THULab/mtitg_so100_01

SO100 01 TsFile

This dataset is an Apache TsFile conversion of mtitg/so100_01, a LeRobot v2.1 SO100 robot dataset for the task "One hot vector test." It contains numeric trajectories, millisecond timing, and episode/task tags. Videos remain in the original repository.

Source Dataset and Attribution

• Original dataset: mtitg/so100_01
• Original author, repository owner, uploader, and sole contributor: Motoi Tanigaki (mtitg)
• License: Apache-2.0
• Task: "One hot vector test"
• Robot: so100_follower; LeRobot version: v2.1
• Split: train; sampling rate: 30 fps
• Scale: 40 episodes, 17,000 frames, 1 task
• Source shards: 40 per-episode Parquet files totaling 968,400 bytes
• 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.

TsFile Contents

• TsFile: data/mtitg_so100_01.tsfile (451,183 bytes)
• Table: mtitg_so100_01
• Rows: 17,000; episodes/devices: 40; tasks: 1
• Time precision: milliseconds
• TsFile/source-Parquet size ratio: 0.466
• meta/ is preserved from the source, with only meta/info.json rewritten to describe the TsFile schema 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_6	FIELD	FLOAT	Flattened from observation.state[7]

The first six state/action dimensions are shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, and gripper.pos. observation_state_6 preserves the source's additional binary state dimension. Dots in source names are replaced by underscores. No numeric row, episode, task, state dimension, or action dimension is dropped.

Encodings and Compression

• FLOAT/DOUBLE: GORILLA + LZ4
• INT32/INT64: TS_2DIFF + LZ4
• Time: TS_2DIFF + LZ4
• BOOLEAN: RLE + LZ4 (the source contains no BOOLEAN field)
• TAG: TsFile table/device TAG storage

The physical table schema, every field codec, TAG roles, and all 17,000 rows were read back with the Apache TsFile Java API.

Videos

Videos are not included in this TsFile repository. The source contains 80 frame-aligned AV1 MP4 files (272,305,407 bytes), 640x480 at 30 fps with no audio, in two streams with 40 episode files each:

• observation.images.above
• observation.images.wrist

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.

Integrity Checks

Source Parquet, staged Parquet, and complete Java TsFile readback all contain 17,000 rows. Scalar values, all flattened vector values, TAGs, episode indexes, Time mapping and monotonicity, physical codecs, size, and SHA-256 were checked locally. Conversion scripts and local check reports are not part of the dataset directory.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/mtitg_so100_01.tsfile")
with reader.query_table(
    "mtitg_so100_01",
    ["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 Motoi Tanigaki (mtitg) when using this conversion.