Research Survey Loop

Research Survey Loop

Survey topic: $ARGUMENTS

Overview

Use this skill for long-running survey work with a primary focus on robotics, embodied AI, computer vision, world models, spatial intelligence, navigation, manipulation, and 3D perception.

The workflow can extend to adjacent computer science topics, but the bundled search priorities, venue choices, and writing templates are optimized for robotics and embodied AI communities rather than for all research domains equally.

The default working style is:

1. Create or resume one task directory under survey_tasks/<topic-slug>/
2. Keep task.md stable unless the user explicitly changes the task itself
3. Append round_log.md every round
4. Rewrite current_task.md every round so the next unattended run has a precise entry point
5. Grow one long survey.md by category paragraphs rather than per-paper cards
6. Move task-absorbed local PDFs out of the root papers/ pool into the task's own sources/papers/

This skill is academic-first. Only include industry_report/ when the user explicitly wants industry or market material mixed into the survey.

Directory Contract

Every survey task lives under:

survey_tasks/<topic-slug>/
  task.md
  round_log.md
  current_task.md
  survey.md
  sources/
    papers/
    supplementary/

Rules:

• Treat the root papers/ directory as an unsorted staging pool, not as the long-term citation location.
• Once a local PDF is truly used by a task, move it into that task's sources/papers/.
• If another task later needs the same paper, copy it from the first task's sources/papers/ into the new task instead of pointing back to the old path.
• Use relative local links from survey.md, such as ./sources/papers/example.pdf.
• If a source cannot be downloaded locally, cite the canonical web URL instead.

Required Resources

Read these bundled files before acting:

• references/source-priority.md for the fixed search order and venue priorities
• references/writing-rules.md for reading depth, citation format, migration rules, and round-closing discipline

Use these bundled templates and scripts instead of improvising new file layouts:

• assets/task-template.md
• assets/round-log-template.md
• assets/current-task-template.md
• assets/survey-template.md
• scripts/init_task.py
• scripts/fetch_sources.py
• scripts/extract_pdf_chunk.py

Workflow

1. Bootstrap or Resume

If the task directory does not exist, run:

python3 scripts/init_task.py "TOPIC" --workspace-root "$PWD"

If the task already exists, read in this order before doing any new work:

1. task.md
2. current_task.md
3. The latest round in round_log.md
4. survey.md

Do not rewrite task.md unless the user explicitly changes the task definition.

2. Search Web First

Default search order is fixed:

1. Nature / Science and related journals
2. Top CV venues
3. Top robotics venues
4. arXiv
5. Local papers/ pool as supplement, de-dup, or migration source

Use the current agent's web search / fetch capabilities for publisher-first searching. Use scripts/fetch_sources.py to normalize downloads and imports into the task directory.

Typical commands:

python3 scripts/fetch_sources.py search "TOPIC" --task-dir "survey_tasks/TOPIC-SLUG" --max-per-source 5
python3 scripts/fetch_sources.py download --task-dir "survey_tasks/TOPIC-SLUG" --arxiv-id 2402.07556

Search rules:

• Prefer official publisher or venue pages over tertiary summaries.
• Prefer formal published versions over arXiv when both exist.
• Keep arXiv as a rapid frontier source and fallback download source.
• The bundled fetch_sources.py search command uses public Semantic Scholar and arXiv APIs for normalized results.
• If one source is unavailable, continue with the others rather than blocking.

3. Import and Migrate Local Papers

Only migrate local PDFs that the current task actually absorbs.

To move a root-level paper into the task:

python3 scripts/fetch_sources.py import-local --task-dir "survey_tasks/TOPIC-SLUG" "papers/file.pdf"

To reuse a paper from another survey task:

python3 scripts/fetch_sources.py reuse-task-paper --task-dir "survey_tasks/TOPIC-SLUG" --from-task "survey_tasks/OTHER-TASK" "sources/papers/file.pdf"

Migration rules:

• Move from root papers/ into the task when the paper becomes part of the task's reading queue or survey evidence.
• Record the original and destination paths in round_log.md.
• Remove migrated items from the current_task.md pending migration list.
• Never keep citing the old root papers/ path in survey.md.

4. Read PDFs in Chunks

Never read more than 10 PDF pages at once.

Use:

python3 scripts/extract_pdf_chunk.py "survey_tasks/TOPIC-SLUG/sources/papers/file.pdf" --start-page 1 --end-page 10 --json

Rules:

• Long papers must be read in consecutive windows such as 1-10, 11-20, 21-30.
• Do not write firm conclusions into survey.md before understanding:
  • the research problem
  • the core method
  • the claimed contribution
  • the evidence supporting the claim
  • the main limitation
• Use survey papers to build taxonomy and terminology first.
• Use original papers to strengthen category paragraphs and comparisons.

5. Write the Survey Incrementally

survey.md is one long Markdown document, not a pile of independent paper cards.

Writing rules:

• First round: build rough categories and coarse survey paragraphs.
• Later rounds: deepen one category at a time by integrating newly understood papers into the relevant paragraph.
• If a paper is not yet fully understood, keep it in current_task.md or note it in round_log.md as pending.
• Preserve terminology consistency across rounds.
• Keep every claim traceable to a relative local path or canonical web link.

6. Close the Round

At the end of every round:

1. Append a new round block to round_log.md
2. Rewrite current_task.md
3. Leave the next round with:
   • explicit search queue
   • explicit reading queue
   • explicit pending migrations
   • explicit next paragraph targets in survey.md

Do not finish a round with only a vague note like "continue later."

Output Shape

The four main files have fixed roles:

• task.md: stable task contract
• round_log.md: append-only history
• current_task.md: editable current execution plan
• survey.md: long-form Chinese survey with category paragraphs

Default language is Chinese, while preserving English titles, venues, and technical terms where useful.

Failure Handling

• If a source cannot be downloaded, keep the best canonical URL and continue.
• If Semantic Scholar or arXiv access is temporarily unavailable, keep working with the remaining sources.
• If a PDF is longer than expected, continue chunked reading rather than skipping it.
• If a paper is relevant but still unclear, mark it as pending instead of writing a fake summary.
• If a task directory already exists, preserve task.md by default and only recreate missing files.