THULab/siyavash_so101_test

SO101 Test TsFile

Apache TsFile edition of siyavash/so101_test, a LeRobot v2.1 SO101 dataset for grasping a servo box and putting it on the zirlivani. The numeric trajectories are stored in one table-model TsFile.

Source and attribution

• Original author, repository owner, and uploader: siyavash. The source card does not provide a separate personal name.
• License: Apache-2.0.
• The source card does not provide a homepage, paper, or completed BibTeX citation.
• Split: train; 50 episodes; 22,350 frame rows; one task; 30 FPS; 50 source Parquet shards.
• Task 0: Grasp a servo box and put it on the zirlivani.

Data layout

The table is siyavash_so101_test and contains 22,350 rows across 50 TAG devices. The source Parquet shards total 1,164,037 bytes; the TsFile is 393,788 bytes (33.8% of the source Parquet size).

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. The vector column names preserve their source prefixes, with dots changed to underscores. No trajectory row, episode, task, action dimension, or state dimension is removed.

Encoding and compression

• TIME, INT32, and INT64: TS_2DIFF + LZ4
• FLOAT and DOUBLE: GORILLA + LZ4
• BOOLEAN, when present: RLE + LZ4
• TAG values: TsFile table-model device/tag storage

Videos and alignment

The 100 source AV1 videos are not included here. They remain under videos/chunk-000 in the original repository:

• observation.images.laptop: 50 files, 241,192,699 bytes
• observation.images.wrist: 50 files, 128,928,559 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.

Read example

from tsfile import TsFileReader

reader = TsFileReader("data/siyavash_so101_test.tsfile")
with reader.query_table(
    "siyavash_so101_test",
    ["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()