---
authors:
- 120ft Factory AB
- chrisvtom
task_categories:
- robotics
tags:
- tsfile
- timeseries
- tabular
- robotics
- lerobot
- gr00t
- unitree-g1
modality:
- timeseries
- tabular
pretty_name: First Thread GR00T TsFile
configs:
- config_name: default
  data_files:
  - split: train
    path: data/120ft_first_thread_gr00t_train.tsfile
size_categories:
- 10K<n<100K
---

# First Thread GR00T TsFile

This dataset is an Apache TsFile conversion of
[`120ft/first-thread-gr00t`](https://huggingface.co/datasets/120ft/first-thread-gr00t), a LeRobot v2.1 / GR00T Unitree G1-D teleoperation
dataset for the task **"thread lamp part"**.

The converted repository contains time-series and tabular numeric robot data.
The three camera streams remain in the original Hugging Face dataset and are
not duplicated here.

## Source Dataset and Provenance

- Original dataset: [`120ft/first-thread-gr00t`](https://huggingface.co/datasets/120ft/first-thread-gr00t)
- Pinned source revision: [`b8a57a601ac099cec83fe0e49589e1032f2d197c`](https://huggingface.co/datasets/120ft/first-thread-gr00t/tree/b8a57a601ac099cec83fe0e49589e1032f2d197c)
- Publisher/original organization: [120ft Factory AB (`120ft`)](https://huggingface.co/120ft)
- Repository and video contributor: [`chrisvtom`](https://huggingface.co/chrisvtom)
- License: **not declared by the source dataset card**
- Paper/citation: not provided by the source dataset card
- Robot type: `g1d`
- LeRobot codebase version: `v2.1`
- Task: `thread lamp part` (`task_index=0`)
- Split: `train`
- Sampling rate: 30 fps
- Scale: 72 episodes, 96,960 frames, 1 task
- Source Parquet layout: `data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet`
- Source Parquet shards: 72

The source card states that the v2.1 GR00T form was converted from a v3.0
dataset in the 120ft data-collection repository. This conversion uses the
pinned published LeRobot v2.1 Parquet files as its authoritative numeric input.

## Converted Files

- TsFile: `data/120ft_first_thread_gr00t_train.tsfile`
- Table: `120ft_first_thread_gr00t_train`
- Rows: 96,960
- Episodes/devices: 72
- Columns: 37 total (1 TIME, 2 TAG, 34 FIELD)
- TsFile size: 6.30 MiB
- Time precision: milliseconds
- Metadata: source `meta/` files are preserved; `meta/info.json` is rewritten
  to describe the TsFile schema, mapping, provenance, and video policy.

## TsFile Schema

`Time` is an INT64 millisecond timestamp computed as
`round(timestamp * 1000)`. It restarts at zero in each episode.

| Role | Columns | TsFile type |
|---|---|---|
| TIME | `Time` | INT64/TIMESTAMP |
| TAG | `episode_index`, `task_index` | STRING device tags |
| FIELD | `frame_index`, `sample_index` | INT64 |
| FIELD | `observation_state_0` ... `observation_state_15` | FLOAT |
| FIELD | `action_0` ... `action_15` | FLOAT |

The 16 state/action dimensions retain the ordering documented in source
`meta/info.json`: seven left-arm joints, seven right-arm joints, then left and
right Dex1 gripper values.

## Conversion Notes

- All 72 source episode Parquet files are merged into one train table-model
  TsFile. Filter by `episode_index` and `task_index` to select a trajectory.
- `observation.state[16]` and `action[16]` are flattened to scalar FLOAT
  fields. Source prefixes are preserved and `.` is replaced with `_`.
- Source `timestamp` is dropped after Time synthesis because it is exactly
  represented by `Time / 1000` seconds at millisecond precision.
- Source `index` is retained as `sample_index`; `frame_index` is unchanged.
- No source rows, episodes, tasks, or numeric vector dimensions are dropped.

## Encoding and Compression

The conversion generated and fully validated both permitted numeric profiles,
then selected the smaller valid result (`zstd`):

- `Time`: `TS_2DIFF + LZ4`
- FLOAT/DOUBLE FIELD values: `GORILLA + ZSTD`
- INT32/INT64 FIELD values: `TS_2DIFF + ZSTD`
- BOOLEAN FIELD values: `RLE + LZ4`
- TAG values: TsFile table device/TAG mechanism (`PLAIN + LZ4` storage)

The final TsFile is 0.5578 times the combined size of the 72 source
Parquet files.

## Videos

Videos are not included in this converted repository. The pinned source has
216 frame-aligned H.264 480x640 MP4 files at 30 fps, split across three streams:

- [`observation.images.head`](https://huggingface.co/datasets/120ft/first-thread-gr00t/tree/b8a57a601ac099cec83fe0e49589e1032f2d197c/videos/chunk-000/observation.images.head) - 72 per-episode MP4 files
- [`observation.images.left_wrist`](https://huggingface.co/datasets/120ft/first-thread-gr00t/tree/b8a57a601ac099cec83fe0e49589e1032f2d197c/videos/chunk-000/observation.images.left_wrist) - 72 per-episode MP4 files
- [`observation.images.right_wrist`](https://huggingface.co/datasets/120ft/first-thread-gr00t/tree/b8a57a601ac099cec83fe0e49589e1032f2d197c/videos/chunk-000/observation.images.right_wrist) - 72 per-episode MP4 files

Source template: `videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4`. Numeric rows remain aligned
with the original videos through `episode_index`, `frame_index`, and source
episode metadata.

## Validation

Local validation checks source and staged row counts, unique
`(episode_index, task_index, Time)` keys, per-episode monotonic Time, table/TAG
schema, full Java query readback, SHA-256, and actual per-column encodings and
compression. Conversion scripts and validation reports are intentionally not
part of the upload-ready dataset.

## Minimal Read Example

```python
from tsfile import TsFileReader

reader = TsFileReader("data/120ft_first_thread_gr00t_train.tsfile")
table_name = "120ft_first_thread_gr00t_train"
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "observation_state_0",
    "action_0",
]

with reader.query_table(table_name, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())
reader.close()
```

## Citation

The source dataset does not provide a paper or completed citation. Cite the
original Hugging Face dataset, 120ft Factory AB, and the pinned source revision
when using this converted artifact.