README.md

---
license: apache-2.0
authors:
- Henry-Ellis
task_categories:
- robotics
tags:
- tsfile
- robotics
- lerobot
- manipulation
- modality:timeseries
- modality:tabular
pretty_name: "G1 Dex3 Object Placement Dataset TsFile"
configs:
- config_name: default
  data_files:
  - split: train
    path: data/g1_dex3_objectplacement_dataset_train.tsfile
size_categories:
- 10K<n<100K
---

# G1 Dex3 Object Placement Dataset TsFile

Apache TsFile conversion of the [Unitree Robotics G1 Dex3 Object Placement
Dataset](https://huggingface.co/datasets/unitreerobotics/G1_Dex3_ObjectPlacement_Dataset),
a LeRobot v3.0 dataset for picking up toothpaste and a trash bag and placing
them in a blue storage container.

Modalities: Time-series, Tabular

## Source and attribution

- Original dataset: [unitreerobotics/G1_Dex3_ObjectPlacement_Dataset](https://huggingface.co/datasets/unitreerobotics/G1_Dex3_ObjectPlacement_Dataset)
- Original uploader/author shown in the repository history: [Henry-Ellis](https://huggingface.co/Henry-Ellis)
- Repository owner: [Unitree Robotics](https://huggingface.co/unitreerobotics)
- License: Apache-2.0; the source card does not provide a paper or formal citation.
- Robot: `Unitree_G1`, dual-arm 7-DOF robot with three-finger dexterous hands
- Sampling: 30 Hz; 210 episodes, 98,266 frames, one task

## Converted artifact

The train split is stored in one table-model TsFile:

`data/g1_dex3_objectplacement_dataset_train.tsfile`

Table name: `g1_dex3_objectplacement_dataset_train`; 98,266 rows and 210
devices (one device per `episode_index`/`task_index` combination). `Time` is
`round(timestamp * 1000)` in milliseconds and restarts at each episode. The
source `timestamp` is dropped because it is redundant; `index` is renamed to
`sample_index`.

TAG columns are `episode_index` and `task_index`. The vector fields are
flattened as follows:

- `observation.state` -> `observation_state_0` ... `observation_state_27`
- `action` -> `action_0` ... `action_27`

Other scalar fields are `frame_index` and `sample_index`.

## TsFile encoding

| Data category | Encoding | Compression |
|---|---|---|
| FLOAT / DOUBLE | GORILLA | LZ4 |
| INT32 / INT64 | TS_2DIFF | LZ4 |
| Time | TS_2DIFF | LZ4 |
| BOOLEAN | RLE | LZ4 |
| TAG | TsFile table/device TAG mechanism | TsFile schema metadata |

The encoded file is 18,234,639 bytes (the previous PLAIN/UNCOMPRESSED file was
25,547,130 bytes). A Java readback check confirmed completeness, row count,
TAG roles, and all listed codecs.

## Videos

Video files are intentionally not copied to this converted repository. They
remain in the original dataset under
`videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4`.
The four source streams are:

- [`observation.images.cam_left_high`](https://huggingface.co/datasets/unitreerobotics/G1_Dex3_ObjectPlacement_Dataset/tree/main/videos/observation.images.cam_left_high)
- [`observation.images.cam_right_high`](https://huggingface.co/datasets/unitreerobotics/G1_Dex3_ObjectPlacement_Dataset/tree/main/videos/observation.images.cam_right_high)
- [`observation.images.cam_left_wrist`](https://huggingface.co/datasets/unitreerobotics/G1_Dex3_ObjectPlacement_Dataset/tree/main/videos/observation.images.cam_left_wrist)
- [`observation.images.cam_right_wrist`](https://huggingface.co/datasets/unitreerobotics/G1_Dex3_ObjectPlacement_Dataset/tree/main/videos/observation.images.cam_right_wrist)

The source revision contains nine physical MP4 shards and 840 logical episode
videos. Numeric rows remain frame-aligned through `episode_index` and
`frame_index`; source `meta/episodes/` metadata is mirrored.

## Reading

```python
from tsfile import TsFileReader

reader = TsFileReader("data/g1_dex3_objectplacement_dataset_train.tsfile")
with reader.query_table(
    "g1_dex3_objectplacement_dataset_train",
    ["episode_index", "task_index", "Time", "frame_index", "action_0"],
    batch_size=65536,
) as result:
    print(result.read_arrow_batch().to_pandas().head())
```