i4h-workflow-dataset-mimic

par nvidia

Développez un enregistrement HDF5 en clonant des trajectoires avec du bruit d'action/état. Utilisez lorsque l'on vous demande de imiter, développer ou augmenter un ensemble de données ; pas pour enregistrer de nouvelles démonstrations…

npx skills add https://github.com/nvidia/skills --skill i4h-workflow-dataset-mimic

i4h Workflow — Mimic Dataset

Purpose

Expand an HDF5 recording by replicating trajectories with small action and state noise. Use when the user asks to mimic, expand, or augment a dataset without recording new episodes. If the same prompt also asks to visualize the dataset, finish mimic first, then compose [[i4h-workflow-dataset-convert]] and [[i4h-lerobot-viz]] on the mimic output.

Base Code

These steps drive the i4h-workflows base code (the workflows/agentic/ tree). To reuse an existing checkout, set I4H_WORKFLOWS to its path (no clone happens). Otherwise this resolves the current repo, or clones to ~/i4h-workflows — pick that default without prompting. Run every command below from the resolved root:

# Resolve the i4h-workflows base code (provides workflows/agentic/).
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"

Basics

  • Env config (source of truth): workflows/agentic/config/environments/<env>.yaml defines the <env> robot and task the mimicked trajectories replay against.
  • Mimic perturbs action/state, not visuals.
  • Default --include-source keeps the original demos in the output.
  • In a chained workflow after teleop/validate, use the HDF5 produced by that chain. Do not silently fall back to an older same-env recording if the latest/current recording is empty or failed; stop and report that source demos are missing.
  • Count HDF5 episodes with the workflow venv, not system Python (h5py may not be installed globally).

Run

Run the steps below in order. Each step is a separate bash call; variables persist in the local agent's tmux session.

Step 1 — setup and resolve input HDF5

REPO_ROOT="${I4H_WORKFLOWS:-$(git rev-parse --show-toplevel 2>/dev/null)}"; [ -d "$REPO_ROOT/workflows/agentic" ] || REPO_ROOT="$HOME/i4h-workflows"
ENV_ID=scissor_pick_and_place
RUNS_ROOT="${REPO_ROOT}/workflows/agentic/runs"

# Point IN at a real recording to expand (absolute path). Recordings come from teleop or
# validate (which writes data/verify.hdf5 under each runs/eval_* dir). List candidates newest-first:
#   find "${RUNS_ROOT}" -name '*.hdf5' -printf '%TY-%Tm-%Td %TH:%TM  %p\n' | sort -r | head
IN="${IN:-}"
if [ ! -f "${IN}" ]; then
  echo "mimic: set IN to an existing .hdf5 (got '${IN:-<unset>}'). Candidates:" >&2
  find "${RUNS_ROOT}" -name '*.hdf5' -printf '%TY-%Tm-%Td %TH:%TM  %p\n' 2>/dev/null | sort -r | head
  exit 1
fi

# Verify the selected input is a real recording before mimic. A tiny HDF5 or
# `0/N episodes succeeded` teleop artifact is not usable source data.
"${REPO_ROOT}/workflows/agentic/mimic/.venv/bin/python" - "${IN}" <<'PY'
import h5py
import sys
path = sys.argv[1]
with h5py.File(path, "r") as f:
    count = len(f["data"]) if "data" in f else 0
print(f"source episodes: {count}")
if count <= 0:
    raise SystemExit("mimic: input HDF5 has no episodes; record successful demos first")
PY

RUN_DIR="${RUNS_ROOT}/mimic_${ENV_ID}_$(date +%Y%m%d_%H%M%S)"
mkdir -p "${RUN_DIR}/data" "${RUN_DIR}/logs"
ln -sfn "${RUN_DIR}" "${RUNS_ROOT}/.latest"
OUT="${RUN_DIR}/data/demo_mimic.hdf5"

When resolving "my dataset" from prior prompts, inspect candidates newest-first and prefer the current chain's latest successful teleop/validate HDF5. If the newest HDF5 for the target env has 0 episodes, do not skip back to an older run unless the user explicitly asks to reuse that older dataset.

Step 2 — mimic expand

"${REPO_ROOT}/workflows/agentic/mimic/run.sh" --env "${ENV_ID}" \
  --input "${IN}" \
  --output "${OUT}" \
  --episodes 3 \
  --noise-std 0.01 \
  --include-source \
  --overwrite \
  2>&1 | tee "${RUN_DIR}/logs/mimic.log"

Verify

uv --directory "${REPO_ROOT}/workflows/agentic/mimic" run python -c \
  "import h5py; print('episodes:', len(h5py.File('${OUT}','r')['data']))"

Confirm the output episode count equals the source demos plus --episodes. Also confirm the input episode count was greater than zero; mimic output from an empty source is invalid.

If the prompt includes visualization (for example, "Mimic 3 more episodes and visualize my dataset"), continue after this verify step:

  1. Load [[i4h-workflow-dataset-convert]] and set HDF5_PATH="${OUT}" so the augmented HDF5 is converted to LeRobot with --video-codec h264.
  2. Load [[i4h-lerobot-viz]] and set DATASET_DIR to the converted dataset directory (${HF_LEROBOT_HOME}/local/${ENV_ID} by default).
  3. Report both the augmented HDF5 and the visualizer URL.

Prerequisites

  • Workflow set up via [[i4h-workflow-setup]] (the .venv must exist).
  • An existing input HDF5 recording (--input).
  • The env id that produced the recording.

Limitations

  • Perturbs action/state only, not visuals.
  • Augments an existing recording rather than recording new episodes.
  • Output state/action dimensions must match the source.

Troubleshooting

  • Error: .venv not found / mimic fails to launch - Cause: workflow not set up. Fix: run [[i4h-workflow-setup]] first.
  • Error: input recording not found - Cause: wrong or missing --input HDF5 path. Fix: point to the existing recording file.
  • Error: input HDF5 has no episodes - Cause: teleop was launched but no successful episodes were saved. Fix: record successful source demos first; do not substitute an older run without explicit user direction.
  • Error: output already exists / write refused - Cause: --output path is occupied. Fix: choose a new path or pass --overwrite.
  • Error: state/action dimension mismatch on inspect - Cause: --env differs from the env that produced the input. Fix: use the same env id as the source recording.

Final Response

Report input path, output path, generated episode count, noise std, whether source demos were included, and the visualizer URL if visualization was requested.

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…