Abstract
Wearable computing devices and extended reality (XR) systems that rely on motion sensors such as low-cost inertial measurement units (IMUs) frequently suffer from tracking degradation due to sensor errors. Consumer-grade silicon often exhibits non-linearity, thermal drift, vibration rectification error (VRE), cross-axis coupling, and manufacturing variances that exceed the modeling capacity of traditional deterministic models and standard parameterization. Furthermore, relying on per-device physical factory calibration requires expensive equipment and thermal chambers, creating cost barriers and production bottlenecks. Moreover, existing software simulators produce motion data that fails to reflect real-world hardware populations.
The disclosed technology describes an AI-driven IMU calibration system that can replace physical factory testing with a synthetic data framework. An integrated pipeline ingests empirical characterization data from physical hardware tests to extract deterministic parameters and noise profiles across a sample device batch. The system calculates statistical means, variances, and crossparameter covariances from the extracted parameters to construct a statistical population model representing manufacturing variability across various units and statistical boundary cases. A simulator is then used to inject deterministic parameters, stochastic noise, and/or higher-order nonlinear error layers, including thermal compensation coefficients, dynamic motion-coupled biases, and programmable transfer functions for structural resonance, directly into the motion data to generate large-scale synthetic training datasets.
Using the generated synthetic training dataset, the system trains deep learning architectures (e.g., temporal convolutional networks or long short-term memory characterize multi-dimensional error residuals between raw sensor outputs and calibrated kinematic states. The trained AI model is then deployed as an active post-factory calibration layer directly on hardware, which can then continuously apply in-field sensor compensation to adapt to environmental changes, temperature fluctuations, and mechanical drift during device operation. This architecture provides a low-cost calibration framework that achieves higher tracking accuracy on low-cost motion sensors while operating within the memory, compute, and power requirements of wearable devices.
networks) configured to
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Brune, Philip Franklin and Hernandez, Joshua Anthony, "System and Method for Machine-Learning Calibration of Inertial Measurement Units Using Synthetic Training Datasets", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12230