Inertial measurement units (IMUs) are used for inertial motion tracking (IMT) in a growing number of applications as sensor fusion methods are being advanced in three directions: magnetometer-free IMT methods that eliminate the effect of magnetic disturbances; sparse IMT approaches that lead to reduced setup complexity; and automatic self-calibration of sensor-to-segment positions or orientations. In this letter, we propose an approach that combines all three achievements and, for the first time, enables plug-and-play, magnetometer-free, and sparse IMT. This is accomplished by training a recurrent neural-network-based observer (RNNo) on just-in-time generated simulated motion data of kinematic chains. We demonstrate that domain-specific training data augmentations lead to a trained RNNo which zero shot generalizes to previously unseen experimental data and, thus, overcomes the sim-to-real gap. The trained RNNo achieves a tracking error of $< 4$ degrees when estimating the relative pose of a three-segment kinematic chain with two hinge joints. The proposed method offers a novel simulation-data-driven approach for solving complex sparse sensing problems while assuring robust and plug-and-play generalizability to experimental data.