PVEP reproduction repository

This repository accompanies Verifier-Enforced Belief Space Planning for Foundation Model Proposals across Industrial Robot Tasks. It contains curated trial-level result tables, deterministic analysis and figure-regeneration scripts, and reference implementations of the planners, verifiers, and domain models used in the released analyses.

The repository is synchronized with the author-identified manuscript.

Reproduce the reported aggregates and figures

Python 3.10 or newer is recommended.

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python paper/figure_scripts/recompute_ariac_q0.py
python paper/figure_scripts/generate_all.py
python scripts/generate_manifest.py
python scripts/verify_release.py

The Q0 command must report 50 trials, 195 subgoals, 65 first-check passes, a trial-level macro mean of 0.345380952381, and a descriptive pooled micro fraction of 0.333333333333.

Figure regeneration writes:

• paper/fig2_ariac.pdf
• paper/fig3_tv.pdf
• paper/fig4_rotor.pdf
• paper/fig5_coverage_scaling.pdf
• paper/fig6_scope_intervention.pdf (Supplementary belief-support intervention)
• paper/figure_scripts/figure_data_audit.json

The representative overview composites used in panel a of Figs. 2--4 are provided as publication assets; every quantitative panel is rebuilt from the released tables.

Repository layout

• results/: curated trial-level CSV files and parameter tables.
• paper/figure_scripts/: deterministic figure and statistical-analysis code.
• ariac/: ROS 2/ARIAC verifier, repair controller, PDDL, and simulator interface.
• pvep/ibpomcp/: shared online belief-space planning implementation.
• pvep/kinova/: fastening-domain model and controller code.
• pvep/sleeve/: thermal interference-fitting model and controller code.
• docs/DATA_AND_CODE.md: claim-to-file, configuration, and figure mapping.
• docs/PROTOCOL_AND_MODEL_PROVENANCE.md: ARIAC protocol and model-provenance ledger, including fields that were not retained in legacy aggregates.
• scripts/verify_release.py: aggregate, file-integrity, and release-hygiene checks.
• MANIFEST.sha256: cryptographic hashes of released non-generated files.

Scope and access restrictions

The tables under results/ are the curated analysis inputs used by the released scripts. Absolute workstation paths in lineage-only fields have been replaced with repository-relative paths or internal:// identifiers; no numerical result fields were changed by this de-identification.

Raw fixture imagery, camera poses, robot-calibration files, and model weights are excluded because they are subject to industrial data-sharing restrictions. The release instead provides de-identified result tables, observation-model parameters, POMDP configurations, representative publication composites, and the code needed to reproduce the reported quantitative analyses. Editors or reviewers may request confidential access from the corresponding author; access requires institutional and data-owner approval, an executed data-use agreement, and no public redistribution, and therefore cannot be guaranteed for every request.

The ARIAC runtime requires an external ROS 2/ARIAC installation. Trial-level analyses and submitted-figure regeneration require only requirements.txt. Foundation-model calls additionally require explicit provider endpoint and model values; copy .env.example and set PVEP_OPENAI_BASE_URL, PVEP_OPENAI_MODEL, and the API key for each new run. The thermal LLM proposal runner accepts the analogous SLEEVE_LLM_* variables. No third-party endpoint or model alias is selected by default.