THULab/leisaac-pick-orange

LeIsaac Pick Orange (TsFile)

This dataset is an Apache TsFile conversion of LightwheelAI/leisaac-pick-orange at source revision fa6e0625d814352b8e6ee1c6d2482194e4da8ed3. The source dataset card states that it was created with LeRobot, and its metadata identifies LeRobot v2.1 data from an so101_follower robot.

Dataset Summary

• Modalities: Time-series
• Split: train
• Episodes: 60
• Frames and TsFile rows: 36,293
• Tasks: 1
• Task: Grab orange and place into plate
• Sampling rate: 30 FPS
• Source frame files: 60 Parquet files
• Converted files: 1 TsFile
• Table name: leisaac_pick_orange_train

Data File

data/leisaac_pick_orange_train.tsfile

TsFile Schema

The table contains one TIME column, two TAG columns, and fourteen FIELD columns.

Column	TsFile role	Type	Source mapping
Time	TIME	INT64	round(timestamp * 1000) milliseconds
episode_index	TAG	STRING	Source episode index
task_index	TAG	STRING	Source task index
frame_index	FIELD	INT64	Retained unchanged
sample_index	FIELD	INT64	Source index, renamed
action_0 ... action_5	FIELD	FLOAT	Six flattened action components
observation_state_0 ... observation_state_5	FIELD	FLOAT	Six flattened robot-state components

The six action and state components retain the source order: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, and gripper.pos.

Conversion Notes

• All 60 episodes are merged into one train table. episode_index and task_index remain available as TAG columns for filtering.
• action and observation.state were flattened into scalar FLOAT fields.
• Source index was renamed to sample_index.
• Source timestamp is not stored separately because it is redundant with Time / 1000 seconds.
• No numeric row was removed, resampled, or interpolated.

Videos

Videos are not downloaded, converted, or uploaded in this repository. The source camera features are observation.images.front and observation.images.wrist. They remain available in the original videos tree. The retained episode, frame, task, and sample indexes provide alignment with the source media.

Python SDK Example

from pathlib import Path

from tsfile import TsFileReader

path = Path("data/leisaac_pick_orange_train.tsfile")
table = "leisaac_pick_orange_train"

reader = TsFileReader(str(path))
try:
    columns = [
        "episode_index",
        "task_index",
        "frame_index",
        "action_0",
        "observation_state_0",
    ]
    with reader.query_table(table, columns, batch_size=1024) as result:
        batch = result.read_arrow_batch()
        if batch is not None:
            print(batch.to_pandas().head())
finally:
    reader.close()

Source and License

• Source: LightwheelAI/leisaac-pick-orange
• Source repository owner: LightwheelAI
• License: Apache-2.0