SutureBot2 Wound Closure Dataset

20 hours of teleoperated ex vivo wound closure demonstrations on the da Vinci Research Kit — the largest public robotic suturing dataset to date.

This dataset accompanies the paper "SutureBot2: A Framework for Autonomous Ex Vivo Suturing" (submitted to CoRL 2026). It is stored in LeRobot v2.1 format and substantially exceeds prior suturing datasets in scale, tissue realism, and annotation richness.

---

📋 At a Glance

Total Hours	20 hours
Semantic Action Labels	7 hours (frame-by-frame)
Point Annotations	7 hours (insertion & exit points)
Wound Segmentation Images	853 images (masks + bounding boxes)
Tissue Type	Ex vivo skin-on pork belly
Robot Platform	da Vinci Research Kit (dVRK) Si
Data Format	LeRobot v2.1
License	CC BY 4.0
Version	1.0
Collection Period	January – April, 2026

---

📖 Dataset Overview

SutureBot2 is the largest public robotic suturing dataset, surpassing JIGSAWS (3 hours, phantom tissue) and SutureBot v1 (6 hours, phantom tissue). It is also the first dataset to provide semantic labels for the complete suturing procedure on biological tissue.

Demonstrations capture the full wound closure pipeline across ~29 subtasks (Figure 2 in the paper), including needle pickup, bimanual handovers, tissue throwing, a complete 2-1-1 surgical knot, and suture cutting — on skin-on pork belly with 2–5 cm linear incisions treated with synthetic blood for visual realism.

Annotations Included

Annotation Type	Coverage	File
Frame-by-frame semantic action labels	7 hours	labels.jsonl per batch
Needle insertion & exit point labels	7 hours	insertion_exit_points.csv per batch
Wound segmentation masks + bounding boxes	853 images	Included in training splits

Demonstration Quality (worst → best)

1. lerobot_fausto_batch_0
2. lerobot_fausto_batch_1
3. lerobot_jesse_batch_0
4. lerobot_jesse_batch_1 ⭐
5. lerobot_jesse_batch_2 ⭐
6. lerobot_dagger* ⭐

Recommended for fine-tuning: lerobot_jesse_batch_1, lerobot_jesse_batch_2, and all DAgger data contain complete frame-by-frame semantic labels and point annotations, making them ideal for finetuning low-level policies.

---

🔬 Data Collection

Robot Platform

The dataset was collected on a da Vinci Research Kit (dVRK) Si platform equipped with:

• Left arm: Large Needle Driver
• Right arm: Mega SutureCut Needle Driver
• Wrist cameras: 7 mm borescopes mounted 35 mm from each wrist (640×480 @ 15 fps)
• Endoscope: 960×540 @ 30 fps

Task Setup

• Skin-on pork belly with 2–5 cm linear incisions
• Synthetic blood applied for visual realism
• Primarily 3-0 braided silk sutures
• Tissue placement, robot configuration, wrist camera placement, bite depth, suture spacing, incision size, and (occasionally) suture/needle type were all randomized before every episode

Operators

Count	2
Skill Levels	Intermediate (trained researcher) and Novice (ML researcher with minimal experience)

Recovery Demonstrations

Data is collected as full procedure episodes. When a mistake is made, the demonstrator immediately recovers from it in place — no episodes are discarded for errors, making the dataset rich in natural recovery behavior.

---

🎯 Subtask Taxonomy (~29 Subtasks)

The complete procedure is decomposed into the following subtasks (see Figure 2 in the paper):

Category	Subtasks
Needle Manipulation	Pickup Needle (L), Handover (L→R), Pass Back (L→R), Pass Back (R→L), Readjust Needle Grip
Tissue Interaction	Open Incision (L), Throw Needle Top (R), Throw Needle Bottom (R), Extract Needle (L)
Suture Handling	Pull Suture Through (R), Adjust Tail (R), Grab Suture (L), Handoff Suture (R→L), Set Down Needle
Knot Tying (2-1-1)	Wrap Suture CCW 2×, Grab Tail (R), Tighten Knot (R+L↑), Wrap Suture CW 1×, Grab Tail (R), Tighten Knot (R+L↑), Wrap Suture CCW 1×, Grab Tail (R), Tighten Knot (R+L↑)
Finishing	Cut Suture End (R), Set Down Suture (L), Grab Suture Tail (L), Cut Suture Tail (R), Set Down Suture (L)

---

🛠️ Action & State Space

Action Space

action: [x, y, z, qx, qy, qz, qw, gripper]

Field	Description
x, y, z	Absolute position in robot base frame (meters)
qx, qy, qz, qw	Absolute orientation as unit quaternion
gripper	Gripper opening angle (radians)

• Representation: Absolute Cartesian
• Orientation: Quaternion (x, y, z, w)
• Reference Frame: Robot Base Frame

State Space

Field	Description
Joint positions	Absolute positions for all articulated joints (radians)
End-effector pose	Cartesian position + quaternion orientation
Gripper state	Current gripper opening

Note: The paper trains the low-level policy with state dropout = 1.0, forcing reliance on visual inputs rather than dVRK kinematics. Kinematics are stored in the dataset but were not used during LLP training.

---

⏱️ Data Synchronization

Timestamps are saved to images and kinematics as they are pulled from their respective ROS topics. During conversion to LeRobot v2.1 format, the left endoscope image timestamp is used as the primary reference, and the closest prior timestamp for each image and kinematic frame is paired. Images and kinematics are captured at 30 Hz and filtered to remove stagnant frames before storage.

---

🤖 How This Dataset Was Used (SutureBot2 Paper)

This dataset trained three components of the SutureBot2 framework:

High-Level Policy (HLP): A fine-tuned Qwen3-VL-4B-Instruct model trained on 4-second temporal sequences of endoscope images. Five annotators provided free-form semantic action labels; embedding-space clustering (k=15, BAAI/bge-large-en-v1.5) was used to balance the heterogeneous label distribution. Achieved 76.7% label accuracy on held-out episodes.

Low-Level Policy (LLP): Fine-tuned on the 7-hour point-labeled subset, with needle insertion and exit points rendered directly onto images as goal conditioning overlays. Also includes 3 hours of DAgger-style error-recovery data. Built on GR00T-H (Open-H-Embodiment pretraining). Achieved 79.8% average subtask success under oracle routing and 72.7% under fully autonomous HLP routing across the complete 29-step procedure.

RL Token (RLT): Online reinforcement learning (TD3) targeting the top needle throw subtask. Seeded with 25 base-policy episodes, followed by 300 online training episodes. Improved throw precision and recovery over the imitation learning baseline.

Suture Planner: A YOLOv26 + SAM2 pipeline trained on the 853 annotated images achieved 88.6% wound detection IoU, 94.4% stitch detection F1, and delivered target points with a 677 ms round-trip latency on an NVIDIA IGX Orin with RTX 6000 Ada.

---

📊 Comparison to Prior Datasets

Dataset	Hours	Tissue	Semantic Labels	Complete Procedure
JIGSAWS	3 hrs	Phantom	No	No
SutureBot v1	6 hrs	Phantom	No	Partial (6 subtasks)
SutureBot2 (ours)	20 hrs	Ex Vivo	Yes	Yes (29 subtasks)

---

🚀 Loading with LeRobot

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("jhaworth/SutureBot2")
print(dataset)

---

📄 Citation

If you use this dataset, please cite:

@misc{suturebot2_2026,
  title     = {SutureBot2: A Framework for Autonomous Ex Vivo Suturing},
  year      = {2026},
  note      = {Submitted to the 10th Conference on Robot Learning (CoRL 2026).
               Dataset: https://huggingface.co/datasets/jhaworth/SutureBot2}
}

Also consider citing the original SutureBot benchmark this work builds upon:

@inproceedings{haworth2025suturebot,
  author    = {Haworth, J. and Chen, J.-T. and Nelson, N. and Kim, J. W. and
               Moghani, M. and Finn, C. and Krieger, A.},
  title     = {SutureBot: A Precision Framework \& Benchmark for Autonomous
               End-to-End Suturing},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS),
               Datasets and Benchmarks Track},
  year      = {2025}
}

---

👥 Contact

Dataset Lead	Jesse Haworth
Institution	Johns Hopkins University
Contact	jhawort2@jhu.edu