LeRobot Dataset Visualizer
Inspect converted LeRobot datasets in a local browser.
The skill mainly creates run directories, logs, and a symlink, then starts a local visualizer; no malware, credential theft, covert exfiltration, or destructive default is shown. Points are deducted because execution confirmation, rollback/cleanup, data-flow, and permission boundaries are incomplete; it may clone a remote repository by default, and the skill card claims an API key is required without explaining its use.
Input validation, prerequisites, port-conflict handling, and common failures are documented, and the benchmark reports limited correctness results. Static review found no committed tests or CI covering this skill's key path; cloning, script dependencies, and stopping behavior may still fail, so the score remains conservative under the static ceiling.
The audience, trigger scenarios, input format, and non-fit HDF5 conversion case are clear. Points are deducted because “latest dataset” has no explicit selection rule, Chinese-language support is not documented, and the GitHub clone dependency may be unreachable or unreliable for mainland-China users.
The skill includes name, version 0.6.0, Apache-2.0 licensing, author, use case, prerequisites, limitations, troubleshooting, and evaluation information. Points are deducted for missing recommended Instructions and Examples sections, no changelog or explicit update path, and inconsistency between the skill card's credential requirement and SKILL.md.
The core output—starting a local LeRobot visualizer and reporting its URL—is directly aligned with the task and does not perform conversion. Benchmark correctness was 90% for Claude and 70% for Codex, but effectiveness was only 44% and 33%; no independently verifiable representative output or explicit stop command is supplied, so static scoring is limited.
The repository provides evals.json, BENCHMARK.md, version metadata, and explicit expected behaviors, giving limited auditability. Points are deducted because evaluation covers only two positive tasks with no negative-trigger cases, the report is not independently reproducible evidence, and no real CI or test suite covers this skill's key paths.
- Before execution, confirm permission to clone the remote i4h-workflows repository and disclose its network, filesystem, and local-port effects.
- Add an explicit stop command, cleanup/rollback procedure, and a consistent explanation of whether an API key is actually required and how it is handled securely.
- Define how “latest converted dataset” is selected and validate GitHub dependency reachability for mainland-China environments.
What does this skill do, and when should you use it?
i4h-lerobot-viz serves an HTML visualizer for a converted LeRobot dataset. It is intended for visualization and inspection, including video and joint-timeline checks, and does not convert HDF5 data. It requires the i4h-workflows environment, a dataset directory containing meta/info.json, and an absolute dataset path. The visualizer prints a local browser URL and writes run logs and visualization state.
Resolves the i4h-workflows root and dataset location, validates that the dataset contains meta/info.json, creates a timestamped run directory with logs and visualization state, and runs workflows/agentic/dataset/viz.sh with the dataset and state directory. The serving command records output to a visualization log and prints a local URL.
- A data engineer validating a converted LeRobot dataset can open it in a browser and inspect its contents.
- A robotics or physical-AI user checking data after video augmentation can load videos and joint timelines.
- A developer with an existing i4h-workflows checkout can point the skill to that checkout through I4H_WORKFLOWS.
- A user diagnosing dataset setup problems can use the meta/info.json validation and startup errors to identify missing prerequisites.
What are this skill's strengths and limitations?
- Focused on browser-based inspection of converted LeRobot datasets.
- Loads videos and joint timelines and preserves logs and visualization state in a run directory.
- Documents common failures involving the workflow environment, dataset metadata, relative paths, and port conflicts.
- Published within NVIDIA's NVIDIA Agent Skills catalog and declares Apache-2.0 licensing for the skill.
- Does not convert HDF5 to LeRobot; conversion must happen through another workflow.
- Requires a configured .venv, an absolute dataset path, and meta/info.json.
- May need network access to clone i4h-workflows from GitHub when no local checkout is available.
- The source provides no platform test matrix, performance measurements, or standalone test suite.
How do you install this skill?
Install the skill through the NVIDIA skills CLI:
npx skills add nvidia/skills --skill i4h-lerobot-viz --yes
The README does not document a standalone package or additional installation command for this skill. The required i4h-workflows environment must be set up separately.
How do you use this skill?
Ensure that i4h-workflows contains workflows/agentic and that its .venv has been set up. Set DATASET_DIR to the absolute path of a converted LeRobot dataset containing meta/info.json; optionally set I4H_WORKFLOWS. From the workflow root, run:
REPO_ROOT="${I4H_WORKFLOWS:-$(git rev-parse --show-toplevel 2>/dev/null)}"; [ -d "$REPO_ROOT/workflows/agentic" ] || REPO_ROOT="$HOME/i4h-workflows"
RUNS_ROOT="${REPO_ROOT}/workflows/agentic/runs"
DATASET_DIR="/absolute/path/to/dataset"
RUN_DIR="${RUNS_ROOT}/viz_$(date +%Y%m%d_%H%M%S)"
mkdir -p "${RUN_DIR}/logs" "${RUN_DIR}/viz_state"
"${REPO_ROOT}/workflows/agentic/dataset/viz.sh" "${DATASET_DIR}" --state-dir "${RUN_DIR}/viz_state" 2>&1 | tee "${RUN_DIR}/logs/viz.log"
Open the local URL printed by the command. The source does not specify an exact stop command; if the port is already in use, stop the existing visualizer process first.