Evaluating Plug-and-Play Applicability of Magnetometer-free IMU-based Joint Kinematics Estimation

Abstract

Magnetometer-free joint kinematics estimation requires plug-and-play (PnP) functionality for clinical use. However, traditional estimators like Multiplicative Extended Kalman Filter approaches (MEKF) and Maximum A Posteriori (MAP) approaches depend on domain-specific hyperparameters, limiting their autonomy. This study evaluates these methods against a parameter-free Recurrent Neural Network (RNN) across robotic and 3D-printed kinematic chains. Results show that while MEKF and MAP perform well under optimal conditions, their accuracy degrades significantly under motion artifacts due to a shift in optimal parameters. In contrast, the RNN, trained solely on simulated data, maintains robust performance across different domains and motion artifacts without any subject-specific tuning. Consequently, data-driven models offer a more scalable PnP promising solution for unsupervised clinical motion tracking.