Multibody dynamic systems as Bayesian networks: Applications to robust state estimation of mechanisms

摘要

This article addresses the problem of robustly estimating the dynamic state of a mechanism from a set of noisy sensor measurements. We start with a rigorous treatment of the problem from the perspective of graphical models, a popular formalism in the fields of statistical inference and machine learning. The modeling power of such a formalism is demonstrated by showing how the sequential estimation of a mechanism state with an extended Kalman filter (EKF), often used in previous works, becomes just one of the possible solutions. As an interesting alternative, we derive the formulation of a sequential Monte Carlo (SMC) filter, also known as a particle filter (PF), suitable for online tracking the state of a rigid mechanism. We validate our ideas with both simulated and real datasets. Moreover, we prove the usefulness of the particle filtering solution for real-work applications due to its unmatched capability of automatically inferring the initial states of the mechanism along with its “assembly configuration” or “branch” if several ones are possible, a feature not matched by any previously proposed state observer in the multibody literature.