Isaac for Healthcare Dataset Replay
Replay HDF5 episodes in Isaac Sim to verify recorded robot behavior visually.
The skill keeps the HDF5 input local and records replay logs under the runs directory; no credential theft, covert exfiltration, or malware is evident. However, it may clone code from GitHub, create directories and a symlink, and launch Isaac Sim without an explicit confirmation step. It does not address potentially sensitive healthcare data, permission boundaries, or rollback, so points are deducted.
The ordered steps, environment variables, file validation, candidate listing, episode-index guidance, and troubleshooting are coherent. Still, availability of i4h-workflows, Isaac Sim, the virtual environment, and the matching environment is not verified by this file; ENV_ID is hard-coded, and no executable tests cover this skill's key path. Static calibration therefore limits the score.
The trigger, HDF5 input, environment identifier, episode index, and non-fit cases are mostly clear, including ordinal-to-zero-based index mapping. The core workflow depends on Isaac Sim and GitHub-hosted code, with no mainland-China reachability or offline alternative guidance and no Chinese-language support, so points are deducted.
The skill provides a name, version, purpose, dependencies, limitations, troubleshooting, final-output requirements, Apache-2.0 licensing, owner information, and an evaluation report. It lacks the recommended Instructions and Examples sections, provides limited maintenance/change-path detail, and repository-level NOASSERTION metadata weakens governance clarity.
The intended task—replaying an HDF5 recording in Isaac Sim for visual verification—has directly usable commands and a defined final report. However, the supplied evaluation covers only two positive tasks, has no negative-trigger cases, reports 41% effectiveness, and includes no replay artifacts or key-output evidence; only limited credit is justified.
The skill includes concrete commands, failure conditions, and a pinned skill version, while BENCHMARK.md supplies a limited third-party evaluation summary. There are no replay logs, committed tests for this skill, or independent reproduction artifacts; the evaluation has only two tasks and no negative cases, so key claims remain weakly corroborated.
- Before execution, verify that I4H_WORKFLOWS, the HDF5 path, and the environment source are trusted; cloning, directory/symlink creation, and simulator launch should require user confirmation.
- HDF5 recordings may contain healthcare or other sensitive data; restrict file and log permissions and avoid uncontrolled disclosure of paths or contents.
- GitHub and Isaac Sim dependencies may be affected by mainland-China network access, licensing, or hardware constraints; provide reachable mirrors or an offline installation path.
- The evaluation covers only two positive tasks and does not establish error recovery, negative triggering, or real visual-result quality.
What does this skill do, and when should you use it?
This skill replays a recorded HDF5 episode inside Isaac Sim for visual verification. It drives the Arena replay command from the i4h-workflows codebase and selects an episode by zero-based index. An existing HDF5 recording is required, and the environment ID must match the environment that produced it. Each invocation replays one episode and does not modify or expand the recording.
Resolves or reuses an i4h-workflows checkout and clones it from GitHub when needed; checks that HDF5_PATH points to an existing absolute-path recording; creates a timestamped replay directory and log directory; runs arena/run.sh --replay with the selected environment, HDF5 path, and episode index; writes command output to replay.log; and reports the environment, recording path, episode index, launch outcome, and visible mismatches.
- An Isaac for Healthcare user wants to visually inspect an HDF5 recording produced by teleoperation, mimic, or validation.
- A robotics data engineer wants to verify a recording before conversion or training.
- A user needs to replay a specific episode, such as the first or second episode, using zero-based indexing.
- A user needs to diagnose an invalid HDF5 path, an out-of-range episode index, a missing workflow setup, or an environment mismatch.
What are this skill's strengths and limitations?
- A focused workflow for visual replay of HDF5 episodes in Isaac Sim.
- Provides explicit zero-based episode selection, including a second-episode example.
- Checks the recording path and creates a separate run and log directory.
- Documents common failures involving setup, paths, episode indexes, and environment IDs.
- An existing HDF5 recording is mandatory; the skill does not provide a default demo.hdf5.
- Requires Isaac Sim, i4h-workflows, and a configured workflow environment; the source gives no platform matrix or test-suite evidence.
- Only one episode is replayed per invocation, and replay is limited to visual verification.
- Requires shell and filesystem access, and may require network access to clone i4h-workflows.
How do you install this skill?
Install the individual skill through the NVIDIA/skills skills CLI:
npx skills add nvidia/skills --skill i4h-workflow-dataset-replay --yes
The README does not describe a separate package for this skill. After installation, an agent can load it when a relevant task is encountered.
How do you use this skill?
First configure the workflow with i4h-workflow-setup, ensure .venv exists, and provide an existing HDF5 recording using an absolute path. Example trigger: "Replay the second episode from this HDF5 recording in Isaac Sim." For the second episode, set EPISODE_INDEX=1; use 0 for the first episode. The skill then sets HDF5_PATH, ENV_ID, and a run directory, and executes:
"${REPO_ROOT}/workflows/agentic/arena/run.sh" --env "${ENV_ID}" --replay "${HDF5_PATH}" --episode-index "${EPISODE_INDEX}"
ENV_ID must be the environment that produced the HDF5 recording.