Work

SignNav: Directional Signage-Cue Guided Visual Navigation in Large-Scale Indoor Environments

Year: 2026
Type: article
Abstract: Humans routinely leverage directional cues from signage to navigate toward destinations in unfamiliar Large-Scale Indoor (LSI) environments, such as hospitals and airport terminals. However, learning how embodied agents can use sparse and target-relevant signage cues for closed-loop navigation remains underexplored. This paper studiesSignNav, a directional signage-cue guided visual navigation task in which an agent navigates to a semantic target using egocentric observations together with directional cues and cue locations provided by an upstream sign parser. This formulation focuses on downstream cue-conditioned policy learning, rather than open-ended signage parsing, OCR, or compound sign reasoning. To facilitate research in this domain, we construct theLSI-Dataset, a large-scale benchmark with signage cues-labeled indoor environments and automatically generated trajectory-action episodes. SignNav presents two key challenges: grounding transient directional cues into the current observation, and maintaining temporal memory when signage is sparse or temporarily unavailable. To address these challenges, we propose the Spatial-Temporal Aware Transformer (START), which combines a spatial-aware module for cue-observation grounding with a temporal-aware module for history-based decision making. Extensive experiments, including cue-representation ablations, sparse-signage evaluations, and real-world robot trials, demonstrate the effectiveness of START for directional signage-cue guided navigation in large-scale indoor environments.
Language: English
Open Access status: green
OpenAlex ID: W7206197903
FWCI: 0.942
Cites: 52
Cited by: 1
Related to: 10
Parts: 1