Bayesian Experimental Design With Application to Dynamical Vehicle Models

In this article, we propose two novel experimental design techniques for designing maximally informative experiments to estimate the parameters of nonlinear dynamical vehicle models. The two techniques include a batch design and a sequential design technique that seek to maximize the expected Shannon information gain of the parameter distribution using either an online or offline approach (respectively). We apply and compare the techniques in both simulation and real-world experiments with a wheeled vehicle. In our simulation experiments, both of our proposed designs provide superior Shannon information gains relative to an unoptimized benchmark technique. In our real-world experiments, our sequential design technique achieves superior expected Shannon information gains relative to our batch design technique and the benchmark technique.

I. Introduction

The rapid rise of robotic vehicle automation and control has driven the need for accurate mathematical models describing their motion. Many of the physical parameters of a vehicle, such as the tire cornering stiffness, are difficult to measure and have to be inferred from sensors, such as GPS and inertial measurement units. The amount of information we gain about the unknown parameters is largely dependent on how the system is excited. Designing an experiment so that it is maximally informative about the unknown parameters is often the goal of experimental design. This article presents novel experimental design techniques for parameter estimation applied to dynamical vehicle systems.