Learning to track environment state via predictive autoencoding

TL;DR

This paper presents a neural autoencoding architecture that learns to model stochastic environments from image sequences, enabling environment state tracking, long-term prediction, and belief sampling, similar to particle filters.

Contribution

The work introduces a novel neural architecture capable of learning forward models from unstructured image observations for environment state tracking and prediction.

Findings

The model accurately tracks environment states in noisy conditions.

It can perform long-term predictions without observation input.

The learned models compare favorably with particle filters in simulation.

Abstract

This work introduces a neural architecture for learning forward models of stochastic environments. The task is achieved solely through learning from temporal unstructured observations in the form of images. Once trained, the model allows for tracking of the environment state in the presence of noise or with new percepts arriving intermittently. Additionally, the state estimate can be propagated in observation-blind mode, thus allowing for long-term predictions. The network can output both expectation over future observations and samples from belief distribution. The resulting functionalities are similar to those of a Particle Filter (PF). The architecture is evaluated in an environment where we simulate objects moving. As the forward and sensor models are available, we implement a PF to gauge the quality of the models learnt from the data.