In high-speed autonomous racing, it is necessary to have an accurate racecar
vehicle dynamics model in order to push the vehicle closer to its limits. The
choice of the dynamics model has to be made by balancing the computational
demands in contrast to model complexity. Learning-based methods, such as
Gaussian processes (GP)-based regression, have shown promise toward
approximating the vehicle dynamics model. In particular, such methods use a
simplified model structure that is easy to tune and then use GP to model the
mismatch between the output of the simple model and observed system dynamics.
However, current GP approaches often oversimplify the modeling process or apply
strong assumptions, leading to unrealistic results that cannot translate to
real-world settings. This article presents a comprehensive GP-based design for
modeling the dynamics of an autonomous racing car. We do so with high-fidelity
simulation data, a 1/10-scale autonomous racing car platform, and a full-scale
autonomous Indy racing car. In the first part of this article, we present a
rigorous empirical analysis highlighting how the open-loop and closed-loop
performance of GP models for autonomous racing is highly sensitive to the choice
of the GP kernel, the data sample size, and track configurations suggesting
there is no single easy choice that always works. We demonstrate this through a
combinatorial simulation setup for 1/10-scale autonomous racing cars. We then
present a novel method called DKL-SKIP, which uses deep kernel learning to
overcome the challenges of kernel selection and scalability for GP modeling. We
evaluate DKL-SKIP on a high-fidelity AutoVerse simulator as well as real data
from an autonomous real-world full-scale Indy racing car. Our results reveal
that DKL-SKIP surpasses scalable GP models and the N4SID algorithm in both
real-world and high-fidelity simulation environments.
