README.md

---
license: cc-by-4.0
task_categories:
- robotics
tags:
- LeRobot
- reassemble
- contact-rich
- manipulation
- assembly
- disassembly
- franka
- force-torque
size_categories:
- 1M<n<10M
---

# REASSEMBLE (LeRobot v3)

A LeRobot Dataset v3 port of **REASSEMBLE** — a multimodal dataset for **contact-rich robotic assembly and disassembly** on the NIST Assembly Task Board, recorded with a Franka arm.

> **This is a reformatted derivative**, not the original release. The original data, full documentation, and canonical DOI are published by the authors at TU Wien:
> **https://researchdata.tuwien.ac.at/records/0ewrv-8cb44** (DOI [`10.48436/0ewrv-8cb44`](https://doi.org/10.48436/0ewrv-8cb44)).
> Paper: [arXiv:2502.05086](https://arxiv.org/abs/2502.05086) · Project: https://tuwien-asl.github.io/REASSEMBLE_page/ · Code: https://github.com/TUWIEN-ASL/REASSEMBLE

## What this is

The original REASSEMBLE ships one HDF5 per recording session (encoded video blobs, raw PCM audio, a sparse event stream, multi-rate proprioception + force/torque, and hierarchical action-segment annotations). This port converts those **149 recordings into 149 LeRobot episodes** (one recording = one long-horizon episode), with all sensors resampled onto a uniform **30 fps** grid whose master clock is the hand-camera timestamps.

- **Episodes:** 149
- **Frames:** 1,433,406 @ 30 fps
- **Robot:** Franka
- **Cameras:** `hand`, `hama1`, `hama2` (RGB 480×640) + `event_cam` (DAVIS render, 260×346)
- **Per-frame task:** the active high-level action segment's language label (e.g. *"Insert USB."*, *"Pick square peg 3."*)

## Features

| key | dtype | shape | notes |
|---|---|---|---|
| `observation.images.hand` / `hama1` / `hama2` | video | 480×640×3 | RGB cameras |
| `observation.images.event_cam` | video | 260×346×3 | DAVIS event-camera render |
| `observation.state` | float32 | (36,) | joint pos/vel/eff (7×3) + gripper (2) + EE pose (7) + EE velocity (6) |
| `observation.state.joint_position` | float32 | (7,) | |
| `observation.state.gripper_position` | float32 | (2,) | |
| `observation.state.ee_pose` | float32 | (7,) | x,y,z + quaternion (w,x,y,z) |
| `observation.force` / `observation.torque` | float32 | (3,) | measured F/T |
| `observation.force.base` / `observation.torque.base` | float32 | (3,) | gravity/bias-compensated base F/T |
| `action` | float32 | (9,) | **next** absolute EE target: pose (7) + gripper (2) |
| `segment.success` | bool | (1,) | whether the active segment succeeded |
| `segment.index` | int64 | (1,) | high-level segment index within the recording |

## Fidelity notes (please read)

LeRobot is fixed-fps and frame-aligned, so this port makes deliberate tradeoffs:

- **Resampling:** every sensor is nearest-neighbour sampled to 30 fps. High-rate force/torque (~1.7 kHz measured, ~0.5 kHz base) is therefore **downsampled** in the frame stream.
- **Audio is not a frame feature.** It's preserved as a per-episode sidecar `audio/episode_XXXXXX/{hand,hama1,hama2}.wav` (16 kHz PCM). The LeRobot dataloader does **not** return audio tensors.
- **Raw events are not a frame feature.** The full sparse event stream is preserved losslessly as a per-episode sidecar `events/episode_XXXXXX.npz` (`events`: N×3 int64 `x,y,polarity`; `timestamps`: N float64). The event-camera *render* is available as the `event_cam` video.
- **Missing cameras:** a few recordings are missing a camera (e.g. `2025-01-10-16-17-40` has no hand cam — a known issue from the source README). Those frames are **black-filled** for the missing stream to keep the schema consistent.

## Splits

Original author splits are preserved in `meta/splits.json` (per-episode `recording` + `split`): **train = 111, test = 37**, 1 unassigned.

## Usage

```python
from lerobot.datasets import LeRobotDataset

ds = LeRobotDataset("robot-lev/reassemble")
frame = ds[0]
print(frame["observation.state"].shape, frame["action"].shape, frame["task"])
```

Sidecars (audio / raw events) live alongside the dataset and can be downloaded with `huggingface_hub`:

```python
from huggingface_hub import hf_hub_download
import numpy as np
ev = np.load(hf_hub_download("robot-lev/reassemble", "events/episode_000000.npz", repo_type="dataset"))
print(ev["events"].shape, ev["timestamps"].shape)  # (N, 3), (N,)
```

## License & attribution

Released under **CC-BY-4.0**, inherited from the original dataset. You must credit the original authors:

> Sliwowski, Daniel Jan; Jadav, Shail; Stanovcic, Sergej; Orbik, Jędrzej; Heidersberger, Johannes; Lee, Dongheui. *REASSEMBLE: A Multimodal Dataset for Contact-rich Robotic Assembly and Disassembly.* TU Wien, 2025. DOI: 10.48436/0ewrv-8cb44.

```bibtex
@misc{sliwowski2025reassemble,
  title         = {REASSEMBLE: A Multimodal Dataset for Contact-rich Robotic Assembly and Disassembly},
  author        = {Sliwowski, Daniel Jan and Jadav, Shail and Stanovcic, Sergej and Orbik, J\k{e}drzej and Heidersberger, Johannes and Lee, Dongheui},
  year          = {2025},
  eprint        = {2502.05086},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  doi           = {10.48436/0ewrv-8cb44},
  url           = {https://researchdata.tuwien.ac.at/records/0ewrv-8cb44}
}
```

**Porting scripts & walkthrough:** https://github.com/lvjonok/reassemble-lerobot-port — reproducible conversion pipeline, design tradeoffs, and operational gotchas.

*This derivative reorganizes and resamples the data; refer to the original record for the authoritative, full-rate source.*