morozovdd/CRCD-dVRK-LeRobot

---
license: cc-by-4.0
task_categories:
- robotics
size_categories:
- n<1K
tags:
- LeRobot
- robotics
- surgical-robotics
- dVRK
- cholecystectomy
configs:
- config_name: default
data_files: data/*/*.parquet
---

# CRCD-dVRK-LeRobot

18 episodes of robot-assisted cholecystectomy on pig liver using the da Vinci Research Kit (dVRK). [LeRobot](https://github.com/huggingface/lerobot) v3.0 format.

Converted from the [CRCD dataset](https://github.com/borgioli/crcd) (University of Verona / Bologna).

| | |
|---|---|
| Episodes | 18 (7 surgeons, 1-3 attempts each) |
| Frames | 377,950 |
| FPS | 30 |
| Robot | dVRK |
| Task | Cholecystectomy (ex vivo, pig liver) |
| Camera | Stereo endoscope, 848x480 |
| State | 16D — PSM1/PSM2 cartesian pose (xyz + quaternion) + jaw |
| Action | 16D — next-timestep absolute cartesian |
| Splits | train: 0-14, test: 15-17 |

## Usage

```python
# With LeRobot
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("morozovdd/CRCD-dVRK-LeRobot")

# With GR00T-H (needs v3 → v2 conversion)
python scripts/lerobot_conversion/convert_v3_to_v2.py \
--input-path CRCD-dVRK-LeRobot --output-path crcd_v2

torchrun --nproc_per_node=1 gr00t/experiment/launch_finetune.py \
--base-model-path nvidia/GR00T-H \
--dataset-path crcd_v2 \
--embodiment-tag HAMLYN_DVRK_30HZ \
--global-batch-size 2 --max-steps 2000
```

## Conversion Details

- Source: CRCD parquet files with embedded WEBP stereo images + 104 kinematic columns
- Extracted left endoscope → H.264 MP4
- Extracted PSM1/PSM2 cartesian pose + jaw → 16D state/action vectors
- Quaternion convention: xyzw (dVRK) → wxyz (LeRobot/GR00T-H)
- Downsampled ~55Hz → 30Hz
- Conversion script included in `scripts/`

## Citation

```bibtex
@article{borgioli2024crcd,
title={The Comprehensive Robotic Cholecystectomy Dataset (CRCD)},
author={Borgioli, Lorenzo and others},
journal={arXiv preprint arXiv:2412.12238},
year={2024}
}
```