Using Attention to Weight Particles in Particle Filters
Keywords: particle filter, robotics, navigation, localization, sensor
TL;DR: Using attention is a fast, training-free particle filter weighting function alternative to evaluating Gaussian PDFs.
Abstract: Particle filters are a set of algorithms for state estimation in dynamical systems. The archetypal usage of particle filters is estimating the position and orientation of a robot based on noisy sensor readings. In many situations, sensor noise is modeled to be Gaussian, where evaluating particle observations using a Gaussian probability distribution function is a reasonable way to weight particles. In this paper, we propose using attention (i.e. softmax dot product) as an alternative particle weighting function. We investigated using attention vs the traditional Gaussian weighting function in physical and temporal localization and navigation tasks, and found time performance advantages, especially when using a GPU. At the same time, we found that attention maintains comparable accuracy to the Gaussian weighting function. Code is publicly available at github.com/anonuser-2023/project2023.
Primary Area: applications to robotics, autonomy, planning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3888
Loading