Rigid Body Movement Prediction Using Dual Quaternion Recurrent Neural Networks
Andreas Schwung, Johannes Pöppelbaum, Pradeep C. Nutakki · 2021
This paper presents a novel approach for the data based prediction of rigid body movements. To this end, we combine data based learning with a physically motivated neural network architecture using the theory of dual quaternions. Particularly, we develop a novel neural network architecture based on dual quaternion algebra which is particular suitable for representing rigid body movements. To account for multi-step predictions and the inherent dynamics of rigid bodies, we particularly focus on recurrent neural networks. As such we propose both dual quaternion recurrent neural networks as well as dual quaternion long short term memories. We apply the approach to a simplified simulation environment developed using the discrete element method which allows for a very detailed simulation of the movements. The obtained results underline the applicability and potential of the approach in terms of improved prediction performance.