README.md

---
license: apache-2.0
---

<center><strong>[RSS'26] One-Shot Real-World Demonstration Synthesis for Scalable Bimanual Manipulation</strong></center>
<center>[<a href="https://arxiv.org/abs/2512.09297">arXiv</a>][<a href="https://www.roboticsproceedings.org/rss22/p001.pdf">Paper</a>][<a href="https://hnuzhy.github.io/projects/BiDemoSyn/">Project</a>]</center>

### ✅️ Datasets Collected using the Proposed BiDemoSyn

<table>
  <tr>
	<td style="text-align: center;"> [Task ⓵ grid cells 6×6] <br> <a href="./plugpen.zip"><strong>plugpen (2 objects / 1944 demos)</strong></a> </td>
	<td style="text-align: center;"> [Task ⓶ grid cells 6×6] <br> <a href="./inserting.zip"><strong>inserting (2 objects / 3888 demos)</strong></a> </td>
	<td style="text-align: center;"> [Task ⓷ grid cells 12×12] <br> <a href="./unscrew.zip"><strong>unscrew (1 object / 576 demos)</strong></a> </td>
  </tr>
  <tr>
    <td style="text-align: center;"><video width=100% autoplay loop muted><source src="https://huggingface.co/datasets/HoyerChou/BiDemoSyn-Data/resolve/main/plugpen_vis-slim.mp4" type="video/mp4" /></video></td>
    <td style="text-align: center;"><video width=100% autoplay loop muted><source src="https://huggingface.co/datasets/HoyerChou/BiDemoSyn-Data/resolve/main/inserting_vis-slim.mp4" type="video/mp4" /></video></td>
    <td style="text-align: center;"><video width=100% autoplay loop muted><source src="https://huggingface.co/datasets/HoyerChou/BiDemoSyn-Data/resolve/main/unscrew_vis-slim.mp4" type="video/mp4" /></video></td>
  </tr>

  <tr>
	<td style="text-align: center;"> [Task ⓸ grid cells 6×6] <br> <a href="./pouring.zip"><strong>pouring (2 objects / 2592 demos)</strong></a> </td>
	<td style="text-align: center;"> [Task ⓹ grid cells 6×6] <br> <a href="./pressing.zip"><strong>pressing (2 objects / 1296 demos)</strong></a> </td>
	<td style="text-align: center;"> [Task ⓺ grid cells 6×6] <br> <a href="./reorient.zip"><strong>reorient (1 object / 720 demos)</strong></a> </td>
  </tr>
  <tr>
    <td style="text-align: center;"><video width=100% autoplay loop muted><source src="https://huggingface.co/datasets/HoyerChou/BiDemoSyn-Data/resolve/main/pouring_vis-slim.mp4" type="video/mp4" /></video></td>
    <td style="text-align: center;"><video width=100% autoplay loop muted><source src="https://huggingface.co/datasets/HoyerChou/BiDemoSyn-Data/resolve/main/pressing_vis-slim.mp4" type="video/mp4" /></video></td>
    <td style="text-align: center;"><video width=100% autoplay loop muted><source src="https://huggingface.co/datasets/HoyerChou/BiDemoSyn-Data/resolve/main/reorient_vis-slim.mp4" type="video/mp4" /></video></td>
  </tr>
</table>

<!-- ⓵ ⓶ ⓷ ⓸ ⓹ ⓺ https://tw.piliapp.com/symbol/number/#google_vignette -->

### ✅️ Data Format

* For bimanual tasks (including **plugpen**, **inserting**, **pouring** and **pressing**) which have two manipulated objects, each demonstration has **9** corresponding files: `{TaskObjStr}-imgL_mask_arr.json`, `{TaskObjStr}-imgL_mask_{ObjA}.jpg`, `{TaskObjStr}-imgL_mask_{ObjB}.jpg`, `{TaskObjStr}-imgL_raw.jpg`, `{TaskObjStr}-imgR_mask_arr.json`, `{TaskObjStr}-imgR_mask_{ObjA}.jpg`, `{TaskObjStr}-imgR_mask_{ObjB}.jpg`, `{TaskObjStr}-imgR_raw.jpg`, `{TaskObjStr}-labels.json`.
    * `{TaskObjStr}` is a long string which is combined via task-related object names, orientations and positions. An example for the task **inserting** is `ordcup01-marker02_orient01-LcupX01Y01-RpenX04Y01`, which means that the left part is a cup and the right part is a marker pen with orientation id1. The `X01Y01` and `X04Y01` represent their coordinates in the defined grid cells.
    * File `{TaskObjStr}-imgL_mask_arr.json` is a Dict, which uses the object's name as the key and object's [bbox, mask] as the value (from **left** image).
    * Files `{TaskObjStr}-imgL_mask_{ObjA}.jpg` and `{TaskObjStr}-imgL_mask_{ObjB}.jpg` are two binary masks of manipulated objects (from **left** image).
    * File `{TaskObjStr}-imgL_raw.jpg` is a original image of the **left** view from the binocular camera.
    * File `{TaskObjStr}-imgR_mask_arr.json` is a Dict, which are results processed from the **right** image.
    * Files `{TaskObjStr}-imgR_mask_{ObjA}.jpg` and `{TaskObjStr}-imgR_mask_{ObjB}.jpg` are two binary masks of manipulated objects (from **right** image).
    * File `{TaskObjStr}-imgR_raw.jpg` is a original image of the **right** view from the binocular camera.
    * File `{TaskObjStr}-labels.json` is a Dict, which contains sparse 7-DoF keyposes/keyframes (the value under each key is a list [position_x, position_y, position_z, rotation_x, rotation_y, rotation_y, gripper]) of the left-arm and right-arm asynchronously.
![example_inserting-files](https://cdn-uploads.huggingface.co/production/uploads/640dd49c3c82bd463ee3987f/ZCpl_ObabztP2JY7rdCzE.jpeg)

* For bimanual (including **unscrew** and **reorient**) which have only one manipulated object, each demonstration has **7** corresponding files: `{TaskObjStr}-imgL_mask_arr.json`, `{TaskObjStr}-imgL_mask_{Obj}.jpg`, `{TaskObjStr}-imgL_raw.jpg`, `{TaskObjStr}-imgR_mask_arr.json`, `{TaskObjStr}-imgR_mask_{Obj}.jpg`, `{TaskObjStr}-imgR_raw.jpg`, `{TaskObjStr}-labels.json`.
    * `{TaskObjStr}` is also a long string which is combined via task-related object names, orientations and positions. An example for the task **unscrew** is `bottle03_orient01_p1-LRbottleX01Y01`, which means that the platform has only one bottle every time. The `X01Y01` represents its coordinates in the defined grid cells.
    * These **7** files have the same format as above mentioned. The two missing files comparing to **9** are due to the absence of an object mask in the left and right images, respectively.
![example_unscrew-files](https://cdn-uploads.huggingface.co/production/uploads/640dd49c3c82bd463ee3987f/naQgnF6_x2eIwHFLJOw2r.jpeg)


### ✅️ Citation
If you use our data in your research, please cite with:
```
@article{zhou2025one,
	title={One-Shot Real-World Demonstration Synthesis for Scalable Bimanual Manipulation},
	author={Huayi Zhou, Kui Jia},
	journal={arXiv preprint arXiv:2512.09297},
	year={2025}
}
```