---
license: cc-by-4.0
language:
- en
tags:
- medical
---
# SutureBot: A Precision Framework and Benchmark for Autonomous End-to-End Suturing

This dataset is part of the **SutureBot** project, a benchmark for developing and evaluating autonomous surgical robotic policies for end-to-end suturing. It contains high-resolution, multi-camera video and robot kinematics from real robot demonstrations on tissue phantoms, supporting research in imitation learning, vision-language-action modeling, and surgical autonomy.

[project website](https://suturebot.github.io/) | [code repository](https://github.com/SutureBot/SutureBot/tree/ACT).

## Dataset Summary

The dataset includes:

- **Stereo endoscope images** (`left_img_dir`, `right_img_dir`)
- **Wrist camera views** (`endo_psm1`, `endo_psm2`)
- **Low-level kinematics** (end-effector pose, joint angles, jaw state, etc.) in `ee_csv.csv`
- **Multiple tissues** and **multiple task types** (e.g., needle pick-up, suturing)
- All data is organized by:
tissue_[id]/[task_name]/[episode_timestamp]/


The following is a visualization of the dataset folder structure:
 ``` 
 $PATH_TO_DATASET
├── [DATASET_NAME]       # the dataset base dir
|   └── tissue_1                      # data subset
|   |   ├── 1_[task_name]             # task name
|   |   |   ├── [episode]             # should be timestamp when the data was recorded
|   |   |   |      ├── left_img_dir   # left endoscope cam images (frame000000_left.jpg)
|   |   |   |      ├── right_img_dir  # right endoscope cam images (frame000000_right.jpg)
|   |   |   |      ├── endo_psm1      # right wrist cam images (frame000000_psm1.jpg)
|   |   |   |      ├── endo_psm2      # left wrist cam images (frame000000_psm2.jpg)
|   |   |   |      └── ee_csv.csv     # kinematics
|   └── tissue_2                      # data subset
|   |   ├── 1_[task_name]             # task name
|   |   |   ├── [episode]             # should be timestamp when the data was recorded
|   |   |   |      ├── left_img_dir   # left endoscope cam images (frame000000_left.jpg)
|   |   |   |      ├── right_img_dir  # right endoscope cam images (frame000000_right.jpg)
|   |   |   |      ├── endo_psm1      # right wrist cam images (frame000000_psm1.jpg)
|   |   |   |      ├── endo_psm2      # left wrist cam images (frame000000_psm2.jpg)
|   |   |   |      └── ee_csv.csv     # kinematics
...
 ``` 

Each episode is a complete trajectory from a real robotic execution using a da Vinci Research Kit (dVRK) platform.

---

## Usage

This dataset is designed for training and evaluating autonomous policies that take as input multi-view RGB images (optionally language-conditioned), and output continuous control actions.

It supports a wide range of applications including:
- Behavioral cloning and offline RL
- Multi-modal policy learning (images + kinematics + language)
- Visual servoing and tool tracking
- Surgical robot benchmarking

---

## File Structure

Each episode contains:
- `left_img_dir/`: Left endoscope camera images
- `right_img_dir/`: Right endoscope camera images
- `endo_psm1/`: Right wrist (PSM1) camera images
- `endo_psm2/`: Left wrist (PSM2) camera images
- `ee_csv.csv`: Kinematic data including joint states, poses, and jaw angles

---

## Format

- Image format: `.jpg`, RGB
- Kinematics: CSV with columns such as:
- `timestamp`
- `psm1_pose.position.{x,y,z}`
- `psm1_pose.orientation.{x,y,z,w}`
- `psm1_js[0-5]`, `psm2_js[0-5]`, etc.
- `ecm_pose.*`, `suj_pose.*`, etc.

A complete list of columns is available in the `dataset.croissant.json` schema.

---

## Citation

If you use this dataset, please cite:

```bibtex
@misc{suturebot2025,
title       = {SutureBot: A Precision Framework and Benchmark for Autonomous End-to-End Suturing},
author      = {Jesse Haworth, Juo-Tung Chen, Nigel Nelson, Ji Woong Kim, Masoud Moghani, Chelsea Finn, Axel Krieger},
year        = {2025},
note        = {Under review at NeurIPS 2025 Datasets and Benchmarks Track},
howpublished = {\url{https://huggingface.co/datasets/jchen396/suturebot}}
}
```
## License
This dataset is licensed under CC-BY-4.0. Please credit the authors when using this dataset.

## Contact
For questions or collaboration inquiries, please contact:

Juo-Tung Chen
Johns Hopkins University
📧 jchen396@jh.edu