i4h-workflow-dataset-teleop

작성자: nvidia

에이전트 환경의 에피소드를 원격 조작(키보드, SO-ARM 리더, 또는 VR)을 통해 HDF5로 기록합니다. 사용자가 원격 조작 또는 인간 데모 녹화를 원할 때 사용하세요.

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

i4h Workflow — Teleop Record

Purpose

Record episodes for an agentic env via teleoperation (keyboard, SO-ARM leader, or VR) into HDF5. Use when the user wants to teleop or record human demos.

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>.yamlrobot.type, zenoh.camera_names, and the task for <env>.
  • Teleop runs through arena/run.sh --teleop.
  • Device support is env-specific. Check arena/run.sh --env <env> --help for valid --teleop-device values.
  • Do not run teleop headless unless the user explicitly asks. Teleop is a GUI/hardware handoff: launching the recorder is not the same as recording successful demos.

Controls

Reserved keys (consistent across devices):

KeyAction
BStart episode
NMark success, save, advance
RDiscard, reset
FReserved by Isaac Sim — do not bind

Device-specific keybindings (move, rotate, gripper, mode switches) are printed by the teleop process at startup and vary by --teleop-device. Report them to the user from the log; they cannot drive the sim without them. See "Surface Device Keybindings". For keyboard_23d, the banner can print late, after scene creation and recorder setup; wait for mode/control markers such as BOTH_HANDS, HAND MODE, BASE NAVIGATION MODE, SPECIAL KEYS, or Current Mode before saying the controls are missing.

Stop from terminal:

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

Known Devices

EnvDevices
scissor_pick_and_placekeyboard, so101_leader
locomanip_tray_pick_and_placekeyboard_23d
locomanip_push_cartkeyboard_23d

For other envs, consult arena/run.sh --env <env> --help.

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

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="${REPO_ROOT}/workflows/agentic/runs"
RUN_DIR="${RUNS_ROOT}/teleop_${ENV_ID}_$(date +%Y%m%d_%H%M%S)"
mkdir -p "${RUN_DIR}/data" "${RUN_DIR}/logs"
ln -sfn "${RUN_DIR}" "${RUNS_ROOT}/.latest"

Step 2 — teleop record

Use the foreground form when the human operator is ready to drive immediately and the agent can stay attached until completion:

"${REPO_ROOT}/workflows/agentic/arena/run.sh" \
  --env "${ENV_ID}" \
  --teleop \
  --teleop-device <device> \
  --episodes 3 \
  --record-to "${RUN_DIR}/data/demo.hdf5" \
  2>&1 | tee "${RUN_DIR}/logs/teleop.log"

If no human operator is ready, do not fake demos and do not leave a teleop process behind. You may launch long enough to confirm the visible GUI, recorder, and controls, then stop cleanly and report that recording is waiting on a human operator.

Surface Device Keybindings

The teleop process prints its keybinding table to stdout shortly after launch (look for sections such as Keybindings, Controls, Key Map, BOTH_HANDS, HAND MODE, BASE NAVIGATION MODE, SPECIAL KEYS, WORKSPACE LIMITS, Current Mode, mode-switch lines, or any block enumerating keys/actions). Wait for the table to appear, extract it from the log, and report it to the user before they need to drive the sim.

# After launching teleop and tailing the log:
for i in $(seq 1 180); do
  if grep -qiE 'Keybind|Controls|Key Map|BOTH_HANDS|HAND MODE|BASE NAVIGATION MODE|SPECIAL KEYS|WORKSPACE LIMITS|Current Mode|Mode\] Switched|run complete|Traceback|Error' "${RUN_DIR}/logs/teleop.log" 2>/dev/null; then
    break
  fi
  sleep 1
done
grep -n -A80 -E 'Keybind|Controls|Key Map|BOTH_HANDS|HAND MODE|BASE NAVIGATION MODE|SPECIAL KEYS|WORKSPACE LIMITS|Current Mode|Mode\] Switched' "${RUN_DIR}/logs/teleop.log" | head -120

If the block is multi-section (e.g. BOTH_HANDS, BASE_NAV, LEFT_HAND, RIGHT_HAND modes), include every mode in the report. Append the reserved keys above so the user has one consolidated reference.

Notes

  • --record-to must be absolute. The recorder resolves relative paths against workflows/agentic/arena (its CWD) and writes to a nested orphan dir. ${RUN_DIR}/data/demo.hdf5 built from ${REPO_ROOT} is absolute.
  • Use --save-all-episodes only when failed attempts must be kept.
  • If the prompt asks "Run teleop for N episodes" and no human/operator input is available, launch the visible recorder only for readiness/control verification, stop it cleanly, and report that no demos were recorded; do not claim N episodes were recorded.

Verify

  • ${RUN_DIR}/data/demo.hdf5 exists.
  • Log contains run complete: N/M episodes succeeded, and N must equal the requested episode count before reporting success. A tiny HDF5 with 0/M episodes succeeded is only a failed/empty recording artifact, not a usable dataset.
  • Before final response, verify no unintended teleop process is left running unless the user explicitly asked to keep it open.

Prerequisites

  • Workflow set up via [[i4h-workflow-setup]] (the .venv must exist).
  • A valid env id and a --teleop-device it supports (see Known Devices or arena/run.sh --env <env> --help).
  • The chosen teleop device available (keyboard, SO-ARM leader, or VR).
  • An absolute --record-to HDF5 path (relative paths resolve against arena's CWD).

Limitations

  • Device support is env-specific; not every device works with every env.
  • --record-to must be absolute or the recording lands in a nested orphan dir.
  • By default only successful episodes are saved; use --save-all-episodes to keep failed attempts.
  • Device-specific keybindings are only printed at startup; they cannot be known before launching.

Troubleshooting

  • Error: .venv not found / teleop fails to launch - Cause: workflow not set up. Fix: run [[i4h-workflow-setup]] first.
  • Error: invalid --teleop-device for env - Cause: device not supported by that env. Fix: pick a value from arena/run.sh --env <env> --help (or Known Devices).
  • Error: HDF5 written to an unexpected/nested location - Cause: --record-to was relative. Fix: pass an absolute path built from ${REPO_ROOT}.
  • Error: keybindings missing / cannot drive the sim - Cause: the device keybinding table was not surfaced. Fix: extract the table from the log before the operator starts driving (see Surface Device Keybindings).
  • Error: run exits with 0/N episodes succeeded - Cause: no human/operator completed episodes. Fix: record successful source demos with an operator; do not continue to mimic/convert/finetune from an empty HDF5.

Final Response

Report env, device, the device-specific keybinding table extracted from the log, requested vs saved episodes, HDF5 path, log path.

nvidia의 다른 스킬

compileiq-debug
nvidia
무언가 잘못되었을 때 사용: Search()가 멈추거나, 모든 평가가 INVALID_SCORE를 반환하거나, 점수가 개선되지 않거나, 모든 설정이 동일한 숫자를 반환하거나, ptxas 오류 등이 발생할 때
create-github-pr
nvidia
gh CLI를 사용하여 GitHub 풀 리퀘스트를 생성합니다. 사용자가 새 PR을 만들거나, 코드 리뷰를 제출하거나, 풀 리퀘스트를 열고자 할 때 사용합니다. 트리거 키워드 -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
다른 열린 이슈들을 스캔하여 주어진 PR이 함께 수정하거나 실수로 망가뜨릴 수 있는 이슈를 찾습니다. 인접 수정 기회와 모순 위험을 file:line…과 함께 출력합니다.
fhir-basics
nvidia
에이전트에게 FHIR R4 API의 작동 방식, 사용 가능한 리소스, 검색 매개변수를 사용한 쿼리 방법, 모든 응답 형식을 올바르게 파싱하는 방법을 가르칩니다…
compileiq-validate-result
nvidia
검색이 완료된 후, 속도 향상을 청구하거나 ACF를 발송하기 전에 사용합니다. dump_results CSV를 로드하고, 상위 K개 후보(단일 목표)를 추출합니다…
changelog-audit
nvidia
릴리스 전에 Warp CHANGELOG.md를 감사합니다: 누락된 항목 복구, 사용자 영향별 정렬, 항목 언어 다듬기, 줄 바꿈, (릴리스 브랜치 모드) 비교 업데이트…
maintain-dynamic-plugins
nvidia
NeMo Relay 동적 플러그인 로더, 매니페스트, Rust 네이티브 SDK, gRPC 워커 프로토콜, Python 워커 SDK, 문서, 테스트 및 릴리스 워크플로 커버리지를 유지 관리합니다.
dgx-diagnose
nvidia
일반적인 DGX Station GB300 문제 진단 — CUDA 충돌, 잘못된 GPU 타겟팅, vLLM/SGLang 컨테이너 버그, MIG 상태 문제, NVLink/Fabric Manager 오류,…