Datasets:

zjt24
/

record-test

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - tsfile
  - timeseries
  - robotics
  - so101
  - format:tsfile
modality:
  - tabular
  - timeseries
pretty_name: Record Test TsFile
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/record_test.tsfile

Record Test TsFile

This dataset is a TsFile conversion of the Hugging Face dataset sattgle/record-test, which was created using LeRobot.

Modalities: Time-series. The original repository also contains front and rear camera video streams. Videos are not included in this repository and remain available at the source.

Source Dataset

Original dataset: sattgle/record-test
License: Apache-2.0
Task category: robotics
Robot type: so101_follower
Task: clean
LeRobot codebase version: v2.1
Sampling rate: 30 fps
Original video features: observation.images.front and observation.images.rear (640 x 360 RGB, AV1, 30 fps)

The source dataset card does not provide a paper, homepage, or detailed task protocol beyond the LeRobot provenance and task metadata above.

Source Consistency Note

The source repository contains inconsistent trailing files. Its meta/info.json declares 43 episodes and 45,113 frames, while the repository tree contains 50 frame Parquet files. Direct inspection found that files 41 and 42 were overwritten with embedded episode indexes 0 and 1, and files 43-49 are outside meta/episodes.jsonl and also duplicate earlier episode indexes.

This conversion therefore uses the contiguous, internally consistent files episode_000000.parquet through episode_000040.parquet: 41 episodes and 44,533 frames. The adjusted upload metadata and meta/info.json document the excluded files. The corresponding aligned source subset has 82 videos; all source videos remain upstream.

Converted Files

TsFile: data/record_test.tsfile
Rows: 44,533
Episodes: 41 (0:41)
Tasks: 1 (task_index = 0)
Table name: record_test
Time precision: milliseconds
Metadata: meta/ is based on the source metadata and filtered to the validated episode range; meta/info.json records the TsFile mapping and source-repository anomalies.

Schema

Column	Role	Type	Notes
`Time`	TIME	INT64	`round(timestamp * 1000)`, in milliseconds; restarts per episode
`episode_index`	TAG	INT64	Source episode identifier
`task_index`	TAG	INT64	Source task identifier
`frame_index`	FIELD	INT64	Source frame index, preserved
`sample_index`	FIELD	INT64	Renamed from source `index`
`action_0` ... `action_5`	FIELD	FLOAT	Flattened from `action[6]`
`observation_state_0` ... `observation_state_5`	FIELD	FLOAT	Flattened from `observation.state[6]`

The action and observation-state element order is: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, and gripper.pos.

episode_index and task_index are TAG columns, so an individual episode can be selected by filtering on episode_index.

Conversion Notes

The 41 validated LeRobot frame Parquet files were merged into one TsFile.
Vector columns were flattened by preserving the source column name, replacing . with _, and appending the element index.
The source timestamp column is dropped because it is redundant with Time / 1000 seconds.
The source index column is renamed to sample_index.
Front and rear videos are omitted from the converted schema and are not uploaded here. Use the original dataset videos: https://huggingface.co/datasets/sattgle/record-test/tree/main/videos

Read Example

from tsfile import TsFileReader

path = "data/record_test.tsfile"
table = "record_test"

with TsFileReader(path) as reader:
    columns = [
        "episode_index",
        "task_index",
        "frame_index",
        "sample_index",
        "action_0",
        "observation_state_0",
    ]
    with reader.query_table(table, columns, batch_size=4096) as result:
        batch = result.read_arrow_batch()
        print(batch)

Citation

The original dataset card does not provide a BibTeX citation.