Longitudinal Vehicle Dynamics: A Comparison of Physical and Data-Driven Models Under Large-Scale Real-World Driving Conditions

Abstract

Mathematical models of vehicle dynamics will form essential components of future autonomous vehicles. They may be used within inverse or forward control loops, or within predictive learning systems. Often, nonlinear physical models are used in this context, which, though conceptually simple (especially for decoupled, longitudinal dynamics), may be computationally costly to parameterise and also inaccurate if they omit vehicle-specific dynamics. In this study we sought to determine the relative merits of a commonly used nonlinear physical model of vehicle dynamics versus data-driven models in large-scale real-world driving conditions. To this end, we compared the performance of a standard nonlinear physical model with a linear state-space model and a neural network model. The large-scale experimental data was obtained from two vehicles; a Lancia Delta car and a Jeep Renegade sport utility vehicle. The vehicles were driven on regular, public roads, during normal human driving, across a range of road gradients. Both data-driven models outperformed the physical model. The neural network model performed best for both vehicles; the state-space model performed almost as well as the neural network for the Lancia Delta, but fell short for the Jeep Renegade whose dynamics were more strongly nonlinear. Our results suggest that the linear data-driven model gives a good trade-off in accuracy and simplicity, whilst the neural network model is most accurate and is extensible to more nonlinear operating conditions, and finally that the widely used physical model may not be the best choice for control design.

A Mental Simulation Approach for Learning Neural-Network Predictive Control (in Self-Driving Cars)

“…This structuring form needs only the little physical insight required to tell which forces act on the system and their causes u i . So, the total number of parameters in (2) is usually significantly smaller than in (1) because the F i operate on reduced dimension subsets of u [36], [37]. A second technique for shaping a neural network model consists of interpolating the output of M parallel models as follows:…”

“…Also, learning individual sub-models F i is less affected by imbalanced data-sets if the learning domain is partitioned. One example of local models with fixed receptive fields is given in [36], [37] where the drive-line sub-models are learned for different gears independently. The local models F i may be linear, i.e.…”

“…We now use the principles above for constructing a neural network capable of learning a Linear Time-Varying (LTV) model of a vehicle's lateral dynamics (see [36], [37] for a different example).…”

Use of Naturalistic Driving Studies for Identification of Vehicle Dynamics

“…It therefore cannot be counted as a practical implementation of NDS. The data taken in this project and used in [8], [9], [10], and [11] to identify controller parameters of an autonomous vehicle cannot fully be compared to field data taken during private vehicle use. It still indicates the potential for similar approaches.…”

“…It still indicates the potential for similar approaches. James [10] notes that data from everyday driving do not meet the requirements for identification (full range of excitation). However, the paper does not investigate what information can be collected in everyday driving, how the information growth proceeds.…”