Localisation-aware fine-tuning for realistic PointGoal navigation

McGhan, Fraser, Ji, Ze ORCID: https://orcid.org/0000-0002-8968-9902 and Grech, Raphael 2024. Localisation-aware fine-tuning for realistic PointGoal navigation. Presented at: 25th Towards Autonomous Robotic Systems (TAROS) Conference, London, UK, 21-23 August 2024. Published in: Huda, M. Nazmul and Wang, Mingfeng eds. Proceedings Towards Autonomous Robotic Systems: 25th Annual Conference TAROS 2024. , vol.1 Heidelberg: Springer-VerlagBerlin,

PDF - Accepted Post-Print Version Download (1MB)

Abstract

Prior research has demonstrated the effectiveness of end-to-end reinforcement learning for PointGoal navigation tasks within indoor environments. Given 2.5 billion frames of experience, a navigation policy can be trained to achieve a success rate of 0.94 when deployed in unseen environments. However, a limitation of this approach is its reliance on perfect localisation, which is unrealistic for real-world deployment scenarios where localisation must be estimated, inevitably introducing errors. In this paper, we present a study on the effectiveness of integrating a traditional vision-based SLAM algorithm with a reinforcement learning-based PointGoal navigation policy. Through our experimentation, we demonstrate how fine-tuning a pre-trained navigation policy on realistic localisation estimates can increase the success rate by 14% (0.71 → 0.85) and SPL by 15% (0.66 → 0.81) when compared to deploying policies in a zero-shot manner.

Item Type:	Conference or Workshop Item - published (Paper)
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering
Publisher:	Springer-VerlagBerlin
ISBN:	978-3-031-72058-1
Date of First Compliant Deposit:	16 September 2024
Date of Acceptance:	21 June 2024
Last Modified:	17 Apr 2026 12:23
URI:	https://orca.cardiff.ac.uk/id/eprint/172154

Actions (repository staff only)

Edit Item