Datasets:

azithromycin
/

Selinaliu1030_egg_0806

license: apache-2.0
authors:
  - Selinaliu1030
task_categories:
  - robotics
tags:
  - tsfile
  - timeseries
  - tabular
modality:
  - timeseries
  - tabular
pretty_name: Egg 0806 TsFile
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/selinaliu1030_egg_0806.tsfile
size_categories:
  - 10K<n<100K

Egg 0806 TsFile

This dataset is an Apache TsFile conversion of Selinaliu1030/egg_0806, a LeRobot v2.1 SO101 robot-manipulation dataset. It contains numeric trajectories, timing, episode/task tags, and source metadata. Videos remain in the original Hugging Face repository.

Source Dataset and Attribution

Original dataset: Selinaliu1030/egg_0806
Original author/publisher: Selinaliu1030
Authorship note: the source card names no separate authors; this README attributes the dataset to its Hugging Face repository publisher.
License: Apache-2.0
Task: "Grasp the egg and put it in the red bin."
Robot: so101; LeRobot version: v2.1
Split: train; sampling rate: 30 fps
Scale: 90 episodes, 49,957 frames, 1 task
Source shards: 90 Parquet files under data/chunk-000/episode_{episode_index:06d}.parquet
Paper/citation: the source card provides neither a paper nor a completed citation.

Schema

TsFile column(s)	Type	Role	Source
`Time`	INT64	TIME	`round(timestamp * 1000)` ms
`episode_index`	STRING	TAG/device	source `episode_index`
`task_index`	STRING	TAG/device	source `task_index`
`frame_index`	INT64	FIELD	source `frame_index`
`sample_index`	INT64	FIELD	source `index`
`action_0` ... `action_5`	FLOAT	FIELD	flattened `action[6]`
`observation_state_0` ... `observation_state_5`	FLOAT	FIELD	flattened `observation.state[6]`

The six action/state dimensions are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper. Dots in source vector names are replaced by underscores. Time restarts at zero for every episode.

Conversion Details

The 90 train episode shards are merged into one table-model TsFile.
Source timestamp is dropped after Time synthesis because it is exactly represented by Time / 1000 seconds.
Source index is retained as sample_index; frame_index is unchanged.
No numeric row, episode, task, state dimension, or action dimension is dropped.
FLOAT/DOUBLE uses GORILLA + LZ4; INT32/INT64 and Time use TS_2DIFF + LZ4; BOOLEAN uses RLE + LZ4; TAG values use the TsFile table/device mechanism.

Videos

Videos are not included in this TsFile repository. The original dataset has 180 frame-aligned AV1 MP4 files (301,079,898 bytes, about 287.1 MiB), 640x480, 30 fps, no audio, in two streams:

Each stream contains 90 files matching videos/chunk-000/{video_key}/episode_{episode_index:06d}.mp4. episode_index, frame_index, and meta/episodes.jsonl preserve alignment.

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/selinaliu1030_egg_0806.tsfile")
with reader.query_table(
    "selinaliu1030_egg_0806",
    ["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()