THULab/Itylso100_test2

Ityl SO100 Test2 TsFile

This dataset is an Apache TsFile conversion of Ityl/so100_test2, a LeRobot SO100 manipulation dataset. The source dataset was created using LeRobot.

Modalities: Time-series. This converted repository contains numeric robot observations, actions, timing information, episode/task tags, and source metadata. Camera videos remain in the original Hugging Face dataset.

Source Dataset

• Source: Ityl/so100_test2
• Publisher: Ityl
• License: Apache-2.0
• Robot type: so100
• LeRobot codebase version: v2.0
• Split: 30 train episodes (0:30)
• Scale: 11,447 frames, 30 episodes, 1 task, 30 fps
• Task: First serious test
• Source frame layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video layout: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

Original Videos

The 60 source videos are not copied into this TsFile repository. They remain available in the original dataset under videos/. The repository contains one chunk-000 directory and two camera streams:

• Side camera: videos/chunk-000/observation.images.realsense_side
• Top camera: videos/chunk-000/observation.images.realsense_top

Each stream stores one MP4 per episode using the path pattern:

videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

Use episode_index and frame_index from the TsFile to align numeric rows with the corresponding original video and frame.

Converted Files

• TsFile: data/tv_dataset_5.tsfile
• Table: tv_dataset_5
• Rows: 11,447
• Episodes: 30
• Tasks: 1
• Source Parquet files merged: 30
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the source, with meta/info.json updated for the TsFile artifact and original video links.

Schema

Time is synthesized as round(timestamp * 1000) milliseconds and restarts from zero in each episode.

TAG columns:

• episode_index
• task_index

Scalar FIELD columns:

• frame_index
• sample_index, renamed from source column index

Flattened FLOAT FIELD groups:

• action[6] -> action_0 ... action_5
• observation.state[6] -> observation_state_0 ... observation_state_5

The six dimensions, in source order, are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper.

Conversion Notes

• All 30 source episode Parquet files are merged into one table-model TsFile. Filter by episode_index and task_index to select an episode or task.
• Vector columns are flattened to scalar TsFile fields. Source column prefixes are preserved, with . replaced by _.
• Source column timestamp is dropped after Time synthesis because it is redundant with Time / 1000 seconds.
• Source column index is renamed to sample_index.
• No numeric rows, episodes, tasks, state dimensions, or action dimensions are dropped.
• Camera pixels are not stored in TsFile. The original videos are linked above.

Validation

The converted TsFile was read back with the Apache TsFile Java SDK. Its table schema contains 16 non-time columns (2 TAG and 14 FIELD columns), and query readback matched the staged Parquet at 11,447 rows.

Usage

from tsfile import TsFileReader

path = "data/tv_dataset_5.tsfile"
reader = TsFileReader(path)

schemas = reader.get_all_table_schemas()
table_name = "tv_dataset_5"
columns = [
    column.get_column_name()
    for column in schemas[table_name].get_columns()
    if column.get_column_name() != "Time"
]

with reader.query_table(table_name, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()