Research map / Graph learning / Knowledge and temporal graphs
UniGoal: Towards Universal Zero-shot Goal-oriented Navigation
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 UniGoal, a training-free universal zero-shot goal-oriented navigation framework that represents scenes and goals (object, image, text) as graphs and uses graph matching to drive a multi-stage LLM-guided exploration policy.
Problem
Zero-shot goal-oriented navigation methods are task-specific and do not generalize across object-category, instance-image, and text-description goals; prior universal methods require RL training.
Approach
A uniform graph representation for both the online-built 3D scene graph and goals of any type, with graph matching scores (node, edge, topology) that switch among three LLM-driven exploration stages (subgraph search, coordinate projection with anchor-pair alignment, and scene graph correction with goal verification) plus a blacklist to avoid repeated failed matches.
Key results
- UniGoal reaches 60.2 SR / 23.7 SPL on HM3D instance-image-goal navigation, versus 56.1 / 23.3 for the prior zero-shot Mod-IIN.
- On HM3D object-goal navigation UniGoal gets 54.5 SR / 25.1 SPL, slightly above SG-Nav (54.0 / 24.9).
- On HM3D text-goal navigation UniGoal reaches 20.2 SR / 11.4 SPL, outperforming supervised GOAT (17.0 / 8.8) and PSL (16.5 / 7.5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add graph matching score computation (ablation) | UniGoal · HM3D IIN | SR: 54.9 -> 60.2 | Helped |
| add the blacklist mechanism (ablation) | UniGoal · HM3D IIN | SR: 50.6 -> 60.2 | Helped |
| retain stage 2 in the multi-stage exploration policy (ablation) | UniGoal · HM3D IIN | SR: 59.0 -> 60.2 | Helped |
| use the proposed stage 1 instead of FBE (ablation) | UniGoal · HM3D IIN | SR: 55.1 -> 60.2 | Helped |
| add goal graph decomposition (ablation) | UniGoal · HM3D IIN | SR: 59.2 -> 60.2 | Helped |
| add frontier selection (ablation) | UniGoal · HM3D IIN | SR: 57.4 -> 60.2 | Helped |
| use LLM-based coordinate projection instead of a random coordinate guess (ablation) | UniGoal · HM3D IIN | SR: 59.1 -> 60.2 | Helped |
| add anchor pair alignment (ablation) | UniGoal · HM3D IIN | SR: 58.9 -> 60.2 | Helped |
| add scene graph correction (ablation) | UniGoal · HM3D IIN | SR: 59.5 -> 60.2 | Helped |
| add goal verification (ablation) | UniGoal · HM3D IIN | SR: 58.2 -> 60.2 | Helped |
Surprising findings
- Universal zero-shot UniGoal outperforms supervised universal methods PSL and GOAT on all three tasks despite requiring no training.
Limitations
- Gains on object-goal navigation are small (0.8 SR over SG-Nav) because single-node goal graphs do not exercise the graph-specific modules.
- Relies on multiple LLM and VLM calls per step, with no reported latency or cost analysis in the main text.
Experimental setup
Datasets: Matterport3D HM3D RoboTHOR
Benchmarks: MP3D ObjectNav HM3D ObjectNav RoboTHOR ObjectNav HM3D InstanceImageNav HM3D TextNav
Metrics: success rate SPL
Baselines: SemEXP ZSON OVRL-v2 ESC OpenFMNav VLFM SG-Nav Mod-IIN PSL GOAT IEVE Krantz et al.
Builds on: SG-Nav Mod-IIN CLIP LightGlue InstructNav
Closest papers by meaning
- BeliefMapNav: 3D Voxel-Based Belief Map for Zero-Shot Object Navigation (NeurIPS 2025 · similarity 0.88)
- SG-Nav: Online 3D Scene Graph Prompting for LLM-based Zero-shot Object Navigation (NeurIPS 2024 · similarity 0.87)
- NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models (AAAI 2024 · similarity 0.86)
- LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation (NeurIPS 2026 · similarity 0.86)
- 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 0.85)
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.