---
license: other
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - deconav
  - embodied-navigation
  - vision-language-navigation
  - multi-robot
---

# DeCoNav Release Assets

This repository contains the external model and perception assets required by
the DeCoNav evaluator. It is designed to be downloaded into the `assets/`
directory of the DeCoNav source repository.

## Contents

```text
models/
  deconav_nav_policy/       # Fine-tuned Qwen3-VL navigation policy
  deconav_eva_clip.pt       # EVA-CLIP visual encoder checkpoint
perception/
  deconav_clip.pt           # Fine-tuned object-recognition checkpoint
  deconav_clip_thresholds.json
  deconav_room_classifier/
    config.json
    best_model.pt
```

Training state, optimizer state, RNG state, scheduler files, and conversion
utilities are intentionally excluded. The source repository validates every
required file before evaluation starts.

## Usage

```bash
hf download Sunyao/deconav --repo-type model --revision REVISION \
  --local-dir assets
```

Replace `REVISION` with the immutable revision recorded in the DeCoNav
`ASSET_MANIFEST.json`. The runtime uses local files only and does not enable
`trust_remote_code`.

## Intended Use

These assets are intended for research reproduction of the five DeCoNav
evaluation modes on the released task manifest. They are not a general-purpose
robot-control system and must not be used as a safety-critical controller.

## License

The navigation checkpoint is a derivative of Qwen model materials and is
distributed under the Qwen Research License included as `LICENSE`. That
license contains non-commercial, redistribution, attribution, and other use
conditions. EVA-CLIP and other perception components also retain their
upstream terms. The MIT license of the DeCoNav source repository does not
relicense these weights.

## Evaluation

The published 100-task Full reference result is BSR 40.0%, OSR 50.0%, ISR
58.5%, and SPL 0.369752. Detailed result provenance is published with the
matching DeCoNav source release.