Research map / Multi-agent systems and games / Human-AI and goal-conditioned
REGNav: Room Expert Guided Image-Goal 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 REGNav, which pre-trains an unsupervised room expert (style encoder + relation network) on self-collected Gibson images and feeds its same-room relation into an RL image-goal navigation policy built on FGPrompt-EF.
Problem
Image-goal navigation fails when the agent is in a different room from the goal image, since visual similarity is hard to judge and the agent wanders; RGB-only agents lack spatial relationship cues.
Approach
Pre-train a room expert without room labels, using clustering with must-link/cannot-link constraints derived from scene IDs, episode identity and episode difficulty, to predict whether two images are in the same room; then fuse the predicted room relation (explicit fusion) into the navigation policy trained with RL.
Key results
- On Gibson, REGNav achieves SPL 67.1% and SR 92.9%, versus FGPrompt-EF's 66.5% and 90.4%.
- On the Gibson hard split, REGNav reaches SPL 59.4% and SR 87.1%, outperforming memory-based methods such as MemoNav (SR 74.7%).
- Cross-domain, REGNav reaches SPL 50.2% / SR 78.0% on MP3D and 44.0% / 75.2% on HM3D, versus FGPrompt-EF's 48.8% / 75.7% on MP3D and 42.1% / 75.2% on HM3D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the room expert with explicit fusion | FGPrompt-EF · Gibson overall test set | SPL: 66.5% -> 67.1% | Helped |
| Add the room expert with explicit fusion | FGPrompt-EF · Gibson overall test set | SR: 90.4% -> 92.9% | Helped |
| Use implicit fusion instead of explicit fusion (ablation) | REGNav · Gibson | SPL: 67.1% -> 47.4% | Hurt |
| Use implicit fusion instead of explicit fusion (ablation) | REGNav · Gibson | SR: 92.9% -> 77.0% | Hurt |
| Use implicit fusion instead of explicit fusion (ablation) | REGNav · MP3D | SPL: 50.2% -> 33.2% | Hurt |
| Use implicit fusion instead of explicit fusion (ablation) | REGNav · MP3D | SR: 78.0% -> 59.9% | Hurt |
| Use implicit fusion instead of explicit fusion (ablation) | REGNav · HM3D | SPL: 44.0% -> 27.7% | Hurt |
| Use implicit fusion instead of explicit fusion (ablation) | REGNav · HM3D | SR: 75.2% -> 55.8% | Hurt |
| Add clean data to room-expert training (ablation) | Room Expert · MP3D room-relation validation | accuracy: 57.0%+0.0% -> 57.4%+0.4% | Helped |
| Add distance refinement using constraints (ablation) | Room Expert · MP3D room-relation validation | accuracy: 57.0%+0.0% -> 57.9%+0.5% | Helped |
| Add both clean data and distance refinement (ablation) | Room Expert · MP3D room-relation validation | accuracy: 57.0%+0.0% -> 58.4%+0.5% | Helped |
Surprising findings
- Explicit fusion of the room relation outperforms implicit fusion of room-style embeddings by a large margin (Gibson SR 92.9% vs 77.0%).
Limitations
- Room expert relies on the heuristic constraints derived from scene IDs and episode difficulty rather than true room labels.
- Cross-domain results on MP3D and HM3D are only comparable to, not clearly better than, the strongest baseline on HM3D SR (75.2% tie with FGPrompt-EF).
Experimental setup
Datasets: Gibson MP3D HM3D ImageNet
Benchmarks: Gibson MP3D HM3D
Metrics: SPL SR relation accuracy
Baselines: ZER ZSON OVRL OVRL-V2 FGPrompt-MF FGPrompt-EF VGM TSGM Mem-Aug MemoNav
Builds on: FGPrompt ZER InfoMap SAM
Closest papers by meaning
- End-to-End (Instance)-Image Goal Navigation through Correspondence as an Emergent Phenomenon (ICLR 2024 · similarity 0.86)
- CaMP: Causal Multi-policy Planning for Interactive Navigation in Multi-room Scenes (NeurIPS 2023 · similarity 0.86)
- Instance-aware Exploration-Verification-Exploitation for Instance ImageGoal Navigation (CVPR 2024 · similarity 0.86)
- Trajectory Diffusion for ObjectGoal Navigation (NeurIPS 2024 · similarity 0.86)
- Distilling LLM Prior to Flow Model for Generalizable Agent’s Imagination in Object Goal Navigation (NeurIPS 2025 · similarity 0.85)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 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.