File size: 7,751 Bytes
40b6e64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9134826
40b6e64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9134826
 
40b6e64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
---
license: apache-2.0
language:
- en
pretty_name: DOMINO Absolute Motion v2
size_categories:
- 1M<n<10M
tags:
- robotics
- robot-learning
- imitation-learning
- video-latents
- optical-flow
- safetensors
- domino
- dynamicwam
---

# DOMINO Absolute Motion v2

`DOMINO_absolute_motion_v2` is the complete packed training corpus for
DynamicWAM's exact-simulator-time motion pipeline. It is a training-ready
derivative of [H-EmbodVis/DOMINO](https://huggingface.co/datasets/H-EmbodVis/DOMINO),
not a copy of the raw RGB dataset.

Every sample was packed as one aligned record containing video latents,
action/state targets, frame indices, a language-group identifier, and four
history intervals of absolute motion descriptors. The alignment is fixed at
conversion time; training does not join independent sidecars at runtime.

## Dataset summary

| Property | Value |
|---|---:|
| Tasks | 35 |
| Episodes | 10,500 (300 per task) |
| Packed samples | 1,428,327 |
| Main tensor shards | 143 |
| Language-bank entries | 3,500 |
| Language tensor shards | 55 |
| Payload files | 406 |
| Payload size | 333,217,057,616 bytes (310.33 GiB) |
| Format | `dynamicwam_absolute_motion_dataset` |
| Format version | 2 |
| Dataset fingerprint | `2feac8645817e3a896afd855341130f23a6d92bfd28917f4cb3e2dbb4eb68154` |

The packed corpus contains all 1,428,327 valid samples. The production
training contract uses
`episode_balanced_without_replacement`, selecting 10 distinct samples from
each episode per epoch with seed 2026 (105,000 samples per epoch).

## Stored tensors

Each main `.safetensors` shard contains up to 10,000 aligned samples.
Shapes below exclude the leading sample dimension.

| Key | Stored dtype | Per-sample shape |
|---|---|---|
| `condition_latents` | `float32` | `[48, 2, 24, 20]` |
| `future_latents` | `float32` | `[48, 2, 12, 10]` |
| `action_sequences` | `float32` | `[16, 14]` |
| `initial_states` | `float32` | `[14]` |
| `absolute_motion_features` | `float32` | `[4, 12]` |
| `absolute_motion_interval_valid_masks` | `bool` | `[4]` |
| `absolute_motion_acceleration_valid_masks` | `bool` | `[4]` |
| `action_indices` | `int64` | `[16]` |
| `video_indices` | `int64` | `[8]` |
| `condition_frame_indices` | `int64` | scalar |
| `episode_indices` | `int64` | scalar |
| `lang_group_ids` | `int64` | scalar |
| `sample_ids` | `int64` | scalar |

The language bank stores 3,500 precomputed 4096-dimensional UMT5-XXL
entries in `bfloat16`, grouped as 100 entries for each of the 35 tasks.

## Absolute-motion contract

Motion is computed from the head view on a configured `64 x 64` flow grid.
Five history frames define four intervals. Interval duration comes from
`domino_schema_v2.sim_time_seconds`, rather than an assumed fixed frame rate.
Each interval has 12 descriptors:

1. mean x displacement;
2. mean y displacement;
3. mean displacement magnitude;
4. p99 displacement magnitude;
5. elapsed time;
6. mean x velocity;
7. mean y velocity;
8. mean speed;
9. p99 speed;
10. mean x acceleration;
11. mean y acceleration;
12. mean acceleration magnitude.

`motion_stats.json` pins normalization statistics and
`action_stats.json` pins action statistics. The precise optical-flow,
timestamp, sampling, and normalization contracts are recorded in
`dataset.json`.

## File layout

```text
.
├── dataset.json
├── stats.json
├── action_stats.json
├── motion_stats.json
├── episodes.jsonl
├── samples.jsonl
├── shards/
│   ├── shard_000000.safetensors
│   ├── shard_000000.json
│   └── ...
├── lang/
│   ├── lang.json
│   └── shards/
│       ├── shard_000000.safetensors
│       ├── shard_000000.json
│       └── ...
├── SHA256SUMS
├── PAYLOAD_MANIFEST.json
└── VALIDATION.md
```

The JSON sidecar beside each tensor shard gives the exact keys, shapes,
dtypes, sample range, and byte size for that shard.

## Download

Install the current Hugging Face CLI with Xet support, then download the
repository:

```bash
hf download KhalilGao/DOMINO_absolute_motion_v2 \
  --repo-type dataset \
  --local-dir DOMINO_absolute_motion_v2
```

The repository is approximately 311 GiB. Verify the immutable payload after
download:

```bash
cd DOMINO_absolute_motion_v2
sha256sum -c SHA256SUMS
```

`SHA256SUMS` covers the 406 source payload files. Repository documentation
files are intentionally outside that payload manifest.
The SHA-256 of `SHA256SUMS` itself is
`705371585c4fdb29440a5a49849a69e94fd4c85acda63c2a4ed005aa865bad2c`.

## DynamicWAM loader

With the DynamicWAM source tree on `PYTHONPATH`:

```python
from dynamicwam.training.data import PackedAbsoluteMotionDataset

dataset = PackedAbsoluteMotionDataset(
    "DOMINO_absolute_motion_v2",
    max_open_shards=8,
)
sampler = dataset.make_sampler(
    samples_per_episode=10,
    seed=2026,
)
```

The loader validates the format version, dataset fingerprint, motion
contract, motion statistics, action statistics, and sampler contract before
training.

## Tasks

The dataset contains the following DOMINO Level-1 tasks:

`adjust_bottle`, `beat_block_hammer`, `click_alarmclock`, `click_bell`,
`dump_bin_bigbin`, `grab_roller`, `handover_block`, `handover_mic`,
`hanging_mug`, `move_can_pot`, `move_pillbottle_pad`,
`move_playingcard_away`, `move_stapler_pad`, `place_a2b_left`,
`place_a2b_right`, `place_bread_basket`, `place_bread_skillet`,
`place_can_basket`, `place_container_plate`, `place_empty_cup`, `place_fan`,
`place_mouse_pad`, `place_object_basket`, `place_object_scale`,
`place_object_stand`, `place_phone_stand`, `place_shoe`, `press_stapler`,
`put_bottles_dustbin`, `put_object_cabinet`, `rotate_qrcode`, `scan_object`,
`shake_bottle`, `shake_bottle_horizontally`, and `stamp_seal`.

## Provenance

- Raw source: [H-EmbodVis/DOMINO](https://huggingface.co/datasets/H-EmbodVis/DOMINO)
- Video latent model: [Wan-AI/Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B)
- Text encoder: [google/umt5-xxl](https://huggingface.co/google/umt5-xxl)
- Conversion-manifest digest recorded by the packer:
  `926ae37efced8fec3f1a59c4fe2930fbb9bca4c65197efe9e9ba80aa9fabdb16`
- Action-statistics SHA-256:
  `730a91e41b6a23c70cec49cd222835373ba9364680da5753f4408c5534319b04`
- Motion-statistics SHA-256:
  `7a2109f0d93418e9bc3b5c1c6c9961ac7afbda61f672a153b999e8df385a01b4`

No Wan2.2 or UMT5 model weights are included in this repository.

## Limitations

- This is packed simulated training data for the 35 DOMINO Level-1 tasks; it
  does not contain real-robot data.
- Raw RGB observations are not included. Use the upstream DOMINO repository
  when raw episodes are required.
- The stored latents and language embeddings are tied to the encoder
  configurations recorded in `dataset.json`; changing encoders requires
  repacking.
- This custom shard format is optimized for training and is not a standard
  tabular Hugging Face `Dataset`, so the web Data Viewer may not render sample
  rows.
- The corpus does not by itself reproduce the complete DynamicWAM software,
  model checkpoints, or evaluation environment.

## License and citation

This processed dataset is released under Apache License 2.0, consistent with
the upstream DOMINO dataset and the encoder repositories listed above. See
`LICENSE` for the full text.

If you use this artifact, cite the accompanying DynamicWAM work and the
upstream DOMINO paper:

```bibtex
@inproceedings{fang2026towards,
  title     = {Towards Generalizable Robotic Manipulation in Dynamic Environments},
  author    = {Fang, Heng and Li, Shangru and Wang, Shuhan and Xi, Xuanyang and Liang, Dingkang and Bai, Xiang},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
```