README.md

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- tsfile
- timeseries
- format:tsfile
pretty_name: Move Objects Multitask (TsFile)
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/move_objects_multitask_train.tsfile
modality:
- tabular
- timeseries
---

# Move Objects Multitask (TsFile)

This dataset is an Apache TsFile conversion of the Hugging Face dataset
[`villekuosmanen/move_objects_multitask`](https://huggingface.co/datasets/villekuosmanen/move_objects_multitask).
The source was created with
[LeRobot](https://github.com/huggingface/lerobot) and contains ARX5 robot
demonstrations for 58 object-manipulation tasks.

Modalities: Time-series. The source repository also contains three synchronized
camera streams; videos are not included in this converted repository.

## Source Dataset

- Original dataset: [`villekuosmanen/move_objects_multitask`](https://huggingface.co/datasets/villekuosmanen/move_objects_multitask)
- License: `apache-2.0`
- Current LeRobot metadata version: `v2.1`
- Robot type: `arx5`
- Split: `train` (`0:140`)
- Scale: `140` episodes, `100,760` frames, `58` tasks
- Sampling rate: `50` fps
- Source frame files: `140` Parquet files
- Source video files present in the repository: `420`
- Camera streams: `3`, each with `140` MP4 files
- Source data layout: `data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet`
- Source video layout: `videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4`

The tasks include picking up or sliding household objects, moving objects to a
shelf, placing objects in an oven dish or cardboard box, and several composed
object interactions. The complete mapping from `task_index` 0 through 57 is
preserved in `meta/tasks.jsonl`.

The camera streams are `observation.images.front`,
`observation.images.left_wrist`, and `observation.images.right_wrist`. Source
videos are 480 x 640 RGB AV1 at 50 fps and contain no audio.

### Source metadata discrepancy

The current source `meta/info.json` declares `total_videos: 432`, while the
repository file tree contains 420 MP4 files: 140 episodes multiplied by three
camera streams. The converted metadata preserves both facts as
`source_metadata_video_count: 432` and `source_video_count: 420`. The source
README also embeds an older v2.0 snapshot with 103,021 frames; this card uses
the current v2.1 `meta/info.json` and the 100,760 rows measured across all 140
source Parquet files.

## Converted File

- TsFile: `data/move_objects_multitask_train.tsfile`
- TsFile table: `move_objects_multitask_train`
- Converted rows: `100,760`
- Episodes: `140`
- Tasks: `58`
- Time precision: milliseconds
- TAG columns: `episode_index`, `task_index`
- File size: `3,355,749` bytes
- SHA-256: `9cb91a8cb77dd54c49590433765f0d52dff3cbf22638cc888ac40d378e91ae7a`

All train episodes are merged into one TsFile. The original `episode_index` and
`task_index` columns are retained as TAG columns, allowing queries by episode,
task, or both without synthetic aliases.

## Schema

`Time` is computed as `Time = round(timestamp * 1000)` in milliseconds and
restarts in each episode. At 50 fps, consecutive frames are approximately 20 ms
apart. The source `timestamp` column is not retained because it is redundant
with `Time / 1000` seconds. No source rows are dropped.

TAG columns:

- `episode_index`
- `task_index`

FIELD columns:

- `frame_index`
- `sample_index` (renamed from source `index`)
- `action_0` through `action_13` (FLOAT)
- `observation_state_0` through `observation_state_13` (FLOAT)

Both source vector features have 14 float32 elements. The source metadata does
not provide element names, so this conversion preserves their original order
without inventing semantic labels. Dots in source feature names are replaced
with underscores before the zero-based element index is appended.

## Video Policy

The three source video features are not converted or uploaded. Use the original
dataset for synchronized videos:
[`villekuosmanen/move_objects_multitask/videos`](https://huggingface.co/datasets/villekuosmanen/move_objects_multitask/tree/main/videos).

Each numeric row retains `episode_index`, `frame_index`, `task_index`, and
`sample_index`, preserving alignment with the original per-episode videos.

## Metadata

The source `meta/` files are mirrored in this repository. `meta/info.json` is
updated so `data_path` points to
`data/move_objects_multitask_train.tsfile`. Its `tsfile_conversion` object
records the source and converted file counts, declared and observed video
counts, table name, Time formula, TAG columns, row count, feature mappings, and
frame/video alignment. The converted `total_videos` value is `0`.

## Validation

The converted file was compared with all 140 source Parquet files:

- source and staged rows: `100,760`
- source tasks represented: `58`
- duplicate `(episode_index, task_index, Time)` rows: `0`
- maximum action-vector difference: `0`
- maximum observation-state-vector difference: `0`
- source `index` to converted `sample_index` mismatches: `0`

The TsFile is also validated with the project pipeline and read back with the
TsFile Python SDK before publication.

## Usage

```python
from tsfile import TsFileReader

path = "data/move_objects_multitask_train.tsfile"
with TsFileReader(path) as reader:
    schemas = reader.get_all_table_schemas()
    table = schemas["move_objects_multitask_train"]
    print([(column.get_column_name(), column.get_category())
           for column in table.get_columns()])
```

## Source & License

The source dataset is maintained at
[`villekuosmanen/move_objects_multitask`](https://huggingface.co/datasets/villekuosmanen/move_objects_multitask)
and is distributed under the Apache License 2.0. The source card does not
provide a paper, author list, or citation entry.