---

name: data-preparation description: "Prepares and validates the nuScenes, CAN-bus, and VAD temporal annotation inputs required for VAD training, evaluation, and visualization." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

VAD data preparation

Use this route when the task is to prepare nuScenes for VAD, generate temporal info files, validate CAN-bus data, or diagnose missing dataset inputs.

Route

1. Read data-contract.md and choose a canonical data root.
2. Run the bundled, read-only layout checker before invoking any converter: python scripts/check_data_layout.py --data-root DATA_ROOT --canbus-root CANBUS_ROOT --require-train --require-val.
3. Acquire and unpack nuScenes and the CAN-bus expansion separately; this skill does not download data.
4. Generate VAD-specific temporal annotations with the command contract in the reference. Do not substitute stock MMDetection3D info PKLs.
5. Re-run the checker, then use training-evaluation only after the required PKLs and map annotation are present.

The converter is data- and dependency-heavy. Treat a successful argument parse as different from a successful conversion. Full conversion was intentionally not run during construction.

Scope boundaries

• Model/plugin and config registration: architecture-configuration.
• Train/evaluate and checkpoint normalization: training-evaluation.
• Render an existing result artifact: visualization.

For symptoms, use troubleshooting.md. The bundled checker is a safe validator, not a downloader or converter.