Inertial measurement units are widely used for motion tracking of kinematic chains in numerous applications. While magnetometer-free sensor fusion enables reliably high accuracy in indoor environments and near magnetic disturbances, the use of sparse sensor setups would yield additional advantages in cost, effort, and usability. However, it is unclear which sparse sensor setups can be used to track which motions of which kinematic chains, since observability of the underlying nonlinear dynamics is barely understood to date. We propose a method that utilises recurrent neural networks (RNNs) and automatically generated training data to assess the observability of the relative pose of kinematic chains in sparse inertial motion tracking (IMT) systems. We apply this method to a range of double hinge-joint systems that perform fully-exciting random motion. Results show how the degree of observability depends on the kinematic structure and that RNN-based observers can achieve small tracking errors in a large range of sparse and magnetometer-free setups. The proposed methods enables systematic assessment of observability properties in complex nonlinear dynamics and represents a key step toward enabling reliably accurate and non-restrictive IMT solutions.