---
license: apache-2.0
authors:
- Reuben Lim Yaw Hui
task_categories:
- robotics
tags:
- tsfile
- timeseries
- tabular
modality:
- timeseries
- tabular
pretty_name: SO101 Ball in Cup TsFile
configs:
- config_name: default
  data_files:
  - split: train
    path: data/reubenlim_so101_ball_in_cup.tsfile
size_categories:
- 10K<n<100K
---

# SO101 Ball in Cup TsFile

This is an Apache TsFile conversion of
[`ReubenLim/so101_ball_in_cup`](https://huggingface.co/datasets/ReubenLim/so101_ball_in_cup), a LeRobot v2.1 SO101
dataset for picking up a ball and placing it in a cup.

## Source and provenance

- Original dataset: [`ReubenLim/so101_ball_in_cup`](https://huggingface.co/datasets/ReubenLim/so101_ball_in_cup)
- Pinned source revision: [`811ff35e5d324f666941ad289c06de4d13980538`](https://huggingface.co/datasets/ReubenLim/so101_ball_in_cup/tree/811ff35e5d324f666941ad289c06de4d13980538)
- Repository owner, uploader, and commit author: [Reuben Lim Yaw Hui (`ReubenLim`)](https://huggingface.co/ReubenLim)
- License: Apache-2.0
- Paper/citation: not provided by the source repository
- Robot: `so101`; LeRobot codebase: `v2.1`
- Split: `train`; sampling frequency: 30 fps
- Task 0: `Pick up the ball and place it in the cup.`
- Scale: 61 episodes, 57,839 rows, one task, 61 source Parquet shards
- Source data path: `data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet`

The source README still embeds an obsolete 23-episode `info.json` example.
The counts above come from the pinned repository `meta/info.json`, all source
Parquet footers, and the complete repository tree.

## Converted artifact

- TsFile: `data/reubenlim_so101_ball_in_cup.tsfile`
- Table: `reubenlim_so101_ball_in_cup`
- Rows/devices: 57,839 / 61
- Time precision: milliseconds
- Source Parquet size: 2,754,715 bytes
- TsFile size: 780,173 bytes
- TsFile/source size ratio: 0.283214
- `meta/` is preserved, with `meta/info.json` updated for this TsFile artifact.

## Schema

| Column or group | TsFile role | Type | Notes |
|---|---|---|---|
| `Time` | TIME | INT64 | `round(timestamp * 1000)` ms; restarts at zero per episode |
| `episode_index` | TAG | STRING | Original integer value stored as a TsFile device/TAG segment |
| `task_index` | TAG | STRING | Original integer value stored as a TsFile device/TAG segment |
| `frame_index` | FIELD | INT64 | Original per-episode frame index |
| `sample_index` | FIELD | INT64 | Renamed from source `index` |
| `action_0` ... `action_5` | FIELD | FLOAT | Flattened source `action[6]` |
| `observation_state_0` ... `observation_state_5` | FIELD | FLOAT | Flattened source `observation.state[6]` |

Vector prefixes preserve their source names, with `.` replaced by `_`.

## Conversion details

- `Time`: `TS_2DIFF + LZ4`
- FLOAT/DOUBLE: `GORILLA + LZ4`
- INT32/INT64: `TS_2DIFF + LZ4`
- BOOLEAN: `RLE + LZ4` when present; this source has no BOOLEAN field
- TAG values use the TsFile table-model device/tag mechanism.
- All train episodes are merged into one table and sorted by TAG columns and `Time`.
- Source `timestamp` is dropped because it is represented by `Time / 1000` seconds.
- No numeric rows, action/state dimensions, frame indexes, episode indexes, or task indexes are dropped.
- Camera pixels are video assets rather than Parquet columns and are not embedded in TsFile.

## Videos

Videos are not included in this TsFile dataset. The original repository stores
183 frame-aligned AV1 MP4 files: 61 episodes for each of the `front`,
`overhead`, and `side` streams under [`videos/chunk-000/`](https://huggingface.co/datasets/ReubenLim/so101_ball_in_cup/tree/811ff35e5d324f666941ad289c06de4d13980538/videos/chunk-000).
The exact template is
`videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4`.
The source videos total 1,585,929,900 bytes and are 640x480 at 30 fps without
audio. Stream sizes are 573,779,857 bytes (`front`), 483,143,898 bytes
(`overhead`), and 529,006,145 bytes (`side`). Use `episode_index` and
`frame_index` to align TsFile rows with all three views.

## Validation

Local validation read all 57,839 rows with the Apache TsFile
Java reader, checked the physical codecs, TAG schema, per-episode Time ordering,
vector widths, duplicate TAG+Time keys, and the source/staged/TsFile row counts.
The conversion script and validation reports stay local and are not part of the
upload-ready dataset directory.

## Minimal read example

```python
from tsfile import TsFileReader

reader = TsFileReader("data/reubenlim_so101_ball_in_cup.tsfile")
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_0",
    "observation_state_0",
]
with reader.query_table("reubenlim_so101_ball_in_cup", columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())
reader.close()
```

## Citation

The source repository provides no formal citation. Cite the original Hugging
Face dataset and its repository author `Reuben Lim Yaw Hui (ReubenLim)` when
using this conversion.