Vehicle Dynamics Learning From Physics Priors With Model-Structured Neural Networks

Modeling the vehicle dynamics near the handling limits is crucial for autonomous driving and racing. However, building accurate models remains challenging due to strong nonlinearities, limited data availability, and changing environmental conditions. Physics-based models offer interpretability but require costly parameter identification, while general-purpose neural networks typically need large datasets to generalize. This paper introduces a new model-structured neural network (MS-NN-full) that learns the coupled lateral-longitudinal vehicle dynamics by embedding physical knowledge into its internal architecture. MS-NN-full combines physics-inspired neuro-fuzzy models with data-driven components to capture the quasi-steady-state and transient behavior, as well as the mutual coupling between the lateral and longitudinal dynamics. Experimental results on a 1:10-scale autonomous vehicle show that MS-NN-full outperforms existing MS-NN baselines and general-purpose neural networks in accuracy and generalization, using less than two minutes of training data. The model also demonstrates rapid adaptation to new tire configurations and increased vehicle mass with minimal fine-tuning. We release our implementation and datasets to support further research in physics-guided learning for autonomous systems.