Isaac for Healthcare Workflow Validator
Run policy or state-machine rollouts and record verification episodes to HDF5.
The skill discloses cloning an external repository, starting a policy daemon and simulator, and writing logs and HDF5 data; it includes cleanup and annotation limits, but default cloning, high-resource process startup, and external effects require no user confirmation. Data-flow disclosure, sensitive-data isolation, dependency pinning, and rollback are incomplete, so points are deducted. No malware, credential theft, or covert exfiltration red line is evident.
The workflow gives fairly complete sequencing, state-machine branching, cleanup traps, log checks, and troubleshooting. However, it depends on an external repository, environment YAML, virtual environments, models, and GPU runtime, and key paths were not executed in this static review. Static calibration therefore caps the score at 10.
Purpose, trigger semantics, inputs, policy evaluation, and state-machine boundaries are reasonably clear, with several evaluation tasks covering intended routing. Non-fit cases, Chinese-language interaction support, and alternatives for unreachable networks are under-specified; the default GitHub clone may be difficult on mainland-China networks, so points are deducted.
The skill provides a name, version 0.6.1, Apache-2.0 license, author, dependencies, limitations, troubleshooting, output description, and evaluation records. It lacks dedicated Instructions and Examples sections, maintenance ownership and changelog detail are incomplete, and repository license metadata is NOASSERTION, leaving governance uncertainty.
The workflow maps environments, episode counts, checkpoints, and timestep caps into policy or simulator runs and produces HDF5 and logs. The committed evaluation summary reports limited correctness and effectiveness, and no directly usable artifact was independently verified in this static review. Static calibration caps effectiveness at 7, with further deduction for incomplete evidence.
Versioned skill files, an evaluation task list, and a dated NVSkills-Eval summary provide some traceability. There are no raw logs, committed key-path tests, CI coverage, or independently reproducible artifacts, so the evidence remains mainly a declarative report and cannot exceed the static-review cap of 5.
- Execution may access or clone an external GitHub repository, start policy and simulator processes, and consume GPU resources; confirm network, resource, and external-effect boundaries in a controlled environment first.
- Handling rules for model, environment, and HDF5 data flows—including potentially sensitive or patient-related data—are incomplete; do not place credentials or sensitive data in prompts, logs, or outputs.
- The evaluation report lacks raw artifacts and reproducible tests, so its PASS verdict and percentages should not be treated as independently verified results.
What does this skill do, and when should you use it?
This skill validates, evaluates, or runs policies, checkpoints, and scripted state-machine controllers in Isaac for Healthcare environments. It reads policy and Arena defaults from an environment YAML file, launches the required processes, and records verification episodes. Outputs include an HDF5 recording plus Arena and policy logs. It requires a prepared i4h-workflows checkout and a valid environment and model source.
Resolves the i4h-workflows checkout and reads workflows/agentic/config/environments/<env>.yaml for the model source, task description, health port, and Arena timestep default; creates an evaluation run directory; starts the policy daemon and Arena, or runs Arena with --state-machine when explicitly requested; executes the requested number of episodes and timesteps; writes verify.hdf5 to an absolute path; and inspects Arena and policy logs. The optional VLM annotator is run only when the user requests success labels.
- An Isaac for Healthcare user needs to validate a default environment policy over one or more episodes and save the rollout data.
- A checkpoint user needs to evaluate a new or latest checkpoint for a specified timestep cap.
- A developer needs a surgical state-machine smoke run using Arena without a policy daemon.
- An operator needs to confirm that the policy daemon, Arena, recorder, and log outputs work together.
What are this skill's strengths and limitations?
- Covers policy, checkpoint, and scripted state-machine validation.
- Reads model and Arena defaults from the environment configuration as the source of truth.
- Produces an HDF5 episode recording and separate policy and Arena logs.
- Documents cleanup of the policy daemon after Arena exits, including failed runs.
- Requires a prepared i4h-workflows checkout, Python virtual environment, matching environment YAML, and a model source.
- Policy evaluation requires both the policy daemon and Arena; Arena opens a simulation window by default.
- The --record-to destination must be absolute to avoid nested orphan directories.
- VLM annotation is outside the default rollout and must be explicitly requested.
How do you install this skill?
Install the skill from NVIDIA's collection with: npx skills add nvidia/skills --skill i4h-workflow-validate --yes. For Codex, use: npx skills add nvidia/skills --skill i4h-workflow-validate --agent codex --yes. Then complete the i4h-workflow-setup prerequisite; its setup procedure is not included here.
How do you use this skill?
Ask the agent, for example, “Validate the scissor_pick_and_place policy for 1 episode” or “Run a state-machine smoke test for 200 timesteps.” Before execution, set I4H_WORKFLOWS when reusing an existing checkout, choose ENV_ID and EPISODES, and set MAX_TIMESTEPS when requested. Follow Step 1 and then Step 2A for policy evaluation or Step 3 for state-machine runs. Policy evaluation must keep policy/run.sh --ensure and arena/run.sh in one controlled shell, then inspect RUN_DIR/data/verify.hdf5, arena.log, and policy.log.