Reproducibility

Deterministic inputs

Use Python 3.12, record the package version, choose a packaged scene, and pass an explicit seed. The minimal example fixes seed 42, one query, two keypoints, and disables pruning:

online-gcs --scene SINGLE_SHELF --iterations 1 --keypoints 2 --prune-interval 0 --seed 42

A seed controls supported random choices; it does not make solver timing, fetched assets, multiprocessing order, or floating-point results identical across every platform.

Visualization asset

Install the optional capture dependency. In one terminal, run a fixed planner scenario that leaves the final Meshcat scene available for inspection:

python -m pip install -e '.[demo]'
online-gcs --scene SINGLE_SHELF --seed 42 --paper-figure --figure-target-idx 1 --keypoints 10 --visualize --figure-hold-seconds 30

While the 30-second hold is active, copy the printed Meshcat URL into this command. The macOS path is optional because the script auto-detects common system Chrome or Chromium installations. If auto-detection finds nothing, omit the option only when a Playwright Chromium browser is already installed:

python scripts/capture_demo.py --url http://localhost:7001 --output-dir artifacts/demo --width 1280 --height 720 --frames 5 --browser-executable "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"

The capture tool accepts only loopback HTTP(S) URLs, removes stale numbered frame files from the selected output directory, and leaves unrelated files untouched. Inspect all five frames. Only if they contain genuine state changes, assemble them with the frame-limited command printed by the script:

ffmpeg -y -framerate 2 -i artifacts/demo/frame-%02d.png -vf "scale=960:-2:flags=lanczos" -frames:v 5 -loop 0 docs/assets/online-expansion.gif

The repository publishes the three inspected scene screenshots, not a generated GIF. Keep locally captured frames and GIFs under ignored paths until every frame has been reviewed and the sequence shows real planner behavior.

Verification

Install development and documentation extras, then run the public-tree, unit, repository, style, and documentation checks:

python -m pip install -e '.[dev,docs]'
python scripts/check_public_tree.py
pytest tests/unit tests/repo
ruff check .
mkdocs build --strict

Drake integration and smoke checks exercise real scene construction:

pytest tests/integration tests/smoke -m "not slow and not visualization"

Slow, multiprocessing, or visualization checks should be run separately when a change affects those paths.

Generated output

Write generated regions, measurements, screenshots, and experiment output only under ignored artifacts/ or results/ directories. Do not commit CSV files, logs, notebooks, reports, or generated experiment results. The repository policy test enforces this release boundary for tracked files.

Reporting a result

Record the commit, Python and dependency versions, platform, scene, complete CLI arguments, seed, solver availability, and whether models were already cached. Publish the command or script needed to reproduce a claim, not only a summary number. The repository itself intentionally contains no raw benchmark dataset.