Visual Result Showcase

Visual Result Showcase

Use this skill when a completed Roboclaws run needs to become a publishable visual artifact: blog GIF, README demo strip, release note contact sheet, or operator-facing result montage.

Boundary

This is post-run artifact work. By default, use only:

• run_result.json
• trace.jsonl
• agent_view.json
• robot_views/*
• public report images such as before.png, after.png, and map previews

Do not run a new robot task, call a VLM, inspect private manifests, or imply that report-only views were agent inputs. If a score is shown, label it as post-run evaluation.

Household Cleanup GIF

For cleanup result showcases, prefer the whole-task story:

1. task and before state;
2. observe sweep progress;
3. every object cleanup chain, preserving the visible tool order;
4. final state and post-run evaluation.

Use the bundled renderer:

python skills/visual-result-showcase/scripts/render_household_cleanup_showcase.py \
  --run-dir output/household/household-world/cleanup/direct-world-public-labels/seed-7 \
  --out-dir docs/blog/assets/household-cleanup-showcase

The script writes:

• showcase.gif
• contact_sheet.png
• frames/*.png
• manifest.json

Visual Rules

• Make FPV the main panel. Prefer .fpv.bbox.png when available.
• Label FPV as the agent-facing view.
• Label chase/RPV and verification/map views as report-only evidence.
• Include an MCP tool trace bar so the viewer sees observe, navigation, pick, place, receptacle open/close, and done.
• Keep claims honest: "Codex agent drives bounded MCP cleanup tools" is fine; "visual-only", "real robot", or "fully solved cleanup" is not unless the source run proves that exact claim.
• For WeChat/blog publication, keep a GIF plus source frames. Use the contact sheet for quick human review before embedding.