THULab/MsJNeko_so101_prac2

SO-101 Practice Run TsFile

This dataset is a compact Apache TsFile conversion of MsJNeko/so101_prac2, a LeRobot v2.1 SO-101 robot-manipulation practice run.

Modalities: Time-series and tabular. Numeric actions, robot state, frame timing, and episode/task metadata are stored in TsFile. Camera video remains in the source Hugging Face dataset.

Source dataset

• Author/uploader: MsJNeko
• License: Apache-2.0
• Task: Pick up a bottle and put it back down.
• Robot: so101; LeRobot codebase v2.1; sampling rate 30 FPS
• Source page: https://huggingface.co/datasets/MsJNeko/so101_prac2
• Scale: 50 episodes, 24,323 frames, 1 task, 1 source chunk
• Source Parquet: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4
• Video stream: observation.images.phone (50 MP4 files, about 339 MB)

Converted files

• TsFile: data/MsJNeko_so101_prac2.tsfile
• Table: MsJNeko_so101_prac2
• Rows: 24,323
• Time precision: milliseconds
• meta/ is mirrored from the source; meta/info.json documents the TsFile mapping.

Schema and encoding

Time = round(timestamp * 1000) milliseconds, restarting at zero per episode.

TAG columns (TsFile table/device identity): episode_index, task_index.

FIELD columns: frame_index (INT64), sample_index (INT64, renamed from index), action_0 ... action_5 (FLOAT), and observation_state_0 ... observation_state_5 (FLOAT).

The two vector columns are flattened from action[6] and observation.state[6], preserving the source prefixes (. becomes _). The redundant source timestamp field is dropped after Time synthesis; no rows or numeric dimensions are dropped. The compact writer uses GORILLA + LZ4 for FLOAT/DOUBLE and TS_2DIFF + LZ4 for INT32/INT64 and Time.

Videos

Videos are not uploaded with this conversion. They remain in the original videos/ tree under videos/chunk-000/observation.images.phone/. Numeric rows align to video frames through episode_index and frame_index.

Validation and usage

The local validation report confirms a non-empty TsFile and row-count equality with the staged Parquet: 24,323 rows.

from tsfile import TsFileReader
reader = TsFileReader("data/MsJNeko_so101_prac2.tsfile")
with reader.query_table("MsJNeko_so101_prac2", ["episode_index", "task_index", "frame_index", "action_0"], batch_size=65536) as result:
    print(result.read_arrow_batch().to_pandas().head())