SiriusYY commited on
Commit
cd8a8aa
·
verified ·
1 Parent(s): e1c55b7

Present RoboDesk as the MR0 pretraining data release

Browse files

Rebuild the dataset homepage with MR0 context and citation, complete release coverage, download and extraction instructions, and verified Parquet examples.

Files changed (1) hide show
  1. README.md +156 -37
README.md CHANGED
@@ -1,67 +1,186 @@
1
  ---
2
  license: cc-by-nc-sa-4.0
 
 
 
 
 
 
 
 
 
 
3
  ---
4
 
5
- <!-- ROBODESK_RELEASE_BEGIN -->
6
- # RoboDesk
 
7
 
8
- RoboDesk is a collection of embodied robotics episodes organized into task-level datasets, with episode tables, camera videos, and associated metadata.
 
 
 
 
 
9
 
10
- > **Release status — 2026-11-19.** Available task datasets: **846 / 846**. This includes **812 task datasets as individual files** and **34 task datasets as complete tar archives**. Published data represents **1,242,445 logical dataset files** and approximately **10.46 TB**. **0 task datasets remain pending archive publication.**
 
 
 
 
 
 
11
 
12
- ## Coverage
13
 
14
- | Source directory | Tasks as individual files | Tasks in tar archives | Planned tasks | Published logical dataset files |
15
- | --- | ---: | ---: | ---: | ---: |
16
- | `agibotworld-beta` | 126 | 34 | 160 | 535,968 |
17
- | `droid` | 159 | 0 | 159 | 235,532 |
18
- | `galaxea-open-world` | 172 | 0 | 172 | 98,148 |
19
- | `let-base` | 18 | 0 | 18 | 39,946 |
20
- | `robocoin` | 179 | 0 | 179 | 206,005 |
21
- | `robomind-v1.2` | 158 | 0 | 158 | 126,846 |
22
- | **Total** | **812** | **34** | **846** | **1,242,445** |
23
 
24
- Counts refer to tasks selected for this release, rather than the coverage of the original source collections. Logical file counts include files inside published archives. Sizes use rounded decimal TB.
 
 
25
 
26
- ## File-count limit and archive format
27
 
28
- The current limits are **1,000,000 files per repository** and **50 GB per uploaded file**. The planned release contains **1,242,445 logical dataset files**, approximately **10.46 TB**. The final 34 AgiBot task datasets are being published as **97 large uncompressed tar parts**, each no larger than **49 GB**. Each task has 2–8 independently extractable parts.
 
 
29
 
30
- - Tar parts are stored as `archives/agibotworld-beta/<task_name>/part-*.tar`.
31
- - All parts for a task and its `manifest.json` are published together. A task marked **published** in [`index.json`](archives/agibotworld-beta/index.json) has its complete archive set available.
32
- - Each manifest lists every part's checksum and all original member paths, sizes, and SHA-256 checksums. The global [`SHA256SUMS`](archives/agibotworld-beta/SHA256SUMS) lists archive and manifest checksums.
33
- - Extracting all parts for a task from the repository root restores the original `datasets/agibotworld-beta/data/<task_name>/` structure. File contents are preserved; tar ownership and timestamp fields are normalized.
34
 
35
- ## Using the archived task datasets
 
 
 
 
 
 
 
 
36
 
37
- 1. Select a task marked `published` in `archives/agibotworld-beta/index.json`.
38
- 2. Download every tar part listed in its manifest and `archives/agibotworld-beta/SHA256SUMS`.
39
- 3. Verify the checksums, then extract all parts from the repository root.
40
 
41
- For example, once the first archived task is published and downloaded:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42
 
43
  ```bash
44
  task_name=480_make_sandwiches_with_salad_dressing_1213_edition
45
- if grep -F " archives/agibotworld-beta/$task_name/part-" archives/agibotworld-beta/SHA256SUMS | sha256sum --check -
 
 
 
46
  then
47
- for archive in "archives/agibotworld-beta/$task_name"/part-*.tar
48
  do
49
- tar -xf "$archive" -C .
50
  done
51
  fi
52
  ```
53
 
54
- ## Directory structure and features
55
 
56
- Task roots are located under `datasets/<source>/data/`, sometimes with additional robot or group directories before the task name. Each task has episode Parquet files under `data/`, camera videos under `videos/`, and associated files under `meta/`.
57
 
58
- Text annotations are stored in the episode Parquet tables. Depending on the task, these include `global_instruction`, `global_instruction_dense`, `current_instruction`, and `scene_description`. Inspect the Parquet schema and the task's `meta/info.json`; camera names, feature names, capitalization, and array dimensions can vary across sources.
59
 
60
- ## Integrity checks and validation scope
61
 
62
- The 812 task datasets published as individual files were checked for every file's existence and byte size, with hash comparisons for Parquet and metadata and sampled hash comparisons for videos.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63
 
64
- For archived task datasets, every tar member is checked against the reviewed release file list and its source-content checksum before upload. Published archive sizes and checksums are checked remotely. All parts for a task are required to restore the complete task dataset.
65
 
66
- These checks establish file-level publication integrity. End-to-end loading and training have not been tested individually for every task dataset. Annotation accuracy, semantic consistency, and downstream pipeline compatibility have not been exhaustively validated.
67
- <!-- ROBODESK_RELEASE_END -->
 
1
  ---
2
  license: cc-by-nc-sa-4.0
3
+ pretty_name: RoboDesk
4
+ tags:
5
+ - robotics
6
+ - embodied-ai
7
+ - robot-learning
8
+ - vision-language-action
9
+ - multimodal
10
+ - parquet
11
+ - video
12
+ - mr0
13
  ---
14
 
15
+ <h1 align="center">RoboDesk</h1>
16
+ <p align="center"><strong>Robot Pretraining Data for Meituan-Robotics-0 (MR0)</strong></p>
17
+ <p align="center">Meituan Robotics</p>
18
 
19
+ <p align="center">
20
+ <a href="https://arxiv.org/abs/2611.08577"><img src="https://img.shields.io/badge/Paper-MR0-b31b1b?logo=arxiv" alt="MR0 technical report"></a>
21
+ <a href="https://github.com/Meituan-Robotics/Meituan-Robotics-0"><img src="https://img.shields.io/badge/Code-MR0-181717?logo=github" alt="MR0 code"></a>
22
+ <a href="https://meituan-robotics.github.io/Meituan-Robotics-0/"><img src="https://img.shields.io/badge/Project-Page-228b22" alt="MR0 project page"></a>
23
+ <a href="/datasets/MeituanRobotics/RoboDesk/tree/main"><img src="https://img.shields.io/badge/Release-Complete-1f6feb" alt="Browse the complete RoboDesk release"></a>
24
+ </p>
25
 
26
+ <p align="center">
27
+ <a href="#overview">Overview</a> ·
28
+ <a href="#dataset-coverage">Dataset coverage</a> ·
29
+ <a href="#download-and-use">Download and use</a> ·
30
+ <a href="#data-format">Data format</a> ·
31
+ <a href="#citation">Citation</a>
32
+ </p>
33
 
34
+ ## Overview
35
 
36
+ **RoboDesk is the robot pretraining data release of [Meituan-Robotics-0 (MR0)](https://github.com/Meituan-Robotics/Meituan-Robotics-0), a vision-language-action foundation model for desktop manipulation.** Released as part of the MR0 work, RoboDesk brings together selected task datasets from six robotics collections, with robot observations, actions, camera videos, and instruction annotations.
 
 
 
 
 
 
 
 
37
 
38
+ | Source collections | Task datasets | Logical dataset files | Dataset size |
39
+ | :---: | :---: | :---: | :---: |
40
+ | **6** | **846** | **1,242,445** | **≈10.46 TB** |
41
 
42
+ **The full release is available:** 812 task datasets are provided as individual files, and 34 additional AgiBot task datasets are provided in 97 tar parts. Logical file counts include files inside archives; sizes use decimal TB.
43
 
44
+ - **Episodic robot data.** Parquet episode tables, MP4 camera videos, and task metadata are organized by source collection and task.
45
+ - **Instructions in the episode tables.** Text fields include `global_instruction`, `global_instruction_dense`, `current_instruction`, and `scene_description`, as described by the task metadata and Parquet schema.
46
+ - **A component of MR0.** For the model, training pipeline, and evaluation, see the [MR0 repository](https://github.com/Meituan-Robotics/Meituan-Robotics-0) and [technical report](https://arxiv.org/abs/2611.08577). To cite RoboDesk, use the [MR0 citation](#citation) below.
47
 
48
+ ## Dataset coverage
 
 
 
49
 
50
+ | Source directory | Task datasets | As individual files | In tar archives |
51
+ | :--- | ---: | ---: | ---: |
52
+ | [`agibotworld-beta`](/datasets/MeituanRobotics/RoboDesk/tree/main/datasets/agibotworld-beta) | 160 | 126 | 34 |
53
+ | [`droid`](/datasets/MeituanRobotics/RoboDesk/tree/main/datasets/droid) | 159 | 159 | 0 |
54
+ | [`galaxea-open-world`](/datasets/MeituanRobotics/RoboDesk/tree/main/datasets/galaxea-open-world) | 172 | 172 | 0 |
55
+ | [`let-base`](/datasets/MeituanRobotics/RoboDesk/tree/main/datasets/let-base) | 18 | 18 | 0 |
56
+ | [`robocoin`](/datasets/MeituanRobotics/RoboDesk/tree/main/datasets/robocoin) | 179 | 179 | 0 |
57
+ | [`robomind-v1.2`](/datasets/MeituanRobotics/RoboDesk/tree/main/datasets/robomind-v1.2) | 158 | 158 | 0 |
58
+ | **Total** | **846** | **812** | **34** |
59
 
60
+ These counts describe the task datasets selected for RoboDesk, rather than the full coverage of each original collection. Some sources include robot or group subdirectories before the task name.
 
 
61
 
62
+ ## Download and use
63
+
64
+ Choose the source collection and task subset you need. The two packaging forms restore the same task directory layout.
65
+
66
+ ### Tasks available as individual files
67
+
68
+ Browse the [dataset directories](/datasets/MeituanRobotics/RoboDesk/tree/main/datasets) and download the complete directory for a task, including its `data/`, `videos/`, and `meta/` files. Preserve the relative paths so episode tables and videos remain associated with the correct task.
69
+
70
+ ### Tasks available as tar parts
71
+
72
+ The 34 archived AgiBot task datasets are listed in the [archive index](/datasets/MeituanRobotics/RoboDesk/blob/main/archives/agibotworld-beta/index.json). **All 34 are published and complete.** Their files are available under the [archive directory](/datasets/MeituanRobotics/RoboDesk/tree/main/archives/agibotworld-beta).
73
+
74
+ 1. Select a task in the index and open its `manifest.json`.
75
+ 2. Download **every tar part for that task**, its manifest, and the global [`SHA256SUMS`](/datasets/MeituanRobotics/RoboDesk/blob/main/archives/agibotworld-beta/SHA256SUMS). Keep their repository-relative paths.
76
+ 3. Verify the checksums and extract all parts from your local dataset root.
77
+
78
+ For example, after downloading the files for `480_make_sandwiches_with_salad_dressing_1213_edition`:
79
 
80
  ```bash
81
  task_name=480_make_sandwiches_with_salad_dressing_1213_edition
82
+ archive_dir="archives/agibotworld-beta/${task_name}"
83
+
84
+ if grep -F " ${archive_dir}/" archives/agibotworld-beta/SHA256SUMS \
85
+ | sha256sum --check -
86
  then
87
+ for archive in "${archive_dir}"/part-*.tar
88
  do
89
+ tar -xf "${archive}" -C . || exit 1
90
  done
91
  fi
92
  ```
93
 
94
+ This restores `datasets/agibotworld-beta/data/<task_name>/`. Each part is an independently extractable tar file; **all parts are required for the complete task dataset**.
95
 
96
+ The release uses 97 uncompressed tar parts, each **≤49 GB**, to stay within the repository's **1,000,000-file** and **50 GB-per-file** limits. A task has 2–8 parts. Each manifest records the paths, sizes, and SHA-256 checksums of its parts and their member files.
97
 
98
+ ## Data format
99
 
100
+ Task datasets use a LeRobot-style layout:
101
 
102
+ ```text
103
+ datasets/<source>/data/[<robot-or-group>/]<task_name>/
104
+ ├── data/
105
+ │ └── chunk-*/episode_*.parquet
106
+ ├── videos/
107
+ │ └── chunk-*/<camera_name>/episode_*.mp4
108
+ └── meta/
109
+ ├── info.json
110
+ ├── episodes.jsonl
111
+ ├── tasks.jsonl
112
+ └── ...
113
+
114
+ archives/agibotworld-beta/<task_name>/
115
+ ├── part-00000.tar
116
+ ├── part-00001.tar
117
+ ├── ...
118
+ └── manifest.json
119
+ ```
120
+
121
+ The `features` section in each task's **`meta/info.json`** describes its observations, actions, camera streams, and text fields. **The instruction annotations themselves are stored in the Parquet episode tables.**
122
+
123
+ | Content | Example fields |
124
+ | :--- | :--- |
125
+ | Episode and frame indexing | `episode_index`, `frame_index`, `timestamp`, `task_index` |
126
+ | Robot observations | `observation.state`, joint positions, Cartesian poses, gripper state |
127
+ | Robot actions | `action`, joint, Cartesian, and gripper action fields |
128
+ | Camera streams | Video features such as `observation.images.head` and wrist cameras |
129
+ | Instruction and scene text | `global_instruction`, `global_instruction_dense`, `current_instruction`, `scene_description` |
130
+
131
+ Field availability, camera names, capitalization, array shapes, frame rates, and numeric types vary across tasks. Inspect both `meta/info.json` and the actual Parquet schema when building a loader.
132
+
133
+ ### Inspect an episode
134
+
135
+ With [PyArrow](https://arrow.apache.org/docs/python/parquet.html) installed, inspect a downloaded or extracted task without loading an entire episode into memory:
136
+
137
+ ```python
138
+ from pathlib import Path
139
+ import pyarrow.parquet as pq
140
+
141
+ task_root = Path(
142
+ "datasets/agibotworld-beta/data/"
143
+ "480_make_sandwiches_with_salad_dressing_1213_edition"
144
+ )
145
+ episode_path = next((task_root / "data").rglob("*.parquet"))
146
+ episode = pq.ParquetFile(episode_path)
147
+ print(episode.schema_arrow)
148
+
149
+ text_fields = [
150
+ "global_instruction",
151
+ "global_instruction_dense",
152
+ "current_instruction",
153
+ "scene_description",
154
+ ]
155
+ columns = [name for name in text_fields if name in episode.schema_arrow.names]
156
+ if columns:
157
+ preview = next(episode.iter_batches(batch_size=3, columns=columns))
158
+ print(preview.to_pylist())
159
+ ```
160
+
161
+ ## Integrity and compatibility
162
+
163
+ The 812 task datasets provided as individual files were checked for file existence and byte size, with hash comparisons for Parquet and metadata and sampled hash comparisons for videos. For the 34 archived task datasets, every tar member was checked against the release file list and its source-content checksum; all 97 published tar parts were verified remotely by size and SHA-256. Tar ownership and timestamp fields are normalized while file contents are preserved.
164
+
165
+ These checks establish publication integrity. End-to-end loading and training have not been tested separately for every task dataset, and annotation accuracy has not been exhaustively validated.
166
+
167
+ ## Citation
168
+
169
+ RoboDesk is part of the MR0 work. If you use RoboDesk in your research, please cite the [MR0 technical report](https://arxiv.org/abs/2611.08577):
170
+
171
+ ```bibtex
172
+ @misc{meituanroboticsteam2026mr0,
173
+ author = {{Meituan Robotics Team}},
174
+ title = {{Meituan-Robotics-0}: A Vision-Language-Action Foundation Model for Desktop Manipulation},
175
+ year = {2026},
176
+ eprint = {2611.08577},
177
+ archivePrefix = {arXiv},
178
+ primaryClass = {cs.RO},
179
+ doi = {10.48550/arXiv.2611.08577},
180
+ url = {https://arxiv.org/abs/2611.08577}
181
+ }
182
+ ```
183
 
184
+ ## License and acknowledgements
185
 
186
+ RoboDesk is released under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/). We thank the creators and contributors of the six source collections listed above.