What does this skill do, and when should you use it?
This skill supports dataset annotation for Isaac for Healthcare workflows. It uses a vision-language model to determine whether each episode satisfies the environment’s task description. The results can label demonstrations, filter unsuccessful data, or gate fine-tuning with a success classifier. It supports offline HDF5 recordings and live judging from an already running policy or Arena session.
Reads the task description from the i4h-workflows environment YAML, selects or accepts a specified HDF5 recording, and sends offline episodes or live camera frames to an OpenAI-compatible vision endpoint. It creates a run directory under workflows/agentic/runs/, writes annotations.jsonl and logs, optionally produces filtered.hdf5, and reports success and failure counts.
- A data engineer needs labels for every episode in one recorded HDF5 file.
- A robotics researcher wants to filter successful teleoperation or imitation-learning demonstrations.
- A training team needs a success-classifier gate before fine-tuning.
- A user has a running policy or Arena session and wants live judgments of recent camera frames.
How do you install this skill?
- Before execution, explicitly confirm repository cloning, VLM start/stop actions, and whether HDF5 images may be sent to an external endpoint; de-identify healthcare data and document the data flow.
- Make ENV_ID, endpoint, model, and input recording explicit user parameters, and document a local deployment path when mainland-China network access is unavailable.
- Treat VLM labels as classifications requiring human review, not as the sole basis for medical decisions or training-data quality gates.
- Shell / CLI
- Network access
- Local filesystem
Giti4h-workflowsHDF5 recordingOpenAI-compatible vision-model endpoint
Install the single skill from NVIDIA’s catalog:npx skills add nvidia/skills --skill i4h-workflow-dataset-annotate --yes
The README does not document the complete installation procedure for i4h-workflows; the skill requires the workflow to be set up with an existing .venv.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/i4h-workflow-dataset-annotate" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Prepare i4h-workflows, an absolute-path HDF5 recording, and a reachable vision endpoint. For offline annotation, use workflows/agentic/annotator/run.sh with offline --hdf5-path <path> --filter <output>; for live judging, use live --count 5 --interval 2.0 --timeout 30.0. Set VLM_BASE_URL, VLM_MODEL, and HDF5_PATH to override the service, model, and recording. Example trigger: Run Annotation on all recorded episodes.
What are this skill's strengths and limitations?
- Supports both offline HDF5 annotation and live camera-frame judging.
- Produces JSONL annotations and an optional filtered HDF5 dataset.
- Uses the environment YAML as the task-description source, with per-run override support.
- Explicitly prevents use of text-only models that cannot process image inputs.
- Requires configured i4h-workflows, an existing `.venv`, an HDF5 recording, and a reachable vision endpoint.
- A text-only model cannot perform the annotation; the default endpoint is localhost:8000/v1.
- Live mode requires an already running policy or Arena session.
- The source provides workflow instructions but no independent test results or broad platform-coverage evidence.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| i4H Dataset Annotation Workflow this page ✓ NVIDIA · Official | 44 · Not recommended | ★ 3.8k | today | Apache-2.0 |
| I4H Mimic Dataset Augmentation ✓ NVIDIA · Official | 52 · Use with care | ★ 3.8k | today | Apache-2.0 |
| I4H Teleoperation Dataset Recorder ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
| I4H Dataset Converter ✓ NVIDIA · Official | 40 · Not recommended | ★ 3.8k | today | Apache-2.0 |
| LeRobot Dataset Visualizer ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
How did FollowSkills review this skill?
The document constrains artifacts to the run directory, distinguishes local and external VLMs, and warns against exposing secrets. However, it can automatically clone a remote repository, start services, and send potentially sensitive healthcare data to a configurable external endpoint without explicit confirmation, minimization, redaction, boundary disclosure, or rollback guidance, so points are deducted.
The workflow gives detailed path resolution, HDF5 selection, artifact checks, and troubleshooting. However, it depends on an external checkout, a running VLM, environment-specific configuration, and unexecuted shell behavior; static review cannot establish reproducibility, so the score remains at or below the static ceiling of 10.
Triggers, offline/live modes, and non-fit cases are reasonably clear. However, ENV_ID is hardcoded to scissor_pick_and_place, input and environment assumptions are narrow, documentation is English-only, and reachability of GitHub and external model services from mainland China is not addressed, so points are deducted.
The skill provides name, version, license, owner, dependencies, limitations, troubleshooting, and final-output guidance. It lacks the recommended Instructions and Examples sections, an explicit changelog, clear maintenance ownership and update path, and fully expands neither the referenced setup skill nor all dependency assumptions, so points are deducted.
It plausibly completes the core annotation, filtering, and summary task and defines output artifacts. However, VLM judgment quality, portability across environments, and abnormal-data behavior are not verified; the benchmark has only two positive tasks and reports effectiveness around 45% to 48%, and static calibration limits this dimension to 7, so points are deducted.
The supplied evals.json, benchmark report, and static-validation findings provide limited audit evidence. There is no skill-specific committed test suite or CI reproducing the key paths, and the benchmark claims cannot be independently reproduced from the supplied files alone, so only limited credit is justified.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →