AVLA-Skills

Evidence-aware skills for Autonomous Driving + Vision-Language-Action research.
Client-neutral workflows for academic writing, code-grounded method analysis, and editable research figures.

AVLA-Skills is an open-source collection of self-contained Agent Skills for AVLA research. It helps researchers turn supplied evidence into calibrated paper sections, auditable Method Cards, and editable method figures while keeping unsupported claims visible instead of silently filling gaps.

Why AVLA-Skills

Evidence-bounded	Standalone	Auditable
Distinguishes supplied, Direct, Derived, Inferred, and missing evidence.	Each skill folder can be copied and used without the repository knowledge base.	Produces claim checks, evidence reports, structured cards, and validated figure specifications.

The repository contains public metadata and concise paraphrases only. It does not redistribute paper text, paper figures, third-party source code, private prompts, local research cards, or model checkpoints.

Skills

Academic writing

Skill	Research task
avla-write-abstract	Build a compact problem-method-result-limit narrative from available evidence.
avla-write-introduction	Develop motivation, the research gap, positioning, and calibrated contributions.
avla-write-related-work	Synthesize a closed bibliography by technical lineage, mechanism, and evidence boundary.
avla-write-method	Explain representations, coordinate frames, components, objectives, and information flow.
avla-write-experiments	Connect claims to protocols, metrics, baselines, ablations, and defensible result language.
avla-write-conclusion	Close the argument with supported findings, limitations, and scoped future work.

Code-to-artifact

Skill	Research task
avla-code-to-method	Trace repository evidence into a Method Card, method outline, and evidence report.
avla-code-to-figure	Convert a validated Method Card into a figure specification and editable SVG.

Quick Start

1. Choose one skill and copy its complete folder into an Agent Skills-compatible client.

   cp -r skills/avla-write-method PATH_TO_CLIENT_SKILLS/

2. Invoke the skill by name and provide only the evidence it may use.

   Use avla-write-method to draft the method section from this Method Card.
   Separate direct evidence from inference and list unresolved technical details.
   

3. Review every citation, number, implementation claim, and unresolved item before using the result in a paper.

For code-grounded workflows, the copied skill includes standard-library scripts:

python scripts/extract_method_card.py REPOSITORY_ROOT \
  --config CONFIG.json --entrypoint train.py --entrypoint infer.py \
  --output OUTPUT_DIRECTORY

python scripts/build_figure_spec.py method_card.json \
  --style assets/style-tokens.json --output OUTPUT_DIRECTORY

Evidence-aware workflow

flowchart LR
    A[Public sources and project evidence] --> B[Evidence ledger]
    B --> C{Choose a standalone skill}
    C --> D[Academic paper section]
    C --> E[Auditable Method Card]
    E --> F[Editable SVG figure]
    D --> G[Claim and citation audit]
    F --> G
    G --> H[Researcher review]

Loading

The skills never turn a missing input into a fact. Citations, numerical results, datasets, coordinate frames, losses, architecture details, and capabilities must remain absent or explicitly unresolved unless the supplied evidence supports them.

Evaluation and transparency

The repository includes a fixed eight-task synthetic comparison that generates each deliverable once without loading a skill and once with the corresponding skill loaded. A blinded output-only evaluation gave the skill-loaded group 75/80 and the no-skill group 74/80, with 3 wins, 1 tie, and 4 losses for the skill-loaded outputs.

That is a slight aggregate edge, not evidence of universal improvement. The comparison is intentionally preserved so users can inspect both conditions, the rubric, individual failures, and the limitations. See the comparison report and blinded evaluator report.

Repository structure

skills/                 Eight standalone public skills
knowledge/              Public metadata and concise AVLA patterns
schemas/                Method, figure, and provenance contracts
pipeline/               Validation, curation, reference, and release tools
tests/evaluations/       No-skill and skill-loaded comparison artifacts
assets/brand/            Hash-bound public README media

Contributing

Keep contributions public, paraphrased, and traceable. Do not add private prompts, local paths, credentials, raw paper cards, copied passages, paper figures, checkpoints, or third-party source code.

Python 3.11 or later is sufficient for the repository runtime:

python -m unittest discover -s tests -v
python pipeline/validate_knowledge.py knowledge
python pipeline/build_skill_references.py --knowledge knowledge --skills skills
git diff --exit-code -- skills
python pipeline/scan_public_release.py . --media-manifest .release-media.json

License

Executable code is licensed under Apache-2.0; see LICENSE. Skill instructions, documentation, and curated paraphrased knowledge are licensed under CC BY 4.0; see LICENSE-DOCS. See NOTICE for the attribution and boundary policy.