A Plug-and-Play Inertial Motion Tracking Method for Magnetometer-free Orientation Estimation of Arbitrary Joints

Inertial motion tracking (IMT) refers to technology that uses wearable inertial sensors to track the motion of robots or humans, and it has become fundamental to various applications. To achieve a detailed, location-independent analysis of movement patterns, IMT of kinematic chains (KC) is performed. Current IMT-KC methods require expert knowledge of the joints' degrees of freedom (DoF) and their corresponding axes, limiting the applicability of IMT-KC to professional settings. This paper introduces a plug-and-play, recurrent neural network (RNN)-based model for IMT-KC. It is designed to estimate the relative orientation between adjacent inertial measurement units (IMUs) mounted on segments that are connected by rotational joints with arbitrary DoF, without requiring prior knowledge of the joint axes or DoF. The model was trained solely on simulated data, and zero-shot estimates the relative orientation of the IMUs placed on 3D-printed real-world KCs with a mean error of less than 9° for all joint types (1 DoF, 2 DoF, and 3 DoF). Although the proposed method may not always match the absolute accuracy of specialized techniques, it stands out for its flexibility, robustness, and ease of deployment. Furthermore, we show that, once it converges, the model stays highly stable and is inherently resilient to variations in IMU placement. By significantly reducing the need for detailed prior knowledge of joint configurations and offering broad adaptability across different joint types, the proposed model simplifies the deployment of inertial motion tracking systems, making them more accessible to non-experts. This represents a significant step toward applying IMT-KC in biomedical fields such as rehabilitation and everyday motion analysis outside of clinical or laboratory environments, where ease of use is crucial.