THULab/so101_grab_red_cube

so101grabred_cube

This dataset converts the numeric time-series from `Tank-123/smolvla_test_0129` to Apache TsFile format.

Modalities: Time-series. The original Hugging Face dataset also includes video streams; videos are not copied into this TsFile repository.

Dataset Description

The source dataset was created using LeRobot. It contains SO-101 follower robot demonstrations for the task Grab the red cube..

• —Original dataset: `Tank-123/smolvla_test_0129`
• —Robot type: so101_follower
• —Task: Grab the red cube.
• —Episodes: 72
• —Frames / TsFile rows: 26,610
• —Sampling rate: 30 fps
• —Tasks: 1
• —Source split: train (0:72)
• —License: apache-2.0
• —Source data path: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• —Source video path: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

Converted Data

The converted repository contains one TsFile:

data/so101_grab_red_cube.tsfile

Conversion details:

• —All 72 source episodes are stored in one TsFile table named so101_grab_red_cube.
• —episode_index and task_index are TsFile TAG columns. Query a single episode with WHERE episode_index=N.
• —Time = round(timestamp * 1000) in milliseconds, restarting per episode.
• —The source timestamp column is not retained because it equals Time / 1000 seconds after conversion.
• —frame_index is kept. Source index is renamed to sample_index.
• —Vector columns preserve the source column name when flattened: . is replaced with _, and the element index is appended.
• —Flattened vector values are stored as single-precision FLOAT fields.
• —Aside from the redundant timestamp column, no numeric time-series columns or rows are intentionally dropped.

Flattened vector features:

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

Schema

• —Time: TsFile time column, integer milliseconds.
• —TAG columns: episode_index, task_index.
• —FIELD columns: frame_index, sample_index, action_0..action_5, observation_state_0..observation_state_5.

Videos

Videos are not included in this TsFile repository. They remain in the original dataset under `videos/`.

The original video features are:

• —observation.images.laptop
• —observation.images.phone

Use episode_index, frame_index, and Time to align rows in the TsFile with the original video frames.

Metadata

The source meta/ files are mirrored into this repository. The uploaded meta/info.json rewrites data_path to data/so101_grab_red_cube.tsfile and adds a tsfile_conversion object documenting the source dataset, row count, time mapping, TAG columns, flattened features, dropped timestamp column, and video policy.

Usage

Read the TsFile with the Apache TsFile Java or Python SDK:

tsfile_path = "data/so101_grab_red_cube.tsfile"
# Example SQL idea:
# SELECT * FROM so101_grab_red_cube WHERE episode_index = 0