Stein Particle Filter for Nonlinear, Non-Gaussian State Estimation

Estimation of a dynamical system’s latent state subject to sensor noise and model inaccuracies remains critical yet difficult problem in robotics. While Kalman filters provide the optimal solution least squared sense for linear Gaussian problems, general nonlinear non-Gaussian case is significantly more complicated, typically relying on sampling strategies that are limited low-dimensional spaces. In this letter, we devise inference procedure filtering nonlinear, systems exploits differentiability both update prediction models scale higher dimensional Our method, Stein particle filter, can be seen as deterministic flow particles, embedded reproducing kernel Hilbert space, from an initial desirable posterior. The particles evolve jointly conform posterior approximation while interacting with each other through repulsive force. We evaluate method simulation complex localization tasks comparing it sequential Monte Carlo solutions.