Research map / Vision-language and multimodal / Robot action and navigation
OVER-NAV: Elevating Iterative Vision-and-Language Navigation with Open-Vocabulary Detection and StructurEd Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OVER-NAV, which uses GPT-3.5 keyword extraction and OWL-ViT open-vocabulary detection to build a structured omnigraph memory that improves iterative VLN agents (HAMT and MAP-CMA) in discrete and continuous environments.
Problem
Iterative Vision-and-Language Navigation (IVLN) requires agents to exploit unstructured long-term navigation memory across episodes of a tour, but supervision is extremely sparse and prior close-vocabulary semantic mapping is limited to a fixed set of categories.
Approach
Use an LLM to extract milestone keywords from each instruction and an open-vocabulary detector to ground them in panoramic views, storing results in a viewpoint graph (omnigraph) whose neighboring keywords are fused with distance and heading embeddings as agent input.
Key results
- On IR2R val-unseen, OVER-NAV reaches t-nDTW 55 versus 50 for HAMT (about 5% gain), and improves val-seen t-nDTW from 58 to 62.
- On IR2R-CE, OVER-NAV gains 4% t-nDTW on val-seen (56 vs 52 for MAP-CMA) and 3% on val-unseen (50 vs 47).
- Ablating open-vocabulary keywords to 12 closed categories (Type-I) yields smaller gains than the open-vocabulary version, and removing distance and view index from the omnigraph lowers nDTW/t-nDTW (Table 5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add open-vocabulary keyword detection and omnigraph memory | HAMT · IR2R Val-Seen | t-nDTW: 58 ±1 -> 62 ±0 | Helped |
| add open-vocabulary keyword detection and omnigraph memory | HAMT · IR2R Val-Unseen | t-nDTW: 50 ±0 -> 55 ±1 | Helped |
| add open-vocabulary keyword detection and omnigraph memory | MAP-CMA · IR2R-CE Val-Seen | t-nDTW: 52 -> 56 ±2 | Helped |
| add open-vocabulary keyword detection and omnigraph memory | MAP-CMA · IR2R-CE Val-Unseen | t-nDTW: 47 -> 50 ±2 | Helped |
| restrict keywords to 12 common object categories instead of all open-vocabulary keywords (ablation) | OVER-NAV · IR2R Val-Seen and Val-Unseen | t-nDTW: 62 ±0 -> 61 ±1; 55 ±1 -> 52 ±1 | Hurt |
| retain attributes for the 12 common keywords (ablation) | Type-I OVER-NAV · IR2R Val-Seen and Val-Unseen | t-nDTW: 61 ±1 -> 62 ±0; 52 ±1 -> 53 ±1 | Helped |
| preload memory with detections for all viewpoints and current-instruction keywords (ablation) | OVER-NAV · IR2R Val-Seen and Val-Unseen | t-nDTW: 62 ±0 -> 63 ±0; 55 ±1 -> 56 ±1 | Helped |
| use neighbour distance dn=1 or dn=2 instead of dn=3 (ablation) | OVER-NAV · IR2R Val-Seen | t-nDTW: 62.3 -> 62.1; 62.3 -> 62.6 | Hurt |
| use neighbour distance dn=1 or dn=2 instead of dn=3 (ablation) | OVER-NAV · IR2R Val-Unseen | t-nDTW: 54.9 -> 53.4; 54.9 -> 54.0 | Hurt |
| increase neighbour distance to dn=4 or dn=5 instead of dn=3 (ablation) | OVER-NAV · IR2R Val-Seen | t-nDTW: 62.3 -> 61.2; 62.3 -> 61.2 | Hurt |
| increase neighbour distance to dn=4 or dn=5 instead of dn=3 (ablation) | OVER-NAV · IR2R Val-Unseen | t-nDTW: 54.9 -> 54.4; 54.9 -> 54.2 | Hurt |
| add distance and view-index information to keywords (ablation) | OVER-NAV · IR2R Val-Seen and Val-Unseen | t-nDTW: 58.6 -> 62.3; 50.6 -> 54.9 | Helped |
Surprising findings
- TourHAMT, which adds previous-episode history, fails to outperform the HAMT baseline on IR2R.
Limitations
- Relies on an external LLM (GPT-3.5) and detector (OWL-ViT), adding API/compute cost and dependence on their accuracy.
- Requires panoramic observations (panoramic camera or 360-degree rotation) at each viewpoint.
- Gains are modest on some metrics and the Type-III oracle setting shows headroom in keyword detection quality.
Experimental setup
Datasets: IR2R IR2R-CE R2R R2R-CE
Benchmarks: IR2R IR2R-CE
Metrics: t-nDTW TL NE OS nDTW SR SPL
Baselines: HAMT TourHAMT CMA TourCMA PoolCMA PoolEndCMA MAP-CMA
Builds on: HAMT MAP-CMA OWL-ViT R2R IR2R
Closest papers by meaning
- Towards Long-Horizon Vision-Language Navigation: Platform, Benchmark and Method (CVPR 2025 · similarity 0.88)
- NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models (AAAI 2024 · similarity 0.88)
- RAO-Nav: Probing Omni-Language Models for Zero-shot Semantic Audio-Visual Navigation (NeurIPS 2026 · similarity 0.87)
- NavBench: Probing Multimodal Large Language Models for Embodied Navigation (NeurIPS 2025 · similarity 0.87)
- GA-VLN: Geometry-Aware BEV Representation for Efficient Vision-Language Navigation (CVPR 2026 · similarity 0.86)
- Vision-and-Language Navigation via Causal Learning (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.