README.md

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- tsfile
- timeseries
- format:tsfile
pretty_name: eval3_90_permutation (TsFile)
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/eval3_90_permutation.tsfile
modality:
- tabular
- timeseries
---

# eval3_90_permutation (TsFile)

This dataset is an Apache TsFile conversion of the Hugging Face dataset
[`robot-learning-group47/eval3_90_permutation`](https://huggingface.co/datasets/robot-learning-group47/eval3_90_permutation).
The source dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

Modalities: Time-series. The original repository also contains a synchronized
front-camera video stream; videos are not included in this converted repository.

## Source Dataset

- Original dataset: [`robot-learning-group47/eval3_90_permutation`](https://huggingface.co/datasets/robot-learning-group47/eval3_90_permutation)
- License: `apache-2.0`
- LeRobot codebase version: `v3.0`
- Robot type: `so_follower`
- Split: `train` (`0:90`)
- Source scale from `meta/info.json`: `90` episodes, `20,221` frames, `3` tasks
- Sampling rate: `15` fps
- Source data layout: `data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet`
- Source video layout: `videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4`

Tasks:

- `Place the coke on Taylor Swift.`
- `Place the coke on Barack Obama.`
- `Place the coke on Yann LeCun.`

## Converted Files

- TsFile: `data/eval3_90_permutation.tsfile`
- Converted rows: `20,221`
- TsFile table: `eval3_90_permutation`
- Time precision: milliseconds
- TAG columns: `episode_index`, `task_index`
- TsFile size: `496,551` bytes

## Schema

`Time` is synthesized as `round(timestamp * 1000)` in milliseconds. The source
`timestamp` column is dropped because it is redundant with `Time / 1000` seconds.
At 15 fps, consecutive frames are spaced by about 67 ms.

TAG columns:

- `episode_index`
- `task_index`

FIELD columns:

- `frame_index`
- `sample_index` (renamed from source `index`)
- `action_0` to `action_5`
- `observation_state_0` to `observation_state_5`

Vector features are flattened by preserving the source feature name and replacing
`.` with `_`. For example, `observation.state` becomes
`observation_state_0` to `observation_state_5`. The 6-element `action` and
`observation.state` vectors use the source joint order:
`shoulder_pan.pos`, `shoulder_lift.pos`, `elbow_flex.pos`, `wrist_flex.pos`,
`wrist_roll.pos`, and `gripper.pos`.

## Attachments

The evaluation permutation map is mirrored as-is from the source dataset:

- `eval3_episode_permutation_map.csv`
- `eval3_episode_permutation_map.json`

These files map each episode to the image triple and left/center/right
permutation used during recording.

## Video Policy

The source video feature `observation.images.front` is not converted into TsFile
and is not uploaded here. Use the original dataset for videos:
[`robot-learning-group47/eval3_90_permutation/videos`](https://huggingface.co/datasets/robot-learning-group47/eval3_90_permutation/tree/main/videos).

## Metadata

The source `meta/` files are mirrored in this repository. `meta/info.json` is
updated so `data_path` points to `data/eval3_90_permutation.tsfile` and includes
a `tsfile_conversion` object documenting the Time mapping, TAG columns,
flattened features, dropped fields, and video policy.

## Validation

The converted TsFile was validated with the project pipeline and read back using
the TsFile Python SDK:

- staged Parquet rows: `20,221`
- TsFile metadata rows: `20,221`
- TsFile query rows: `20,221`

## Usage

```python
from tsfile import TsFileReader

path = "data/eval3_90_permutation.tsfile"
with TsFileReader(path) as reader:
    schemas = reader.get_all_table_schemas()
    print(schemas.keys())
```