sim_real_data

UPLOAD IN PROGRESS. The complete collection is not yet available. Do not treat partial files as a complete dataset.

600 original robot demonstration episodes: each task has 100 successful simulation recordings and 100 original real-robot recordings. Lossless archives preserve the original files, timing, and units. This is a raw-data distribution, not a ready-to-train unified LeRobot dataset. No trained models are included.

Task	Simulation episodes / frames	Real episodes / frames
Adjust_Bottle	100 / 41,049	100 / 78,591
Grab_Roller	100 / 16,041	100 / 53,235
Stack_Bowls_Two	100 / 66,648	100 / 78,245

The uncompressed collection is 121,754,669,760 bytes (about 113.4 GiB). Allow at least 200 GiB of free disk space to download and restore all tasks with temporary archives. Simulation archives are split into parts of at most 1.8 GB. They are consecutive byte ranges, not independently extractable archives.

Download and restore

Download download_hf.py from this repository, install the Hub client, and run:

pip install huggingface_hub
python download_hf.py --output /path/to/sim_real_data

Select one task or domain to reduce download size:

python download_hf.py --output /path/to/data --task Grab_Roller --kind sim

The script refuses incomplete manifests, verifies individual parts and combined archives with SHA256, and restores sim/Task and real/Task. It does not require custom dataset loading code or a token for this public repository. The HF dataset viewer is not configured for these raw archive files. verification_report.json describes the local source verification; dataset_manifest.json records the uploaded files and completion status.

Recording formats and provenance

• Simulation: original 10 Hz JSONL and synchronized front / left_wrist / right_wrist PNG images. Each episode includes metadata, frames, and diagnostics. Only complete recordings with success=true were selected. No lossy video conversion or resampling was applied. Grab_Roller combines 64 original and 36 continuation episodes.
• Real: original LeRobot v3 Parquet, metadata and videos/images at 30 Hz. Count trajectories using episode metadata/indexes, not Parquet or MP4 file counts. Real recordings are the original collected demonstrations; no new simulation-based success audit is claimed.
• Inventories contain file sizes, provenance and real-file hashes. Local machine paths in these reports are provenance only and are not required to use the data.

Interface and limitations

18 dimensions: right arm joints 0-5, right gripper 6, left arm joints 7-12, left gripper 13, base velocity 14-16, lift 17. Left/right refer to the robot itself.

Simulation grippers store total two-finger opening in meters; real grippers store motor radians. The existing V2 mapping is real_gripper_rad = sim_opening_m / 0.10 * (-4.5). Lift mapping is real_height_mm = sim_height_m * 1000. These archives retain original units; they have NOT been transformed. reference/ contains the mapping code and document from an Adjust_Bottle experiment, not proof of full calibration for the other tasks. Arm conventions, cameras, limits, initial poses, and dynamics still require checking.

Stack_Bowls_Two simulation has 61 left-first and 39 right-first trajectories; bowl0 is handled by the right arm and bowl1 by the left arm, with the designated bottom bowl moved first. The 100 real trajectories are right-first. Do not assume identical demonstration ordering across domains.

Simulation success does not certify real deployment safety. No hardware-control permission is granted by this dataset. License terms have not yet been specified by the publisher; do not assume a particular open-source license. Contact the repository owner to clarify reuse permissions.

中文摘要

三个任务各有仿真100条、真机100条，共600条。数据保持原始格式和单位，以无损分卷发布。下载脚本校验SHA256后还原目录。仿真与真机的夹爪/升降单位、相机及示范顺序并未自动统一，使用前请阅读映射参考和上述限制。上传完成以 dataset_manifest.json 的 status=verified_complete 为准。