---
license: other
task_categories:
  - reinforcement-learning
tags:
  - deconav
  - embodied-navigation
  - habitat
  - hm3d
  - multi-robot
size_categories:
  - 1K<n<10K
---

# DeCoNav Dataset

This repository contains the complete compact DeCoNav task release, split into
1,113 training tasks and 100 fixed evaluation tasks. It includes the task
annotations and compact evaluator trajectories required by the DeCoNav data
and evaluation pipeline. RGB/depth frame dumps, simulator scenes, debug logs,
and machine-specific files are intentionally excluded.

## Splits

| Split | Tasks | Purpose |
| --- | ---: | --- |
| `train/` | 1,113 | DeCoNav training-task pool after excluding the fixed test tasks. |
| `test/` | 100 | Fixed task order used by the released DeCoNav evaluation. |

Each task directory contains:

```text
single_config.json
collab_json/collab_config_00.json
success/trial_*/collab_00/task.json
```

The test directory names are identical to the original 100-task release. Some
training instructions occur in multiple scenes or source batches. Those
collisions receive a deterministic `.__batch_XX` suffix so that no task is
overwritten. `splits.json` records the original task name, source batches,
selected successful trial, file sizes, and SHA-256 hashes.

## Download

```bash
hf download Sunyao/deconav --repo-type dataset --local-dir assets/datasets/deconav
```

Use `assets/datasets/deconav/test` as the evaluator dataset directory. The
DeCoNav source release may pin an older immutable dataset revision for exact
paper-result reproduction; that revision remains available in repository
history.

## Data Scope

- Simulator: Habitat-Sim
- Scene identifiers: HM3D v0.2
- Scene files: not included
- Released modalities: JSON task annotations and compact trajectories
- Trial policy: preserve the original released trial for test tasks; use the
  lexicographically latest successful trial for training tasks when multiple
  successful trials exist

Evaluator-only positions, geodesic distances, and reference paths stored in
the compact trajectory are used only for initialization and metric
calculation. They are not policy inputs in the released DeCoNav evaluator.

## Integrity

`DATA_SHA256SUMS` contains hashes for every task file. `splits.json` contains
the same per-file hashes plus split-level and whole-dataset tree hashes. Tree
hashes are computed from sorted UTF-8 lines of the form
`<relative_path>\t<size_bytes>\t<file_sha256>\n`.

## License

See `LICENSE`. HM3D scenes are not redistributed and remain subject to their
original terms. Access to this dataset does not grant a license to HM3D.