README.md

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- TsFile
- time-series
- robotics
- koch
- modality:timeseries
- format:tsfile
configs:
- config_name: default
  data_files:
  - split: train
    path: data/jackvial_koch_test_chunk_v2.tsfile
  - split: train_legacy
    path: data/jackvial_koch_test_train_legacy.tsfile
---

# koch_test

This dataset converts all non-video Parquet time-series data found in
[`jackvial/koch_test`](https://huggingface.co/datasets/jackvial/koch_test) to
[Apache TsFile](https://tsfile.apache.org/) format.

Modalities: Time-series. The original Hugging Face dataset also includes video
streams; videos are not copied into this TsFile repository.

## Dataset Description

The source dataset was created using
[LeRobot](https://github.com/huggingface/lerobot). The current LeRobot metadata
describes a Koch robot dataset with the task `test_description`.

- **Original dataset:** [`jackvial/koch_test`](https://huggingface.co/datasets/jackvial/koch_test)
- **Robot type:** `koch`
- **Task:** `test_description`
- **Sampling rate:** 30 fps
- **License:** apache-2.0
- **Current LeRobot metadata:** v2.0, 2 episodes, 1,192 frames, 1 task
- **Legacy Parquet layout also present:** 3 episodes, 880 rows
- **Source video path:** `videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4`

## Source Data Audit

The source repository contains two distinct non-video Parquet layouts. They are
not duplicates: their row counts, episode counts, and schemas differ. Both are
converted.

| Source layout | Source files | Episodes | Rows | Converted file |
|---|---:|---:|---:|---|
| LeRobot v2 chunk layout | `data/chunk-000/episode_000000.parquet`, `data/chunk-000/episode_000001.parquet` | 2 | 1,192 | `data/jackvial_koch_test_chunk_v2.tsfile` |
| Legacy train layout | `data/train-00000-of-00001.parquet` | 3 | 880 | `data/jackvial_koch_test_train_legacy.tsfile` |

Video files are not converted. Source metadata directories `meta/` and
`meta_data/` are mirrored for traceability.

## Converted Data

### `jackvial_koch_test_chunk_v2`

- Output file: `data/jackvial_koch_test_chunk_v2.tsfile`
- Source rows: 1,192
- Table name: `jackvial_koch_test_chunk_v2`
- TAG columns: `episode_index`, `task_index`
- FIELD columns: `frame_index`, `sample_index`, `next_reward`,
  `action_0..action_5`, `observation_state_0..observation_state_5`
- Time: `Time = round(timestamp * 1000)` milliseconds, restarting per episode

### `jackvial_koch_test_train_legacy`

- Output file: `data/jackvial_koch_test_train_legacy.tsfile`
- Source rows: 880
- Table name: `jackvial_koch_test_train_legacy`
- TAG column: `episode_index`
- FIELD columns: `frame_index`, `sample_index`, `next_done`,
  `observation_state_0..observation_state_5`, `action_0..action_5`
- Time: `Time = round(timestamp * 1000)` milliseconds, restarting per episode

## Conversion Notes

- `timestamp` is converted to the TsFile `Time` column and is not retained as a
  separate field.
- `index` is renamed to `sample_index`.
- Source column names containing `.` are normalized by replacing `.` with `_`.
  For example, `next.reward` becomes `next_reward`.
- Vector columns are flattened by preserving the source column name with `.`
  replaced by `_`, then appending the element index.
- `action[6]` is flattened to `action_0..action_5`.
- `observation.state[6]` is flattened to `observation_state_0..observation_state_5`.
- The legacy train column `observation.images.laptop` is a video path/timestamp
  struct and is omitted from TsFile. The corresponding videos remain in the
  original dataset.
- Aside from the redundant `timestamp` column and the video path struct noted
  above, no numeric source columns or rows are intentionally dropped.

## Videos

Videos are not included in this TsFile repository. They remain in the original
dataset under
[`videos/`](https://huggingface.co/datasets/jackvial/koch_test/tree/main/videos).

Use `episode_index`, `frame_index`, and `Time` to align TsFile rows with the
original video frames where applicable.

## Metadata

The uploaded metadata includes:

```text
meta/
meta_data/
tsfile_conversion.json
```

`tsfile_conversion.json` records the source files, row counts, episode coverage,
converted TsFile paths, dropped timestamp handling, and video policy.

## Usage

Read the TsFile files with the Apache TsFile Java or Python SDK:

```python
chunk_v2_path = "data/jackvial_koch_test_chunk_v2.tsfile"
legacy_path = "data/jackvial_koch_test_train_legacy.tsfile"

# Example SQL idea:
# SELECT * FROM jackvial_koch_test_chunk_v2 WHERE episode_index = 0
# SELECT * FROM jackvial_koch_test_train_legacy WHERE episode_index = 0
```