azithromycin/lerobot_svla_so100_sorting

SVLA SO100 Sorting TsFile

This is an Apache TsFile conversion of lerobot/svla_so100_sorting, a LeRobot v3.0 SO100 sorting dataset used with the SmolVLA work.

Source dataset and attribution

• Source: lerobot/svla_so100_sorting
• License: Apache-2.0; robot: SO100; split: train; sampling rate: 30 fps
• Task: Put the red cube in the right box and the blue cube in the left box.
• Repository owner: LeRobot
• Repository contributors recorded in the source history include Dana Aubakirova, Michel Aractingi, Caroline Pascal, and Andito.
• Related paper: SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics.

The current source card does not retain a separate dataset-author list. Its history links the dataset to the SmolVLA paper, whose authors are Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, Simon Alibert, Matthieu Cord, Thomas Wolf, and Remi Cadene.

Converted scale

The single source Parquet shard is merged into one table-model TsFile. The source meta directory is retained as metadata only; no old data Parquet is placed under meta.

TsFile schema

Time = round(timestamp * 1000) milliseconds and restarts at zero for every episode.

Role	Columns
Time	Time (TIMESTAMP)
TAG/device	episode_index, task_index
Scalar FIELD	frame_index, sample_index (renamed from index)
FLOAT FIELD	action[6] to action_0 through action_5
FLOAT FIELD	observation.state[6] to observation_state_0 through observation_state_5

The six vector dimensions are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper. The source timestamp field is dropped after Time synthesis because it equals Time / 1000. No rows, frame indexes, action dimensions, or state dimensions are dropped.

Encoding and compression

• FLOAT and DOUBLE: GORILLA + LZ4
• INT32 and INT64: TS_2DIFF + LZ4
• Time: TS_2DIFF + LZ4
• BOOLEAN: RLE + LZ4 (no BOOLEAN source fields occur)
• TAG: TsFile table/device TAG mechanism

Videos and alignment

Videos are not included in this TsFile dataset. They remain at the pinned source videos directory:

• videos/observation.images.top/chunk-000/file-000.mp4 (524,012,810 bytes)
• videos/observation.images.top/chunk-000/file-001.mp4 (64,082,916 bytes)
• videos/observation.images.wrist/chunk-000/file-000.mp4 (242,711,573 bytes)

The three AV1 MP4 files total 830,807,299 bytes. Both streams are 640 x 480 at 30 fps without audio. episode_index, frame_index, and the preserved meta/episodes/chunk-000/file-000.parquet offsets maintain numeric/video alignment.

Citation

@misc{shukor2025smolvlavisionlanguageactionmodelaffordable,
  title={SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics},
  author={Mustafa Shukor and Dana Aubakirova and Francesco Capuano and
          Pepijn Kooijmans and Steven Palma and Adil Zouitine and
          Michel Aractingi and Caroline Pascal and Martino Russi and
          Andres Marafioti and Simon Alibert and Matthieu Cord and
          Thomas Wolf and Remi Cadene},
  year={2025},
  eprint={2506.01844},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2506.01844}
}

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/lerobot_svla_so100_sorting.tsfile")
with reader.query_table(
        "lerobot_svla_so100_sorting",
        ["episode_index", "task_index", "frame_index", "sample_index",
         "action_0", "observation_state_0"],
        batch_size=65536) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()