Deep learning of vehicle dynamics

Recent advances in deep-learning-based identification of dynamic systems have resulted in a new generation of approaches utilizing state-space neural models with innovation noise structure, improved reformulation of multiple shooting, batch optimization, and a subspace identification-inspired form of encoders. The latter is used to learn a reconstructability map to estimate model states from past inputs and outputs. By using the SUBNET approach, which belongs to the state-of-the-art of these methods, we show how to effectively use these approaches to identify reliable vehicle models from data both in continuous and discrete time, respectively. We showcase the approach on the identification of the dynamics of a Crazyflie 2.1 nano-quadcopter and an F1tenth electric car both in a high-fidelity simulation environment, and in case of the electric car, on real measured data. The results indicate that new-generation of deep-learning methods offer efficient system identification of vehicle dynamics in practice.