Datasetpublic

THULab/Dangvi_soarm100_data2

SOARM100 Data2 TsFile Apache TsFile edition of Dangvi/soarm100_data2, a LeRobot v2.1 SO100 dataset with demonstrations for lifting orange or green 2x4 blocks above an 8x8 blue block. Source and attribution Original author, repository owner, and uploader: DANBI KIM (Dangvi). License: Apache-2.0. The source card provides no homepage, paper, or completed BibTeX citation. Split: train; 75 episodes; 33,440 frame rows; 2 tasks; 30 FPS; 75 source Parquet shards. The… See the full description on the dataset page: https://huggingface.co/datasets/THULab/Dangvi_soarm100_data2.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes37downloads

Dataset Card

SOARM100 Data2 TsFile

Apache TsFile edition of Dangvi/soarm100_data2, a LeRobot v2.1 SO100 dataset with demonstrations for lifting orange or green 2x4 blocks above an 8x8 blue block.

Source and attribution

—Original author, repository owner, and uploader: DANBI KIM (Dangvi).
—License: Apache-2.0.
—The source card provides no homepage, paper, or completed BibTeX citation.
—Split: train; 75 episodes; 33,440 frame rows; 2 tasks; 30 FPS; 75 source Parquet shards.
—The source card's embedded metadata example says 25 episodes, but the current meta/info.json, episode metadata, and repository file tree consistently contain 75 episodes.
—Task 0: Pick up the 2x4 orange block and go up the 8x8 blue block.
—Task 1: Pick up the 2x4 green block and go up the 8x8 blue block.

Data layout

The table is dangvi_soarm100_data2 and contains 33,440 rows across 75 episode/task devices.

Column	TsFile role	Type	Meaning
`Time`	TIME	INT64 milliseconds	`round(timestamp * 1000)`, restarting at zero per episode
`episode_index`	TAG	STRING from source INT64	Source episode identity
`task_index`	TAG	STRING from source INT64	Source task identity
`frame_index`	FIELD	INT64	Frame position within the episode
`sample_index`	FIELD	INT64	Source `index`, renamed for clarity
`action_0` ... `action_5`	FIELD	FLOAT	Flattened `action[6]`
`observation_state_0` ... `observation_state_5`	FIELD	FLOAT	Flattened `observation.state[6]`

timestamp is not retained as a separate FIELD because it is represented by Time / 1000 seconds. Vector names preserve their source prefixes, with dots changed to underscores. No trajectory row, episode, task, action dimension, or state dimension is removed.

Videos and alignment

The 150 source AV1 videos remain in the original repository under `videos/chunk-000`:

—`observation.images.webcam`: 75 MP4 files, 256,529,229 bytes
—`observation.images.wristcam`: 75 MP4 files, 133,437,099 bytes

The source template is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Use episode_index and frame_index to align each numeric row with both 30 FPS video streams. The videos are not included in this TsFile dataset.

Read example

python

from tsfile import TsFileReader

reader = TsFileReader("data/dangvi_soarm100_data2.tsfile")
with reader.query_table(
    "dangvi_soarm100_data2",
    ["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"],
    batch_size=1024,
) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())
reader.close()