---
language:
  - en
pretty_name: "Vacuame/train4 (LeRobot SO-100) - TsFile"
tags:
- time-series
- tsfile
- robotics
- lerobot
- so100
- manipulation
- timeseries
modality: timeseries
task_categories:
  - robotics
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/vacuame_train4.tsfile
---

# Vacuame/train4 (LeRobot SO-100) - TsFile

This dataset converts the numeric frame time-series from
[`Vacuame/train4`](https://huggingface.co/datasets/Vacuame/train4) to Apache
TsFile format.

Modalities: Time-series. The original camera videos are not included in this
converted repository; they remain in the source Hugging Face dataset.

## Source Dataset

- Original dataset: [`Vacuame/train4`](https://huggingface.co/datasets/Vacuame/train4)
- LeRobot version: v2.0
- Robot type: `so100`
- Scale: 2 episodes, 119 frames, 1 task (`try`)
- Sampling rate: 30 fps
- Source frame data: `data/chunk-000/episode_000000.parquet` and
  `data/chunk-000/episode_000001.parquet`
- Source videos: 2 camera streams, `observation.images.laptop` and
  `observation.images.phone`, available in the source dataset
  [`videos/`](https://huggingface.co/datasets/Vacuame/train4/tree/main/videos)
  tree
- License: the source dataset does not declare an explicit license

## Converted Layout

```text
README.md
data/vacuame_train4.tsfile
meta/info.json
meta/tasks.jsonl
meta/episodes.jsonl
meta/stats.json
```

The converted TsFile contains all 119 numeric frame rows in one table named
`vacuame_train4`.

## Schema

| Role | Column(s) | Type | Notes |
|------|-----------|------|-------|
| Time | `Time` | INT64 ms | `round(timestamp * 1000)` |
| TAG | `episode_index`, `task_index` | INT64 | Source columns used as TsFile device tags |
| FIELD | `frame_index`, `sample_index` | INT64 | `index` is renamed to `sample_index` |
| FIELD | `action_0` ... `action_5` | FLOAT | Flattened from `action[6]` |
| FIELD | `observation_state_0` ... `observation_state_5` | FLOAT | Flattened from `observation.state[6]` |

## Conversion Notes

- The generic `lerobot` converter was used.
- `timestamp` is dropped because it is redundant with `Time / 1000` seconds.
- `frame_index` is kept so rows can be aligned back to source video frames.
- Vector columns preserve the source feature name with `.` replaced by `_`:
  `observation.state` becomes `observation_state_0..observation_state_5`, and
  `action` becomes `action_0..action_5`.
- Video features are omitted from the converted repository. Use the original
  dataset's `videos/` tree for camera data.
- Source `meta/` files are mirrored; `meta/info.json` is updated with a
  `tsfile_conversion` object documenting the converted path, schema roles,
  flattened features, dropped features, omitted videos, and row count.

## Usage

```python
from tsfile import TsFileReader

reader = TsFileReader("data/vacuame_train4.tsfile")
schemas = reader.get_all_table_schemas()
table_name = next(iter(schemas))

columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "observation_state_0",
    "action_0",
]

with reader.query_table(table_name, columns, batch_size=65536) as result:
    while (batch := result.read_arrow_batch()) is not None:
        df = batch.to_pandas()
        # Process rows here.

reader.close()
```

## Citation

```bibtex
@misc{vacuame_train4,
  title     = {train4 (LeRobot SO-100)},
  author    = {Vacuame},
  url       = {https://huggingface.co/datasets/Vacuame/train4},
  publisher = {Hugging Face}
}
```