File size: 5,368 Bytes
7bc0acb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
---
license: cc-by-nc-4.0
pretty_name: "Bonsai Vineyard (Sample)"
task_categories:
  - text-to-video
  - image-to-video
tags:
  - agriculture
  - vineyard
  - robotics
  - cosmos
size_categories:
  - n<1K
---

# Bonsai Vineyard (Sample)

Front-facing camera video from a farm robot driving in one vineyard, with per-frame segmentation and depth control
videos and a text caption for each clip, plus four clips from an orchard and a strawberry field for out-of-domain
checks. It is the training, validation and test data for the Cosmos3-Nano
multi-control LoRA recipe in [NVIDIA/cosmos-framework](https://github.com/NVIDIA/cosmos-framework).

## Contents

| Split | Clips | Use |
|---|---|---|
| train | 50 | LoRA training |
| val | 32 | validation loss, checkpoint choice |
| test | 50 | held-out evaluation, recorded on a different day from train and val |
| stretch | 4 | out-of-domain check: 2 orchard, 2 strawberry-bed clips (crops not in train) |

Each clip is 150 frames (10 s) at 15 fps, 1280x720, H.264.

```
clips/<id>/rgb.mp4      camera video
clips/<id>/seg.mp4      semantic segmentation, colour-coded (legend below)
clips/<id>/depth.mp4    depth, encoded as below
clips/<id>/meta.json    frame count, fps, size, depth encoding
{,val_,test_,stretch_}{vision,seg,depth}.jsonl   cosmos-framework dataset files (train has no prefix);
                        clip paths are relative to this folder, so point DATASET_PATH here
metadata.csv            id, split, crop, sequence, motion_class, place, eval_panel, num_frames, fps, caption
```

- `sequence`: clips cut from the same recording share a sequence id. Clips in one sequence can overlap in time, so
  keep a sequence within one split if you re-split the data.
- `crop`: `vineyard`, or `orchard` / `strawberry` for the stretch clips.
- `motion_class`: `straight`, `turn` or `curve`. `place`: `row` (between vine rows), `edge` (headland) or `yard`.
- `eval_panel`: the six test clips used for the recipe's before/after evaluation.

### Depth encoding

`depth.mp4` stores inverse depth in 8 bits (grey level `v`): `v = 1 + 254 * (1/z - 1/z_far) / (1/z_near - 1/z_far)`,
with `z_near = 0.75 m`, `z_far = 15 m`; `v = 0` means no depth. The video is lossy (H.264), so treat recovered metric
depth as approximate.

### Segmentation

Each pixel of `seg.mp4` is drawn in its class colour. The video is lossy, so map each pixel to the nearest colour in
the table. The table is shared with Bonsai's other crops; in this dataset almost every pixel is dirt, canopy, sky,
unassigned, dirt lane, tree trunk (vine trunks and posts), gravel, fence or grass.

| id | colour (RGB) | class |
|---|---|---|
| 0 | 0, 0, 0 | unassigned |
| 5 | 220, 20, 60 | human |
| 6 | 255, 140, 0 | vehicle |
| 7 | 128, 64, 128 | building |
| 8 | 153, 153, 153 | utility_pole |
| 9 | 0, 60, 100 | water_valve |
| 10 | 110, 70, 20 | tree_trunk |
| 11 | 34, 139, 34 | canopy |
| 12 | 230, 230, 230 | bed (plastic-mulch raised bed) |
| 13 | 150, 110, 70 | furrow |
| 14 | 135, 206, 235 | sky |
| 15 | 255, 105, 180 | mammal |
| 16 | 255, 182, 193 | bird |
| 17 | 255, 215, 0 | farm_equipment |
| 18 | 218, 165, 32 | bee_box |
| 19 | 250, 170, 30 | sign |
| 20 | 70, 130, 180 | tank |
| 21 | 0, 0, 142 | pump |
| 23 | 190, 153, 153 | fence |
| 24 | 102, 102, 156 | gate |
| 25 | 255, 0, 255 | irrigation_line |
| 26 | 107, 142, 35 | bush |
| 27 | 0, 255, 127 | crop |
| 28 | 196, 164, 132 | road (dirt lane) |
| 29 | 169, 169, 169 | gravel |
| 30 | 139, 90, 43 | dirt |
| 31 | 124, 252, 0 | grass |
| 32 | 101, 67, 33 | mud |
| 33 | 160, 82, 45 | berm |
| 34 | 0, 191, 255 | standing_water |
| 35 | 105, 105, 105 | train_tracks |
| 36 | 72, 61, 139 | drainage |
| 37 | 190, 190, 190 | concrete |

## How the data was made

Recorded by Bonsai Robotics farm robots. The vineyard clips come from one vineyard over a few days: train and
validation clips share recording days, and all test clips come from a separate day. The stretch clips come from an
orchard and a strawberry field. Clips were chosen to balance straight driving, turns and curves. Segmentation is produced automatically by a
segmentation model (not hand-labelled), and depth comes from the robot's stereo camera; both are reprojected into the same 1280x720 pinhole
view as the camera video. For the two strawberry clips the segmentation and depth exist for every 5th frame; the
frames in between take the nearest labelled frame, shifted to match the camera motion with optical flow. Captions are
generated from the robot's motion and the sun position at recording time.

## Limitations

- This is a small sample dataset (50 training clips from one vineyard, daylight only; 4 stretch clips). Models trained on it show
  what a small dataset can do; they will not reach the quality a larger, more varied dataset gives.

## Licence

Released under the [Creative Commons Attribution-NonCommercial 4.0 International licence (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/).
You may share and adapt the data for non-commercial purposes, with credit to Bonsai Robotics. Commercial use is not
permitted.

## Citation

```bibtex
@misc{bonsai_vineyard_sample_2026,
  title        = {Bonsai Vineyard (Sample)},
  author       = {{Bonsai Robotics}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/BonsaiRobotics/bonsai-vineyard-sample}}
}
```