Range-based 6-DoF Monte Carlo Localization with GPU-Accelerated Stein Particle Filter

IEEE International conference on Robotics and Automation (ICRA2024)

Masashi Yokozuka1

Atsuhiko Banno1

1National Institute of Advanced Industrial Science and Technology (AIST), Japan

Supplementary video

Closeup view

4K resolution snapshots

Experimental results

Trajectory smoothing

Simple trajectory smoothing is used to filter out pose jitters:

Particle update schemes

A comparison of SVGD (Stein Variational Gradient Descent) and particle filter (resampling).

  • Efficient sampling with fewer samples
  • Free from the sample impoverishment problem
PF (resampling):
  • Needs many particles
  • Suffers from the sample impoverishment problem

SVGD : 64 particles

PF (Resampling) : 1024 particles


Dataset can be found at this Zenodo repository DOI