THULab/lerobot_stanford_hydra_dataset

Stanford HYDRA Dataset TsFile

Apache TsFile representation of lerobot/stanford_hydra_dataset. It stores numeric robot trajectories as a table-model time series. The original videos remain at the source dataset.

Source Dataset

• Research project: https://sites.google.com/view/hydra-il-2023
• Original research authors: Suneel Belkhale, Yuchen Cui, and Dorsa Sadigh
• Hugging Face repository publisher: LeRobot
• License: MIT
• Paper: HYDRA: Hybrid Robot Actions for Imitation Learning (https://arxiv.org/abs/2306.17237)
• DOI: https://doi.org/10.48550/arXiv.2306.17237
• Project robot: Franka Emika Panda
• LeRobot metadata robot type: unknown
• LeRobot codebase version: v3.0
• Train split: 570 episodes, 358,234 frames, 3 tasks at 10 fps
• Source frame layout: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet (1 Parquet shard)

HYDRA uses hybrid sparse waypoint and dense low-level actions for imitation learning. The trajectories cover coffee making, dish placement, and toast making.

Tasks

task_index	Task	Episodes	Frames
0	make a cup of coffee with the keurig machine	250	144,484
1	palce dishes in the dish rack	110	60,847
2	make a piece of toast with the oven	210	152,903

Data Layout

The train split is one table-model TsFile at data/stanford_hydra_dataset.tsfile. meta/ preserves source LeRobot metadata, and meta/info.json changes data_path to the TsFile path.

Role	Columns	Type
TIME	Time	INT64 milliseconds
TAG	episode_index, task_index	TsFile table device/tag metadata
FIELD	frame_index, sample_index	INT64
FIELD	next_reward	FLOAT
FIELD	next_done	BOOLEAN
FIELD	observation_state_0 through observation_state_7	FLOAT
FIELD	action_0 through action_6	FLOAT

Time = round(timestamp * 1000) in milliseconds and restarts at zero within every episode. frame_index is preserved. timestamp is omitted because Time / 1000 represents it exactly; no source rows, task labels, state dimensions, action dimensions, reward values, or done values are otherwise omitted. index is retained as sample_index; dots in source names are replaced with underscores.

Storage Encoding

• Time: TS_2DIFF + LZ4
• FLOAT fields: GORILLA + ZSTD
• INT64 fields: TS_2DIFF + ZSTD
• BOOLEAN fields: RLE + LZ4
• TAG fields: TsFile device/tag storage, PLAIN + LZ4

Videos

The original repository has two 240x320 AV1 streams at 10 fps:

• videos/observation.images.image/ (4 MP4 shards): https://huggingface.co/datasets/lerobot/stanford_hydra_dataset/tree/main/videos/observation.images.image
• videos/observation.images.wrist_image/ (3 MP4 shards): https://huggingface.co/datasets/lerobot/stanford_hydra_dataset/tree/main/videos/observation.images.wrist_image

Video layout: videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4. Video files are not included. episode_index, frame_index, and meta/episodes retain frame alignment with the source video streams.

Citation

@inproceedings{belkhale2023hydra,
  title={HYDRA: Hybrid Robot Actions for Imitation Learning},
  author={Belkhale, Suneel and Cui, Yuchen and Sadigh, Dorsa},
  booktitle={Proceedings of the 7th Conference on Robot Learning (CoRL)},
  year={2023}
}

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/stanford_hydra_dataset.tsfile")
with reader.query_table(
    "stanford_hydra_dataset",
    ["episode_index", "task_index", "frame_index", "next_done", "action_0"],
    batch_size=65536,
) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()