112-EDIE-candy_from_drawer (TsFile)

This dataset is an Apache TsFile conversion of LeRobot-worldwide-hackathon/112-EDIE-candy_from_drawer.

Modalities: Time-series.

The source is a LeRobot robot-manipulation dataset recorded on an SO-100 (so100) 6-DoF arm for the task: "Take a blue candy from the drawer and give it to me." The converted TsFile stores the numeric robot state, action, frame, time, and task data. Camera videos remain in the original Hugging Face dataset and are not uploaded here.

Dataset Scale

• Converted file: data/edie_candy_from_drawer.tsfile
• Table name: edie_candy_from_drawer
• Episodes converted: 120
• Frame rows converted: 69,146
• Task count: 1
• Sampling rate: 30 fps
• Time range per episode: starts at 0 ms and advances by about 33 ms per frame
• Robot type: so100
• Source Parquet files converted: 120 files under data/chunk-000/

Note: the downloaded source data/ directory contains 120 episode Parquet files, while the source meta/info.json declares 50 episodes and 28,503 frames. This conversion keeps all 120 episodes present in the source data/ directory. No rows were dropped.

TsFile Schema

Time is the TsFile time axis. It is derived from the source timestamp column as:

Time = round(timestamp * 1000)

The source timestamp column is not stored as a FIELD because it is redundant with Time / 1000. The source frame_index column is retained.

TAG columns:

• episode_index: episode id, ranging from 0 to 119
• task_index: task id, always 0 in this dataset

FIELD columns:

• frame_index: frame number within the episode
• sample_index: source global index column
• action_0 through action_5: flattened source action[6]
• observation_state_0 through observation_state_5: flattened source observation.state[6]

Vector columns preserve the source column name, replacing . with _ and appending the element index. Values are stored as single-precision FLOAT.

Conversion Notes

• Converter: generic scripts/converters/lerobot.py
• Conversion mode: merged LeRobot table
• Time precision: milliseconds
• TAG columns: episode_index, task_index
• Output granularity: all episodes are stored in one TsFile using the TsFile table model
• Videos are not included. The original video streams are available in the source dataset under videos/.
• Source meta/ files are mirrored under meta/; meta/info.json is updated with a tsfile_conversion object describing this conversion.

Validation completed before upload:

• TsFile exists and is non-empty.
• TsFile metadata row count matches staged Parquet row count: 69,146.
• Python SDK query readback row count matches staged Parquet row count: 69,146.
• Distinct episode_index values: 120, from 0 through 119.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/edie_candy_from_drawer.tsfile")
table_name = next(iter(reader.get_all_table_schemas()))

columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "action_0",
    "observation_state_0",
]

with reader.query_table(table_name, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    if batch is not None:
        df = batch.to_pandas()
        print(df.head())

Source

• Original dataset: https://huggingface.co/datasets/LeRobot-worldwide-hackathon/112-EDIE-candy_from_drawer
• Publisher: LeRobot-worldwide-hackathon
• License: not declared in the source dataset metadata at conversion time