i4h-workflow-validate

Validate Isaac for Healthcare policies by running rollout episodes and recording verification HDF5 data.

3.2k|370|Updated Feb 25, 2026
One-click install
npx skills add https://github.com/NVIDIA/skills --skill i4h-workflow-validate
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: i4h-workflow-validate
Source: https://github.com/NVIDIA/skills/tree/main/skills/i4h-workflow-validate
Command: npx skills add https://github.com/NVIDIA/skills --skill i4h-workflow-validate

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Validating a trained robot policy or checkpoint in Isaac for Healthcare (i4h) requires coordinating a policy daemon and the Arena simulator, configuring environment YAMLs, and capturing verification episodes to HDF5 — a multi-step process that is easy to get wrong, especially in non-interactive agent sessions.

Core Features & Use Cases

  • Policy and Checkpoint Rollouts: Launch the policy daemon and Arena together to evaluate a policy or a specific checkpoint-NNNN/ directory against an i4h environment, recording episodes to verify.hdf5.
  • State-Machine Smoke Runs: Run scripted surgical state-machine controllers (e.g., surgical_reach_psm) via Arena --state-machine without starting a policy daemon.
  • Non-Interactive Session Handling: Provides a controlled single-shell rollout (Step 2A) for Claude Code --print and Codex exec sessions so the policy daemon is not orphaned between tool calls.
  • Use Case: Ask your agent to "run eval using new checkpoint for 300 timesteps" and it resolves the latest checkpoint, sets MAX_TIMESTEPS=300, runs the rollout, and reports episodes saved plus log paths.

Quick Start

Ask your agent to evaluate scissor pick and place for 2 episodes and it will run the i4h validation workflow and record the verification HDF5.

Frequently Asked Questions about i4h-workflow-validate

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I evaluate an Isaac for Healthcare policy checkpoint?▼

Set ENV_ID to the environment YAML id and MODEL_PATH to a checkpoint-NNNN directory containing safetensors weights, experiment_cfg, and processor files. The skill starts the policy daemon, runs Arena for the requested episodes, and records results to verify.hdf5.

How do I run a surgical state-machine smoke test in i4h?▼

Set STATE_MACHINE=true and run Arena with the --state-machine flag for environments like surgical_reach_psm or surgical_lift_needle. This path does not start a policy daemon and is intended for quick scripted smoke runs.

Can I run i4h policy evaluation in Claude Code or Codex non-interactive sessions?▼

Yes, but you must use the controlled Step 2A flow, which runs the policy daemon and Arena in one foreground bash command with a cleanup trap. Launching them in separate tool calls or background tasks can orphan the daemon before Arena connects.

Why does verify.hdf5 end up in a nested orphan directory?▼

This happens when --record-to is given a relative path, because the recorder resolves it against the workflows/agentic/arena working directory. Always pass an absolute path under the run directory's data folder.

What are the limitations of i4h workflow validation?▼

Both the policy daemon and Arena are required for policy evaluation, and the daemon must be stopped afterward to avoid leaking GPU memory. The assemble_trocar environment is inference-only, and the VLM annotator runs only when explicitly requested.