Research map / Vision-language and multimodal / Robot action and navigation
VoroNav: Voronoi-based Zero-shot Object Navigation with Large Language Model
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 VoroNav, a training-free zero-shot object navigation framework that builds a Reduced Voronoi Graph from a real-time semantic map and uses GPT-3.5 to pick waypoints from combined path and farsight text descriptions.
Problem
Zero-shot object navigation (ZSON) in unfamiliar household environments requires efficient exploration and commonsense reasoning, while existing map-based methods choose waypoints at fixed intervals and miss informative intersections.
Approach
Extract a Reduced Voronoi Graph from the semantic map to get intersection waypoints and exploratory paths, describe those paths and BLIP captions of panoramic farsight views as text for GPT-3.5, and select the next waypoint with a hierarchical reward (exploration > efficiency > semantic).
Key results
- VoroNav reaches 42.0 Success and 26.0 SPL on HM3D, versus 39.2 Success and 22.3 SPL for ESC (training-free, GPT-3.5).
- On HSSD, VoroNav achieves 41.0 Success and 23.2 SPL, improving over the best baseline by +2.6% Success and +3.8% SPL.
- Ablation: removing farsight or path descriptions lowers HM3D Success to 41.2 and 40.0 respectively, versus 42.0 for full VoroNav.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine path and farsight descriptions | VoroNav · HM3D | Success: +2.8% | Helped |
| combine path and farsight descriptions | VoroNav · HM3D | SPL: +3.7% | Helped |
| combine path and farsight descriptions | VoroNav · HSSD | Success: +2.6% | Helped |
| combine path and farsight descriptions | VoroNav · HSSD | SPL: +3.8% | Helped |
| add semantic guidance to Voronoi navigation (ablation) | VoroNav · HM3D and HSSD | Success: HM3D: 38.7 -> 42.0; HSSD: 40.3 -> 41.0 | Helped |
| add semantic guidance to Voronoi navigation (ablation) | VoroNav · HM3D and HSSD | SPL: HM3D: 23.3 -> 26.0; HSSD: 22.2 -> 23.2 | Helped |
| add farsight descriptions to path descriptions (ablation) | VoroNav · HM3D and HSSD | Success: HM3D: 40.0 -> 42.0; HSSD: 40.6 -> 41.0 | Helped |
| add farsight descriptions to path descriptions (ablation) | VoroNav · HM3D and HSSD | SPL: HM3D: 24.2 -> 26.0; HSSD: 23.0 -> 23.2 | Helped |
| add path descriptions to farsight descriptions (ablation) | VoroNav · HM3D and HSSD | Success: HM3D: 41.2 -> 42.0; HSSD: 40.8 -> 41.0 | Helped |
| add path descriptions to farsight descriptions (ablation) | VoroNav · HM3D and HSSD | SPL: HM3D: 25.2 -> 26.0; HSSD: 22.7 -> 23.2 | Helped |
Surprising findings
- The Voronoi planner without any semantic guidance already outperforms the semantic baselines L3MVN and ESC in SPL on HM3D (23.3 vs 16.5 and 22.3).
Limitations
- Relies on GPT-3.5 API calls for each waypoint decision.
- Most failures on HM3D are attributed to detection errors, and on HSSD to planning abnormalities.
Experimental setup
Datasets: HM3D HSSD
Benchmarks: HM3D HSSD
Metrics: Success SPL SCA SEA
Baselines: Random Exploration Frontier Voronoi L3MVN Pixel-Nav ESC
Builds on: Grounded-SAM BLIP Fast Marching Method Frontier-based exploration L3MVN
Closest papers by meaning
- BeliefMapNav: 3D Voxel-Based Belief Map for Zero-Shot Object Navigation (NeurIPS 2025 · similarity 0.88)
- GAMap: Zero-Shot Object Goal Navigation with Multi-Scale Geometric-Affordance Guidance (NeurIPS 2024 · similarity 0.86)
- Hierarchical Semantic-Augmented Navigation: Optimal Transport and Graph-Driven Reasoning for Vision-Language Navigation (NeurIPS 2025 · similarity 0.86)
- Enhancing Multi-Robot Semantic Navigation Through Multimodal Chain-of-Thought Score Collaboration (AAAI 2025 · similarity 0.86)
- Affordances-Oriented Planning Using Foundation Models for Continuous Vision-Language Navigation (AAAI 2025 · similarity 0.86)
- ImagineNav: Prompting Vision-Language Models as Embodied Navigator through Scene Imagination (ICLR 2025 · 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.