mcpbeat

I4h Workflow Dataset Teleop

nvidia/i4h-workflow-dataset-teleop

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.

5k tokens
context cost
the whole folder, loaded on every use
5
files
instructions only
0
copies elsewhere
how many repositories repackaged it
2778
stars on the repo
on the repository, not the skill itself

Install

one command, takes just this skill from the repository
npx skills add https://github.com/NVIDIA/skills --skill i4h-workflow-dataset-teleop

What comes with it

13 418 bytes besides the instruction
BENCHMARK.md
evals/evals.json
skill-card.md
skill.oms.sig

The instruction itself

16 sections, as written by the author

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):

| Key | Action |

|---|---|

| B | Start episode |

| N | Mark success, save, advance |

| R | Discard, reset |

| F | Reserved 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

| Env | Devices |

|---|---|

| scissor_pick_and_place | keyboard, so101_leader |

| locomanip_tray_pick_and_place | keyboard_23d |

| locomanip_push_cart | keyboard_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.

How to use it

Copy the folder

Take nvidia/i4h-workflow-dataset-teleop from the repository into ~/.claude/skills for personal use, or into .claude/skills inside a project.

Check the name does not clash

The agent identifies a skill by the name field in its header. Two skills with the same name cannot sit side by side — one of them will be ignored.