File size: 3,001 Bytes
cb36872
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0c2a491
cb36872
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
---
license: mit
task_categories:
- video-text-to-text
- robotics
language:
- en
tags:
- vision-language-navigation
- VLN
- navigation
- embodied-ai
- envdrop
size_categories:
- 100K<n<1M
---

# VLNCE-EnvDrop

Synthetic **Vision-Language Navigation (VLN)** data-augmentation set, derived from the
EnvDrop augmentation used in VLN-CE / NaVILA-style training. Each of the **146,304**
samples pairs a short first-person navigation video with the natural-language
instruction the agent was following and the discrete action sequence it executed.

This dataset provides the visual + motion supervision for training a GRU-augmented
Qwen3-VL navigation model: the language conditions the backbone, while the per-step
`motion` sequence feeds a GRU whose output is projected into the LLM embedding space.

## Contents

| File | Size | What it is |
|------|------|------------|
| `envdrop_videos_00.tar``envdrop_videos_14.tar` | ~270 GB | The raw first-person navigation videos, one `<video_id>.mp4` per sample, sharded into 15 tarballs. |
| `envdrop_motion.json` | 392 MB | **Primary training annotation.** One record per sample: instruction, decoded frame paths, and the per-step action (`motion`) sequence. |
| `annotations.json` | 22.5 MB | Lightweight `video_id → instruction` index (powers the dataset preview). A subset of the info in `envdrop_motion.json`. |

## Record schema — `envdrop_motion.json`

```json
{
  "video_id": "34300",
  "q": "Walk forward and stop at the end of the aisle.",
  "frames": ["34300/frame_0.jpg", "34300/frame_1.jpg", "..."],
  "motion": [3, 3, 1, 1, 3, 1, 1, 2, 2, 1, "..."]
}
```

- **`video_id`** — key into the tarballs (`<video_id>.mp4`).
- **`q`** — the natural-language navigation instruction.
- **`frames`** — decoded frame paths for the clip (frames are extracted from the
  corresponding `.mp4` at load time; they are **not** stored separately).
- **`motion`** — the discrete action taken at each step (small action vocabulary,
  e.g. forward / turn-left / turn-right / stop). This is the GRU input.

## Layout

```
VLNCE-EnvDrop/
├── envdrop_videos_00.tar        # <video_id>.mp4 clips
│   ...                          # (15 shards, ~270 GB total)
├── envdrop_videos_14.tar
├── envdrop_motion.json          # primary training annotation (146,304 records)
└── annotations.json             # video_id -> instruction index / preview
```

## Usage

```python
from huggingface_hub import snapshot_download

# annotations only (small)
snapshot_download("Rithvik762/VLNCE-EnvDrop", repo_type="dataset",
                  allow_patterns=["*.json"])

# full dataset incl. video tars (~270 GB)
snapshot_download("Rithvik762/VLNCE-EnvDrop", repo_type="dataset")
```

After download, extract the shards (e.g. `for f in envdrop_videos_*.tar; do tar xf "$f"; done`).
Videos and `envdrop_motion.json` must be kept together — the JSON references
`video_id`s that live inside the tarballs.

## License

Released under the MIT license.