the comparison results.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the multi-dimensional embodied data quality verification and processing method as described in any of the preceding claims.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multi-dimensional embodied data quality verification and processing method described in any of the preceding claims.

[0014] The present invention provides a multi-dimensional embodied data quality verification and processing method, system, device, and medium. This method receives raw multimodal data streams collected by an embodied intelligent robot; performs temporal alignment on the raw multimodal data streams in a pre-collection buffer, aggregating them into joint data samples; extracts multi-dimensional feature vectors from the joint data samples in real time, the multi-dimensional feature vectors including at least temporal features, spatial geometric features, and ontological state features; calculates multi-dimensional quality indicators based on the temporal features, spatial geometric features, and ontological state features; weights and aggregates the multi-dimensional quality indicators to generate a comprehensive quality score for the joint data samples; compares the comprehensive quality score with preset high-quality and low-quality thresholds, and makes storage routing decisions based on the comparison results. Through real-time quality control at the source, data availability is improved, fundamentally reducing the cost of collecting and labeling massive amounts of invalid data; and by calculating the comprehensive quality score through multi-dimensional feature vectors, the credibility and robustness of the data are ensured, realizing a shift from passive cleaning to proactive intelligent governance. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the multi-dimensional embodied data quality verification and processing method provided in the embodiments of the present invention; Figure 2 This is a functional structure diagram of the embodied data pre-disk quality decision gate system provided in an embodiment of the present invention; Figure 3 This is a functional structure diagram of the electronic device provided in the embodiments of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention 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 invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Figure 1 A flowchart of the multi-dimensional embodied data quality verification and processing method provided in the embodiments of the present invention is shown below. Figure 1 As shown, the multi-dimensional embodied data quality verification and processing method provided in this embodiment of the invention includes: Step 101: Receive the raw multimodal data stream collected by the embodied intelligent robot; Step 102: The original multimodal data stream is time-aligned in the pre-acquisition buffer pool and aggregated into a joint data sample; In this embodiment of the invention, the pre-ingestion buffer pool is used to receive multimodal data streams at high speed and perform high-precision timing alignment (sub-millisecond to nanosecond level) to aggregate the data into joint data samples.

[0019] Step 103: Extract multi-dimensional feature vectors from the joint data samples in real time. The multi-dimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features. In this embodiment of the invention, the multidimensional feature vector is used as a quantitative feature vector for quality assessment, including time-series features ( Deviation), spatial / geometric features (calibration reference point), and body state features (kinematic error).

[0020] Step 104: Based on the temporal features, spatial geometric features, and ontological state features, calculate multidimensional quality indicators, perform weighted aggregation on the multidimensional quality indicators, and generate a comprehensive quality score for the joint data sample. In this embodiment of the invention, the comprehensive quality score of the joint data sample includes a calibrated health score. ) and cross-modal physical consistency ( ), and according to preset weights Aggregated into a comprehensive quality score .

[0021] Step 105: Compare the comprehensive quality score with the preset high quality threshold and low quality threshold, and make a storage routing decision based on the comparison result.

[0022] In this embodiment of the invention, based on With multi-level thresholds Establish a quality decision gate to perform three-level real-time decisions (pass, downgrade storage, reject) on joint data samples. Perform automatic quality verification and correction during data acquisition, and feed the results back to the acquisition process to form an adaptive optimization loop.

[0023] Traditional embodied data is directly stored after collection, lacking the ability to perform real-time quality assessment and intervention before disk write-to-disk. This results in a large amount of invalid or low-quality data entering downstream processes, causing a huge waste of storage and computing resources and shifting high costs to subsequent data cleaning and labeling stages. Embodied intelligence tasks have stringent requirements for the synchronization accuracy of multimodal data, often needing to reach millisecond or even nanosecond levels. However, existing general-purpose data processing systems struggle to perform real-time online verification of this high-precision temporal consistency. Small misalignments between multi-source data streams can accumulate to produce significant temporal errors, thereby corrupting data processing results. Furthermore, existing technologies are mostly fragmented solutions addressing single problems. For example, for frame loss, high-speed caching and multi-threaded processing are used to reduce data loss; for duplicate frame detection, content quality inspection systems use inter-frame similarity thresholds. However, these solutions lack a systematic closed-loop verification framework, failing to uniformly cover multiple key quality dimensions such as time synchronization, and thus failing to meet the stringent data quality requirements of large-scale embodied intelligence models.

[0024] The multi-dimensional embodied data quality verification and processing method provided in this invention receives raw multimodal data streams collected by an embodied intelligent robot; performs temporal alignment on the raw multimodal data streams in a pre-collection buffer pool, aggregating them into joint data samples; extracts multi-dimensional feature vectors from the joint data samples in real time, the multi-dimensional feature vectors including at least temporal features, spatial geometric features, and ontological state features; calculates multi-dimensional quality indicators based on the temporal features, spatial geometric features, and ontological state features, performs weighted aggregation on the multi-dimensional quality indicators, and generates a comprehensive quality score for the joint data samples; compares the comprehensive quality score with preset high-quality thresholds and low-quality thresholds, and makes storage routing decisions based on the comparison results. Through real-time quality control at the source, data availability is improved, fundamentally reducing the cost of collecting and labeling massive amounts of invalid data; by calculating the comprehensive quality score through multi-dimensional feature vectors, the credibility and robustness of the data are ensured, realizing a shift from passive cleaning to proactive intelligent governance.

[0025] Based on any of the above embodiments, the step of comparing the comprehensive quality score with preset high-quality thresholds and low-quality thresholds, and making storage routing decisions based on the comparison results, includes: When the overall quality score is greater than or equal to the high quality threshold, it is determined to pass, and the joint data sample is stored in the main training database; When the overall quality score is between the high quality threshold and the low quality threshold, it is determined to be degraded for storage, and the joint data sample is stored in the archive library; When the overall quality score is less than the low quality threshold, it is determined to be rejected.

[0026] This invention implements a three-level decision-making process (pass, downgrade, reject). The downgrade storage strategy allows the system to retain medium-quality data for robust training or debugging, thereby maximizing the use of all collected information and avoiding wasteful one-size-fits-all approaches.

[0027] For rejected samples, an action instruction (re-sampling prompt) is immediately generated and sent to the robot control system; simultaneously, the system dynamically adjusts its operation based on feedback from downstream annotation and model training. The weights and thresholds are used to enable online self-learning and updating of rules.

[0028] When a rejection is detected, a re-collection prompt is generated in real time and sent to the control system of the embodied intelligent robot to trigger the re-execution of the data collection task in the current scenario. Through real-time quality control and re-collection mechanisms at the source, the data availability rate is significantly improved to 90%-95%.

[0029] Based on any of the above embodiments, the step of performing time-series alignment on the original multimodal data stream in the pre-acquisition buffer pool and aggregating it into a joint data sample includes: Based on a hardware clock or precision clock synchronization protocol, timestamps with a unified time base are applied to the raw data frames of all modes. In the pre-acquisition buffer pool, triggered by a time window or key event, each modal data frame falling within the same time window is resampled onto a unified time series using a timestamp-based linear or spline interpolation algorithm to generate joint data samples.

[0030] In this embodiment of the invention, the data acquisition layer receives raw data streams from the embodied robot's multimodal sensors (e.g., RGB-D cameras, torque sensors, IMUs, joint encoders, etc.). This layer ensures that all data acquisitions are accompanied by high-precision, hardware or kernel-level timestamps.

[0031] The raw data then enters the pre-acquisition buffer, a high-performance queue or circular buffer used to decouple the high-speed data stream from complex quality verification logic and perform a high-precision timing alignment mechanism. A high-precision clock synchronization algorithm is used to calculate and minimize the maximum deviation of the timestamps for each modality frame. This ensures that the aggregated joint data samples meet the stringent requirements of embodied AI training for sub-millisecond or even nanosecond-level synchronization.

[0032] Based on any of the above embodiments, the step of extracting a multidimensional feature vector from the joint data sample in real time, wherein the multidimensional feature vector includes at least temporal features, spatial geometric features, and ontological state features, includes: Calculate the maximum timestamp deviation between each modal data frame within the joint data sample, and calculate the instantaneous frame rate of at least one key sensor and its stability deviation from the nominal frame rate to extract temporal features; Extract the image coordinates or 3D coordinates of a pre-defined calibration reference from visual or depth data, and calculate the residual between the reference and the calibration position; at the same time, calculate the sparsity ratio of the depth map or the optical distortion measure of the image to extract spatial geometric features. Based on the robot's kinematic model, the expected pose of the end effector is calculated according to the joint encoder readings, and the actual pose of the end effector is identified from the visual data. The instantaneous spatial error between the two is calculated. The smoothness feature score is quantized according to the instantaneous spatial error to extract the body state features.

[0033] In this embodiment of the invention, the multidimensional feature vector includes temporal features, including the maximum timestamp deviation within the joint data sample. Frame rate stability of key sensor data. Spatial / geometric features: coordinates of calibration feature points or reference objects used for real-time drift detection; image distortion and depth map sparsity. Body state features include the smoothness of robot joint trajectories; instantaneous error between kinematic model predictions and actual encoder readings (i.e., low-level features for joint-endpoint posture consistency verification). Device metadata features include device health information such as sensor temperature and network latency.

[0034] In this embodiment of the invention, drift refers to the dynamic degradation of the calibration parameters of a robot's multi-sensor system due to physical shock, environmental pressure, or long-term wear.

[0035] Based on any of the above embodiments, the comprehensive quality score of the joint data sample includes a calibrated health score, and the method for generating the calibrated health score includes: For at least one target sensor in the joint data sample, its current reading is obtained, and the corresponding predicted reading is calculated based on the calibration model of the target sensor; Calculate the residual between the current reading and the predicted reading, and store the residual in a residual sequence; The distribution of the residual sequence is monitored in real time using a statistical estimation algorithm to obtain the mean or variance of the residuals; The calibrated health score is generated based on the changes in the mean or variance of the monitored residual sequence.

[0036] Calibration drift error The inverse index. Using statistical estimation algorithms (such as the adaptive sliding window algorithm ADWIN or a dynamic calibration curve based on minimum mean square error (MMSE), the residual distribution between sensor readings and the predicted output of its calibration model is monitored in real time.

[0037] When the mean or variance of the residuals shows a statistically significant increase, it is determined that... Downgrade. This continuous online monitoring is crucial to ensuring data accuracy.

[0038] In this embodiment of the invention, by calibrating a health score, the dynamic spatial geometric consistency is ensured in real time, thus solving the problem of traditional methods relying on lengthy offline calibration. This is achieved through real-time calibration of health scores. The evaluation mechanism can detect and quantify the calibration drift of sensors during long-term operation online, ensuring the continuous reliability of the geometric accuracy of multi-sensor fusion in dynamic environments.

[0039] Based on any of the above embodiments, the overall quality score of the joint data samples includes a cross-modal physical consistency score, and the method for generating the cross-modal physical consistency score includes: Obtain predefined physical association rules, and extract first modal features and second modal features corresponding to the physical association rules from the joint data samples; wherein, the first modal features include at least visual features, and the second modal features are robot body state features; The first modal feature and the second modal feature are input into a lightweight feature association model to obtain the predicted value of the target physical quantity; From the joint data sample, extract the actual observed values ​​corresponding to the target physical quantity; calculate the residual between the actual observed values ​​and the predicted values; The residual is compared with a dynamic threshold, and the cross-modal physical consistency score is generated based on the comparison result.

[0040] Define and test physical association rules (PCR) between multiple sensors. For example, train a lightweight feature association model in a grasping task. The model receives visual features and the robot's body state Predict the expected torque characteristics .

[0041] Residual analysis: Calculating actual observations With model prediction residuals between .like Exceeding the dynamic threshold If so, the physical consistency check is deemed to have failed, resulting in... This reduces the likelihood of physical discrepancies when visual and torque signals describe the same event.

[0042] This invention introduces cross-modal physical consistency verification (...) This paper utilizes a lightweight model to verify the logical consistency of different sensing modalities (such as vision and torque) in describing the same embodied event in real time, effectively filtering out data containing physical conflicts or internal sensor anomalies.

[0043] In this embodiment of the invention, It is a multidimensional quality indicator (such as) , , , Weighted aggregation results of (etc.) Weight It is configurable and learnable, dynamically reflecting the current task's dependence on different metrics. In some embodiments of this invention, a metadata analysis method based on a large language model (LLM) can also be used to perform zero-shot or few-shot anomaly detection using sensor-derived metadata, assisting in... Complex correlation analysis.

[0044] The multi-dimensional embodied data quality verification and processing method provided in this invention can be deployed on the robot body or a near-end server to fully utilize high-throughput computing power and minimize the impact of network latency on the real-time verification process.

[0045] Based on any of the above embodiments, the multi-dimensional embodied data quality verification and processing method further includes: Receive feedback information from downstream model training tasks; Based on the feedback information, the preset weights of the multidimensional quality indicators and / or the high-quality threshold and low-quality threshold are dynamically adjusted.

[0046] In this embodiment of the invention, the quality decision gate is based on With the preset high quality threshold and low quality threshold Perform three-level real-time routing decisions: (1) Pass: The data is written to the main training data lake, along with detailed quality metadata.

[0047] (2) Degraded Storage: The data is written to low-priority archive storage, which may be used for robust training.

[0048] (3) Reject: The data is immediately discarded to prevent it from being written to disk.

[0049] When the decision gate rejects data ( The feedback module immediately generates an action instruction (re-sampling prompt) and sends it directly to the robot control system or data acquisition scheduler. This instruction requires the robot to immediately repeat the most recent task steps in the current scenario, greatly improving data acquisition efficiency.

[0050] Continuously track the false positive rate in downstream annotation and training. When the downstream error rate corresponding to a specific quality metric increases significantly, the system automatically increases the value of that metric. Penalty weights in aggregation (Adaptive weighting based on error rate). Furthermore, unsupervised mechanisms such as Hebbian association learning can be employed to dynamically discover new quality failure modes and optimize rule thresholds. If In case of an anomaly, the feedback module can also trigger an alarm or an automated self-calibration process.

[0051] Rejection feedback can trigger not only simple resampling, but also further AI-based error diagnosis and correction processes. For example, if it is determined that the data was rejected due to calibration drift, it will automatically initiate a model-based self-calibration procedure to attempt to correct the sensor parameters before resampling.

[0052] The multi-dimensional embodied data quality verification and processing method provided by this invention achieves end-to-end, real-time, and adaptive quality control of embodied data streams, ensuring that the output dataset meets the stringent requirements of embodied intelligent large-scale models for high-precision and high-availability data. The online rule self-learning update mechanism integrated into the decision gate (e.g., error rate-based weight adaptation) enables the system to dynamically adjust its verification rules according to the needs of downstream tasks and the slow time-varying characteristics of the environment, guaranteeing the accuracy and long-term robustness of the quality decision gate system before embodied data is written to disk.

[0053] The following describes the method for quality verification and processing of multi-dimensional embodied data provided by the present invention. The method for quality verification and processing of multi-dimensional embodied data described below can be referred to in correspondence with the method for quality verification and processing of multi-dimensional embodied data described above.

[0054] Figure 2 This is a functional structure diagram of the Embodied Quality Decision-Making System provided in an embodiment of the present invention. The Embodied Quality Decision-Making System is used to perform real-time quality verification before data is persistently stored. Figure 2 As shown, the embodiment of the present invention provides a quality decision gate system for embedded data before disk storage, which includes: The receiving module 201 is used to receive the raw multimodal data stream collected by the embodied intelligent robot; Alignment module 202 is used to perform time-series alignment of the original multimodal data stream in the pre-acquisition buffer pool and aggregate it into a joint data sample; The extraction module 203 is used to extract multi-dimensional feature vectors from the joint data samples in real time. The multi-dimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features. The generation module 204 is used to calculate multi-dimensional quality indicators based on the temporal features, spatial geometric features and ontological state features, and to perform weighted aggregation of the multi-dimensional quality indicators to generate a comprehensive quality score for the joint data sample. The decision module 205 is used to compare the comprehensive quality score with preset high quality thresholds and low quality thresholds, and make storage routing decisions based on the comparison results.

[0055] In this embodiment of the invention, the modular architecture of the system naturally supports plug-in expansion, and can adapt to future additions of tactile or auditory sensing modalities.

[0056] The embodied data pre-disk quality decision gate system provided in this invention aims to solve the problem that existing technologies in embodied intelligent multimodal data acquisition and processing cannot perform real-time, systematic, and embodied characteristic-specific quality verification before data enters persistent storage, resulting in low data availability, high annotation costs, and limited model training effects. It employs a "pre-acquisition buffer + multi-dimensional quality scoring model". The innovative pipeline architecture of "+ Decision Gate" enables real-time assessment of the unique health status of embodied data. ) and cross-modal physical consistency ( The system implements three-level decision-making and real-time resampling prompts, ultimately achieving the technical effect of transforming the quality control paradigm from passive offline cleaning to proactive real-time source control, significantly improving data availability to 90%-95%, and greatly reducing downstream annotation and training costs.

[0057] Figure 3 An example is a schematic diagram of the physical structure of a communication device, such as... Figure 3As shown, the communication device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The memory 330 includes a computer program, an operating system, and acquired data. The processor 310 can call logical instructions in the memory 330 to execute a multi-dimensional embodied data quality verification and processing method. This method includes: receiving raw multimodal data streams collected by an embodied intelligent robot; performing time-series alignment on the raw multimodal data streams in a pre-collection buffer pool to aggregate them into joint data samples; extracting multi-dimensional feature vectors from the joint data samples in real time, wherein the multi-dimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features; calculating multi-dimensional quality indicators based on the temporal features, spatial geometric features, and ontological state features; weighting and aggregating the multi-dimensional quality indicators to generate a comprehensive quality score for the joint data samples; comparing the comprehensive quality score with preset high-quality thresholds and low-quality thresholds, and making storage routing decisions based on the comparison results.

[0058] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a portion 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 described in the various embodiments of the present invention. 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.

[0059] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for performing the multi-dimensional embodied data quality verification and processing method provided by the above-described methods. The method includes: receiving a raw multimodal data stream collected by an embodied intelligent robot; performing time-series alignment on the raw multimodal data stream in a pre-collection buffer pool to aggregate it into a joint data sample; extracting multi-dimensional feature vectors from the joint data sample in real time, the multi-dimensional feature vectors including at least temporal features, spatial geometric features, and ontological state features; calculating multi-dimensional quality indicators based on the temporal features, spatial geometric features, and ontological state features; weighting and aggregating the multi-dimensional quality indicators to generate a comprehensive quality score for the joint data sample; comparing the comprehensive quality score with preset high-quality thresholds and low-quality thresholds, and making storage routing decisions based on the comparison results.

[0060] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0061] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A method for multi-dimensional embodied data quality verification and processing, characterized in that, include: Receive raw multimodal data streams collected by embodied intelligent robots; The original multimodal data stream is time-aligned in the pre-acquisition buffer pool and aggregated into a joint data sample; Multidimensional feature vectors are extracted in real time from the joint data samples. The multidimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features. Based on the temporal features, spatial geometric features, and ontological state features, multidimensional quality indicators are calculated, and the multidimensional quality indicators are weighted and aggregated to generate a comprehensive quality score for the joint data sample. The overall quality score is compared with preset high-quality thresholds and low-quality thresholds, and a storage routing decision is made based on the comparison results.

2. The multi-dimensional embodied data quality verification and processing method according to claim 1, characterized in that, The step of comparing the comprehensive quality score with preset high-quality thresholds and low-quality thresholds, and making storage routing decisions based on the comparison results, includes: When the overall quality score is greater than or equal to the high quality threshold, it is determined to pass, and the joint data sample is stored in the main training database; When the overall quality score is between the high quality threshold and the low quality threshold, it is determined to be degraded for storage, and the joint data sample is stored in the archive library; When the overall quality score is less than the low quality threshold, it is determined to be rejected.

3. The multi-dimensional embodied data quality verification and processing method according to claim 1, characterized in that, The step of performing time-series alignment on the original multimodal data stream in the pre-acquisition buffer pool and aggregating it into a joint data sample includes: Based on a hardware clock or precision clock synchronization protocol, timestamps with a unified time base are applied to the raw data frames of all modes. In the pre-acquisition buffer pool, triggered by a time window or key event, each modal data frame falling within the same time window is resampled onto a unified time series using a timestamp-based linear or spline interpolation algorithm to generate joint data samples.

4. The multi-dimensional embodied data quality verification and processing method according to claim 1, characterized in that, The step of extracting multidimensional feature vectors from the joint data samples in real time, wherein the multidimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features, including: Calculate the maximum timestamp deviation between each modal data frame within the joint data sample, and calculate the instantaneous frame rate of at least one key sensor and its stability deviation from the nominal frame rate to extract temporal features; Extract the image coordinates or 3D coordinates of a pre-defined calibration reference from visual or depth data, and calculate the residual between the reference and the calibration position; at the same time, calculate the sparsity ratio of the depth map or the optical distortion measure of the image to extract spatial geometric features. Based on the robot's kinematic model, the expected pose of the end effector is calculated according to the joint encoder readings, and the actual pose of the end effector is identified from the visual data. The instantaneous spatial error between the two is calculated. The smoothness feature score is quantized according to the instantaneous spatial error to extract the body state features.

5. The multi-dimensional embodied data quality verification and processing method according to claim 1, characterized in that, The overall quality score of the joint data sample includes a calibrated health score, and the method for generating the calibrated health score includes: For at least one target sensor in the joint data sample, its current reading is obtained, and the corresponding predicted reading is calculated based on the calibration model of the target sensor; Calculate the residual between the current reading and the predicted reading, and store the residual in a residual sequence; The distribution of the residual sequence is monitored in real time using a statistical estimation algorithm to obtain the mean or variance of the residuals; The calibrated health score is generated based on the changes in the mean or variance of the monitored residual sequence.

6. The multi-dimensional embodied data quality verification and processing method according to claim 1, characterized in that, The overall quality score of the joint data samples includes a cross-modal physical consistency score, and the method for generating the cross-modal physical consistency score includes: Obtain predefined physical association rules, and extract first modal features and second modal features corresponding to the physical association rules from the joint data samples; wherein, the first modal features include at least visual features, and the second modal features are robot body state features; The first modal feature and the second modal feature are input into a lightweight feature association model to obtain the predicted value of the target physical quantity; From the joint data sample, extract the actual observed values ​​corresponding to the target physical quantity; calculate the residual between the actual observed values ​​and the predicted values; The residual is compared with a dynamic threshold, and the cross-modal physical consistency score is generated based on the comparison result.

7. The multi-dimensional embodied data quality verification and processing method according to claim 1, characterized in that, Also includes: Receive feedback information from downstream model training tasks; Based on the feedback information, the preset weights of the multidimensional quality indicators and / or the high-quality threshold and low-quality threshold are dynamically adjusted.

8. A quality decision gate system for embodied data before disk storage, characterized in that, include: The receiving module is used to receive the raw multimodal data stream collected by the embodied intelligent robot; The alignment module is used to perform time-series alignment of the original multimodal data stream in the pre-acquisition buffer pool and aggregate it into a joint data sample; The extraction module is used to extract multi-dimensional feature vectors from the joint data samples in real time. The multi-dimensional feature vectors include at least temporal features, spatial geometric features, and ontological state features. The generation module is used to calculate multi-dimensional quality indicators based on the temporal features, spatial geometric features and ontological state features, and to perform weighted aggregation of the multi-dimensional quality indicators to generate a comprehensive quality score for the joint data sample. The decision module is used to compare the comprehensive quality score with preset high quality thresholds and low quality thresholds, and make storage routing decisions based on the comparison results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-dimensional embodied data quality verification and processing method as described in any one of claims 1 to 7.

10. A non-transitory readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the multi-dimensional embodied data quality verification and processing method as described in any one of claims 1 to 7.

Images