Robotwin Create Benchmark

Trigger Patterns

Use this skill when the user asks for any of the following:

• Create a new RoboTwin benchmark env.
• Automate writing envs/<task_name>.py.
• Use /robotwin cb [task name] [task description] [other demands].
• Create a task class from a benchmark description.
• Validate object placement by rendering and visually inspecting the scene.

Inputs

Normalize the request into:

• task_name
• task_description
• other_demands

If the user gives a slash-like command, parse it first and then normalize with scripts/normalize_request.py. For raw slash-style parsing, prefer quoted task_name and task_description, or use || as a separator between task_name, task_description, and other_demands.

Required Workflow

1. Normalize the request.
2. Discover RoboTwin roots with scripts/discover_robotwin.py.
3. Run scripts/check_robotwin_health.py on the selected root, including render checks.
4. Stop and ask the user if there are multiple strong candidates, missing assets, or a failed render environment.
5. Read references/task_contract.md.
6. If you need examples, only glance at the fixed files listed in references/fixed_examples.md.
7. Generate envs/<task_name>.py directly. Do not do a repo-wide similarity search for the closest task.
8. Validate the task file with scripts/validate_task_env.py.
9. Render initial-scene previews with scripts/render_preview.py.
10. Build a preview montage with scripts/build_preview_montage.py when multiple images exist.
11. Review the rendered image or montage with the model's vision ability using references/vision_review_checklist.md.
12. If placement is unreasonable, patch load_actors or related setup logic and repeat the preview loop.
13. Optionally write task_config/<task_name>.yml with scripts/write_task_config_stub.py.
14. Before any asset download or real collection run, stop and ask the user.

Generation Rules

• The generated file must be envs/<task_name>.py.
• The class name must exactly match task_name.
• The class must inherit Base_Task.
• The file must implement setup_demo, load_actors, play_once, and check_success.
• setup_demo must call super()._init_task_env_(**kwags).
• Prioritize correctness of load_actors, initial placement, and check_success.
• play_once may be approximate and does not need to be data-collection-ready.
• Prefer a small number of stable references over repo-wide similarity search.

Render And Vision Review Rules

• Render review is mandatory before calling the task creation complete.
• Environment readiness is also mandatory. If SAPIEN or the renderer cannot initialize, do not claim the task is verified.
• The first required preview is the initial scene after setup_demo, not after a full successful play_once.
• Review the image for floating objects, penetration, off-table placement, unreachable positions, bad target poses, edge-risk placement, and embodiment mismatch.
• If the review result is warn or fail, keep iterating or stop to ask the user when the issue is preference-sensitive.

When To Pause And Ask The User

• More than one RoboTwin root looks plausible.
• No healthy RoboTwin root is found.
• Default assets appear missing and downloading may be needed.
• The render environment cannot initialize.
• The task description is too vague to determine placement or success criteria.
• The preview looks wrong but the fix depends on user preference.
• The user has not confirmed whether to continue into data collection.

Scripts

• scripts/normalize_request.py
• scripts/discover_robotwin.py
• scripts/check_robotwin_health.py
• scripts/generate_task_env.py
• scripts/validate_task_env.py
• scripts/render_preview.py
• scripts/build_preview_montage.py
• scripts/write_task_config_stub.py

References

• references/task_contract.md
• references/fixed_examples.md
• references/vision_review_checklist.md

Output Contract

The usual outputs are:

• envs/<task_name>.py
• task_config/<task_name>.yml when requested
• .codex/robotwin_cb/<task_name>/preview_seed_*.png
• .codex/robotwin_cb/<task_name>/preview_montage.png
• .codex/robotwin_cb/<task_name>/vision_review.json

If you cannot complete render verification, say so clearly and explain whether the blocker is project health, missing assets, or Python/render environment failure.