jchen396/cao_cautery

Dataset and Imitation Learning Framework for Autonomous Tumor Resection

We present cao_cautery, the first long-horizon dataset for autonomous tumor resection on the da Vinci Research Kit (dVRK). This dataset captures a challenging dual-arm resection workflow where one arm retracts tissue while the other performs electrocautery-based cutting. It is designed to support research in robotic surgery, imitation learning, and visual policy learning under realistic intraoperative conditions.

---

Dataset Summary

• 3,640+ dual-arm demonstrations of resection tasks
• Recorded on anatomically constrained phantom tumors
• Each demonstration segmented into structured subtasks
• Includes multi-view video and synchronized robot kinematics
• Demonstrations reflect realistic intraoperative challenges:
  • Soft-tissue deformation
  • Smoke-induced occlusion
  • Shifting illumination and camera perspectives

The dataset supports imitation learning, vision-language-action modeling, and multi-view surgical policy training under varying data regimes.

---

File Structure

Each demonstration episode is stored under:

cao_cautery/ tissue_/ 1_resection/ # Task / left_img_dir/ # Left endoscope images right_img_dir/ # Right endoscope images endo_psm1/ # Wrist camera (right arm) endo_psm2/ # Wrist camera (left arm) ee_csv.csv # Kinematics log

• Images are in .jpg format
• ee_csv.csv contains timestamped poses, joint angles, jaw positions, and control targets for PSM1, PSM2, ECM, and SUJ arms

---

Tissue and Cut Index Mapping

The following table summarizes the mapping between tissue samples and associated cut indices used in the dataset.

	Tissue 1	Tissue 2	Tissue 3	Tissue 4	Tissue 5	Tissue 6	Tissue 7	Tissue 8
Cut #	1, 2, 3, 4, 5	6, 7, 8, 9, 10, 11	12, 13, 14, 15, 16, 17	18, 19, 20, 21, 22	23, 24	25, 26, 27, 28	29, 30, 31	32

Applications

This dataset is designed for:

• Visual imitation learning and offline reinforcement learning
• Learning dual-arm coordination and tool-tissue interaction
• Studying robustness under partial occlusion and dynamic visual changes
• Subtask-conditioned policy learning and benchmarking

---

Format

• RGB image resolution: varies (synchronized across views)
• Kinematics format: CSV with ~100 columns
  • End-effector position/orientation
  • Joint states and commands
  • Jaw and RCM pose
  • PSM1, PSM2, ECM, SUJ poses
• Data regimes supported for training:
  • Full dataset
  • 60% subset
  • 3% few-shot subset

---

Citation

If you use this dataset, please cite:

@misc{cao2025,
  title       = {Dataset and Imitation Learning Framework for Autonomous Tumor Resection},
  author      = {Nural Yilmaz, Juo-Tung Chen, Mariana Smith, Ji Woong Kim, Brendan Burkhart, Axel Krieger},
  year        = {2025},
  note        = {Under review},
  howpublished = {\url{https://huggingface.co/datasets/jchen396/cao_cautery}}
}

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 jchen396@jh.edu Johns Hopkins University