THULab/task1dataset_all_trimmed

Task 1 Dataset All Trimmed (TsFile)

This is a time-series conversion of ETHrobotlearning/task1dataset_all_trimmed, a LeRobot v3.0 dataset containing banana-placement demonstrations from a so_follower robot.

• Modalities: Time-series
• Robot: so_follower
• Tasks: place the banana in the blue, red, or green colored bowl
• Source license: not declared

Dataset scale

Split	Episodes	Frames / TsFile rows	Sampling rate	Tasks	Source Parquets	Final TsFiles
train	120	19,095	10 fps	3	20	1

Each task has 40 episodes. The blue-bowl task contains 6,326 frames, the red-bowl task 6,384 frames, and the green-bowl task 6,385 frames.

All numeric frame data is merged into exactly one file, data/task1dataset_all_trimmed.tsfile, containing the table task1dataset_all_trimmed.

TsFile schema

Column	TsFile role	Type	Description
Time	TIME	INT64	Milliseconds, computed as Time = round(timestamp * 1000); time restarts within each episode.
episode_index	TAG	INT64	Original episode identifier (0 through 119).
task_index	TAG	INT64	Original task identifier: 0 blue bowl, 1 red bowl, and 2 green bowl.
frame_index	FIELD	INT64	Original frame position within the episode.
sample_index	FIELD	INT64	Original global index, renamed to avoid ambiguity.
observation_state_0 ... observation_state_5	FIELD	FLOAT (float32)	Complete six-element source observation.state vector: shoulder pan, shoulder lift, elbow flex, wrist flex, wrist roll, and gripper position.
action_0 ... action_5	FIELD	FLOAT (float32)	Complete six-element source action vector with the same actuator ordering.

The source timestamp column is dropped after generating Time because it is the redundant seconds representation of the same time coordinate (Time / 1000). No rows or other numeric source columns are dropped. Both vectors are flattened without truncation, index is renamed, and the episode, task, and frame identifiers remain available.

Videos and metadata

The source contains one 480 x 640 RGB AV1 stream, observation.images.front, stored in 15 MP4 files. Video files are intentionally omitted from this time-series repository and remain in the original dataset's videos/ tree.

Numeric rows retain episode_index, frame_index, and the 10 fps time coordinate, so they remain aligned with the corresponding source video frames. The source meta/ files are mirrored, except that meta/info.json is rewritten to describe the single TsFile path, TIME/TAG/FIELD schema, flattened and renamed features, omitted video feature, and source-video alignment. No videos/ directory is included here.

Reading the data

Install the Apache TsFile Python package, then query the actual table. query_table returns the time column automatically in addition to the requested TAG and FIELD columns.

from tsfile import TsFileReader

reader = TsFileReader("data/task1dataset_all_trimmed.tsfile")
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "observation_state_0",
    "action_0",
]

with reader.query_table(
    "task1dataset_all_trimmed", columns, batch_size=65536
) as result:
    batch = result.read_arrow_batch()
    if batch is not None:
        print(batch.to_pandas().head())

Source and citation

The source dataset is published under the ETHrobotlearning Hugging Face organization. Its repository does not include a dataset card, license declaration, paper, or BibTeX citation; consult the source repository for future metadata updates.