VIBTARVE

VLM-Initiated Behavior Tree Auditing of Robot VLA Execution.

RoboCasa GR1 tabletop task recordings of NVIDIA Isaac GR00T N1.6 execution, human-review annotations, and VLM auditor outputs for robot simulation, behavior-tree auditing, and vision-language-action policy evaluation. The simulation stack uses RoboCasa GR1 tabletop tasks, robosuite, and MuJoCo.

Public dataset release snapshot. Hugging Face repository: https://huggingface.co/datasets/jmberhold/VIBTARVE. The maintainer confirmed redistribution rights and reviewer consent for this snapshot. That confirmation is recorded, not an independent legal review of upstream assets. CC BY 4.0 applies to the released dataset; it does not relicense upstream software, model weights, simulator assets, or the implementation repository.

Scope

488 recordings: 24 prompt-development recordings and 464 evaluation recordings. These are related scenario/run trials, not a random train/test split. 120 evaluation recordings have confirmed completed human review. Workbook presence and empty completion cells alone do not establish review. Human completion flags are null outside confirmed available node review. Evaluation confirmation comes from the completed review schedule. Prompt-development confirmation comes from the explicitly reviewed episode-1 analysis CSV; the evaluation schedule does not cover that separate cohort. review_basis records the distinction.

Layout and loading

from datasets import load_dataset
episodes = load_dataset("jmberhold/VIBTARVE", "episodes", split="evaluation")
annotations = load_dataset("jmberhold/VIBTARVE", "human_annotations", split="evaluation")

The dataset is public and ungated; no Hugging Face login or access token is required. Locally use pandas.read_parquet("data/episodes/evaluation.parquet"). Each configuration is a separate relational table, joined by episode_dir and, where applicable, node_id, sheet/row, model name, or frame step. Only nonempty cohorts are advertised as loadable splits; in this snapshot progress annotations have an evaluation split but no prompt-development split.

• data/: Parquet tables, plus CSV copies of normalized human/progress annotations.
• summaries/: all source-root BTAudit2 CSVs, with original filenames and column layouts. Paths/contact fields are sanitized. provenance/summary_exports.json distinguishes current canonical exports from historical analyses.
• episodes/: portable BT/recording metadata, simulator-state evidence, and readable videos. Absolute source paths are rebased or explicitly redacted.
• media/frames/: uncompressed TAR shards of both saved camera views. JPEG pixels are unchanged. frames identifies each archive/member, step, camera, and checksum. Archive members are relative episodes/<episode>/frames/<camera>/frame_<step>.jpg. Extract chosen shards into the release root to use rebased image paths. Extract only verified archives and reject absolute/traversal member paths.
• provenance/prompts/: saved replay prompt snapshots, content-addressed after sanitization. Original and released hashes are distinguished.
• examples/: representative scene galleries, not independent annotations.
• provenance/: source hashes, cohort lists, attempt configurations, exclusions, export policy, and privacy transformations. Raw credentials and debug logs are excluded.

The workbook_rows table preserves all nonempty cells and hyperlinks in every worksheet, including historical model sheets, Sheet0 progress, Sheet1 human review, headers, notes, and formulas as text. Coordinates and workbook hashes retain provenance; workbook styling, macros, document properties, and original XLSX files are not distributed. human_annotations additionally exposes Sheet1 node rows as columns. Partial/unconfirmed annotations remain available but are not automatically scored as ground truth.

Measurement and limitations

Primary selected auditor observations come from the canonical schema-5 tables; historical conditions remain in workbook rows and replay-attempt records. Synthetic BT milestones are labeled separately from mission-engine responses. Synthetic milestones do not supply scoring judgments. Unreached positions within a known model contract and missing model episodes must not be conflated.

Ordinal reconciliation (October 25, 2026): comparisons use Node 1, Node 2, etc., preserving raw IDs and full model-specific contracts in model_node_contracts_json. Observation rows retain node_position, node_position_basis, and model_node_count. Only shared positions are scored; extra positions are not comparable, not false incomplete. Whole-mission accuracy excludes unequal full node counts. This measures positional agreement, not semantic equivalence of historical BT versions.

Among 120 reviewed episodes, full counts differ in 44 Qwen3-VL, 20 original-prompt Qwen3.5, and 17 tailored-Qwen3.5 cases. Seven reviewed Qwen3 cases (22 across the evaluation corpus) use explicitly authorized first-seen response order beginning at frame zero, with the full count recovered from recorded BTs; these bindings are labeled as inferred. Five unreviewed Qwen3.5 model/episode contracts remain unresolved and are excluded from comparable scoring. See paper/tables/d21_node_positions.csv for each episode/model's binding basis, full count, and shared/extra positions. The revised results replace exact-ID scoring, which had understated legacy coverage. Results for Molmo2, Gemma 4, Qwen3.8, and Luna are unchanged.

Updated figures, tables, supporting data, and a browseable index are under paper/; download the complete bundle as vibtarve_paper.zip. The update changes derived analysis and portable exports only, not original recordings or human workbooks.

Machine/simulation time differs from review-video time. Luna's derived video times use frame-index/duration scaling; original timestamps and timing-basis metadata are retained. Historical Qwen3.8 uses eight-frame contact sheets; Luna uses ten separate images. These are not controlled input-matched model comparisons. The historical Qwen3-VL clock discrepancy remains unresolved; ordinal node reconciliation does not resolve source-clock differences. Do not silently recalibrate it.

Unreadable videos are excluded from playable media, not from episode/node/annotation tables or available saved frames. Consult provenance/quality_issues.json for explicit exclusions and workbook-lock warnings. Video checks measure duration and decode the first frame; they are not a guarantee that every encoded frame is undamaged.

Human progress is distinct from binary node completion. Original model explanations are claims, not independent visual ground truth. Missing labels remain missing. Consult the flagged artifacts when assessing dataset quality. Automated privacy checks do not prove that free-text notes contain no personal information.

Attribution and reproducibility

Authors: J. Mark Berhold, Bruce A. Cox, Neil C. Ranly, and Nicholas J. Yielding. Implementation and upstream attribution: https://github.com/jmberhold/dabtroll and its docs/research_attribution.md. The original snapshot's source code/worktree state is recorded in release_manifest.json; its derived_revision identifies the newer scoring code and provenance/ordinal_reconciliation.json records the derived-only refresh. Historical source inventories and publication receipts remain unchanged. A dirty worktree is not an immutable software release. No model weights or upstream simulator assets are redistributed as standalone assets.

Use SHA256SUMS to verify released files. Per-frame hashes verify archive contents. Source fingerprints describe the immutable on-disk inputs, not unsaved reviewer edits. This exporter never uploads or changes repository visibility.