Hybrid mapping for object goal navigation

Sun, Jingwen 2025. Hybrid mapping for object goal navigation. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (29MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (474kB) | Request a copy

Abstract

Object Goal Navigation (ObjectNav) is a fundamental task in embodied artificial intelligence, requiring an autonomous agent to navigate to an object specified by its semantic category. Achieving robust performance in this task demands effective integration of spatial reasoning, semantic understanding, and perception under partial observability. This thesis presents a comprehensive study of ObjectNav from both analytical and methodological perspectives. It begins with a systematic survey of the field, covering its definition, simulators, datasets, and evaluation metrics, and classifies existing methods into end-to-end, modular, and zero-shot methods. The review identifies several limitations in current ObjectNav systems, such as the lack of comprehensive semantic representations, limited use of 3D spatial information, and challenges in ensuring reliable perception. To address these issues, three key contributions are introduced. First, RSMP Net, a relationship-guided semantic map prediction network, is proposed to enhance map completeness and accuracy by exploiting semantic and spatial relationships through graph reasoning. Second, the Fourier Unsigned Distance Field (FUDF) is presented as a compact representation that enriches 2D semantic maps with 3D geometric information, enabling a hybrid semantic–spatial understanding of the environment. Third, several strategies are developed to improve semantic reliability, including category-specific thresholds, semantic verification using large language models, and active multi-view observation, which collectively improve detection robustness and consistency. Experimental evaluations demonstrate that the proposed methods achieve better performance in semantic map prediction and significantly enhance navigation efficiency and semantic accuracy in complex 3D environments. Overall, this thesis contributes to advancing ObjectNav by bridging the gap between semantic reasoning and spatial perception, offering both theoretical insights and practical frameworks for developing more intelligent, perceptually grounded, and generalisable embodied agents

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA76 Computer software
Date of First Compliant Deposit:	6 May 2026
Date of Acceptance:	1 May 2026
Last Modified:	06 May 2026 11:06
URI:	https://orca.cardiff.ac.uk/id/eprint/186816

Actions (repository staff only)

Edit Item