---
license: mit
task_categories:
- robotics
tags:
- LeRobot
- tsfile
- format:tsfile
- timeseries
size_categories:
- 10K<n<100K
---



# lerobot_pusht (TsFile format)

> This dataset is a conversion of the Hugging Face dataset [**lerobot/pusht**](https://huggingface.co/datasets/lerobot/pusht) to [Apache TsFile](https://tsfile.apache.org/) format.
> **Original dataset**: https://huggingface.co/datasets/lerobot/pusht

## Dataset Description

Push-T is a robot-manipulation benchmark task introduced by **Diffusion Policy** (Chi et al. 2023): the agent must push a T-shaped block to a target pose. This dataset was collected with [LeRobot](https://github.com/huggingface/lerobot) (codebase v2.0) and is its **low-dimensional state version** — time series of 2-D end-effector state and 2-D action.

- **Homepage**: https://diffusion-policy.cs.columbia.edu/
- **Paper**: https://arxiv.org/abs/2303.04137v5
- **License**: MIT
- **Size**: 206 episodes / 25,650 frames / 1 task
- **Sampling rate**: 10 fps (100 ms per frame)

## Original Data Structure

| Column | Type | Description |
|----|------|------|
| `observation.state` | float32[2] | End-effector state (x, y) |
| `action` | float32[2] | Action (x, y) |
| `episode_index` | int64 | Episode index |
| `frame_index` | int64 | Frame index within the episode |
| `timestamp` | float32 | Seconds elapsed within the episode |
| `next.reward` | float32 | Reward |
| `next.done` / `next.success` | bool | Termination / success flags |
| `index` | int64 | Global sample index |
| `task_index` | int64 | Task index |
| `observation.image` | video[96×96×3] | (**not included** — this low-dimensional version has no images) |

## TsFile Conversion Notes

Conversion uses the "script preprocessing + Apache TsFile Java tool (schema mode)" path:

- **Array expansion**: `observation.state[2]` → `state_0`, `state_1`; `action[2]` → `action_0`, `action_1` (kept as float32 → TsFile `FLOAT`).
- **Column-name cleanup**: `.` is replaced with `_` (`next.reward` → `next_reward`, etc.).
- **Time axis**: `Time = frame_index × 100 ms` (10 fps), millisecond precision.
- **Tag columns (device dimension)**: `episode_id`, `task_id` are declared as TsFile **TAG**, so each episode is an independent device with its own time axis starting at 0.
- **⚠️ Dropped columns**: the original `timestamp` (per-episode elapsed seconds, float) is dropped because it repeats across episodes; an integer-millisecond time axis (`frame_index × 100 ms`) is used instead. All other columns are retained.

The converted table is named `pusht`, in a single file `lerobot_pusht.tsfile` with 25,650 rows.

## Usage

```python
# Read lerobot_pusht.tsfile with the Apache TsFile SDK
from tsfile import TsFileReader
reader = TsFileReader("lerobot_pusht.tsfile")
# table "pusht": tag columns episode_id / task_id, remaining columns are field measurements
```

## Citation

```bibtex
@article{chi2024diffusionpolicy,
    author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
    title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
    journal = {The International Journal of Robotics Research},
    year = {2024},
}
```