Particle Gibbs with Ancestor Sampling for Identification of Tire-Friction Parameters


Particle Gibbs with Ancestor Sampling (PGAS) is a particle Markov chain Monte Carlo method (PMCMC) for Bayesian inference and learning. PGAS conditions on a referencestate trajectory in the underlying particle filter using ancestor sampling. In this paper, we leverage PGAS for identification of cornering-stiffness parameters in road vehicles only using production-grade sensors. The cornering-stiffness parameters are essential for describing the motion of the vehicle. We show how PGAS can be adapted to efficiently learn the stiffness parameters by conditioning on the noise-input trajectory instead of the state trajectory. We verify on a three-minute long experimental test drive that our method correctly identifies the tire-stiffness parameters.