Towards Differentiable Resampling

TL;DR

This paper introduces a neural network-based resampler for particle filters, enabling end-to-end differentiability and improved performance over traditional methods in simulated localization tasks.

Contribution

It proposes the particle transformer, a novel neural network architecture for differentiable resampling in particle filters, allowing joint optimization of all components.

Findings

Learned resampler outperforms traditional resampling methods on synthetic data.

The model achieves better localization accuracy in simulated robot tasks.

End-to-end training improves overall particle filter performance.

Abstract

Resampling is a key component of sample-based recursive state estimation in particle filters. Recent work explores differentiable particle filters for end-to-end learning. However, resampling remains a challenge in these works, as it is inherently non-differentiable. We address this challenge by replacing traditional resampling with a learned neural network resampler. We present a novel network architecture, the particle transformer, and train it for particle resampling using a likelihood-based loss function over sets of particles. Incorporated into a differentiable particle filter, our model can be end-to-end optimized jointly with the other particle filter components via gradient descent. Our results show that our learned resampler outperforms traditional resampling techniques on synthetic data and in a simulated robot localization task.