Experimental Evaluation of Deep Neural Networks for Vehicle Model Identification

In light of the rapid evolution of Artificial Intelligence (AI), a growing number of researchers are investigating the use of Artificial Neural Networks (ANNs) to enhance first-principle Vehicle Models (VMs) or potentially replace them altogether. This paper investigates how AI can be used optimally to identify a VM in the context of a specific case study based on a small-scale experimental vehicle. To this end, three different VMs, each based on a distinct approach, are implemented and compared: (1) a Kinematic Vehicle Model (KVM), (2) a Deep Neural Network (DNN) based VM, and (3) a coupled approach of DNN with KVM, namely Improved KVM (IKVM), where the DNN is used to learn any unmodeled errors produced by the KVM. In the context of the DNN-based approaches, four types of DNNs are implemented based on different configurations of layers (fully connected, convolution, and long short-term memory). For DNN training and evaluation, a custom dataset of driving data is created by driving an F1tenth model car for around nine and a half hours on an indoor track while recording all motions using a motion tracking system. The experiments examine the VMs based on multiple performance metrics: the sampling period, 12 different scenarios, and the number of prediction steps the VMs are able to regressively predict without receiving updates regarding extrinsic vehicle states before the error grows too large, i.e., above 1 cm. Our findings are that DNN can increase KVM fidelity substantially. The optimal use of VMs, however, depends on the problem parameters and the vehicle states to be predicted.