Research map / Vision-language and multimodal / Robot action and navigation
MapDream: Task-Driven Map Learning for Vision-Language Navigation
Guoxin Lian, Shuo Wang, Yucheng Wang, Yongcai Wang, Maiyue Chen, Kaihui Wang, Bo Zhang, Zhizhong Su, Deying Li, Zhaoxin Fan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Stage 1 supervised generative BEV map conditioning to the VLN policy (ablation) | NVILA-2B VLN policy · R2R-CE Val-Unseen | SR: 37.7 -> 42.2 | Improved |
| add Stage 2 joint reinforcement fine-tuning (GRPO) of map module and policy (ablation) | MapDream after SPT · R2R-CE Val-Unseen | SR: 42.2 -> 45.6 | Improved |
| add Stage 2 RFT, effect on SPL and NE (ablation) | MapDream after SPT · R2R-CE Val-Unseen | SPL / NE: SPL 36.5 -> 40.5; NE 7.03 -> 6.35 | Improved |
| add RFT when map is initialized with Distance channel only (ablation) | MapDream · R2R-CE Val-Unseen | SR: 39.4 -> 44.9 | Improved |
| add RFT when map is initialized with Landmark channel only (ablation) | MapDream · R2R-CE Val-Unseen | SR: 40.4 -> 44.5 | Improved |
| add RFT when map is initialized with Occupancy channel only (ablation) | MapDream · R2R-CE Val-Unseen | SR: 39.4 -> 43.6 | Improved |
| use all three channels instead of a single channel for map supervision (ablation) | MapDream after RFT · R2R-CE Val-Unseen | SR: final SR 43.6-45.6 across channel choices | No clear effect |
| increase BEV map size from 96x96 (36 tokens) to 384x384 (576 tokens) (ablation) | MapDream map generator · R2R-CE Val-Unseen | SR: 42.2 -> 43.1 | No clear effect |
| increase BEV map size from 96x96 to 384x384 (ablation) | MapDream map generator · R2R-CE Val-Unseen | inference latency per step (s): 1.3 -> 12.7 | Hurt |
| use MapDream with monocular RGB instead of panoramic baselines | MapDream · R2R-CE Val-Unseen | SPL: 54.4 vs BEVBert-FSTTA 51.0 | Improved |
| compare MapDream against BEVBert-FEEDTTA on success rate | MapDream · R2R-CE Val-Unseen | SR: 59.8 vs 61.0 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- MonoDream — MonoDream: Monocular Vision-Language Navigation with Panoramic Dreaming (AAAI 2026) · better on 3 of 3 shared results
- Dynam3D — Dynam3D: Dynamic Layered 3D Tokens Empower VLM for Vision-and-Language Navigation (NeurIPS 2025) · better on 1 of 1 shared result
- Aux-Think — Aux-Think: Exploring Reasoning Strategies for Data-Efficient Vision-Language Navigation (NeurIPS 2025) · better on 1 of 1 shared result
- ETPNav — ETPNav: Evolving Topological Planning for Vision-Language Navigation in Continuous Environments (TPAMI 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GridMM: Grid Memory Map for Vision-and-Language Navigation (ICCV 2023 · similarity 0.90)
- Iterative Vision-and-Language Navigation (CVPR 2023 · similarity 0.89)
- Scene Map-based Prompt Tuning for Navigation Instruction Generation (CVPR 2025 · similarity 0.89)
- Hierarchical Semantic-Augmented Navigation: Optimal Transport and Graph-Driven Reasoning for Vision-Language Navigation (NeurIPS 2025 · similarity 0.89)
- LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation (NeurIPS 2026 · similarity 0.88)
- NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models (AAAI 2024 · similarity 0.88)
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.
