Autonomous vehicles have to meet high safety standards in order to be commercially viable. Before real-world testing of an autonomous vehicle, extensive simulation is required to verify software functionality and to detect unexpected behavior. This incites the need for accurate models to match real system behavior as closely as possible. During driving, planing and control algorithms also need an accurate estimation of the vehicle dynamics in order to handle the vehicle safely. Until now, vehicle dynamics estimation has mostly been performed with physics-based models. Whereas these models allow specific effects to be implemented, accurate models need a variety of parameters. Their identification requires costly resources, e.g., expensive test facilities. Machine learning models enable new approaches to perform these modeling tasks without the necessity of identifying parameters. Neural networks can be trained with recorded vehicle data to represent the vehicle's dynamic behavior. We present a neural network architecture that has advantages over a physics-based model in terms of accuracy. We compare both models to real-world test data from an autonomous racing vehicle, which was recorded on different race tracks with high- and low-grip conditions. The developed neural network architecture is able to replace a single-track model for vehicle dynamics modeling.     «

Autonomous vehicles have to meet high safety standards in order to be commercially viable. Before real-world testing of an autonomous vehicle, extensive simulation is required to verify software functionality and to detect unexpected behavior. This incites the need for accurate models to match real system behavior as closely as possible. During driving, planing and control algorithms also need an accurate estimation of the vehicle dynamics in order to handle the vehicle safely. Until now, vehicl...     »