---
license: apache-2.0
authors:
- Leo Lin
task_categories:
- robotics
tags:
- tsfile
- timeseries
- tabular
modality:
- timeseries
- tabular
pretty_name: ZBot Pick Cube35 TsFile
configs:
- config_name: default
  data_files:
  - split: train
    path: data/leolin6_zbot_pick_cube35_train.tsfile
size_categories:
- 10K<n<100K
---

# ZBot Pick Cube35 TsFile

This is an Apache TsFile conversion of [leolin6/zbot_pick_cube35](https://huggingface.co/datasets/leolin6/zbot_pick_cube35), a LeRobot v2.1
ZBot Inspire teleoperation dataset.

## Source Dataset and Author

- Original dataset: [leolin6/zbot_pick_cube35](https://huggingface.co/datasets/leolin6/zbot_pick_cube35)
- Pinned revision: [7e9b39936d42ac1bba9201f960a02e6651325df9](https://huggingface.co/datasets/leolin6/zbot_pick_cube35/tree/7e9b39936d42ac1bba9201f960a02e6651325df9)
- Creator and uploader: [Leo Lin (leolin6)](https://huggingface.co/leolin6)
- License: Apache-2.0
- Task: ZBot teleoperation recording for cube-picking demonstrations
- Robot: zbot_inspire; LeRobot: v2.1
- Split and rate: train episodes 0:35 at 30 fps
- Scale: 35 episodes, 20,644 rows, 1 task, 35 Parquet shards
- Source data layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
- Paper, homepage, and completed citation: not provided by the source card

## Converted File

- TsFile: data/leolin6_zbot_pick_cube35_train.tsfile
- Table: leolin6_zbot_pick_cube35_train
- Rows/devices: 20,644 / 35
- Time precision: milliseconds
- Size: 1,811,413 bytes
- Source meta is preserved; meta/info.json documents the conversion

## Schema

Time is INT64 milliseconds computed as round(timestamp * 1000) and restarts
within each episode.

| Role | Columns | Meaning |
|---|---|---|
| TIME | Time | INT64 milliseconds |
| TAG | episode_index, task_index | TsFile table device dimensions |
| FIELD | frame_index, sample_index | INT64 indexes |
| FIELD | action_0 through action_15 | 16 flattened FLOAT actions |
| FIELD | observation_state_0 through observation_state_25 | 26 flattened FLOAT states |

The table has 47 columns: 1 TIME, 2 TAG, and 44 FIELD columns.

## Conversion

- All 35 episode Parquet files are merged into one train TsFile.
- Source episode_index and task_index are TsFile TAG columns.
- action[16] and observation.state[26] are flattened; dots become underscores.
- Source index becomes sample_index and frame_index is unchanged.
- timestamp is dropped after Time synthesis because timestamp = Time / 1000.
- No numeric rows, episodes, tasks, action values, or state values are dropped.

## Encoding and Compression

- Time: TS_2DIFF + LZ4
- FLOAT/DOUBLE: GORILLA + ZSTD
- INT32/INT64: TS_2DIFF + ZSTD
- BOOLEAN: RLE + LZ4 (none are present)
- TAG: TsFile table device/TAG storage

Validated TsFile/source-Parquet ratio: 0.8635.

## Videos

Videos are not included. The original repository contains 35 frame-aligned MP4
files in [videos/chunk-000/observation.images.front/](https://huggingface.co/datasets/leolin6/zbot_pick_cube35/tree/7e9b39936d42ac1bba9201f960a02e6651325df9/videos/chunk-000/observation.images.front), named
episode_000000.mp4 through episode_000034.mp4. The source template is
videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Alignment is preserved by
episode_index, frame_index, and source episode metadata.

## Validation

Apache TsFile Java 2.2.1 readback returned all 20,644 rows.
TAG roles, monotonic Time, vector widths, encodings, compression, and row counts
were checked locally. Scripts and reports are excluded from upload files.

## Minimal Read Example

    from tsfile import TsFileReader

    reader = TsFileReader("data/leolin6_zbot_pick_cube35_train.tsfile")
    table_name = "leolin6_zbot_pick_cube35_train"
    with reader.query_table(
        table_name,
        ["episode_index", "task_index", "frame_index", "action_0"],
        batch_size=65536,
    ) as result:
        print(result.read_arrow_batch().to_pandas().head())

## Citation

The source card has no completed citation. Cite the original dataset and Leo Lin.