Inference-Enabled Information-Theoretic Exploration of Continuous Action Spaces

Shi Bai, Jinkun Wang, Kevin Doherty, Brendan Englot

Research output: Chapter in Book/Report/Conference proceedingChapterpeer-review

7 Scopus citations

Abstract

We consider an autonomous exploration problem in which a mobile robot is guided through an a priori unknown environment by a controller that chooses the most informative action within a local region. We propose a novel approach to efficiently evaluate information gain over the continuous action space that leverages supervised learning, with the anticipated mutual information achieved by a discrete set of action primitives serving as training data. We describe an autonomous exploration algorithm that uses this approach to cover a priori unknown environments. Computational results demonstrate that the method offers an improved rate of entropy reduction, surpassing a baseline approach that selects from the discrete action set, which in some instances requires more computational effort and yields less information.

Original languageEnglish
Title of host publicationSpringer Proceedings in Advanced Robotics
Pages419-433
Number of pages15
DOIs
StatePublished - 2018

Publication series

NameSpringer Proceedings in Advanced Robotics
Volume3
ISSN (Print)2511-1256
ISSN (Electronic)2511-1264

Fingerprint

Dive into the research topics of 'Inference-Enabled Information-Theoretic Exploration of Continuous Action Spaces'. Together they form a unique fingerprint.

Cite this