CRCD-dVRK-LeRobot

18 episodes of robot-assisted cholecystectomy on pig liver using the da Vinci Research Kit (dVRK). LeRobot v3.0 format.

Converted from the CRCD dataset (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

# 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

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