BURGER: Painter Robot Drawing Datasets

Project Homepage: https://siyingzhao.github.io/BURGER/

Robot drawing trajectories from the Painter/BURGER project, including simulation and real-robot recordings. This repository contains six collections of LeRobot v2.1 datasets for robot policy training.

Release status: complete. All six groups have been uploaded and their remote file paths and sizes verified.

Repository organization

The four ManiSkill collections share one maniskill/ directory in this same Hugging Face dataset repository. They retain independent episode indices, metadata and training formats.

BURGER/
  svg_samples/      # 132 original SVG examples
  painter-libero-single-v6/
  painter-real/
  maniskill/
    single/       # Single-object drawing, category TAR shards
    left_right/   # Left/right relations, task TAR shards
    up_down/      # Above/below relations, native LeRobot folders
    in_out/       # Inside/contains relations, task TAR shards

Original SVG sample

A small, byte-preserving sample of the original drawing inputs is available in svg_samples: 132 SVG files, one per single-object category, with source filenames and SHA-256 checksums. These are drawing assets; they are separate from the trajectory collections below.

Collections

Directory	Contents	Subdatasets	Episodes	Frames
painter-libero-single-v6	LIBERO-based simulation, 132 single-object categories	132	14,751	17,386,870
painter-real	Real-robot recordings, 34 distinct object categories	4	200	272,551
maniskill/single	ManiSkill simulation, 132 single-object categories	132	14,506	13,668,603
maniskill/left_right	Simulation, left/right spatial relations	111	12,198	19,406,658
maniskill/up_down	Simulation, above/below spatial relations	50	500	475,934
maniskill/in_out	Simulation, inside/contains spatial relations	50	5,600	6,049,039

These counts describe the selected release, not all raw recordings. Simulation variants can share drawing inputs; collections should not be assumed statistically independent. The LIBERO-based collection is generated drawing data, not a redistribution of the official LIBERO benchmark. Its release format is LeRobot, not the original simulator HDF5 format.

File layout and use

Each unpacked subdataset is an independent LeRobot v2.1 root:

<collection>/<object-or-task>/
  meta/info.json
  meta/tasks.jsonl
  meta/episodes.jsonl
  meta/episodes_stats.jsonl
  data/chunk-000/episode_000000.parquet
  videos/chunk-000/<camera-key>/episode_000000.mp4
  episode_mapping.jsonl

The four real-robot subdatasets are real_10class_100, real_20class_60, real_3class_30, and real_burger_10. Each collection also has a manifest.json listing its subdatasets, exact features, episode counts, and release exclusions.

Parquet contains observations, actions, timestamps and indices. Video files contain camera observations. Task instructions are recorded in meta/tasks.jsonl; per-episode lengths and tasks are in meta/episodes.jsonl. episode_mapping.jsonl maps release episode indices to original episode indices.

The repository root is a collection container, not one directly loadable LeRobot dataset. Download a subdataset, then use its directory as the dataset root with a loader supporting v2.1:

from pathlib import Path
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Zhaosying66/BURGER",
    repo_type="dataset",
    allow_patterns=["painter-libero-single-v6/baseball/**"],
    local_dir="./BURGER",
)
dataset_root = Path("./BURGER/painter-libero-single-v6/baseball")

All selected datasets declare 20 Hz and 320 × 240 RGB video. The LIBERO-based collection has front and wrist cameras and 8-dimensional actions. The other five collections have a single camera and 7-dimensional actions. Observation states are 8-dimensional. Consult each meta/info.json for its exact schema. Action and state conventions differ between simulation and real-robot sources; do not concatenate them into a training stream without the appropriate project-specific transforms and normalization.

Category TAR downloads

painter-real, maniskill/up_down, and painter-libero-single-v6 are distributed as native folders. The other three collections are distributed as 293 uncompressed category/task TAR shards:

Collection	TAR shards
maniskill/single	132
maniskill/in_out	50
maniskill/left_right	111

For these collections, use <collection>/shards/<category>.tar. Each archive contains <collection>/<category>/meta, data, videos and the episode mapping, preserving the LeRobot v2.1 directory structure. The included _archive_manifest.json records every original member's size and SHA-256. No video re-encoding or Parquet conversion is performed.

Each collection's archives.json lists all shards, sizes and SHA-256 values; SHA256SUMS provides standard checksum verification. The repository viewer configurations cover only the three native-folder collections. TAR collections should be downloaded and extracted before use.

Example for one ManiSkill category:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="Zhaosying66/BURGER",
    repo_type="dataset",
    allow_patterns=[
        "maniskill/single/shards/baseball.tar",
        "maniskill/single/SHA256SUMS",
        "maniskill/single/archives.json",
    ],
    local_dir="./BURGER",
)

(cd ./BURGER/maniskill/single && sha256sum --ignore-missing -c SHA256SUMS)
tar -xf ./BURGER/maniskill/single/shards/baseball.tar -C ./BURGER

The extracted dataset root is ./BURGER/maniskill/single/baseball. Extract into the repository root, not into the shards directory.

Quality control and exclusions

• Only episodes registered in the source metadata are included; stray files and duplicate directory aliases are excluded.
• The known corrupt original baseball episode 76 in the LIBERO-based collection is excluded in full, including both cameras and its Parquet record. Remaining episode indices, global frame indices and affected per-episode index statistics are updated in the release copy.
• The single-episode real-robot apple overfit set is excluded.
• Uploaded collections pass Parquet row count, finite action/state, shape, index, task-index and timestamp checks, plus complete CPU video decoding with frame-count checks. Prior full video-audit results are reused only when file size, modification time and expected frame count match.
• The original source datasets are unchanged. Environment-specific caches, pickles, logs, checkpoints and derived GR00T caches are not included.

The supplied split is train; this release does not establish a held-out benchmark. Construct evaluation splits by drawing source or object/task as appropriate to avoid leakage.

License and attribution

This dataset release is licensed under Creative Commons Attribution–NonCommercial 4.0 International (CC BY-NC 4.0). See LICENSE and the official license.

Credit the Painter/BURGER dataset contributors, link to this repository and the license, and indicate modifications. The license does not grant permission for commercial use. Third-party rights, where applicable, remain with their respective owners. The dataset license does not relicense the ManiSkill or LIBERO simulator software.

Citation

If you use BURGER in your research, please cite:

@misc{xu2026burger,
  title = {{BURGER}: Benchmarking and Understanding Robotic Generalization via Embodied dRawing},
  author = {Xu, Xiangyu and Peng, Bangshuai and Liu, Ziyi and Zhao, Mengjie and Lu, Chenhao and Fan, Zhengyang and Xu, Huazhe and Hu, Kaizhe},
  year = {2026},
  eprint = {2610.40532},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.40532}
}

中文说明

ManiSkill 单物体、包含关系、左右关系三组采用按类别/任务打包的 TAR 分片；下载并校验后，解压到仓库根目录即可恢复原始 LeRobot 结构。

本仓库包含 Painter/BURGER 项目的六组机器人绘画训练数据，覆盖真机、单物体仿真及空间关系仿真。每个类别或任务目录都是独立的 LeRobot v2.1 数据集。请根据各子目录的 meta/info.json 和项目训练代码处理动作、坐标系及归一化，不应直接混用不同来源的动作表示。

数据采用 CC-BY-NC-4.0：要求署名、注明修改，不授予商业使用许可。六组数据已全部上传并完成远端文件清单核验。