---
license: apache-2.0
pretty_name: ABC fold_the_paper_box (LeRobot v2.1)
language:
  - en
tags:
  - robotics
  - manipulation
  - imitation-learning
  - bimanual
  - lerobot
  - yam
task_categories:
  - robotics
size_categories:
  - 1M<n<10M
---

# ABC fold_the_paper_box — LeRobot v2.1

A **LeRobot v2.1** conversion of the `fold_the_paper_box` task from
[XDOF/ABC-130k](https://huggingface.co/datasets/XDOF/ABC-130k), prepared for
fine-tuning [pi0.5](https://github.com/Physical-Intelligence/openpi) (`pi05_base`)
on the bimanual YAM platform.

## Contents

| Attribute | Value |
| --- | --- |
| Episodes | 500 (first 500 of 2,387 train episodes, sorted by episode uuid) |
| Frames | 2,260,990 (~21 hours) |
| FPS | 30 |
| Robot | Bimanual station, 2× 6-DoF YAM arms, parallel-jaw grippers |
| Cameras | `head_camera`, `left_wrist_camera`, `right_wrist_camera` — 640×480 h264 |
| Task instruction | `fold the paper box` |
| Codebase version | LeRobot `v2.1` |

### State / action format

Both `observation.state` and `action` are **14-D float32**:

```
[left_joint_0..5, left_gripper, right_joint_0..5, right_gripper]
```

- Joint values are absolute positions in **radians**, base → wrist.
- Gripper is the normalized aperture from ABC-130k (**0 = closed, 1 = open**).
- `action` holds the **commanded** joint positions (ABC `/{side}-arm-action` +
  `/{side}-ee-action`); `observation.state` holds the **measured** ones
  (`/{side}-arm-state` + `/{side}-ee-state`).
- Actions are absolute, not deltas. (openpi's pi0.5 recipe applies
  `DeltaActions` internally at train time.)

## How this was converted

Source episodes are MCAP files. In ABC-130k the streams run on independent
clocks — the action stream at ~200 Hz, state at ~265 Hz, cameras at 30–60 Hz
depending on station type — so the conversion resamples everything onto a
**fixed 30 Hz tick clock** using causal floor matching (the latest message at or
before each tick), over the overlap window of all streams. This matches the tick
rate used by the official ABC exporter.

Both station types present in the source are handled: RealSense (mono top
camera) and ZED-X (stereo top camera, where one eye is picked deterministically
per episode via `sha1(episode_id)` parity). All camera streams are letterboxed
to 640×480 preserving aspect ratio.

Conversion script:
[`convert_abc_mcap_to_lerobot_v21.py`](https://github.com/Avant-US/openpi) —
`scripts/convert_abc_mcap_to_lerobot_v21.py`.

`meta/episode_ids.json` maps each `episode_index` back to its original ABC-130k
episode uuid for traceability.

## Usage

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Sichang0621/abc_fold_box_v21")
print(ds.num_episodes, ds.num_frames, ds.fps)
item = ds[0]  # observation.state (14,), action (14,), 3× images (3,480,640)
```

## Attribution and license

This dataset is **derived from [XDOF/ABC-130k](https://huggingface.co/datasets/XDOF/ABC-130k)**
(Apache-2.0), released alongside the ABC project ([abc.bot](https://abc.bot/),
[code](https://github.com/amazon-far/abc)). All robot trajectories and imagery
originate from ABC-130k; this repository contributes only the format conversion
described above. Please cite the ABC project when using this data.

Note that the upstream ABC-130k dataset is access-gated on the Hub. Licensed
under Apache-2.0, consistent with the source.