THULab/HarrisonLee24_so100_test

SO100 Test TsFile

This dataset is an Apache TsFile conversion of HarrisonLee24/so100_test, a LeRobot v2.1 SO100 robot-manipulation dataset for grasping a LEGO block and placing it in a bin. It contains numeric trajectories, frame timing, and episode/task tags. Videos remain in the original Hugging Face repository.

Source Dataset and Attribution

• Original dataset: HarrisonLee24/so100_test
• Original author, repository owner, uploader, and commit author: Harrison Lee (HarrisonLee24)
• License: Apache-2.0
• Task: "Grasp a lego block and put it in the bin."
• Robot: so100; LeRobot version: v2.1
• Split: train; sampling rate: 30 fps
• Scale: 92 episodes, 80,091 frame rows, 1 task
• Source shards: 92 Parquet files totaling 3,926,293 bytes
• Episode lengths: 868 to 873 frames
• Source frame layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Paper, external homepage, and completed citation: not provided by the source card.

The source README embeds an older two-episode copy of meta/info.json. The current meta/info.json, 92 source Parquet files, meta/episodes.jsonl, and meta/episodes_stats.jsonl consistently describe the 92-episode dataset and are used here.

Data Layout

• TsFile: data/harrisonlee24_so100_test.tsfile (1,372,574 bytes)
• Table: harrisonlee24_so100_test
• Rows: 80,091; episodes/devices: 92; tasks: 1
• Time precision: milliseconds
• Source metadata is retained under meta/; meta/info.json points to the TsFile data path and records the field mapping and source-video alignment.

Schema and Mapping

Time = round(timestamp * 1000) milliseconds. Time starts at zero and is strictly increasing inside every episode. The source timestamp is dropped afterward because it is represented by Time / 1000 seconds at the selected precision.

TsFile column	Role	Type	Source mapping
Time	TIME	TIMESTAMP	round(timestamp * 1000) ms
episode_index	TAG	STRING	Original INT64 episode index
task_index	TAG	STRING	Original INT64 task index
frame_index	FIELD	INT64	Preserved
sample_index	FIELD	INT64	Renamed from index
action_0 ... action_5	FIELD	FLOAT	Flattened from action[6]
observation_state_0 ... observation_state_5	FIELD	FLOAT	Flattened from observation.state[6]

The six vector dimensions are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper. Dots in source names are replaced by underscores. No numeric row, episode, task, state dimension, or action dimension is dropped.

Videos

Videos are not included in this TsFile dataset. Current meta/info.json references 184 frame-aligned H.264 MP4 files (565,481,091 bytes), 640x480 at 30 fps without audio, in two streams with 92 episode files each:

• observation.images.global_cam - 92 files, 363,719,494 bytes
• observation.images.hand_eye_cam - 92 files, 201,761,597 bytes

The source layout is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Use episode_index, frame_index, and meta/episodes.jsonl to align numeric rows with the original video frames.

The repository also retains 65 files in five older video directories that are not referenced by the current meta/info.json: action_cam (20), laptop (10), osmo (5), phone (10), and pocket_cam (20). They remain accessible under the source videos/ tree, but this card does not claim frame alignment for those historical streams.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/harrisonlee24_so100_test.tsfile")
with reader.query_table(
    "harrisonlee24_so100_test",
    ["episode_index", "task_index", "frame_index", "sample_index",
     "action_0", "observation_state_0"],
    batch_size=65536,
) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())
reader.close()

Citation

The source card provides no paper or completed citation. Cite the original Hugging Face dataset and Harrison Lee (HarrisonLee24) when using this data.