An IMU wearing error simulation and high-precision data acquisition method for embodied intelligence

Abstract

The application relates to an IMU wearing error simulation and high-precision data acquisition method for embodied intelligence, which is executed in a test environment containing a multi-degree-of-freedom industrial robot arm and a variable-damping installation adaptation device and comprises the following core steps: (S1) system initialization and motion trajectory library loading; (S2) installation deviation calibration; (S3) time synchronization calibration; (S4) variable-damping physical contrast collection for embodied artifacts; (S5) spatiotemporal true value alignment and multi-modal data set generation. The application fills the embodied intelligence data blank, truly restores human demonstration dynamics, realizes sub-millimeter absolute motion true value and microsecond time synchronization, realizes multi-level motion data set and quantitative wearing artifact simulation, realizes full-process automatic collection, is low in cost, easy to deploy, supports multi-type algorithm research, and is high in engineering practicability. The functions of the modules of the system are independent, the modules are high in cooperativity, the calibration and collection method steps are clear and standardized, and the system supports user-defined trajectory extension.

Description

Technical Field

[0001] This invention relates to the fields of embodied intelligence, robot sensing, and human teaching technology, specifically to a method for simulating IMU wearable errors and acquiring high-precision data for embodied intelligence. Background Technology

[0002] Inertial measurement units (IMUs), as core sensing components for robot attitude perception and motion measurement, are widely used in robot teleoperation, UAV navigation, and industrial robotic arm attitude detection. However, IMUs face numerous technical bottlenecks in practical engineering applications, severely limiting their measurement accuracy and engineering practicality. These problems mainly stem from the IMU's own hardware characteristics, its installation method with the carrier, and external environmental interference.

[0003] On the one hand, the gyroscopes and accelerometers of IMUs inherently possess zero-bias drift and temperature drift characteristics, leading to accumulated errors in the integrated attitude over time. Simultaneously, the measurement data is mixed with various noise components such as random walk and quantization noise, and current technologies lack effective calibration and filtering schemes, making it difficult to eliminate the impact of these noises on measurement accuracy.

[0004] On the other hand, the installation method of the IMU and the measured carrier has a significant impact on measurement accuracy. Both rigid and flexible installations introduce different error characteristics. However, current technologies lack systematic research and quantitative analysis methods for the IMU installation effect, failing to provide a scientific basis for engineering installation selection. Especially in wearable IMU applications, the relative sliding of human skin during movement, sensor position shifts caused by muscle contraction and relaxation, and high-frequency vibrations generated during rapid limb movement introduce a large amount of nonlinear noise unrelated to real motion into the IMU-acquired signals—i.e., wearable artifacts. Existing methods often employ low-pass filtering and Kalman filtering for post-processing, but these methods struggle to separate "real motion" from "wearable artifacts," often resulting in signal delay and loss of high-frequency details, failing to meet the demands of embodied intelligence for high-fidelity motion data.

[0005] In IMU calibration, traditional methods rely on high-precision turntables or optical motion capture systems to provide ground truth values. However, these devices are expensive and complex to operate, making it difficult to achieve automated, large-scale data acquisition with ground truth labels. Although some studies have used robotic arms for IMU calibration, these studies have not solved the problem of high-precision time synchronization between robotic arm ground truth data and IMU measurement data, lack precise calibration methods for the IMU coordinate system and robotic arm tool coordinate system, and have low levels of automation in data acquisition, making it impossible to generate dedicated datasets for practical applications such as industrial remote operation.

[0006] Crucially, in the field of embodied AI, training effectiveness heavily relies on large-scale, high-quality human teaching data. This data not only needs to contain rich visual and proprioceptive information but also requires motion signals with high precision, low latency, and realistic noise distribution to support accurate model imitation and generalization in complex physical environments. However, current technologies lack a systematic method for generating paired datasets that "have absolute ground truth and include real-world wear noise." Traditional IMU calibration methods cannot reflect the complex error characteristics in wearable scenarios, while direct data collection from the human body cannot obtain absolute motion ground truth, resulting in a lack of effective supervised training data for deep learning denoising networks and wear error compensation models. Existing research often uses simulated data or artificially synthesized noise for training, which differs significantly from the real-world wear error distribution, leading to poor generalization ability of the models in practical deployments.

[0007] In summary, existing technologies struggle to achieve low-cost, high-precision, and automated calibration and data acquisition of IMUs, and cannot provide high-quality, truth-labeled data for IMU installation effect research, drift suppression algorithm development, and embodied intelligence model training. Therefore, a novel IMU data acquisition and truth calibration scheme is urgently needed to address these technical challenges and promote the widespread application and development of IMUs in robotics and embodied intelligence fields. Summary of the Invention

[0008] This invention provides a method for simulating IMU wearability errors and acquiring high-precision data for embodied intelligence. It aims to solve the problems of low accuracy, missing ground truth, and uncontrollable noise in IMU teaching data caused by wearability artifacts in the prior art. By simulating human motion trajectory with an industrial robotic arm and simulating skin slippage, muscle deformation, and motion tremors with a controllable physical damping device, a paired dataset containing both real wearability errors and high-precision motion ground truth is generated. This provides high-quality data support for imitation learning of embodied intelligence models, training of motion denoising networks, and optimization of teleoperation accuracy.

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] A method for simulating wearable IMU errors and acquiring high-precision data for embodied intelligence, which relies on a test environment including a multi-degree-of-freedom industrial robotic arm and a variable damping mounting adapter, includes the following core steps:

[0011] (S1) System initialization and motion trajectory library loading:

[0012] Configure basic parameters such as IMU sensor sampling frequency and range, and establish a communication link between the host computer and the robotic arm controller. Load the preset "motion trajectory library," which contains two main categories: basic reference motion and humanoid complex motion, covering motion modes from simple to complete. Specifically, it includes:

[0013] Basic reference exercises:

[0014] Static state: Used to quantify the static drift characteristics of the IMU and acquire zero-bias reference data.

[0015] Uniform linear motion: Moving at a constant speed along the X, Y, and Z axes respectively, used to analyze the steady-state response of the IMU at a constant speed.

[0016] Sinusoidal oscillation motion: The end of the instrument reciprocates sinusoidally around a certain axis, which is used to analyze the dynamic response characteristics of the IMU under periodic variable acceleration motion.

[0017] Circular motion: The end effector performs uniform circular motion in a plane, used to analyze the effect of centripetal acceleration on IMU measurements.

[0018] Humanoid complex motion:

[0019] The trajectory simulates the rapid start and stop motion of a human grasping intent: It includes instantaneous start and stop movements with high jerk (Jerk) to simulate the sudden speed change of the end effector when a human grasps or places an object, thereby stimulating transient artifacts caused by inertial impact on the wearable IMU.

[0020] Superimposed high-frequency micro-vibration trajectory to simulate human physiological tremors: 5-15Hz high-frequency small-amplitude sinusoidal micro-vibrations are superimposed on the baseline motion trajectory to simulate human muscle contraction, physiological tremors, or high-frequency shaking when holding tools, quantifying the dynamic response error of the IMU under such conditions.

[0021] Simulating a six-DOF composite spatial reversal motion for large-scale obstacle avoidance and handling: This trajectory includes rapid reversal, curved rotation, and continuous attitude changes at the end point in three-dimensional space. It is used to simulate the coordinated full-body movements of humans in complex operations such as handling and obstacle avoidance, and to test the error accumulation characteristics of the IMU under complex spatial motion.

[0022] (S2) Installation deviation calibration:

[0023] Since the IMU sensor is fixed to the end effector flange of the robotic arm via a variable damping mounting adapter, a fixed rotational relationship exists between its coordinate system and the end effector tool coordinate system. By controlling the robotic arm to move to multiple known postures, the posture data output by the IMU sensor in each posture is read. The fixed rotational deviation matrix between the IMU coordinate system and the end effector tool coordinate system is solved using the least squares method. This deviation matrix is ​​stored as a mounting parameter for subsequent spatial coordinate transformation of the IMU measurement data, unifying all data to the robotic arm's base coordinate system.

[0024] (S3) Time synchronization calibration:

[0025] The host computer software synchronization module is activated, employing an architecture that separates the acquisition thread from the control thread. The control thread sends control commands to the robotic arm at a first preset frequency to ensure smooth trajectory reproduction. The acquisition thread runs at a second preset frequency, immediately reading the latest IMU value and the end-effector pose value calculated by the robotic arm controller at each sampling moment, and assigning the same millisecond-level timestamp to this data pair by a high-precision timer on the host computer. This mechanism achieves microsecond-level time synchronization accuracy without additional hardware triggering, ensuring precise alignment of the motion truth and sensor signals in the time domain.

[0026] (S4) Physical comparison of variable damping for embodied artifacts:

[0027] For each baseline trajectory in the aforementioned motion trajectory library (including stationary state, uniform linear motion, sinusoidal oscillation, circular motion, rapid start / stop, high-frequency micro-vibration, and compound reciprocating motion), synchronous data was collected under the following three control installation modes. Each mode was repeated multiple times to ensure data repeatability and statistical validity.

[0028] Mode A (Rigid Reference Acquisition): A rigid mounting bracket is used to directly fix the IMU sensor to the flange at the end of the robotic arm, achieving a rigid connection between the IMU and the moving platform. The data acquired in this mode is ideal sensor measurement reference data without wearable artifacts, serving as the "absolute true value" for subsequent error analysis.

[0029] Mode B (Human Soft Tissue Elasticity Simulation Acquisition): A flexible pad with a specific elastic coefficient is clamped between the IMU and the end flange of the robotic arm to simulate the low-frequency buffering error caused by the sliding and deformation of human skin and subcutaneous fat during movement. This mode aims to quantify the filtering and hysteresis effects of soft tissue on low-frequency motion signals (such as large-range swinging).

[0030] Mode C (Simulation of Strap Loosening and High-Frequency Impact): This mode uses an asymmetric, loose, porous material pad and a partially constrained installation method to simulate random high-frequency contact artifacts generated when wearable devices age, straps loosen, or clothing collide with them. This mode aims to quantify high-frequency impact noise and instantaneous sensor attitude changes caused by poor device fixation.

[0031] (S5) Spatiotemporal truth alignment and multimodal dataset generation:

[0032] The raw IMU data (acceleration, angular velocity, quaternions) with different physical artifacts collected under modes A, B, and C are precisely time-aligned with the high-precision end-effector pose calculated from the forward kinematics of the robotic arm at the same moment, using the timestamp established in step three as the reference. Simultaneously, the IMU measurement data is transformed to the robotic arm's base coordinate system using the rotational deviation matrix calibrated in step (S2), achieving spatial coordinate system unification.

[0033] The aligned and transformed data is stored in a structured format (such as CSV or Parquet) to construct an embodied intelligence pairing dataset with real physical noise labels. Data fields include: millisecond-level timestamps, IMU three-axis acceleration, IMU three-axis angular velocity, robotic arm end-effector 3D position, robotic arm end-effector 3D pose (Euler angles or quaternions), acquisition mode labels (A / B / C), trajectory type labels, and damping module parameters, among other metadata. This dataset can be directly used to train IMU wearable error compensation networks, motion denoising models, and motion decoding modules for large-scale embodied intelligence models.

[0034] The beneficial effects of this invention are as follows: Compared with the prior art, this invention has the following beneficial effects:

[0035] (1) It fills the data gap of embodied intelligence and truly restores the dynamics of human teaching: For the first time, it proposes to actively simulate human wearing errors by using an industrial robotic arm combined with a variable damping physical control experiment (rigid installation / high elastic silicone / loose foam). While obtaining the absolute motion truth value, it generates real wearing noise and constructs an accurate matching dataset of ideal benchmark and wearing noise. At the same time, it designs a humanoid motion trajectory library, including trajectories that simulate irregular human movements such as rapid start and stop, physiological tremor, and six-degree-of-freedom compound reversal. The generated dataset can better reflect the dynamic complexity of real teleoperation scenarios and provide high-quality data support for the training of anti-interference embodied large models and IMU denoising networks.

[0036] (2) Submillimeter-level absolute motion truth and microsecond-level time synchronization: Using the forward kinematics calculation results of the industrial robotic arm as the truth label, the repeatability accuracy reaches ±0.02mm, which is far higher than the millimeter-level accuracy of the optical motion capture system. A software synchronization mechanism that separates the acquisition thread and the control thread is adopted to assign the same millisecond-level timestamp to the IMU data and the robotic arm pose data, achieving microsecond-level synchronization accuracy without the need for additional hardware triggering, ensuring the precise alignment of the truth and the signal in the time domain.

[0037] (3) Multi-level motion dataset and quantitative simulation of wearable artifacts: The motion trajectory library covers basic reference motions such as stillness, uniform linear motion, sinusoidal oscillation, and circular motion, as well as complex human-like motions such as rapid start and stop, physiological tremor, and compound return, forming a multi-level data system from simple to complete. By changing the damping modules with different hardness, thickness, and viscoelastic properties, the error characteristics under different wearing conditions such as soft tissue slippage and strap slack can be quantitatively simulated, providing experimental support for the study of the physical mechanism of wearable artifacts and error modeling.

[0038] (4) Fully automated data acquisition, low cost and easy deployment: The host computer programs and controls the robotic arm to move automatically along a preset trajectory, supporting repeated execution under multiple working conditions and continuous data acquisition 24 hours a day. The asynchronous writing mechanism avoids disk I / O blocking, greatly improving data acquisition efficiency and repeatability. The system is based on the expansion of the existing industrial robotic arm platform, without the need for a high-precision turntable or optical motion capture system. The installation and adaptation device is simple in structure and low in cost, and can be quickly deployed in laboratories or industrial sites with robotic arms.

[0039] (5) Supports research on multiple algorithms and has strong engineering applicability: The generated structured pairing dataset can be directly used for IMU wearable error compensation network training, remote operation teaching signal optimization, and pre-training of embodied large model action modalities such as VLA, supporting end-to-end policy learning from human teaching to robot execution. The system's modules are functionally independent and highly collaborative, the calibration and acquisition methods are clear and standardized, and it supports user-defined trajectory extensions, possessing good engineering applicability and scalability. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention;

[0041] Figure 2 This is an assembly diagram of the data acquisition system according to an embodiment of the present invention.

[0042] Figure 3 This is a structural assembly diagram of the IMU installation adapter device according to an embodiment of the present invention.

[0043] Figure 4 This is a flowchart of the IMU data acquisition and truth calibration method according to an embodiment of the present invention.

[0044] Figure 5 This is a schematic diagram of the data acquisition timing according to an embodiment of the present invention.

[0045] The markings in the diagram are as follows: 1. First industrial robotic arm; 2. Second industrial robotic arm; 3. Third industrial robotic arm; 4. Fourth industrial robotic arm; 5. Fifth industrial robotic arm; 6. Sixth industrial robotic arm; 7. Base; 8. Connecting base; 9. IMU sensor; 10. Sensor mounting plate; Detailed Implementation

[0046] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. These embodiments are merely illustrative and are not intended to limit the scope of protection of the invention.

[0047] Example

[0048] The method for embody intelligence-oriented IMU wearable error simulation and high-precision data acquisition in this embodiment specifically includes the following steps:

[0049] (1) System setup:

[0050] like Figure 1 The diagram shown is a schematic representation of the system structure of an embodiment of the present invention. This embodiment constructs an IMU wearable error simulation and high-precision data acquisition system for embodied intelligence, employing a six-degree-of-freedom collaborative robot as an industrial robotic arm, with an end-effector repeatability accuracy of ±0.02mm. An industrial-grade MEMS IMU is selected, with a measurement range of ±2000° / s (gyroscope) and ±8g (accelerometer), and a sampling frequency of 200Hz. The wearable error simulation device includes a rigid mounting base and replaceable physical damping modules: a high-elasticity silicone pad (Shore hardness A20, thickness 3mm) for simulating soft tissue slippage, and a loose polyurethane foam pad (thickness 5mm, asymmetrical structure) for simulating strap loosening and high-frequency impact.

[0051] like Figure 2 , Figure 3 As shown. The six joints of the first industrial robotic arm 1, the second industrial robotic arm 2, the third industrial robotic arm 3, the fourth industrial robotic arm 4, the fifth industrial robotic arm 5, and the sixth industrial robotic arm 6 are fixed to a base 7 via flanges, and a mounting adapter consisting of a connecting base 8, an IMU sensor 9, and a sensor mounting plate 10. The mounting adapter includes the connecting base 8, the IMU sensor 9, and the sensor mounting plate 10. The connecting base 8 is used to fix the robotic arm end effector; the IMU sensor 9 is used to collect the end effector's acceleration, angular velocity, and end effector attitude; the sensor mounting plate 10 is used to support the IMU sensor, and a variable damping physical module can be optionally installed between the IMU sensor 9 and the sensor mounting plate 10 to simulate physical artifacts caused by the elasticity of human soft tissue or the loosening of straps.

[0052] The robotic arm controller is electrically connected to the industrial robotic arm, with the robotic arm's IP address being 192.168.1.100. It is responsible for sending motion commands and recording joint data. The host computer communicates with the robotic arm controller via Ethernet (IP address 192.168.1.112) and with the IMU sensor via USB, running self-developed data acquisition software that integrates installation deviation calibration, software synchronization, and data preprocessing functions. A data storage server is used to store structured datasets and metadata.

[0053] 2. Implementation Process

[0054] like Figure 4 As shown, this embodiment is executed according to steps one through five of the technical solution:

[0055] Step 1: System Initialization and Motion Trajectory Library Loading:

[0056] The host computer establishes communication with the robotic arm controller and IMU, configuring the acquisition frequency to 100Hz and the acquisition time for a single trajectory to be 60 seconds. A preset motion trajectory library is loaded, including: stationary (60 seconds), uniform linear motion on the X-axis (speed 0.2m / s), sinusoidal oscillation on the Y-axis (amplitude 10°, frequency 1Hz), planar circular motion (radius 0.1m, angular velocity 10° / s), rapid start / stop (accelerating from stationary to 0.5m / s and then immediately stopping), oscillation trajectory superimposed with 10Hz high-frequency micro-vibration, and six-degree-of-freedom composite reciprocating motion (continuously changing spatial curves).

[0057] Step 2: Installation Deviation Calibration

[0058] Control the robotic arm to move to 3 different postures (each posture is held for 5 seconds), and simultaneously record the posture of the robotic arm end effector and the posture data output by the IMU. Use the least squares method to solve the fixed rotation deviation matrix between the IMU coordinate system and the robotic arm end effector coordinate system, and store it for subsequent coordinate transformation.

[0059] The robotic arm is controlled to move to three different postures using a hand-eye calibration principle. Model solution. Let the robot arm's base coordinate system be... The coordinate system of the end flange is The IMU sensor coordinate system is .

[0060] The attitude transformation matrix of the robotic arm's end effector is:

[0061]

[0062] The attitude transformation matrix of the IMU sensor is:

[0063]

[0064] In the formula: Indicates that the end flange of the robotic arm starts from the first... The movement reaches the first posture. The relative pose transformation matrix for each pose; Indicates that the robotic arm is in the first position. At any given orientation, the end flange coordinate system Relative to the base coordinate system The homogeneous pose transformation matrix; Indicates that the robotic arm is in the first position. At any given orientation, the end flange coordinate system Relative to the base coordinate system The homogeneous pose transformation matrix; This indicates the operation of inverting a matrix. Indicates that the IMU sensor is from the first The movement reaches the first posture. The relative pose transformation matrix for each pose; Indicates that the robotic arm is in the first position. At each pose, the IMU sensor coordinate system Relative to the base coordinate system The homogeneous pose transformation matrix; Indicates that the robotic arm is in the first position. At each pose, the IMU sensor coordinate system Relative to the base coordinate system The homogeneous pose transformation matrix.

[0065] Construct a least-squares objective function to solve for the optimal fixed installation deviation matrix X* (i.e.):

[0066]

[0067] In the formula: This represents the optimal fixed installation deviation matrix obtained by solving the least squares objective function. This represents the matrix that minimizes the objective function. ; Represents the coordinate system of the end flange to be solved. With IMU sensor coordinate system The fixed installation deviation matrix between them; The Frobenius norm is used to measure the magnitude of matrix error, and the resulting rotation matrix will serve as the benchmark for subsequent spatial transformations.

[0068] Step 3: Time Synchronization Calibration

[0069] Start the software synchronization module to allow the control thread (125Hz) and the acquisition thread (100Hz) to run independently. Set the global high-precision timer timestamp to [value missing]. The time for the acquisition thread to read IMU data is The time to read the robotic arm pose is For any k-th frame sample, define the data alignment time difference. :

[0070]

[0071] In the formula: Represents any number of The alignment time difference between IMU data and robotic arm pose data during frame sampling; Indicates that the acquisition thread reads the first... Timestamps of frame IMU data; Indicates that the acquisition thread reads the first... Timestamps of frame robotic arm pose data.

[0072] The system determines in real time: if and only if (Tolerance threshold in this embodiment) When a pair is determined to be a valid synchronization pair, a unified global timestamp is assigned to it. ;like This triggers a linear interpolation compensation algorithm to perform phase alignment on the sensor data.

[0073] In the formula: This indicates the preset data alignment time tolerance threshold; Indicates assigning the first A unified global timestamp for valid frame synchronization pairs; Indicates the global high-precision timer at the 1st The timestamp value output at the frame sampling time.

[0074] Step 4: Variable Damping Physical Comparison Data Acquisition for Embossed Artifacts:

[0075] For each trajectory in the motion trajectory library, synchronous data is collected in sequence under the following three installation modes, with each mode repeated twice:

[0076] Mode A (Rigid Reference): The IMU is directly fixed to the end flange of the robotic arm via a rigid mounting bracket.

[0077] Mode B (Soft Tissue Elasticity Simulation): A silicone pad with a Shore A20 hardness and a thickness of 3mm is clamped between the IMU and the rigid mount.

[0078] Mode C (Strap slack and high-frequency impact simulation): An asymmetric loose polyurethane foam gasket is clamped between the IMU and the rigid mounting base, and a non-fully constrained fixing method is used.

[0079] During the data acquisition process, an asynchronous write mechanism is enabled. The acquisition thread stores the data pairs into a memory queue, and an independent write thread writes them to disk asynchronously, thus avoiding I / O blocking.

[0080] Step 5: Spatiotemporal truth alignment and multimodal dataset generation:

[0081] like Figure 5 As shown, using timestamps as a reference, the raw IMU data acquired in the three modes are precisely aligned with the corresponding true values ​​of the robotic arm's end effector pose. The rotational deviation matrix calibrated in step two is used to transform the IMU data to the robotic arm's base coordinate system. Finally, the data is stored in .csv format, with fields including: timestamp (milliseconds), IMU three-axis acceleration, IMU three-axis angular velocity, IMU three-dimensional attitude (quaternion), robotic arm end effector three-dimensional position, robotic arm end effector three-dimensional attitude (quaternion), acquisition mode label (A / B / C), trajectory type label, and damping module parameters. A JSON metadata file is also generated to record the experimental conditions and calibration parameters.

[0082] After establishing coordinate system one using the calibration matrix, in order to directly match the imitation learning requirements of the Embodied Intelligence Large Model (VLA model), the data preprocessing module encapsulates the collected alignment data into a sequence of teaching trajectories in a Markov Decision Process (MDP):

[0083]

[0084] In the formula: Represents the complete teaching trajectory sequence in a Markov decision-making process; Let represent any discrete time step in the teaching trajectory sequence, and ; This indicates the total time step (i.e., total number of frames) contained in the teaching trajectory sequence. They represent the first The observed state feature vector of the step; Indicates the first The feature vector of expert action labels for each step.

[0085] The observation state The specific expression is:

[0086]

[0087] In the formula: Indicates the first The observation state feature vector at each time step; Indicates the first Each time step is a three-axis acceleration data collected by the IMU sensor and unified to the robot arm's base coordinate system after spatial coordinate transformation; Indicates the first Each time step is a three-axis angular velocity data collected by the IMU sensor and unified to the robot arm's base coordinate system after spatial coordinate transformation; This represents the acquisition mode environment variable, which is a one-hot encoded environment variable characterizing the three physical artifact mounting modes: rigid reference, silicone elasticity, and foam relaxation. .

[0088] The expert action tag Specifically defined as the increment of the end-effector pose true value calculated with high precision by the robotic arm, its expression is:

[0089]

[0090] In the formula: Indicates the first Expert action label feature vectors at each time step; Indicates the first Each time step is the true value of the end-effector's three-dimensional position (translation) increment calculated by the high-precision forward kinematics of the robotic arm; Indicates the first Each time step is used to calculate the true value of the end-effector three-dimensional attitude (rotation) increment from the high-precision forward kinematics of the robotic arm. The three-dimensional attitude increment is represented by Euler angles or quaternions to construct a high-fidelity embodied intelligent pairing dataset.

[0091] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for simulating wearable errors and acquiring high-precision data using an IMU (Integrated Mutual Assistance Unit) for embodied intelligence, characterized in that: The method is implemented in a test environment that includes a multi-degree-of-freedom industrial robotic arm and a variable damping mounting adapter, and includes the following core steps: (S1) System initialization and motion trajectory library loading: Configure basic parameters such as IMU sensor sampling frequency and range, and establish a communication link between the host computer and the robotic arm controller; load the preset "motion trajectory library", which contains two main categories: basic reference motion and humanoid complex motion, covering motion modes from simple to complete. (S2) Installation deviation calibration: Since the IMU sensor is fixed to the end flange of the robotic arm via a variable damping mounting adapter, there is a fixed rotational relationship between its coordinate system and the end-tool coordinate system of the robotic arm. By controlling the robotic arm to move to multiple known postures, the posture data output by the IMU sensor in each posture is read, and the fixed rotational deviation matrix between the IMU coordinate system and the end-tool coordinate system of the robotic arm is solved using the least squares method. This deviation matrix is ​​stored as an installation parameter for subsequent spatial coordinate transformation of the IMU measurement data, unifying all data to the robotic arm base coordinate system. (S3) Time synchronization calibration: The host computer software synchronization module is started, adopting an architecture that separates the acquisition thread and the control thread; the control thread sends control commands to the robotic arm at a first preset frequency to ensure the smoothness of trajectory reproduction; the acquisition thread runs at a second preset frequency, immediately reading the latest IMU value and the end pose value calculated by the robotic arm controller at each sampling moment, and the host computer's high-precision timer assigns the same millisecond-level timestamp to this set of data pairs; (S4) Physical comparison of variable damping for embodied artifacts: For each baseline trajectory in the motion trajectory library, synchronous data is collected under the following three comparative installation modes; each mode is repeated multiple times to ensure the repeatability and statistical validity of the data. (S5) Spatiotemporal truth alignment and multimodal dataset generation: The raw IMU data (acceleration, angular velocity, quaternion) with different physical artifacts collected under the above three modes are precisely time-aligned with the high-precision end-effector pose true value calculated by the forward kinematics of the robotic arm at the same time, based on the timestamp established in step (S3); at the same time, the IMU measurement data are transformed to the robotic arm base coordinate system using the rotation deviation matrix calibrated in step (S2), realizing the unification of the spatial coordinate system. The aligned and transformed data is stored in a structured format (such as CSV or Parquet) to build an embodied intelligent pairing dataset with real physical noise labels; The data fields include: millisecond-level timestamps, IMU three-axis acceleration, IMU three-axis angular velocity, 3D position of the robotic arm end effector, 3D pose of the robotic arm end effector (Euler angles or quaternions), acquisition mode labels (A / B / C), trajectory type labels, and damping module parameters, etc. This dataset can be directly used to train the IMU wearable error compensation network, motion denoising model, and motion decoding module of embodied intelligent large model.

2. The method for simulating wearable IMU errors and acquiring high-precision data for embodied intelligence according to claim 1, characterized in that: In the method described above, the base reference motion includes: (1) Static state: used to quantify the static drift characteristics of the IMU and collect zero-bias reference data; (2) Uniform linear motion: Uniform motion is performed along the X, Y, and Z axes respectively, which is used to analyze the steady-state response of the IMU at a constant speed; (3) Sinusoidal oscillation motion: The end moves in a sinusoidal reciprocating motion around a certain axis, which is used to analyze the dynamic response characteristics of the IMU under periodic variable acceleration motion; (4) Circular motion: The end moves in uniform circular motion in the plane, which is used to analyze the effect of centripetal acceleration on IMU measurement.

3. The method for simulating wearable errors and acquiring high-precision data for IMUs oriented towards embodied intelligence according to claim 1, characterized in that: The method described includes: (1) Simulating the rapid start and stop motion of human grasping intention: The trajectory includes instantaneous start and stop movements with high jerk (Jerk) to simulate the sudden speed change of the end effector when humans grasp and place objects, thereby stimulating the transient artifacts generated by the wearable IMU due to inertial impact. (2) Superimposed high-frequency micro-vibration trajectory to simulate human physiological tremor: 5-15Hz high-frequency small-amplitude sinusoidal micro-vibration is superimposed on the reference motion trajectory to simulate human muscle contraction, physiological tremor or high-frequency shaking when holding a tool, and to quantify the dynamic response error of the IMU under such working conditions. (3) Simulate the six-degree-of-freedom composite spatial reversal motion of large-scale obstacle avoidance and transportation: The trajectory includes rapid reversal, curve rotation and continuous attitude change in the three-dimensional space at the end point, which is used to simulate the whole-body coordinated movement of humans in complex operations such as transportation and obstacle avoidance, and to test the error accumulation characteristics of IMU under complex spatial motion.

4. The method for simulating wearable errors and acquiring high-precision data for embodied intelligence using IMUs according to claim 1, characterized in that: In step (S4), the reference trajectory includes static state, uniform linear motion, sinusoidal oscillation, circular motion, rapid start and stop, high-frequency micro-vibration, and compound reciprocating motion.

5. The method for simulating wearable errors and acquiring high-precision data for embodied intelligence using IMUs according to claim 1, characterized in that: In step (S4), the three modes are as follows: (1) Mode A (rigid reference acquisition): The IMU sensor is directly fixed to the end flange of the robotic arm using a rigid mounting bracket to achieve a rigid connection between the IMU and the motion carrier; the data acquired in this mode is the ideal sensor measurement reference data without wearable artifacts, which serves as the "absolute true value" for subsequent error analysis. (2) Mode B (Human soft tissue elasticity simulation acquisition): A flexible pad with a specific elastic coefficient is clamped between the IMU and the end flange of the robotic arm to simulate the low-frequency buffering error caused by the sliding and deformation of human skin and subcutaneous fat during movement; this mode aims to quantify the filtering and hysteresis effect of soft tissue on low-frequency motion signals. (3) Mode C (Simulation of strap slack and high frequency impact): Asymmetric loose porous material pads are used and a non-fully constrained installation method is adopted to simulate random high frequency contact artifacts generated when wearable devices are aged, straps are loose, or clothing is collided; this mode aims to quantify the high frequency impact noise and instantaneous attitude change of the sensor caused by poor wear fixation.

Images