---
license: cc-by-4.0
tags:
- LeRobot
- piper
- piper-h
- pick-and-place
- imitation-learning
- real-robot
- realsense
task_categories:
- robotics
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# Piper-H Pick-and-Place XY Random 50

This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=Smokepord/piper_pick_place_xy_random_50_v1">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
</a>

## Dataset Description

This dataset contains 50 successful real-robot demonstrations collected with an AgileX Piper-H robotic arm. The task is to pick up a blue 3D-printed container twist-lock part and place it inside a black container.

- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)

## Dataset Structure

- 50 episodes and 18,884 frames at 20 Hz
- Two synchronized RGB observations from wrist-mounted and fixed Intel RealSense D435i cameras
- 640 x 480 AV1 video without audio
- 7-dimensional state and action vectors ordered as six joint positions followed by gripper position
- Absolute position actions expressed in the same normalized coordinate system as the state

The `piper_follower` value in [`meta/info.json`](meta/info.json) is the LeRobot plugin robot type. The physical robot used to collect the dataset was an AgileX Piper-H.

## Collection Scope

The object was randomly translated within a small planar XY region. Object height and orientation, container pose, cameras, lighting, and the wider workspace arrangement remained effectively fixed.

This dataset supports an in-region imitation-learning baseline. It should not be interpreted as covering changes in object height or orientation, lighting, camera pose, container geometry, or larger workspace regions.

## ACT Baseline

An ACT policy selected at checkpoint 25,000 achieved 16 successes in 20 real-robot trials (80%; approximate 95% Wilson interval: 58.4%-91.9%). Evaluation used four in-region object start positions with five trials per position; each position produced four successes and one failure. Failure stages and trial durations were not separately recorded.

## Citation

```bibtex
@misc{smokepord_piper_pick_place_xy_random_50_v1,
  author       = {Smokepord},
  title        = {Piper-H Pick-and-Place XY Random 50 Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Smokepord/piper_pick_place_xy_random_50_v1}}
}
```