---
pretty_name: ReCoNav Benchmark
language:
  - en
license: other
license_name: hm3d-hm3dsem-academic-use
license_link: https://huggingface.co/datasets/psiang/ReCoNavBench/blob/main/DATA_TERMS.md
size_categories:
  - n<1K
tags:
  - vision-and-language-navigation
  - embodied-ai
  - navigation
  - benchmark
  - hm3d
extra_gated_prompt: "Read DATA_TERMS.md and the Matterport Academic Use Agreement at https://matterport.com/legal/matterport-end-user-license-agreement-academic-use-model-data before requesting access."
extra_gated_fields:
  Name: text
  Institution: text
  I accept the Matterport Academic Use Agreement and the dataset terms: checkbox
---

# ReCoNav Benchmark

**192 constrained navigation tasks across nine HM3DSem scenes**, with scene ASGs, physical navigation geometry and evaluator annotations. Tasks cover goals, ordered visits, avoidance, counts, preferences and combinations of these requirements. All tasks are evaluation tasks.

[Code and running instructions](https://github.com/psiang/ReCoNav) · [Data terms](DATA_TERMS.md)

## Citation

```bibtex
@misc{peng2026reconav,
  title  = {{ReCoNav}: Reliable Constrained Long-Horizon Vision-and-Language Navigation via a Determination Engine},
  author = {Peng, Sixiang and Yu, Zehuan and Liu, Wenyi and Xie, Zhensheng and Zhang, Charles and Jiang, Chenhan},
  year   = {2026},
  url    = {https://github.com/psiang/ReCoNav}
}
```

Also cite [HM3D](https://aihabitat.org/datasets/hm3d/) and [HM3DSem](https://aihabitat.org/datasets/hm3d-semantics/) when using the underlying scenes.

## Download

Accept the access terms on [this dataset's page](https://huggingface.co/datasets/psiang/ReCoNavBench), then download:

```bash
python -m pip install huggingface_hub
hf auth login
export RECONAV_BENCHMARK_REPO=psiang/ReCoNavBench
export RECONAV_DATA_ROOT=/absolute/path/to/data
hf download "$RECONAV_BENCHMARK_REPO" reconav-benchmark-v1.0.0.tar.gz \
  --repo-type dataset --revision v1.0.0 --local-dir downloads
mkdir -p "$RECONAV_DATA_ROOT"
tar -xzf downloads/reconav-benchmark-v1.0.0.tar.gz -C "$RECONAV_DATA_ROOT"
```

Version: **1.0.0**. The archive extracts one `benchmark/` directory. Data can be stored anywhere, independently of the code checkout.

## Contents

| Path under `benchmark/` | Contents |
|---|---|
| `tasks/hm3dsem/<scene>/<task>.json` | Instruction/start, task category, gold symbolic requirements, target metadata and reference route |
| `scenes/hm3dsem/<scene>/asg/` | Entity, geometry and topology JSON files |
| `scenes/hm3dsem/<scene>/navigation/graph.json.gz` | Physical free-space geometry and connectivity for route planning and evaluation |
| `splits/` | `all.json` (192 tasks) and `hard44.json` (44 tasks) |
| `vocabularies/` | Object and region category mappings |
| `manifest.json` | Version, scene IDs, task counts and coordinate conventions |

Hard44 contains 44 hard-constraint-resolving tasks supported by every compared method. Each task's `uniqueness_proof.identity_class` records whether target identity is resolved by a local qualifier, direct reference, hard constraint or preference.

Prediction inputs are the instruction, start and scene information. Gold target IDs, symbolic contracts and route witnesses are evaluator annotations. Route points use Habitat x/y/z, with y vertical; task `robot_position` uses x/y horizontally and `floor_height` vertically. ASG footprints use x/z.

## Use

Install the code repository, then inspect the benchmark:

```bash
reconav benchmark
```

Follow the repository README for model setup and a one-task ReCoNav run. Choose `--split all`, `--split hard44` or `--task-ids ID [ID ...]` independently of the selected method. A new method can use the task instructions/starts and ASGs with the repository's common evaluator.

Follow the code repository for method-specific environments, checkpoints and resources.

## Original scene meshes

ReCoNav and its ablations use the packaged ASGs and physical graphs. Simulator and image-rendering methods also need HM3D v0.2 validation scene files. Sign in to Matterport and follow the [official download/access instructions](https://github.com/matterport/habitat-matterport-3dresearch). The relevant upstream packages are:

| Download | Scene files |
|---|---|
| [hm3d-val-habitat-v0.2.tar](https://api.matterport.com/resources/habitat/hm3d-val-habitat-v0.2.tar) | Habitat `.basis.glb` geometry |
| [hm3d-val-glb-v0.2.tar](https://api.matterport.com/resources/habitat/hm3d-val-glb-v0.2.tar) | Original JPEG-textured `.glb` for rendering |

ReCoNav uses these nine scene directories, also listed in `benchmark/manifest.json`:

```text
00824-Dd4bFSTQ8gi
00829-QaLdnwvtxbs
00843-DYehNKdT76V
00848-ziup5kvtCCR
00849-a8BtkwhxdRV
00853-5cdEh9F2hJL
00861-GLAQ4DNUx5U
00873-bxsVRursffK
00891-cvZr5TUy5C5
```

Keep each selected scene's files together and point `RECONAV_HM3D_ROOT` to the `val/` directory:

```bash
export RECONAV_HM3D_ROOT=/absolute/path/to/hm3d/val
```

For example:

```text
$RECONAV_HM3D_ROOT/
└── 00824-Dd4bFSTQ8gi/
    ├── Dd4bFSTQ8gi.basis.glb
    └── Dd4bFSTQ8gi.glb
```

Use the same directory pattern for the other eight scenes. See the [baseline manual](https://github.com/psiang/ReCoNav/blob/main/docs/baselines.md#habitat-navigation-resources) for the separate Habitat navmesh resource.

## Terms

HM3D/HM3DSem-derived data is subject to the upstream academic, non-commercial terms in [DATA_TERMS.md](DATA_TERMS.md). The code repository has an Apache-2.0 license; this dataset retains its data terms.