What does this skill do, and when should you use it?
This skill expands HDF5 recordings used in Isaac for Healthcare workflows. It clones existing trajectories and adds small action and state perturbations to create additional episodes. It requires an existing input HDF5 containing successful episodes and the environment ID that produced the recording. It does not record new demonstrations or alter visual data.
The skill resolves the i4h-workflows base code, checks that the input HDF5 exists and contains episodes, creates a mimic run directory, and invokes mimic/run.sh to expand the dataset. The documented example targets scissor_pick_and_place, adds 3 episodes with a noise standard deviation of 0.01, and preserves the source episodes. It then uses the workflow virtual environment to verify that the output episode count equals the source count plus the requested additions.
- An Isaac for Healthcare robotics dataset developer wants more training trajectories from existing HDF5 demonstrations.
- A user who has completed teleoperation or validation and has successful recordings wants to augment the dataset.
- A user needs simulated action and state variation without recording new demonstrations.
- A user who also requests visualization can mimic first, then convert the augmented HDF5 and launch the visualization workflow.
How do you install this skill?
- The default commands clone an unpinned workflow repository from GitHub when absent; confirm network access, source revision, and user authorization before execution.
- --overwrite can replace the specified output; use an explicit new output path and preserve the input as a backup.
- Robot-trajectory HDF5 files may contain sensitive or regulated data; the skill does not specify anonymization, access control, or log-data handling.
- The benchmark has only two positive tasks, no negative tasks, and no raw execution evidence; it should not be treated as comprehensive validation.
- Shell / CLI
- Network access
- Local filesystem
GitBashPython with h5pyuvi4h-workflows
Install the skill from NVIDIA's collection with: npx skills add nvidia/skills --skill i4h-workflow-dataset-mimic --yes. Alternatively, run npx skills add nvidia/skills and select the skill interactively. The source does not document a separate package installation for i4h-workflows; the skill reuses the current checkout or clones it to ~/i4h-workflows when needed.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/i4h-workflow-dataset-mimic" ~/.claude/skills/
rm -rf "$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?
Once installed, send your agent any of these to trigger it:
- augment my HDF5 dataset
- mimic 3 episodes with 0.01 noise,
Ask the agent to do something such as “augment my HDF5 dataset” or “mimic 3 episodes with 0.01 noise,” and provide an existing input HDF5 plus its environment ID. The input must exist and contain at least one episode. The documented command uses ENV_ID=scissor_pick_and_place, --episodes 3, --noise-std 0.01, --include-source, and --overwrite. Do not use this skill to record new demonstrations.
What are this skill's strengths and limitations?
- Documents a complete flow from input validation through augmentation and episode-count verification.
- Keeps source demonstrations by default, allowing original and generated data to be used together.
- Explicitly rejects empty HDF5 files and failed recordings as invalid sources.
- Can be chained with documented conversion and visualization skills when requested.
- Requires an existing HDF5 recording with successful episodes.
- The environment ID, state dimensions, and action dimensions must match the input recording.
- Perturbs actions and states only; it does not augment or change visual data.
- The source provides no standalone test suite, platform coverage, or performance benchmarks.
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 Mimic Dataset Augmentation this page ✓ 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 Annotation Workflow ✓ NVIDIA · Official | 44 · Not recommended | ★ 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 |
The skill explicitly distinguishes augmenting an existing dataset from recording new demonstrations; new demonstrations should use i4h-workflow-dataset-teleop. The source provides no other verifiable competitor or alternative comparison.
How did FollowSkills review this skill?
The skill documents inputs, outputs, noise parameters, and source-data checks, and avoids silently falling back from empty data; however, it clones from GitHub, creates directories and a symlink without explicit user confirmation, and uses --overwrite without a rollback procedure. Permission scope, data-flow risks, and sensitive-data handling are incomplete, so points are deducted.
The ordered workflow, input validation, environment consistency check, and troubleshooting guidance are useful; however, the dependency repository is not version-pinned, key scripts and runtime details are not included in the selected skill path, and static review cannot verify execution. The score is conservatively bounded by the static cap.
The audience, intended augmentation use cases, non-fit recording scenario, HDF5 input, environment ID, and optional visualization chain are clear; however, the skill is narrowly tied to the i4h workflow, trigger boundaries are only partially specified, and missing dependencies may require GitHub access that can be unreliable or unreachable from mainland China, so points are deducted.
Version 0.6.0, Apache-2.0, author, dependencies, limitations, troubleshooting, and final reporting fields are documented. Points are deducted for missing recommended Instructions and Examples sections, no explicit changelog or maintenance/update path, and limited concrete parameter examples. Repository license metadata is NOASSERTION, which further weakens governance clarity.
The skill provides a directly actionable path from an existing recording to an augmented HDF5, episode-count verification, and optional visualization. The benchmark reports passing two positive tasks and some correctness uplift; however, Codex correctness was 65% and effectiveness 27%, the dataset had only two positive tasks and no negative cases, and no directly inspectable output is included. The static ceiling therefore applies.
The pinned source contains auditable instructions, evaluation tasks, and a results summary; however, it lacks committed key-path tests, raw execution logs, or independently reproducible artifacts. The benchmark summary alone supports only limited verification, so the score remains low.
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 →