Skip to content

SKILL.md

Latest commit

 

History

History
109 lines (87 loc) · 4.51 KB

SKILL.md

File metadata and controls

109 lines (87 loc) · 4.51 KB
Raw
name run-data-collection
description Launch a data_collection automated trajectory-collection task on a GPU host using the `geniesim autocollect run` CLI verb (which wraps scripts/run_data_collection.sh: docker run -d + in-container server+client). Trigger: when the user asks to "采集数据", "跑数据采集", "run data collection", "collect a task", "生产轨迹", "launch a tasks/geniesim_2025/<...>.json", or wants to produce agibot-format episodes from a data_collection task template.
license MPL-2.0
metadata
author version
genie-sim
1.0

When to Use

  • User wants to produce trajectory episodes from a data_collection task template on a workstation with Docker + an NVIDIA GPU.
  • User references a task under source/data_collection/tasks/.

Do not use for:

  • Running a benchmark/evaluation task → run-benchmark.
  • Just listing/inspecting tasks → geniesim autocollect list directly.

Critical Patterns

  1. run is host-orchestrated, not an in-container exec. It shells out to scripts/run_data_collection.sh, which does docker run -d against geniesim3-data-collection:latest and the entrypoint launches two processes (Isaac Sim server + task client). Don't treat it like benchmark run.
  2. Collect the inputs first: the task (basename / path / unique substring) and the run flags (--headless, --no-record, --standalone, --container-name). Use --dry-run to confirm resolution before launching.
  3. Prerequisites: Docker + NVIDIA GPU; the image registry.agibot.com/genie-sim/geniesim3-data-collection:latest built/pulled; geniesim_assets pip-installed (editable) on the host — the CLI discovers it via find_spec and bind-mounts it at /geniesim_assets.
  4. Unattended works. run_data_collection.sh grants uid 1234 access preferring sudo setfacl, degrading to chmod -R a+rwX when sudo isn't usable — so headless/background runs work without a tty. (The fallback world-writes the output dirs on the host.)
  5. Confirm before launching. A real run spawns a GPU container, takes minutes, and writes ~1.5 GB per episode. Ask before kicking it off.

Workflow

Step 1 — Resolve the task

geniesim autocollect list --robot=g2 <substr>     # discover
geniesim autocollect run <TASK> --headless --standalone --dry-run   # preview

--dry-run prints the resolved task path + the exact run_data_collection.sh command without launching. Disambiguate if it reports multiple matches.

Step 2 — Check prerequisites

docker images | grep geniesim3-data-collection      # image present?
nvidia-smi                                             # GPU free?
python3 -c "import importlib.util as u; print('geniesim_assets OK' if u.find_spec('geniesim_assets') else 'NOT INSTALLED')"   # assets pkg editable-installed?

Step 3 — Launch

Interactive terminal (sudo can prompt):

pip install -e /path/to/geniesim_assets   # once on the host (editable)
geniesim autocollect run <TASK> --headless --standalone

Unattended / detached (no tty) — works directly (the script degrades to chmod when sudo is unavailable; no PTY trick needed):

cd <repo-root>
PYTHONPATH=source/geniesim_cli/src \
  nohup python3 -m geniesim_cli autocollect run <TASK> --headless --standalone \
  > /tmp/dc-run.log 2>&1 &

(Use python3 -m geniesim_cli … if the geniesim console script isn't on PATH.)

Step 4 — Monitor & verify

tail -f source/data_collection/logs/<TASK>/data_collector_server.log   # Isaac Sim startup
tail -f source/data_collection/logs/<TASK>/run_data_collection.log     # stages / TASK SUCCESS / job done
docker ps | grep data_collection                                       # container up
ls source/data_collection/recording_data/                              # episodes landing

Success looks like job done in the client log, the container auto-removed (EXIT trap), and one recording_data/[{TASK}_{INDEX}]/ dir per episode with aligned_joints*.h5, observations/videos/*, state.json, data_info.json.

Notes

  • --no-record disables recording (drops --publish_ros + --use_recording); omit it to record.
  • Recording produces ~1.5 GB/episode — watch disk; clean recording_data/ after validating.
  • The container is ephemeral; only the mounted recording_data/, logs/, saved_task/ and the Isaac cache survive a run.
  • Full task-config authoring: source/data_collection/TASK_CONFIG_GUIDE.md. Module reference: source/data_collection/AGENTS.md.