NVIDIA logo

Skill

i4h-workflow-create

create new agentic environments

Published by NVIDIA Updated Jul 23
Covers Robotics NVIDIA Engineering

Description

Create a new agentic env by forking an existing env. Use for new env/task scaffolding, not scene edits or baking.

SKILL.md

i4h Workflow - Create Env

Purpose

Create the first runnable version of a new workflows/agentic environment by forking the closest existing env. Keep this skill focused on env scaffolding: YAML, assets, task, env class, runtime, and validation. Do not use this skill to polish an existing scene, add optional props/cameras, or bake bridge edits; use [[i4h-workflow-scene-edit]] for that.

Base Code

Resolve and work from the i4h-workflows root:

ROOT="${I4H_WORKFLOWS:-$(git rev-parse --show-toplevel 2>/dev/null)}"
if [ ! -d "$ROOT/workflows/agentic" ]; then
  ROOT="${I4H_WORKFLOWS:-$HOME/i4h-workflows}"
  [ -d "$ROOT/workflows/agentic" ] || git clone https://github.com/isaac-for-healthcare/i4h-workflows "$ROOT"
fi
export I4H_WORKFLOWS="$ROOT"; cd "$ROOT"

What to Load

Before editing, load only the references that match the request:

  • Always load skills/i4h-workflow/references/repo-map.md.
  • Always load skills/i4h-workflow-create/references/create-contract.md.
  • Load skills/i4h-workflow-create/references/env-authoring-patterns.md to choose the source env, robot owner, policy stack, and YAML pattern.
  • Load skills/i4h-workflow-create/references/hybrid-layout-rules.md only for hybrid scene+robot envs, G1 footprint/table-height work, catalog USD scale, or support-surface layout.
  • Load a recipe file only when the prompt matches it exactly enough to avoid re-deciding components.

Recipe routing:

Prompt shapeReference
Surgical tool sorting using G1 based on scissor_pick_and_placeskills/i4h-workflow-create/references/g1-surgical-tool-sort.md

Create Contract

Create the env shell first. A normal create task should produce exactly these surfaces unless the chosen existing pattern requires an explicit exception:

SurfacePath
Env YAMLworkflows/agentic/config/environments/<env>.yaml
Assetsworkflows/agentic/arena/arena/assets/<env>.py
Taskworkflows/agentic/arena/arena/tasks/<env>.py
Env classworkflows/agentic/arena/arena/environments/<env>_environment.py
Runtimeworkflows/agentic/arena/arena/runtimes/<env>.py

Keep all paths repo-root relative and keep the workflows/agentic/ prefix. Do not create a new policy package, README, docs, or shared-module edits unless the user explicitly asks and the policy stack truly needs it.

Workflow

  1. Resolve components: env id, source env, scene/assets source, robot owner, policy stack, model/checkpoint, cameras, objects/destinations, success rule.
  2. Inspect the selected source env YAML, env class, assets, task, runtime, and policy stack files before writing code.
  3. Fork the closest working pattern. Preserve inline-scene vs registry-asset style; do not invent a new architecture.
  4. Keep the first version minimal and runnable. Optional visual polish, extra props, new cameras, and layout changes happen through scene-edit after the env exists.
  5. Run validation and fix source until the env builds and the rendered scene is visually sane.

For an exact recipe match, do not re-derive the architecture after loading the recipe. Read only the source files the recipe names, skip large source runtime files when the recipe says the runtime is a re-export, then write the five contract files immediately. Do not inspect third-party framework internals unless a validation error requires it.

Plan shape:

Env id:
Source env / recipe:
Scene/assets source:
Robot owner:
Policy stack + model/checkpoint:
Objects/destinations:
Success rule:
Files to create:
Validation:

Hard Rules

  • Env YAML is the source of truth for robot, policy, cameras, task text, dataset mapping, and train defaults.
  • Fork from the nearest existing implementation. For hybrid envs, robot integration comes from the robot owner and scene construction comes from the scene source.
  • Sorting tasks need at least two object types, at least two destinations, and a success rule that fails swapped placements.
  • Static/kinematic destination props may not expose .data.root_pos_w; success checks must fall back to entity.get_world_poses().
  • Functions referenced by func= inside config classes must be defined above those classes.
  • G1 WBC envs need footprint clearance and matched ground/base-height values; load hybrid-layout-rules.md before choosing those numbers.
  • Additional fixed cameras are normally scene-edit/bake work. If a create prompt explicitly requires a policy/dataset camera, load the scene-edit camera and bake references and wire every surface in one pass.

Validation

Run these static checks:

python -m py_compile <changed-python-files>
workflows/agentic/policy/run.sh --list-envs
workflows/agentic/arena/run.sh --env <env> --dry-run
workflows/agentic/policy/run.sh --env <env> --dry-run

Then run the real build/visual gate from skills/i4h-workflow-create/references/create-validation.md. --dry-run is shallow: it does not prove the task/assets instantiate or the scene is visually usable. Do not report a new env as ready until the bridge reaches ready, key objects are valid, a viewport capture has been inspected, and the bridge is stopped with:

workflows/agentic/arena/stop.sh --env <env>

If Isaac Sim cannot launch on the host, report that as a blocker and include the static check results; do not present static-only as success.

Hand Off to Scene Edit

After the env shell passes create validation, use [[i4h-workflow-scene-edit]] for:

  • Adding, moving, resizing, or replacing objects/assets.
  • Adding fixed room/overhead/wrist cameras.
  • Live bridge edits.
  • Baking live changes into source.
  • Running local-agent/validate-bake.sh.

The scene-edit workflow owns object snippets, camera snippets, bridge endpoint details, and bake checklists so this create workflow stays small.

Prerequisites

  • Workflow setup has completed via [[i4h-workflow-setup]]; .venv, third-party checkouts, and Isaac Sim launch support are present.
  • The source env and target env id are known, or the prompt provides enough information to choose them from existing patterns.
  • Bridge validation runs on a GPU host that can launch Isaac Sim.

Limitations

  • Create one env per invocation.
  • Fork existing patterns; do not invent a new policy stack or shared framework.
  • Use scene-edit for post-create polish, optional props, camera baking, and source persistence of live edits.
  • --dry-run is not a substitute for the bridge build and visual validation gate.

Troubleshooting

  • Missing setup: if .venv, third-party checkouts, or run.sh entrypoints are missing, run [[i4h-workflow-setup]] first.
  • Env not listed: check that the YAML exists at workflows/agentic/config/environments/<env>.yaml and all files use the full workflows/agentic/ repo-root prefix.
  • Build fails after static checks pass: inspect the first Isaac/arena stack trace; common causes are bad cfg kwargs, a helper defined below a config class, or an observation pointing at a missing sensor.
  • Scene looks wrong: fix source or use scene-edit to live-adjust and bake; do not report success from static checks alone.

Final Response

Report the env id, source choices, files created, validation commands and results, capture paths, and any blocker. Do not commit changes unless the user explicitly asks.

© 2026 YourAI.tools. Every skill from an identity-verified publisher.

Independent catalog. Not affiliated with, endorsed by, or sponsored by Anthropic or any listed publisher. All trademarks belong to their respective owners.