Research map / Vision-language and multimodal / Robot action and navigation
FloNa: Floor Plan Guided Embodied Visual 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.
Introduces FloNa, a floor-plan-guided visual navigation task, and FloDiff, a diffusion-policy framework with a localization module, evaluated on a 20k-episode iGibson/Gibson dataset.
Problem
Visual navigation to goals marked on a floor plan using only RGB observations, where floor plans are abstract and mismatched with actual furniture-laden layouts and observations must be aligned to the plan without depth or multi-sensor input.
Approach
A diffusion policy that fuses RGB observation features with floor-plan features via transformer attention, with an explicit pose-localization component (either predicted jointly or supplied by GT/F3Loc) feeding the denoising policy that generates action sequences.
Key results
- Loc-FloDiff (GT) reaches 68.4% SR and 53.4% SPL at tau_c=50, tau_d=0.3m, outperforming Loc-A* (GT) at 29.4% SR under the same threshold.
- Loc-FloDiff (F3) achieves 30.8% SR at tau_c=50... versus Loc-A* (F3) at 0.6% SR at tau_d=0.3m.
- Loc-FloDiff (GT) achieves 22.45 mean collisions vs 25.65 for Loc-A* (GT) at tau_c=50 among successful episodes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Loc-FloDiff with ground-truth poses instead of Loc-A* with ground-truth poses | FloNa navigation · FloNa test set (τc=50, τd=0.30m) | SR (%): 29.4 -> 68.4 | Helped |
| use Loc-FloDiff with ground-truth poses instead of Loc-A* with ground-truth poses | FloNa navigation · FloNa test set (τc=50, τd=0.30m) | SPL (%): 29.3 -> 53.4 | Helped |
| use ground-truth poses instead of F³Loc-predicted poses | Loc-FloDiff · FloNa test set (τc=50, τd=0.30m) | SR (%): 20.6 -> 68.4 | Helped |
| use modular localization instead of learning pose prediction end-to-end | FloDiff · FloNa test set (τc=50, τd=0.30m) | SR (%): 15.6 -> 68.4 | Helped |
| use Loc-FloDiff instead of Loc-A* with the same ground-truth localization module | FloNa navigation · FloNa successful cases (τd=0.3m, τc=50) | mean collision count: 25.65 -> 22.45 | Helped |
| retrain F³Loc using images collected at 80cm height (ablation) | Loc-FloDiff (F³) · FloNa test set (τc=50, τd=0.30m) | SR (%): 20.6 -> 21.8 | Helped |
| remove the floor plan embedding by masking it during training (ablation) | Loc-FloDiff (GT) · FloNa test set (τc=50, τd=0.30m) | SR (%): 68.4 -> 63.0 | Hurt |
| add Gaussian noise with variance 0.1m to the ground-truth pose input (ablation) | Loc-FloDiff (GT) · FloNa test set (τc=50, τd=0.3m) | SR (%): 68.4 -> 26.0 | Hurt |
| add Gaussian noise with variance 0.3m to the ground-truth pose input (ablation) | Loc-FloDiff (GT) · FloNa test set (τc=50, τd=0.3m) | SR (%): 68.4 -> 20.0 | Hurt |
| add Gaussian noise with variance 0.5m to the ground-truth pose input (ablation) | Loc-FloDiff (GT) · FloNa test set (τc=50, τd=0.3m) | SR (%): 68.4 -> 24.0 | Hurt |
Surprising findings
- Naive FloDiff, which predicts pose end-to-end, underperforms the decoupled localization variants, suggesting encoders struggle to serve localization and planning simultaneously.
- The pre-trained F3Loc, trained on 175cm-height data, degrades localization for an 80cm agent, causing Loc-A* (F3) to nearly fail (0.6% SR at tau_d=0.3m).
Limitations
- Only RGB images are fused with floor plan images; hand-drawn sketches and other floor plan forms are not handled.
- Performance degrades with noisy localization input, and the policy sometimes collides or moves away from goal.
- Collisions occur because the floor plan omits furniture obstacles.
Experimental setup
Datasets: Gibson iGibson
Benchmarks: FloNa dataset (50 test scenes, 500 episodes)
Metrics: SR SPL mean collision count
Baselines: Loc-A* (F3Loc) Loc-A* (GT) Naive-FloDiff Loc-Diff (GT)
Builds on: diffusion policy ViNT F3Loc A*
Closest papers by meaning
- RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation (CVPR 2025 · similarity 0.86)
- Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator (NeurIPS 2026 · similarity 0.86)
- Scene Map-based Prompt Tuning for Navigation Instruction Generation (CVPR 2025 · similarity 0.85)
- SAVVY: Spatial Awareness via Audio-Visual LLMs through Seeing and Hearing (NeurIPS 2025 · similarity 0.84)
- Task Planning for Object Rearrangement in Multi-Room Environments (AAAI 2024 · similarity 0.84)
- Affordances-Oriented Planning Using Foundation Models for Continuous Vision-Language Navigation (AAAI 2025 · similarity 0.84)
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.