THULab/Ting_grip_box

Ting Grip Box (TsFile)

This dataset is an Apache TsFile conversion of the Hugging Face dataset ethanCSL/Ting_grip_box. The source dataset was created with LeRobot and contains demonstrations recorded with a Koch follower robot.

Modalities: Time-series. The original repository also contains synchronized camera videos; videos are not included in this converted repository.

Source Dataset

• Original dataset: ethanCSL/Ting_grip_box
• License: apache-2.0
• LeRobot codebase version: v2.1
• Robot type: koch_follower
• Split: train (0:270)
• Scale from meta/info.json: 270 episodes, 114,441 frames, 4 tasks
• Source video count: 540 (270 episodes x 2 camera streams)
• Sampling rate: 30 fps
• Source data 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

The four source tasks are:

• task_index=0: grip the green block and put into the box
• task_index=1: grip the red block and put into the box
• task_index=2: grip the black block and put into the box
• task_index=3: grip the white block and put into the box

Converted File

• TsFile: data/ting_grip_box_train.tsfile
• TsFile table: ting_grip_box_train
• Converted rows: 114,441
• Episodes: 270
• Time precision: milliseconds
• TAG columns: episode_index, task_index
• File size: 2,336,981 bytes

All source episodes in the train split are merged into one TsFile. The TAG columns preserve episode and task identity, so an episode can be selected with episode_index while the task can be selected with task_index.

Schema

Time is computed as round(timestamp * 1000) in milliseconds and restarts in each episode. The source timestamp column is dropped because it is redundant with Time / 1000 seconds. No source rows are dropped.

TAG columns:

• episode_index
• task_index

FIELD columns:

• frame_index
• sample_index (renamed from source index)
• action_0 to action_5 (single-precision FLOAT)
• observation_state_0 to observation_state_5 (single-precision FLOAT)

The six elements in both action and observation.state use the source joint order:

1. shoulder_pan.pos
2. shoulder_lift.pos
3. elbow_flex.pos
4. wrist_flex.pos
5. wrist_roll.pos
6. gripper.pos

Vector names preserve the full source feature name: . is replaced with _ and the element index is appended. For example, observation.state becomes observation_state_0 through observation_state_5.

Video Policy

The source video features observation.images.front and observation.images.top are not converted or uploaded. Each stream contains 480 x 640 RGB AV1 video at 30 fps. Use the original dataset for the synchronized videos: ethanCSL/Ting_grip_box/videos.

Numeric frame rows retain frame_index, episode_index, task_index, and sample_index, preserving their alignment with the original videos.

Metadata

The source meta/ files are mirrored in this repository. meta/info.json is updated so data_path points to data/ting_grip_box_train.tsfile; its tsfile_conversion object records the actual table name, Time mapping, TAG columns, source episode-file count, converted TsFile count, flattened features, renamed and dropped fields, row count, and frame/video alignment. The converted total_videos value is 0; the original count of 540 is preserved as tsfile_conversion.source_video_count.

Validation

The converted file was validated with the project pipeline and read back with the TsFile Python SDK:

• staged Parquet rows: 114,441
• TsFile metadata rows: 114,441
• TsFile query rows: 114,441
• duplicate (episode_index, task_index, Time) rows: 0
• TsFile size: 2,336,981 bytes

Usage

from tsfile import TsFileReader

path = "data/ting_grip_box_train.tsfile"
with TsFileReader(path) as reader:
    schemas = reader.get_all_table_schemas()
    table = schemas["ting_grip_box_train"]
    print([(column.get_column_name(), column.get_category())
           for column in table.get_columns()])