RoboDesk / README.md
SiriusYY's picture
Open Code badge directly at the MR0 source branch
fe729cc verified
|
Raw History Blame Contribute Delete
9.61 kB
metadata
license: cc-by-nc-sa-4.0
pretty_name: RoboDesk
tags:
  - robotics
  - embodied-ai
  - robot-learning
  - vision-language-action
  - multimodal
  - parquet
  - video
  - mr0

RoboDesk

Open-Source Real-World and Simulation Data Used in MR0 Training

Meituan Robotics

MR0 technical report MR0 code MR0 project page Browse the complete RoboDesk release

Overview · Dataset coverage · Download and use · Data format · Citation

Overview

RoboDesk provides the open-source real-world and simulation data subset used in training Meituan-Robotics-0 (MR0), 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.

Source collections Task datasets Logical dataset files Dataset size
6 846 1,242,445 ≈10.46 TB

All 846 task datasets in this RoboDesk release are 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.

  • Episodic robot data. Parquet episode tables, MP4 camera videos, and task metadata are organized by source collection and task.
  • 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.
  • A component of MR0. For the model, training pipeline, and evaluation, see the MR0 repository and technical report. To cite RoboDesk, use the MR0 citation below.

Dataset coverage

Source directory Task datasets As individual files In tar archives
agibotworld-beta 160 126 34
droid 159 159 0
galaxea-open-world 172 172 0
let-base 18 18 0
robocoin 179 179 0
robomind-v1.2 158 158 0
Total 846 812 34

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.

Download and use

Choose the source collection and task subset you need. The two packaging forms restore the same task directory layout.

Tasks available as individual files

Browse the dataset directories 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.

Tasks available as tar parts

The 34 archived AgiBot task datasets are listed in the archive index. All 34 are published and complete. Their files are available under the archive directory.

  1. Select a task in the index and open its manifest.json.
  2. Download every tar part for that task, its manifest, and the global SHA256SUMS. Keep their repository-relative paths.
  3. Verify the checksums and extract all parts from your local dataset root.

For example, after downloading the files for 480_make_sandwiches_with_salad_dressing_1213_edition:

task_name=480_make_sandwiches_with_salad_dressing_1213_edition
archive_dir="archives/agibotworld-beta/${task_name}"

if grep -F "  ${archive_dir}/" archives/agibotworld-beta/SHA256SUMS \
    | sha256sum --check -
then
    for archive in "${archive_dir}"/part-*.tar
    do
        tar -xf "${archive}" -C . || exit 1
    done
fi

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.

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.

Data format

Task datasets use a LeRobot-style layout:

datasets/<source>/data/[<robot-or-group>/]<task_name>/
├── data/
│   └── chunk-*/episode_*.parquet
├── videos/
│   └── chunk-*/<camera_name>/episode_*.mp4
└── meta/
    ├── info.json
    ├── episodes.jsonl
    ├── tasks.jsonl
    └── ...

archives/agibotworld-beta/<task_name>/
├── part-00000.tar
├── part-00001.tar
├── ...
└── manifest.json

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.

Content Example fields
Episode and frame indexing episode_index, frame_index, timestamp, task_index
Robot observations observation.state, joint positions, Cartesian poses, gripper state
Robot actions action, joint, Cartesian, and gripper action fields
Camera streams Video features such as observation.images.head and wrist cameras
Instruction and scene text global_instruction, global_instruction_dense, current_instruction, scene_description

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.

Inspect an episode

With PyArrow installed, inspect a downloaded or extracted task without loading an entire episode into memory:

from pathlib import Path
import pyarrow.parquet as pq

task_root = Path(
    "datasets/agibotworld-beta/data/"
    "480_make_sandwiches_with_salad_dressing_1213_edition"
)
episode_path = next((task_root / "data").rglob("*.parquet"))
episode = pq.ParquetFile(episode_path)
print(episode.schema_arrow)

text_fields = [
    "global_instruction",
    "global_instruction_dense",
    "current_instruction",
    "scene_description",
]
columns = [name for name in text_fields if name in episode.schema_arrow.names]
if columns:
    preview = next(episode.iter_batches(batch_size=3, columns=columns))
    print(preview.to_pylist())

Integrity and compatibility

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.

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.

Citation

RoboDesk is part of the MR0 work. If you use RoboDesk in your research, please cite the MR0 technical report:

@misc{meituanroboticsteam2026mr0,
  author        = {{Meituan Robotics Team}},
  title         = {{Meituan-Robotics-0}: A Vision-Language-Action Foundation Model for Desktop Manipulation},
  year          = {2026},
  eprint        = {2611.08577},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  doi           = {10.48550/arXiv.2611.08577},
  url           = {https://arxiv.org/abs/2611.08577}
}

License and acknowledgements

RoboDesk is released under CC BY-NC-SA 4.0. We thank the creators and contributors of the six source collections listed above.