Research map / Vision-language and multimodal / Robot action and navigation
LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary 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.
Introduces HieraNav, a four-level (scene/room/region/instance) open-vocabulary goal navigation task, the human-verified LangMap benchmark on HM3D with 18K+ tasks, and PlaNaVid, an RGB-only planner-plus-policy baseline with Bounded Diverse Memory.
Problem
Existing language-conditioned goal navigation benchmarks focus on category-level goals or use noisy VLM-generated instance descriptions, and they lack region- and room-level disambiguation, which limits reliable evaluation.
Approach
Contrastive, human-written discriminative descriptions for regions and instances across four semantic goal levels, plus a decoupled RGB-only method that uses a VLM planner over diverse bounded memory to prime a reactive policy for multi-goal navigation.
Key results
- PlaNaVid-7B reaches 42.8% multi-goal SR and 31.7% single-goal SR on LangMap, the top RGB-only result and comparable to panoramic RGB-D MTU3D (41.2%).
- LangMap descriptions reach about 80% one-to-many text-to-view matching accuracy versus 57% for GOAT-Bench, with about one quarter as many words.
- An independent human audit yields 92.5% unique-and-correct matches across 200 judgments.
Surprising findings
- Uni-NaVid attains slightly higher single-goal SR than MTU3D despite lacking depth and explicit high-level planning, yet trails it by 6.8 points in multi-goal SR.
- About 39.8% of GOAT-Bench instance-level descriptions in ten inspected scenes contain semantic errors or ambiguities.
Limitations
- Human verification raises annotation cost and may limit scaling to more scenes.
- Descriptions reflect annotator choices about which attributes to emphasize.
- Covers only static indoor HM3D scenes; no outdoor or dynamic environments.
Experimental setup
Datasets: HM3D HM3D-Sem GOAT-Bench
Benchmarks: LangMap
Metrics: SR SPL SeqSR@k OSR latency
Baselines: 3D-Mem MTU3D PSL SenseAct-M Uni-NaVid
Builds on: Uni-NaVid Qwen2.5-VL GOAT-Bench 3D-Mem MTU3D
Closest papers by meaning
- Hierarchical Semantic-Augmented Navigation: Optimal Transport and Graph-Driven Reasoning for Vision-Language Navigation (NeurIPS 2025 · similarity 0.91)
- Distilling LLM Prior to Flow Model for Generalizable Agent’s Imagination in Object Goal Navigation (NeurIPS 2025 · similarity 0.90)
- NavBench: Probing Multimodal Large Language Models for Embodied Navigation (NeurIPS 2025 · similarity 0.89)
- Towards Long-Horizon Vision-Language Navigation: Platform, Benchmark and Method (CVPR 2025 · similarity 0.89)
- Scene Map-based Prompt Tuning for Navigation Instruction Generation (CVPR 2025 · 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.