chrisvoncsefalvay/openembodiment-dvrk-subset

Open-H dVRK convenience subset

This dataset is a convenience subset of nvidia/PhysicalAI-Robotics-Open-H-Embodiment, curated for dVRK/da Vinci surgical robotics experiments and staged so training runs can stream from one place instead of rehydrating the upstream corpus file-by-file.

The NVIDIA dataset is the canonical source: a CC-BY-4.0, LeRobot v2.1 corpus of paired healthcare robotics video and kinematics, with video stored as MP4 and kinematics stored as Parquet. All substantive credit for the source data belongs with NVIDIA and the Open-H Embodiment contributors. Please cite and review the upstream dataset card, licence and terms before using this snapshot. This repository exists to make reproducible dVRK-domain Cosmos post-training less theatrical.

Media and tensors are stored under data/openh_snapshot/; manifests and snapshot metadata are under metadata/.

Intended use

• Domain-specific post-training of action-conditioned video models.
• dVRK action-space validation and short-horizon surgical world-model research.
• Streaming/restart-friendly training on ephemeral GPU instances.

This snapshot is not a clinical dataset release, a new annotation layer or a replacement for the upstream Open-H Embodiment corpus.

Streaming the manifest

from datasets import load_dataset

repo_id = "chrisvoncsefalvay/openembodiment-dvrk-subset"
manifest = load_dataset(
    "json",
    data_files=f"hf://datasets/{repo_id}/metadata/openh_dvrk_manifest.train.jsonl",
    split="train",
    streaming=True,
)
first = next(iter(manifest))

Rehydrating the curated snapshot

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="chrisvoncsefalvay/openembodiment-dvrk-subset",
    repo_type="dataset",
    allow_patterns=["data/openh_snapshot/**", "metadata/**"],
)

The upload path uses Hugging Face Hub Xet storage when available, so resumed uploads/downloads avoid the upstream API-rate-limit bottleneck.

Citation

If you use this convenience subset, cite both this dataset repository and the upstream Open-H Embodiment source corpus. This subset mirrors selected dVRK/da Vinci roots from nvidia/PhysicalAI-Robotics-Open-H-Embodiment; all source data credit remains with Open-H and the original contributors.

@misc{voncsefalvay2026openembodimentdvrksubset,
  title = {Open-H dVRK convenience subset},
  author = {von Csefalvay, Chris},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/chrisvoncsefalvay/openembodiment-dvrk-subset},
  note = {Convenience dVRK/da Vinci subset of nvidia/PhysicalAI-Robotics-Open-H-Embodiment}
}

@misc{nvidia2026physicalairoboticsopenh,
  title = {PhysicalAI-Robotics-Open-H-Embodiment},
  author = {{NVIDIA Corporation} and {Open-H Embodiment community}},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Open-H-Embodiment},
  note = {CC-BY-4.0 Open-H Embodiment LeRobot v2.1 source corpus}
}

@misc{openh2026medicalrobotics,
  title = {Open-H-Embodiment: A Large-Scale Dataset for Enabling Foundation Models in Medical Robotics},
  author = {{Open-H-Embodiment Consortium}},
  year = {2026},
  eprint = {2604.21017},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2604.21017}
}