I4H Dataset Converter
Convert Isaac for Healthcare HDF5 recordings into training-ready LeRobot datasets.
Commands, input paths, output locations, and the main data flow are visible, with no evidence of malware or credential theft; however, the default clone is unpinned, uses the home directory, executes external code, and enables --overwrite plus symlink creation without confirmation, backup, or rollback. HDF5 may contain sensitive healthcare data, but sanitization, log exposure, and access boundaries are not addressed, so points are deducted.
The workflow is clearly sequenced and provides feedback for missing HDF5, setup failures, and AV1 problems; however, the conversion implementation is outside the supplied evidence, dependencies are not locked, success verification checks only a few files and log text, and abnormal inputs, partial failures, permissions, and recovery are not covered. The benchmark has only two positive attempts, so static evidence does not support a higher score.
The audience, HDF5-to-LeRobot use case, trigger phrases, and non-fit for visualization are reasonably clear; however, the example hard-codes scissor_pick_and_place, the claimed latest-recording case only lists candidates rather than selecting the latest, and supported behavior depends on external environment YAML files. There is no evidence of Chinese-language interaction or mainland-China network reachability, so points are deducted.
The skill provides version, license, author, purpose, prerequisites, limitations, troubleshooting, and final-output guidance, with generally readable structure; however, it lacks the recommended Instructions and Examples sections, changelog, explicit maintenance responsibility, and update path. The skill card describes Markdown output while SKILL.md requests a result report, creating documentation inconsistency.
The core task, commands, and expected LeRobot artifacts are explicit, and the static benchmark reports two positive tasks passing with 79% Codex correctness and 52% effectiveness; however, the sample is only two positive tasks with no negative cases, and no verifiable artifacts or key-path tests are supplied. Users still must validate environment, paths, and outputs, so the score remains below the static ceiling.
Versioned evaluation files, two evaluation tasks, and an external evaluation summary provide limited auditability; however, the selected skill has no committed dedicated tests or CI coverage, and the benchmark lacks raw logs and output artifacts. Key conversion claims therefore cannot be independently reproduced from the supplied files, warranting only limited credit.
- Confirm the external repository source and revision before execution; the default git clone is not version-pinned.
- --overwrite and symlink operations have no backup or rollback; confirm the target directory and protect original HDF5 and existing LeRobot data.
- Healthcare robot recordings may contain sensitive data; review sanitization, log contents, permissions, and storage location.
- The claimed latest-recording scenario does not automatically select the latest HDF5; manually set HDF5_PATH and verify that ENV_ID matches the recording source.
- AV1 can break downstream decord reading; inspect metadata, videos, and skipped or failed episodes after conversion.
What does this skill do, and when should you use it?
This skill converts an agentic HDF5 recording into a LeRobot dataset containing Parquet data, metadata, and videos. It uses the same environment configuration that produced the recording to obtain robot, task, camera, action, state, and dataset defaults. The result is written under HF_LEROBOT_HOME for the selected repo-id. It is intended for conversion and training preparation, not recording visualization.
Takes an existing HDF5 recording, reads robot, task, camera, and dataset settings from the environment YAML, invokes workflows/agentic/dataset/run.sh, and produces Parquet files, meta files, and per-episode videos. It verifies meta/info.json, video presence, and the saved episode count. Conversion uses H.264 video and resizes camera streams according to the environment configuration.
- An Isaac for Healthcare user needs to turn teleop, mimic, or validate HDF5 recordings into LeRobot training data.
- A robotics researcher needs to export a scissor SO-ARM recording with modality.json generated from YAML splits.
- A G1 locomanip or assemble-trocar user needs conversion driven by a modality template specified in the environment YAML.
- A data engineer needs to normalize mixed-resolution camera streams such as head and overview cameras.
What are this skill's strengths and limitations?
- Produces the Parquet, metadata, and video components expected by a LeRobot dataset.
- Uses the environment YAML as the source of truth for robot, task, camera, and modality settings.
- Allows camera output size overrides with --image-size.
- Documents troubleshooting for missing inputs, codecs, and modality configuration.
- Requires an existing HDF5 recording and does not replace the recording workflow.
- The environment must match the one used to create the recording.
- H.264 is required because the default AV1 output breaks GR00T's decord reader during fine-tuning.
- It depends on configured i4h-workflows and a .venv; the supplied material does not document supported platforms or an independent test suite.
How do you install this skill?
Install through the skills CLI: npx skills add nvidia/skills --skill i4h-workflow-dataset-convert --yes. The skill also requires workflow setup through i4h-workflow-setup and an existing .venv; detailed setup steps are not provided here.
How do you use this skill?
Prepare an existing HDF5 file at an absolute path and use the same ENV_ID that produced it. Set I4H_WORKFLOWS to an i4h-workflows checkout containing workflows/agentic/, define HDF5_PATH and HF_LEROBOT_HOME, then run: "${REPO_ROOT}/workflows/agentic/dataset/run.sh" --env "${ENV_ID}" --hdf5-path "${HDF5_PATH}" --repo-id "local/${ENV_ID}" --video-codec h264 --overwrite. Afterward, check meta/info.json, per-episode videos, and the episode count in the log.
How does this skill compare with similar options?
The source explicitly separates conversion from visualization: use i4h-lerobot-viz for viewing recordings.